Mechanomyography Sensor Design and Multisensor Fusion For Upper-Limb Prosthesis Control

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/267776027
Mechanomyography Sensor Design and Multisensor Fusion for Upper-Limb

Prosthesis Control
Article
CITATIONS READS
10 663
1 author:
Jorge Silva
Edge Impulse
14 PUBLICATIONS 338 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Activity Detection View project
Brain-Computer Interfaces View project
All content following this page was uploaded by Jorge Silva on 29 October 2015.
The user has requested enhancement of the downloaded file.

Mechanomyography Sensor Design and Multisensor Fusion
for Upper-Limb Prosthesis Control
by
Jorge Silva
A thesis submitted in conformity with the requirements

for the degree of Master of Applied Sciences
Graduate Department of Mechanical & Industrial Engineering
University of Toronto
c 2004 by Jorge Silva

Copyright
Abstract
Mechanomyography Sensor Design and Multisensor Fusion for Upper-Limb Prosthesis
Control
Jorge Silva
Master of Applied Sciences
Graduate Department of Mechanical & Industrial Engineering

University of Toronto
2004
Current electromyography (EMG) sensors have not been successfully combined with more
comfortable and functional silicone soft sockets in externally powered prostheses due to
contact, wire breakage and daily use issues. To overcome those issues, this thesis pro-
poses a novel control framework based on the measurement of the mechanical activity
of remnant muscles (mechanomyography or MMG). This document details the technical
implementation, design optimization and practical tests that lay the groundwork for the
development of a new generation of MMG-driven upper-limb externally powered prosthe-
sis. Fundamental contributions include a systematic characterization and optimization
of silicone-embedded sensors for MMG signal recording, the novel design of a coupled
MMG sensor pair for dynamic noise reduction and the first ever reported use of a MMG
sensor array for the generation of multiple signals for prosthesis control. Tests with am-
putees demonstrate that MMG-driven prostheses may advance the frontier of externally
powered prostheses research.
ii
Dedication
To my friends and colleagues, my best wishes for all your future endevours.
A mis padres, mi amor y respeto incondicionales.
To my supervisor, my endless admiration and gratitude.
To my partner, my most valuable possession: my life.
iii
Acknowledgements
I would like to acknowledge the Bloorview MacMillan Childrens Foundation, the Nat-
ural Sciences Engineering Research Council of Canada, the Hospital for Sick Children
Foundation and the Bloorview Childrens Hospital Foundation for their financial support.
Special thanks to Dr. Steve Naumann and the researchers of the Gait Laboratory at
the Bloorview MacMillan Children’s Centre; to Dr. Parham Aarabi in the Artificial Per-
ception Laboratory; and to Dr. Z. W. Zu in the Vibration Laboratory at the University

of Toronto for their useful suggestions, time and generousity facilitating the necessary
resources for the development of this work.
I would also like to acknowledge the unconditional support, respectful trust, endless
dedication, and clever guidance of Dr. Tom Chau, my thesis supervisor. Working with
him has been, and will always be, a continuously rewarding and satisfying experience.
iv
Contents
1 Introduction 1
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Silicon-Embedded Accelerometers 5
2.1 Quantifying the Signal-to-Noise Ratio of Silicon-Embedded Sensors for
Mechanomyography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.3 Materials & Methods . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Systematic Characterisation of Silicon-Embedded Accelerometers for Mechano-
myography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.3 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
v
2.2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3 Silicon-Embedded Microphones 33
3.1 Optimization of the Signal-to-Noise Ratio of Silicon-Embedded Micro-
phones for Mechanomyography . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.3 Mechanomyography . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.4 MMG Sensors & Prosthesis Control . . . . . . . . . . . . . . . . . 35
3.1.5 Materials & Methods . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4 Coupled MMG Sensor Pair 42
4.1 Coupled Microphone-Accelerometer Sensor Pair for Dynamic Noise Re-
duction in MMG Signal Recording . . . . . . . . . . . . . . . . . . . . . 43
4.1.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.3 Sensor Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.1.4 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Coupled MMG Sensor Pair: Model Validation and Optimization . . . . . 49
4.2.1 Rationale: Model-Based Source Separation . . . . . . . . . . . . . 49
4.2.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2.3 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
vi
5 Multisensor Fusion 72
5.1 MMG-Based Multisensor Data Fusion for Prosthesis Control . . . . . . . 73
5.1.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.1.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.2 MMG-Based Multisensor Fusion Strategy for Detection and Classification
of Muscular Activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2.1 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.2 Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2.3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.4 Algorithm Integration . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2.5 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.2.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6 Conclusions 105
6.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2.1 The Prototype . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

6.2.3 Pattern Recognition . . . . . . . . . . . . . . . . . . . . . . . . . 111
Bibliography 112
vii
List of Tables
1.1 Section numbers with relevant novel content for each chapter of the thesis. 4
2.1 Embedding properties of the prepared samples. . . . . . . . . . . . . . . 9
2.2 Mean values of SNR obtained for the hardest silicon type (shore 65A). . 12
2.3 Rank sum values obtained for samples reported on Table 2.2. Significant
differences (p 0.025) are highlighted in bold. . . . . . . . . . . . . . . . 13
2.4 Top-layer (TT L ) and bottomlayer (TBL ) thickness values of samples prepared 19
2.5 SNR values (dB) measured in embedded and non-embedded transducers.
Increasing applied forces are labelled F1 -F4 . . . . . . . . . . . . . . . . . 25
3.1 SNR values obtained for each combination of embedding properties tested 39
3.2 Normalized sensitivity values obtained for each combination of embedding

properties tested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1 Performance measures for different sets of values of γ . . . . . . . . . . . 81
5.2 Possible pre-processing combinations for MMG-based detection and clas-
sification of muscle contractions. . . . . . . . . . . . . . . . . . . . . . . . 92
viii
List of Figures
1.1 Thesis roadmap. The corresponding section numbers identifying the novel
information in each chapter are listed in Table 1.1 . . . . . . . . . . . . . 3
2.1 Schematic diagram of the silicon-embedded accelerometers prepared. “TT L ”
stands for Top Layer Thickness, “TBL ” stands for Bottom Layer Thickness. 9
2.2 Schematic diagram of the mechanical stimulator. Samples were placed

over a closed-cell foam to assure physical contact. . . . . . . . . . . . . . 10
2.3 Schematic diagram of externally powered below-elbow prosthesis. Cur-
rently, EMG sensors are fastened in a hard socket wall. The hard shell
that holds terminal element is adjusted over the hard socket . . . . . . . 16
2.4 Comparison of (a) laminated hard socket and ( b) soft silicon socket or
roll-on sleeve. Rectangular openings in hard socket are intended for EMG
sensor attachment. Note that silicon socket has been rolled up to empha-
sise its flexibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Schematic diagram of silicon-embedded accelerometer samples. TT L =top
layer thickness; TBL =bottom layer thickness . . . . . . . . . . . . . . . . 19
2.6 Block diagram of instrumented system for stimulating and recording re-
sponse from sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.7 (a) Top view and (b) schematic diagram of mechanical stimulator built
for measurements. Placement of embedded samples is shown in schematic
diagram. All dimensions are in millimetres . . . . . . . . . . . . . . . . . 21
ix
2.8 Mechanical coupling of tested samples to vibrating surface of stimulator.
(a) Top view. (b) Front view . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.9 Typical results obtained from most representative sensor groups for each
variable tested. (a) Silicon hardness test; R2 =0.909. ( b) Contact surface
test; R2 =0.978. (c) Top layer thickness test; R2 =0.849. (d) Bottom layer
thickness test; R2 =0.808. Each graph shows linear regression. . . . . . . 27
2.10 Sample bode plot of one of samples tested. Reference for each frequency
(0 dB, 0 degrees) is non-embedded transducer response . . . . . . . . . . 28
2.11 General form of embedded accelerometers relative magnitude. A: Low-
pass band. B: Peak frequency band. C: Reject band. D: High-pass band.
R1 =reject band ripple. R2 =highpass band ripple. Dashed line indicates
maximum deviation from general form. In region A, this deviation was
due to silicon hardness and contact surface variations . . . . . . . . . . . 29
2.12 Example of actual MMG signal recorded using silicon-embedded accelerom-
eter. Vertical axis is in units of gravitational acceleration g . . . . . . . . 30
3.1 Schematic diagram of the MMG sensor design used in experiments. H
stands for height and D for diameter of the enclosed cylindrical air chamber. 37
3.2 Normalized SNR surface obtained from interpolation of sensitivity masked
(Table 3.2) and normalized SNR values (Table 3.1) as a function of air
chamber height and diameter. . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Magnitude response of silicon-embedded microphones with optimal air
chamber dimensions and different silicon hardness values. The legends

of individual graphs are labeled with two letters, the first being the hard-
ness of the silicon case and the second being the hardness of the driving
membrane. S stands for “soft” (shore 20A) and H stands for “hard” (shore
65A). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
x
4.1 Schematic diagram of MMG sensor pair, placed against the skin. . . . . . 45
4.2 Sample measurements of extension (left) and random limb movement (right).
Note the considerable differences in the accelerometer signal (bottom pair
of graphs). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Comparison of the mean RMS values and standard deviations among dif-
ferent activities for the accelerometers. Note the increased RMS value in
the accelerometer during random limb movement. . . . . . . . . . . . . . 48
4.4 Schematic representation of the mechanical effect produced by the silicone
membrane vibration in the microphone’s air chamber. An ideal linear
spring model may be assumed. . . . . . . . . . . . . . . . . . . . . . . . . 52
4.5 Reaction forces due to membrane displacement calculated for the air cham-
ber (Fa ) and silicone membrane (Fm ). The total force FT is also shown
for comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.6 Local linearization of the reaction force FT exerted by the microphone’s
air chamber as a function of relative membrane displacement l(t) = x.
The fitted equation and the corresponding R2 value are shown. . . . . . . 55
4.7 Electromechanical model of the coupled MMG sensor pair in Figure 4.1.
Corresponding electrical equivalents for mass, stiffness and damping coef-
ficients of the mechanical variables are shown. . . . . . . . . . . . . . . . 56
4.8 Electrical equivalent circuit (a) and impedance circuit (b) of the model
proposed in figure 4.7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.9 Schematic representation of the experimental setup for model validation
measurements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
xi
4.10 Empirical magnitude and phase responses of microphone to an external
force Fe (t). Each point is calculated from the estimated force applied by
the shaker platform Fe (t) to the measured relative membrane position l(t)
in the coupled MMG sensor. The corresponding fitted polynomials are
also shown as the solid lines. . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.11 Empirical magnitude and phase responses of accelerometer to an external
force Fe (t). Each point is calculated from the estimated force applied
by the shaker platform Fe (t) to the measured sensor acceleration A(t) in

the coupled MMG sensor. The corresponding fitted polynomials are also
shown as the solid lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.12 Empirical (solid line) and optimum model (dashed line) responses for the
transfer functions involved in the model optimization. . . . . . . . . . . . 68
4.13 Sample source localization test for simultaneous active sources. M(t) and
A0 (t) are the microphone and accelerometer signals in Volts and lM (t) and
lMd
(t) are the measured and estimated skin vibration signals. . . . . . . . 70
5.1 MMG sensor pair designed and used for the experiments. (a) Schematic
diagram of the design. (b) Accelerometer (posterior) view of MMG sensor
pair before embedding. (c) Microphone (anterior) view of MMG sensor
pair after embedding; the silicon contact membrane was added in the final
embedding stage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2 Schematic diagram of sensor placement around the distal end of the resid-
ual forearm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.3 Sample detection of an extension (“opening” contraction) signal from the

forearm and the discrimination of movement artifact. . . . . . . . . . . . 81
5.4 Schematic diagram of equation 4.20. The relative membrane position and
acceleration signals (l(t) and A(t), respectively) are individually filtered
and added to obtain an estimate of skin displacement lM (t). . . . . . . . 84
xii
5.5 Typical gain surface (Phase-based TFM) in the direction of sensor i as a
function of the phases, φk and φj , of the other two accelerometers in the
array, at a single input noise frequency. . . . . . . . . . . . . . . . . . . . 86
5.6 Feature space for muscle activity classification. Every point P is defined by
three RMS values derived from the microphone signals. A linear classifier
can be defined for the differentiation between wrist flexions and extensions. 89
5.7 Flow chart for the proposed integrated data/sensor fusion strategies. . . . 91
5.8 Sample soft silicone socket built for MMG signal recording. Note the
embedded multisensor array containing three coupled MMG sensors at
equi-distant angles around the distal end of the stump. . . . . . . . . . . 94
5.9 Experimental setup for real MMG signal recordings from an EMG-driven
prosthesis user. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.10 Original EMG signals (top) and derived targets (middle and bottom) for
the optimization of the proposed MMG-based strategy. . . . . . . . . . . 96
5.11 Mean and standard deviation values of the cross-validated accuracy of
detection for each subject and each proposed strategy of Table 5.2. . . . 98
5.12 Sample MMG-based detection (top) and EMG-based detection (bottom)
of muscle contraction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.13 Sample MMG-based estimation of muscle activity (top) and its corre-
sponding EMG-based counterpart (bottom). . . . . . . . . . . . . . . . . 100
5.14 Sample MMG-based detection of muscle activity (top) and its correspond-
ing EMG-based counterpart (bottom) during random limb movement. The
middle graph shows the accelerometer signals in the three coupled MMG
sensors in the array. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.1 Schematic diagram of a cross-sectional MMG measurement and the asso-
ciated parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
xiii
a
Index of Symbols
Symbol Description
MMG Mechanomyography
EMG Electromyography
SNR Signal-to-noise ratio
MVC Maximum voluntary contraction
PCB Printed circuit board

LTI Linear time-invariant
MSE Mean square error
TL Top Layer of silicone-embedded accelerometer samples
BL Bottom Layer of silicone-embedded accelerometer samples
RMS Root mean square value
MBF Model-based filtering

ASF Accelerometer Sensor Fusion
RBD RMS-based detection of muscle contraction
RAC RMS-based activity classifier
χ2 Chi squared statistical index / distribution
RSV Rank Sum Value(s)
F1 − F4 Increasing distributed position dependent forces

R2 Correlation coeficient
α Significance level for hypothesis testing
p Probability value for hypothesis testing
xiv
Symbol Description
G1 Stimulus signal gain
G2 Accelerometer signal gain
region A Low-pass band of typical silicone-embedded sensor response

region B Peak of typical silicone-embedded sensor response
region C Reject band of typical silicone-embedded sensor response
region D High-pass band of typical silicone-embedded sensor response
R1 Reject band ripple of typical silicone-embedded sensor response
R2 High-pass band ripple of typical silicone-embedded sensor response
l Relative silicone membrane displacement

l0 Filtered relative silicone membrane displacement
lM Relative silicone membrane displacement due to muscle contraction
lc
M Model estimated relative silicone membrane displacement due to
muscle contraction
dlM (t)/dt Relative silicone membrane vibration due to muscle contraction
Fe External force
Pa Air chamber pressure
∆Pa Air chamber acoustic pressure
P0 Air chamber pressure at equilibrum (l(t) = 0)
FT Total restoring force
Fa Restoring force due to air chamber compression/rarefaction
Fm Restoring force due to silicone membrane elasticity

Am Silicone membrane surface area
r Air chamber radius
Va Air chamber volume
V0 Air chamber volume at equilibrum (l(t) = 0)
xv
Symbol Description
VCON E Silicone membrane cone-shaped volume
Tm Silicone membrane thickness
Ym Silicone membrane Young’s modulus

ka Air chamber spring constant
La Air chamber equivalent inductance
Ra Air chamber equivalent resistor
Ll Silicone sleeve equivalent inductance
Rl Silicone sleeve equivalent resistor
Cs Sensor mass differential

Cf Forearm mass
Vm Silicone membrane relative velocity
Vs Coupled MMG sensor velocity
A Coupled MMG sensor acceleration
A0 Filtered coupled MMG sensor acceleration
Za Air chamber impedance (electrical equivalent)

Zf Forearm impedance (electrical equivalent)
Zl Silicone sleeve impedance (electrical equivalent)
Zs Coupled MMG sensor impedance (electrical equivalent)
M Microphone signal
A0 Accelerometer signal
SM Microphone sensitivity
SA Accelerometer sensitivity
~θ Model parameter vector ~θ = {θ1 , θ2 , θ3 } = {L−1 −1 −1
l , Ra , Rl }
ŷ Model’s magnitude or phase response
y Empirical magnitude or phase response
xvi
Symbol Description
mic
MSEmag Objective function for optimization of microphone’s magnitude
response
mic
MSEphase Objective function for optimization of microphone’s phase response
acc
MSEmag Objective function for optimization of accelerometer’s magnitude
response
acc
MSEphase Objective function for optimization of accelerometer’s phase response
J Overall objective function for model optimization
m Phase response slope
∆t Time delay
γ Dynamic threshold
Hm Microphone signal filter
Ha Accelerometer signal filter
φ Accelerometer phase for a single interfering noise frequency
H1 Hypothesis: observed activity is due to muscle contraction
H0 Hypothesis: observed activity is not due to muscle contraction

Hf Hypothesis: observed activity is due to wrist flexion
He Hypothesis: observed activity is due to wrist extension

g
RMS Normalized RMS value of microphone-derived signal
P Point in the 3-D feature space
θ̂ Classification plane defining vector (θ̂ = {θ1 , θ2 , θ3 , θ4 })
d Distance from point P to classification plane in the 3-D feature space

dT Target distance to classification plane in the 3-D feature space
xvii
Preface
This thesis documents research on an alternative technique for muscle activity mon-
itoring for practical upper-limb prosthesis control. This technique, known as mechano-
myography (MMG), is the measurement of the minute mechanical vibrations (“sound”)

produced by contracting muscles. Although the methodologies, results and conclusions
for every stage of the research are fully documented, it is necesary to alert the reader to
the unconventional format of the thesis. This thesis is largely a compilation of reprinted
scientific papers which address specific aspects of the ultimate project’s goal, namely,
to determine the feasibility of controlling below-elbow, externally powered prostheses by
means of MMG.
Details on the introduction, problem statement and rationale for the project are given
in every reprinted paper since each is a self-contained document. For this reason each
chapter begins by refering the reader to the novel information contained therein.
It is also important to note that, since the background for the project is detailed
throughout the entire document, the introductory chapter presents only a broad de-
scription of the problem statement and objectives. The introduction also explains the
sequential order order in which issues were addressed throughout the research project,
by way of a detailed roadmap.
xviii
Chapter 1
Introduction
1.1 Problem Statement
Every year, an estimated 130 people become upper-limb prosthesis users across Canada
due to congenital or acquired amputations1 . Depending on a number of factors such as
the functionality of remaining muscles and user preference, most of these patients will
be fitted with externally powered prostheses. Current commercially available externally
powered prostheses for upper-limb amputees measure the electric impulses produced dur-
ing the contraction of remnant muscles on the patient’s residual limb. Those signals are
then used as commands for the control algorithms. New socket technology has facili-
tated the creation of more comfortable soft silicone sockets or roll-on sleeves [16, 19].
These soft sockets also have the potential to enhance the functionality of externally pow-
ered prosthesis. However, due to coupling, wire breakage and daily use issues, current
electromyography (EMG) sensors have not been succesfully used with roll-on sleeves in
externally powered prosthesis [16, 27]. The main objective of this research is to overcome
these issues by providing a novel control strategy based on the measurement of mechanical
rather than the electrical activity of the remnant muscles. The epidermal measurement
1
Based on data from the Amputee Statistical Database for the UK [28]. Similar incidence rates were
assumed for Canada
1
Chapter 1. Introduction 2
of the axial mechanical vibrations produced by contracting muscles is called mechano-

myography (MMG) [23]. The rationale for the selection of MMG as an alternative control
strategy for prosthesis control is twofold: First, the vibrations propagate through soft
tissue allowing us to record mechanical activity distal to the contracting muscles. This
distal recording is required to avoid wire breakage issues when using soft silicon sockets
[33]. Secondly, MMG shares the advantage as EMG-based systems for the monitoring
of muscular activity [25, 23, 37], namely, a linear relationship between the amplitude
of the MMG signal and the strength of the contraction. The present thesis reports so-
lutions towards the technical implementation of MMG driven prostheses, addressing in
particular, sensor design optimization, data fusion and empirical test results. This thesis
lays the groundwork for the development of a new generation of below-elbow externally
powered prostheses. These prostheses have the potential to directly impact the quality of
life of approximately 150 current clients of the musculoskeletal program at the Bloorview
MacMillan Children’s Centre [11] and, eventually, patients with upper-limb amputations
in clinics locally and abroad. For further background information please refer to section
2.2.2.
1.2 Objectives
The overall objective of the thesis is to design and instrument a reliable control system for
externally powered prosthesis for below-elbow amputees by means of mechanomyography.
The proposed design should be comparable with current technologies in terms of function-
ality, reliability, efficiency and cost. If MMG is proven feasible, the advantages provided
by the use of roll-on sleeves may trigger the development of a new generation of externally
powered prosthesis. Specific objectives for this research are:
1. To design and implement a reliable MMG sensor for practical use in prosthesis
control. Due to wire breakage issues, the sensor must facilitate distal recording of
Figure 1.1: Thesis roadmap. The corresponding section numbers identifying the novel
information in each chapter are listed in Table 1.1
MMG signals. It is also desirable that the sensor can be embedded into the silicone
soft socket to ease the wearing process for the user.
2. To design and implement a robust signal processing and data fusion strategy to
provide enhancement of the SNR and activity discrimination for prosthesis control.
Due to the complexity of the noise sources, dynamic noise elimination and sound
localization may be needed. Multiple sensor combinations should be investigated

as well as the use of different sensor technologies.
1.3 Roadmap
Figure 1.1 depicts the order in which the main issues of this research project were ad-
dressed. Chapters 2 and 3 present a systematic evaluation of novel designs of silicone-

embedded accelerometers and microphones, respectively, for practical MMG signal record-
ing for prosthesis control. These chapters provide an in-depth analysis of the transducers’
characteristics and suggest optimal embedding parameters for the maximization of their
SNR. This information is then used in chapter 4 to justify the proposed design of a
Table 1.1: Section numbers with relevant novel content for each chapter of the thesis.
Chapter 1 Chapter 2 Chapter 3 Chapter 4 Chapter 5 Chapter 6
3.1.1 4.1.1 5.1.1
3.1.5 4.1.3 5.1.3
All All 3.1.6 4.1.4 5.1.4 All
3.1.7 4.1.5 5.1.5

4.1.6 5.1.6
4.2 5.2
microphone-accelerometer sensor pair for MMG signal recording. A mechanical model
for the proposed MMG sensor design, together with its validation and optimization are
also included in this chapter. The proposed model represents one of the evaluated ap-
proaches to dynamic noise elimination and is essential to the first information fusion
stage (data fusion) where single sensor pair signals are combined to provide significant
SNR enhancement under extreme noise conditions. Chapter 5 proposes and compares
different data fusion (combination of information from sensors withing a single sensor
pair) and sensor fusion (combination of information from an array of multiple sensor
pairs) techniques. The practical advantages and disadvantages of each alternative are
discussed. Finally, chapter 6 summarizes main findings of the thesis, its contributions
to the field of Rehabilitation Engineering, specifically in the enhancement of prosthetics
technology, and future work.
Table 1.1 shows the section numbers identifying new content within each chapter. It
is recommended that the reader directs his/her attention exclusively to these sections.
These section numbers are reprinted at the beginning of each chapter for the reader’s con-
venience.
Chapter 2
Silicon-Embedded Accelerometers
This chapter details research on (1) the feasibility of using silicone-embedded accelerom-
eters for MMG signal recording, and (2) the determination of their optimum embedding
characteristics and manufacturing techniques. The results were summarized and pub-
lished in two papers reprinted by permission in the following sections. Although some
overlap can be observed between the publications, it is important to distinguish between
the approach and scope of each. The first paper focuses exclusively on the comparison
between silicon-embedded and non-embedded accelerometers in terms of their signal-to-
noise ratios (SNR) providing some initial observations of the sensors’ behaviour. The
second paper reports the detailed effects of the silicon, the optimum embedding charac-
teristics and its implications on MMG signal recording. In particular, the second paper
supplements the initial observations by providing a systematic characterization of the
silicon-embedded accelerometer.
The reading of the whole chapter is recommended as most of the information is
presented for the first time. However, the project’s background and methodology for this
stage are more extensively explained in sections 2.2.2 and 2.2.3, respectively. Therefore,
these sections deserve special attention.
5
Chapter 2. Silicon-Embedded Accelerometers 6
2.1 Quantifying the Signal-to-Noise Ratio of Silicon-
Embedded Sensors for Mechanomyography
c
2002 CMBES. Reprinted, with permission, from Silva J., Chau T., and Naumann
S., “Quantifying the Signal-to-Noise Ratio of Silicon-Embedded Sensors for Mechano-
myography,” in Proceedings of the 27th Canadian Medical and Biological Engineering
Conference, Ottawa ON, Canada, November 2002.
2.1.1 Abstract
Experiments consisted of systematic measurements of the signal-to-noise ratio (SNR) of
signals acquired from a mechanical stimulator, using silicon-embedded accelerometers.
The objective of using the latter was to determine the combination of embedding prop-
erties which provide the highest SNR for mechanomyography (MMG) signal recording.
Variations in silicon hardness and geometry were tested. Two important conclusions can
be derived from the experiments: (1) It is possible to acquire MMG signals using silicon-
embedded sensors; and (2) The embedded sensor’s performance is affected by changes in
the geometry of the embedding. The intended application of this study is the use of soft
silicon suction sockets with embedded sensors as a more comfortable and functional al-
ternative to current hard-socket powered prostheses for below-elbow amputees currently
using electromyography as the control signal(s).
2.1.2 Introduction
Mechanomyography
Mechanomyography (MMG)1 is the epidermal measurement of the mechanical activity

(“muscular sound”) of contracting muscles. This phenomenon is a superficial summation
1
Term suggested by Orizio (1993) [23]
of propagated motor unit twitches, and shape changes in muscular fibres during voluntary
contractions [24]. It has been consistently stated that the MMG signal amplitude can
be related to muscle strength in non-fatiguing contractions at fractions of maximum
voluntary contraction (MVC) [23, 24, 25, 26, 17, 14]. Previous studies show that most
of the power of the MMG signal is located in the 0-45 Hz bandwidth [23, 24, 25, 26, 17].
Furthermore, no significant frequency components have been reported beyond 100 Hz
for arm muscles. Muscle mechanical activity has been measured and recorded with
piezoelectric microphones [25], condenser microphones [14, 35] and accelerometers [37].
In order to filter out external “noise”, some researchers have developed different methods
to isolate the recording site. Courteville et al. (1998) used a complex fixing method
to couple a microphone and a silicon support to a foam armband. They used a silicon
membrane to convert the displacement of the skin due to the muscle vibration into
pressure waves for detection by microphone [14]. Goldenberg et al. (1991) reported the
use of surgical cement for fixing a microphone to the skin over the abductor digiti minimi
(ADM) muscle of the subject’s dominant hand [17].
Previous Theoretical Analysis
A theoretical study showed that the fundamental oscillation frequency of the MMG de-
creases as the mass of the transducer increases [38]. Since this study is formulated on
the basis of an ideal situation where the mass of the sensor is uniformly distributed along
the muscle surface, it provides valuable information about the expected influence of the
silicon-embedded sensor on the signal itself. On the other hand, Barry et al. (1986) sug-
gested that muscle-molded transducers would increase the SNR in MMG measurements
[13].
MMG & Prosthesis Control
Although MMG signals have been described and studied for many years, they have never
been used for practical prosthesis control. Barry et al. (1986) demonstrated that reliable
control signals can be obtained from muscular mechanical activity, having tested a MMG-
based control system to open and close a prosthetic hand [13]. While soft silicon suction
sockets are a more comfortable and functional alternative for powered prostheses, sensor
attachment issues have delayed their practical use [16]. Research on MMG as alternative
channels for prosthesis control may cement the union between silicon sleeves and powered
prostheses.
2.1.3 Materials & Methods
Preparation of Samples
Three types of silicon with different hardness values (shores 20A, 35A and 65A) were
chosen to prepare fifteen samples (five samples for each silicon type) of silicon-embedded
sensors using BU-7135 accelerometers from Emkay Innovative Products. Table 2.1 shows
the dimensions of the samples for each silicon type. An initial silicon layer was placed on
a flat surface before curing. The wired accelerometer was placed over the silicon layer.
This initial silicon layer constituted the bottom layer (BL) of the embedding. A second
silicon layer, the top layer (TL), was placed over both the first layer and the sensor. The
whole system was then cut down to its final dimensions. Finally, the sample was placed
into an oven for the silicon to cure. Figure 2.1 shows a detailed diagram of the samples.
Stimulus Generation
A mechanical stimulator based on a WM-R57A card-type speaker from Panasonic (Figure
2.2), was used to transmit vibrations to the prepared samples. The “signal” was a pre-
defined vibration pattern with no frequency components beyond 100 Hz. The peak
Table 2.1: Embedding properties of the prepared samples.
Sample No. TT L [mm] TBL [mm]
1 3.0 1.0
2 2.0 1.0
3 1.0 1.0
4 2.0 2.0
5 2.0 3.0
Figure 2.1: Schematic diagram of the silicon-embedded accelerometers prepared. “TT L ”
stands for Top Layer Thickness, “TBL ” stands for Bottom Layer Thickness.
Figure 2.2: Schematic diagram of the mechanical stimulator. Samples were placed over
a closed-cell foam to assure physical contact.
frequency (20 Hz) amplitude was approximately the same value as that of previously
reported MMG signals (0.3 m/s2 ) [35]. “Noise” consisted of a uniformly distributed
random signal whose maximum amplitude was 15% of the “signal” peak amplitude. Both
signals were generated electronically with a PC 16-bit soundcard using the MATLAB
environment.
Test Methodology
Two different stimuli were separately applied to each sample, one being a pure noise
signal and the other, a vibration signal with additive noise. In order to determine the
effect of contact surface changes on SNR, two more sets of recordings were acquired
after cutting down each dimension (width and length) by a 10 mm decrement. The
three surfaces tested were: 32×42mm, 22×32mm, and 12×22mm (width and length,
respectively). Root-mean-square (RMS) values of voltages obtained from the sensors

were automatically calculated and stored for further analysis. Data from transducers were
acquired through a pair of LM 386 audio amplifiers using a PC 16-bit soundcard. Two
additional measurements with non-embedded transducers were used as control references
for the experiments. From the RMS values obtained, we calculated the SNR of the sample
using equation 2.1.
2
RMSSIGN AL+N OISE
SNR = 10 log −1 (2.1)
RMSN OISE
Analysis and Validation Tests
Forty repeated measurements for each of the 15 samples were recorded. The Lilliefors test
for normality [9] was applied to the SNR data of each sample. Since some data did not
follow a normal distribution, we used the Wilcoxon-Mann-Whitney rank sum one-sided
test to determine whether or not the SNR values obtained for a certain combination of
variables were higher than those of the non-embedded sensors. To test the hypothesis
of independence between SNR and the other variables, the data were first organized
into sensor groups. A sensor group is defined as a collection of samples which differ in
only one variable (silicon type, top layer thickness, bottom layer thickness, or surface
area). Subsequently, we constructed a contingency table with the observed frequencies
of increasing, decreasing or non-monotonic trends in SNR. A Monte-Carlo simulation
with 106 groups of random numbers provided the expected frequencies under the null
hypothesis. A χ2 test for goodness-of-fit was used to test the null hypothesis that there
was no monotonic relationship between the SNR and the other variables. Finally, an
adjusted residuals matrix [3] was computed to determine which types of trends were
significantly different from the expected values. For all statistical tests, a 5% significance
level (α = 0.05) was used.
2.1.4 Results
Table 2.2 shows the mean values of the SNR obtained for the hardest silicone type (shore
65A). Similar tables were obtained for each of the two remaining silicone types. In the
shaded columns, the samples are denoted as x.y, where x is the sample number in Table
2.2, and y identifies the surface area, with y=1 signifying the largest surface. The Control
Table 2.2: Mean values of SNR obtained for the hardest silicon type (shore 65A).
Sample No. SNR [dB] Sample No. SNR [dB] Sample No. SNR [dB]
Control 16.0 Control 16.0 Control 16.0
1.1 13.6 1.2 25.9 1.3 22.2
2.1 8.3 2.2 19.8 2.3 20.4
3.1 16.5 3.2 17.3 3.3 16.9
4.1 12.4 4.2 28.3 4.3 20.6
5.1 16.9 5.2 23.2 5.3 26.0
(non-embedded) sample is included for comparison. Note that most of the SNR values for
the embedded samples are higher than the ones obtained for the non-embedded samples.
Table 2.3 shows the rank sum values (RSV) obtained for the samples reported in Table
2.2. The values that are significantly higher than the control at a 5% level of significance
(RSV < 82) are highlighted in bold. From the Monte-Carlo simulation we obtained
probability values of 0.16, 0.16 and 0.66 for the probabilities of increasing, decreasing
and non-monotonic trends, respectively. From the χ2 test, we determined that SNR
is dependent on the bottom layer thickness, top layer thickness, contact surface area
and silicon hardness. Finally, from residual analysis, the significant trends in the above
dependencies were as follows: The SNR increased as:
1. the bottom layer thickness increased
2. the top layer thickness increased
3. the contact surface area decreased
4. the hardness of the silicon decreased

Table 2.3: Rank sum values obtained for samples reported on Table 2.2. Significant
differences (p 0.025) are highlighted in bold.
Sample No. RSV Sample No. RSV Sample No. RSV
1.1 155 1.2 55 1.3 55
2.1 155 2.2 55 2.3 55
3.1 55 3.2 55 3.3 65∗
4.1 155 4.2 55 4.3 55
5.1 55 5.2 55 5.3 55
∗
p = 0.025
2.1.5 Conclusions
Although the results suggest statistically significant trends, further experiments are
needed to determine the exact nature of these dependencies. For example, in the case
of contact surface area measurements, there may have been a mechanical damping effect
exclusively due to the stimulator characteristics. This effect would not be present in real
MMG measurements. Furthermore, the propagation of the MMG signals through the
tissues may differ from the propagation of the vibration signal through the closed-cell
foam (Figure 2.2). Two main conclusions can be derived from the above results.
1. It is possible to record MMG signals using silicon-embedded sensors without ap-
preciable loss of SNR.
2. Each embedded sensor performance is dependent on the geometry and silicon hard-
ness of the embedding.
The demonstration of the feasibility to record MMG signals with silicon-embedded
sensors is a required step towards the development of MMG-based control systems for
soft silicon socket powered prosthesis.

2.2 Systematic Characterisation of Silicon-Embedded
Accelerometers for Mechanomyography
c
2003 IFMBE. Reprinted, with permission, from Silva J., Chau T., Naumann S., and
Heim W., “Systematic Characterization of Silicon-Embedded Accelerometers for Mechano-
myography,” Medical & Biological Engineering & Computing, Vol. 41, No. 3, pp. 290-
295, May 2003.
2.2.1 Abstract
Silicon soft suction sockets (roll-on sleeves) currently used in passive prostheses for below-
elbow amputees could also be used in externally powered prostheses, enhancing their
functionality and comfort. However, as it is extremely difficult to hold currently used
electromyography (EMG) sensors in place reliably within a silicon socket, an alternative
measurement of muscular activity as the control input is necessary. Mechanomyography
(MMG) is the epidermal measurement of the low-frequency vibrations produced by a
contracting muscle. MMG sensors do not have to be in direct contact with the skin.
Moreover, the embedding of sensors in the roll-on sleeve may also solve attachment is-
sues, making sensor placement flexible. Therefore the objective was to determine the
feasibility of recording MMG signals using silicon-embedded, micro-machined accelerom-
eters. Fifteen embedded accelerometers were excited with predefined vibration patterns.
The signal-to-noise ratio (SNR) and frequency response of each sample were measured
and compared with those of non-embedded accelerometers. The SNR of embedded sam-
=19 dB) was significantly higher than that of non-embedded samples (∼

ples (∼ =12 dB),
owing to the considerable mechanical damping effect of the silicon in the 300-900 Hz
bandwidth (p=0.0028). This has implications for the application of silicon-embedded
accelerometers for externally powered prosthesis control.

2.2.2 Introduction
Soft as Against Rigid Sockets
A schematic diagram of the constitutive parts of an externally powered below-elbow pros-
thesis is shown in Fig. 2.3. The socket is the part of the prosthesis in direct contact
with the patients residual limb. Therefore its shape and size depend on the particular
characteristics of the amputees stump. Current, externally powered prostheses are dis-
pensed with rigid supercondylar sockets encompassing the elbow condyles to suspend
the total weight of the prosthesis. An example is shown in Fig. 2.4a. EMG sensors
are fastened flush in the socket wall and positioned precisely over the muscle bulk. A
hard shell, which holds the terminal element and acts as the actual forearm, covers the
hard socket. Advances in socket design have led to the creation of mouldable silicon
suction sockets or roll-on silicon sleeves, as shown in Fig. 2.4b. These silicon sleeves
provide significant advantages over rigid supercondylar sockets. Comfort is improved,
because the total weight of the prosthesis is distributed along the residual limb instead
of resting exclusively over the elbow condyles. Functionality is also enhanced, because
roll-on sleeves preserve natural pronation and supination of the forearm in below-elbow
amputees. However, silicon sleeves are not currently prescribed with externally powered
prostheses, as EMG sensors cannot be reliably held in place over the muscle bulk [16].
Research into alternative channels for prosthesis control may allow silicon sleeves to be
used in place of hard sockets.
Mechanomyography
Mechanomyography (MMG) is the epidermal measurement of the low-frequency vibra-
tions (‘muscular sound’) produced by contracting muscles owing to a non-linear summa-
tion effect of propagated motor unit twitches and shape changes in muscular fibres during
voluntary contraction [24]. The model of a vibrating string has been consistently used
Figure 2.3: Schematic diagram of externally powered below-elbow prosthesis. Currently,
EMG sensors are fastened in a hard socket wall. The hard shell that holds terminal
element is adjusted over the hard socket
to describe the resonant vibration caused by the effect just described. This model fits
well with experimental observations, such as the half sine-like mode of vibration found
by Ouamer et al. [26, 38]. The magnitude of the MMG signal is linearly related to
muscle strength in non-fatiguing contractions that are between 20 and 80% of the maxi-
mum voluntary contraction (MVC) [25, 23, 37]. Most of the power of the MMG signal is
located in the 0-45 Hz bandwidth for forearm muscles. No significant components have
been found beyond 100 Hz [25, 23, 26, 17].
MMG and Prosthesis Control
MMG signals can be measured using microphones [25, 14, 35] or accelerometers [35, 37].
Some authors suggest microphones for MMG signal recording, as they are less susceptible
to motion artifacts. However, higher sensitivity values have been reported for accelerom-
eters [35]. Through use of multiple accelerometers, it may be possible to eliminate motion
a b
Figure 2.4: Comparison of (a) laminated hard socket and ( b) soft silicon socket or roll-on
sleeve. Rectangular openings in hard socket are intended for EMG sensor attachment.
Note that silicon socket has been rolled up to emphasise its flexibility
artifacts using simple differential recording techniques commonly applied to EMG mea-
surements [7, 2]. MMG signals recorded by a standard microphone have been used to
operate a free-standing prosthetic hand in a controlled, laboratory setting [13]. Using a
single-channel, full-wave rectified MMG signal and a tristate control strategy, Barry et al.
demonstrated that two below-elbow amputees could use intentional MMG to open and
close a prosthetic hand. The reported system successfully discriminated between MMG
signals arising from wrist flexion and extension and exhibited robustness to changes in
skin impedance and microphone placement. However, unresolved sensor attachment and
low frequency noise elimination issues limited its practical usefulness as an alternative
prosthesis control signal. Our objective was to determine whether MMG signals can be
recorded using silicon embedded transducers (microphones and accelerometers). This
investigation is a step towards the use of soft silicon suction sockets in externally pow-
ered prostheses. The experiments discussed here focus only on the characterisation of
silicon-embedded accelerometers.
2.2.3 Materials and Methods
To quantify the signal-to-noise ratio (SNR) and frequency responses of silicon-embedded
accelerometers, predefined vibration patterns were generated using a custom-made me-
chanical stimulator. These were applied to standardise excitation to the accelerometers.

For comparison, the same measurements were conducted using non-embedded accelerom-
eters as controls.
Preparation of Samples
Single-axis micro-machined accelerometers2 , with a mass of 0.28 g and dimensions of
5.59×2.28×8.56mm (width, height and length, respectively), were used for the tests.
Fig. 2.5 shows a schematic diagram of the samples prepared, indicating their embedding
2
BU-7135, Emkay Innovative Products
Figure 2.5: Schematic diagram of silicon-embedded accelerometer samples. TT L =top
layer thickness; TBL =bottom layer thickness
Table 2.4: Top-layer (TT L ) and bottomlayer (TBL ) thickness values of samples prepared
Sample TT L [mm] TBL [mm]
1, 6, 11 3±0.5 1±0.5
2, 7, 12 2±0.5 1±0.5
3, 8, 13 1±0.5 1±0.5
4, 9, 14 2±0.5 2±0.5
5, 10, 15 2±0.5 3±0.5
parameters, namely, silicon hardness, contact surface, top layer thickness and bottom-
layer thickness. Three groups of five accelerometers each were embedded into silicon
squares with different hardness values (shores 20A, 35A and 65A) for a total of 15 pre-
pared samples. These were numbered 1-15. Samples 1-5 were embedded using the hardest
silicon type (shore 65A), samples 6-10 used the medium silicon type (shore 35A), and
samples 11-15 used the softest silicon type (shore 20A). Table 2.4 shows the numerical
values of the top-layer thickness TT L and bottom-layer thickness TBL of each sample (see
Fig. 2.5).
Figure 2.6: Block diagram of instrumented system for stimulating and recording response
from sensors
Hardware Set-Up
Fig. 2.6 shows a block diagram of the system used for the experiments. A 16-bit PC
soundcard3 was used simultaneously to generate the stimulation pattern and record the
sensor response. Stimulus and response signals were amplified, respectively, by gains of
G1 = 20 and G2 = 400 using LM 386 audio amplifiers. The mechanical stimulator4 ,
shown in Fig. 2.7, was a card-type speaker, mounted in a wooden resonance box to
ensure good, low-frequency signal reproduction. As we were concerned with relative
responses, i.e. the response of the embedded accelerometers relative to the non-embedded
accelerometers, the typical requirement for a stimulator with a flat frequency response
over the frequencies of interest was not necessary. The speaker has the advantage of
being a nonuniform vibrating surface in which the maximum amplitude is at the centre
of the speaker. This non-uniform response corresponds to previous reports of real MMG
measurements, in which the muscle was described as a vibrating string with a half-sine-
like mode of vibration [26].
3
SB Live!, Creative Labs
4
WM-R57A, Panasonic
Figure 2.7: (a) Top view and (b) schematic diagram of mechanical stimulator built for
measurements. Placement of embedded samples is shown in schematic diagram. All
dimensions are in millimetres

Generation of Vibration Patterns
Two different patterns were electronically generated for SNR determination using the
MATLAB environment5 . The ‘signal’ was a summation of scaled in-phase sinusoidal

signals from 1 to 100 Hz (ten frequencies per decade, in equally spaced intervals). Once
amplified and sent to the stimulator, the magnitude of the peak frequency (20 Hz) resem-
bled that of real MMG signals at approximately 75% of MVC (0.3 ms−2 ). This estimate
was obtained by linearly extrapolating from previously reported measurements, where an
MMG signal amplitude of approximately 0.08 ms−2 corresponded to 20% of MVC [35].
The ‘noise’ consisted of band-limited white noise, the maximum magnitude of which was
15% of the peak ‘signal magnitude.
Test Methodology
We tested different embedding properties to determine if there were significant trends
in sensor performance. For SNR quantification, two different stimuli were separately
applied to each sample, one being a pure noise signal and the other a signal with added
noise. The measurement of the signal with added noise presented a more accurate way to
obtain the SNR value (rather than measuring the signal by itself), as it was impossible
completely to eliminate the background noise during measurements. SNR values were
calculated as follows:
2
RMSSIGN AL+N OISE
SNR = 10 log −1 (2.2)
RMSN OISE
The subtraction of 1 is a correction factor that allowed us to obtain the actual value of
SNR for the worst case in which the noise is completely additive (or in phase with the
signal). Fig. 2.7 shows the vector representing the static forces that were carefully ap-
plied directly over each sample, emulating the force of the skin against the sensor within
an externally powered prosthesis. Owing to the nature of the applied force and the ge-
5
MathWorks Inc.
ometrical differences among samples, the exact amount of force could not be precisely
determined. However, it was possible to ensure that all forces were in the 0.2-0.5 N range.
Fig. 2.8 shows the actual mechanical coupling of the samples to the vibrating surface. A
weight was placed over the supporting ring. This applied an evenly distributed force on
the foam along the ring’s perimeter. In turn, this force compressed the foam, generating
a distributed position-dependent force over the sample. A similar effect would be present
within a silicon sleeve in externally powered prostheses. Four values of increasing force
(F1 -F4 ) were produced by gradually incrementing the weight on top of the supporting
ring by a constant amount. These forces were deemed to be comparable with the real
forces experienced by sensors within a soft silicon socket, pressing against the skin. Three
sets of recordings (including force variations) were acquired using different contact sur-
face values, to evaluate their effects on the SNR. The three surfaces tested were 32×42
mm2 , 22×32 mm2 and 12×22 mm2 (width and length, respectively). The independent
variables tested were thus silicon hardness, the thickness of both the top and bottom
layers of embedding, the contact surface and the applied force. For frequency response
determination, the samples were systematically excited with scaled sinusoidal signals
ranging from 1 to 4000 Hz (ten frequencies per decade in equally spaced intervals), with
magnitudes close to 0.3 ms−2 . The lower limit of this range (1 Hz) is defined by the low-
est frequency signal that can be reproduced by the stimulating system. The upper limit
is defined by the maximum frequency that can be reliably measured (approximately 11
points per cycle), given the maximum sampling frequency of the soundcard. Root mean
square (RMS) and phase values of signals obtained from the sensors were automatically
calculated and stored for further analysis.
Test Validation
Ten repeated measurements for each of the 15 samples with each contact surface and each
force applied were recorded. The Lilliefors test for normality [9] was applied to the set of
a b
Figure 2.8: Mechanical coupling of tested samples to vibrating surface of stimulator. (a)
Top view. (b) Front view
SNR calculations of each sample. As some data did not follow a normal distribution, we
used the Wilcoxon-Mann-Whitney rank sum one-sided test to determine whether or not
the SNR values obtained for a certain combination of variables were significantly higher
than those of the non-embedded sensors. To test the hypothesis of independence be-
tween the SNR and the other variables, the data were first organised into sensor groups.
A sensor group is defined as a collection of samples that differ in only one variable (sil-
icon type, top layer thickness, bottom layer thickness, surface area or applied force).
Subsequently, we constructed a contingency table with the observed frequencies of in-
creasing, decreasing or nonmonotonic trends in SNR. A Monte-Carlo simulation with 106
groups of random numbers provided the expected frequencies under the null hypothesis
of no monotonic relationships between SNR and the independent variables. A χ2 test
for goodness-of-fit was used to test this null hypothesis. Finally, an adjusted residuals
matrix [3] was computed to determine which types of trend were significantly different
from the expected values. For all statistical tests, a 5% significance level (α=0.05) was
used. For the frequency behaviour determination, ten repeated measurements were taken
for each frequency tested. The relative response in both magnitude and phase was cal-
Table 2.5: SNR values (dB) measured in embedded and non-embedded transducers.
Increasing applied forces are labelled F1 -F4
Force
Samples F1 F2 F3 F4
Non-Embedded 12.42±0.03 10.54±0.13 11.73±0.02 9.21±0.03
Embedded 19.86±3.91 19.18±3.98 19.07±3.74 18.70±4.31
culated. The reference in this case (0 dB and 0 degrees, respectively) was always the
non-embedded accelerometer response. This method ensured that differences observed
were due strictly to the embedding and not to the combined effects of the amplifier, the
speaker or the soundcard itself.
2.2.4 Results
Table 2.5 shows the mean values and standard deviations of SNR obtained for embedded
and non-embedded accelerometers for each force applied. Forces are labelled as F1 -F4
identifying increasing forces (increasing weights on top of the supporting ring, shown in
Fig. 2.8). Note that the mean SNR values of embedded accelerometers were significantly
higher than those of their non-embedded counterparts ( p < 0.01 for each group from F1
to F4 , and p=0.0028 for the whole set). With the exception of a few samples (mainly the
medium silicon hardness samples), this behaviour was consistently repeated also within
sensor groups.
From the Monte-Carlo simulation, we obtained values of 0.16, 0.16 and 0.66 for the
probabilities of increasing, decreasing and non-monotonic trends, respectively. From the
χ2 test, we determined that the SNR is dependent on all of the variables tested. Finally,
from residual analysis, the significant trends in the above dependencies were as follows.
The SNR increased as

(i) the bottom layer thickness increased
(ii) the top layer thickness increased
(iii) the contact surface area decreased
(iv) the hardness of the silicon decreased
No significant trend was found for variations in applied force. Fig. 2.5 shows the
typical SNR values obtained for the most representative sensor groups for each variable
tested. These graphs clearly reflect the trends described above. Note the relatively high
R2 values for each regression. Fig. 2.10 shows a bode plot of one of the samples tested
(hardest silicon, medium contact surface, lowest applied force). Note that there were no
considerable phase shifts within the 5-100 Hz bandwidth. None of the samples showed a
considerable deviation from this behaviour.
Fig. 2.11 shows the general form of the relative magnitude response for silicon em-
bedded accelerometers. Compared with their non-embedded counterparts, the embedded
accelerometers behaved as a band-stop filter whose rejection band C fell an average of
20 dB in the 300-900 Hz bandwidth, with a ripple R1 from about 10 to 20 dB wide. The

low-pass band A extended from 5 to 100 Hz. This is the bandwidth of interest for the
proposed application. Note that, here, the response is relatively flat.
There were no significant changes in the frequency response due to changes in the
applied force. However, the silicon hardness and contact surface of embedding did con-
siderably affect the relative magnitude response of the sample in this region. A peak
occurred in region B at 200 Hz. The high-pass band D began at 1000 Hz and presented
a ripple R2 of up to 20 dB wide. In most cases, within this band D, a peak occurred
at 1000 or 2000 Hz, rising about 5 dB above the low-pass band level A. The broken
lines in the low-pass band represent the maximum deviation from the general form of the
response. Deviation in section A was attributed to the medium silicon hardness (shore
Figure 2.9: Typical results obtained from most representative sensor groups for each
variable tested. (a) Silicon hardness test; R2 =0.909. ( b) Contact surface test; R2 =0.978.
(c) Top layer thickness test; R2 =0.849. (d) Bottom layer thickness test; R2 =0.808. Each
graph shows linear regression.

Figure 2.10: Sample bode plot of one of samples tested. Reference for each frequency (0
dB, 0 degrees) is non-embedded transducer response
Figure 2.11: General form of embedded accelerometers relative magnitude. A: Low-
pass band. B: Peak frequency band. C: Reject band. D: High-pass band. R1 =reject
band ripple. R2 =highpass band ripple. Dashed line indicates maximum deviation from
general form. In region A, this deviation was due to silicon hardness and contact surface
variations
Figure 2.12: Example of actual MMG signal recorded using silicon-embedded accelerom-
eter. Vertical axis is in units of gravitational acceleration g
35A). Deviation in section B was elicited by changes in the contact surface such that, as
the contact surface increased, the peak at 200 Hz tended to disappear.
As an aside, Fig. 2.12 depicts a real MMG signal recorded from the wrist flexors using
a silicon-embedded accelerometer. This verifies that silicon-embedded accelerometers can

indeed be used to record real MMG signals.
2.2.5 Discussion
It is possible that some of the variations in the measured responses were influenced by the
mechanical characteristics of the system. This is especially true for the contact surface
tests and, possibly, for the thickness tests (of both top and bottom layers). Changes in
the mass of the complete system were inherent to both tests. Increments in mass added
static forces to the system. In some samples, especially the ones with the highest mass
values, these forces were comparable with the externally applied forces tested. Hence,
rather than gauging relationships between the SNR and various independent variables,
a more precise observation might be that the SNR of embedded samples increased with
increasing, but concentrated, mass and applied force. This statement is consistent with
the nature of the mechanical stimulator. As the stimulator was based on a vibrating
membrane mechanically grounded along its perimeter, the magnitude of the vibrations
would be higher towards the centre of the speaker. As mentioned earlier, this behaviour
simulates previously reported longitudinal amplitude losses in real MMG measurements.
Nonetheless, in the future, it is necessary to confirm the observed sensor performance
in real MMG measurements. Recall that MMG signals have no significant frequency
components beyond 100 Hz. Therefore frequency components occurring in the C and D
bands in Fig. 2.11 can be considered noise. The mechanical properties of silicon naturally
provide considerable noise damping within the C band. Further noise suppression within
the C and D bands can be achieved using either hardware or software filtering.
2.2.6 Conclusions
From the results we can state that
(a) The SNR of embedded accelerometers is consistently higher than that of their non-
embedded counterparts making the former suitable for MMG signal recording.
(b) The reject band in the frequency response of embedded samples is wide enough to
eliminate higher-frequency noises.
(c) The softest or the hardest silicon types (shores 20A and 65A) are recommended for
MMG measurements, as they exhibited the least variability during the tests.
(d) Although there seem to be clear relationships between sensor performance and the
variables tested, experimentation with real MMG recording is required to further
validate these findings.

The reported findings do show however, that it is feasible to embed accelerometers in

a soft silicon suction socket for MMG signal acquisition for externally powered prosthesis
control.
Chapter 3
Silicon-Embedded Microphones
Continuing with the task of characterising embedded sensors for MMG signal recording,
the present chapter deals with the evaluation of silicon-embedded microphones. The
same tests that were used with silicon-embedded accelerometers in the previous chapter
are employed here. A published paper, which presents the main results from this stage of
research, is reprinted with permission in the following section. Special attention should
be paid to the innovative manufacturing process, and to the increased complexity of
coupling silicon-embedded microphones with the signal source.
Suggested sections to read with relevant new content are 3.1.1, 3.1.5, 3.1.6 and 3.1.7.
3.1 Optimization of the Signal-to-Noise Ratio of Silicon-
Embedded Microphones for Mechanomyography
c
2003 IEEE. Reprinted, with permission, from Silva J., Chau T., Naumann S., Heim
W., and Goldenberg A., “Optimization of the Signal-to-Noise Ratio of Silicon-Embedded
Microphones for Mechanomyography,” in Proceedings of the Canadian Conference on
Electrical and Computer Engineering, Montreal QB, Canada, Vol. 3, pp. 1493-1496,
May 2003.
33
Chapter 3. Silicon-Embedded Microphones 34
3.1.1 Abstract
Soft silicon sockets enhance mobility and comfort of artificial limbs for below-elbow
amputees compared to acrylic resin laminated hard sockets. However, standard elec-
tromyographic (EMG) sensor attachment issues prevent their use with electrically pow-
ered prostheses. Mechanomyography (MMG) may facilitate the use of soft silicon sockets
with electrically powered prostheses. Unlike EMG sensors, MMG sensors can be reliably
placed distally over the patients stump, by embedding them into the silicon socket. Our
experiments focused on maximizing the signal-to-noise ratio (SNR) of silicon-embedded
microphones for MMG. Optimal dimensions were obtained for the cylindrical air chamber
enclosed within the embedding capsule (13 mm diameter, 2 mm height). Furthermore,
shores 20A and 65A were identified as the optimal silicon hardness values for the em-
bedding case and driving membrane, respectively. Embedded microphones provided, on

average, a 6 dB SNR enhancement over previously reported embedded accelerometers.
3.1.2 Introduction
Several advantages have been consistently reported for soft silicon sockets or roll-on
sleeves, over acrylic resin laminated hard sockets for prostheses for below-elbow amputees.
With soft sockets, suspension is improved as the weight of the prosthesis is distributed
along the residual limb instead of resting exclusively over the elbow condyles [19]. Mobil-
ity of the proximal joint is also improved since lower trim lines are needed for suspension
[27]. The benefits for the patient arise from the enhanced comfort and functionality of
the system. However, wire breakage and the lack of reliability in the attachment and
coupling of electromyography (EMG) sensors, prevent the use of soft silicon sockets in
electrically powered prostheses for below-elbow amputees [27, 16]. Therefore, alternative
techniques to measure muscular activity at the distal end of the stump are necessary in
order to facilitate the combination of roll-on sleeves with electrically powered prostheses.
3.1.3 Mechanomyography
When a muscle contracts, mechanical oscillations can be detected superficially [23, 25, 14].
The transduced signal is termed the surface mechanomyogram (MMG) [23]. Although
the physiological origin of the MMG is not completely understood, three main causes
have been related to this phenomenon: (i) a summation of motor unit twitches causing
a sudden and high amplitude vibration (compared to the rest of the signal) at the onset
of the contraction [13]; (ii) shape changes and non-linear summation of vibrating motor
units causing a resonant vibration which depends on the size, shape and mass of the
muscle [24]; and (iii) the lack of uniformity in the tissue causing bending waves which
add acoustic energy to the lateral vibrations. The overall effect can be modelled as a
vibrating spherical body (i.e. the muscle vibrates as a whole) with the most significant
frequency components around the mechanical resonant frequency of the muscle [24, 12]
and a predominant half sine-like mode of vibration which suggests that the amplitude of
MMG signals is largest at intermediate muscle lengths [12, 26]. Most of the power of the
MMG signal falls within the 0-30Hz bandwidth and no significant frequency components
beyond 100Hz have been reported for arm muscles [23, 25, 38, 17].
3.1.4 MMG Sensors & Prosthesis Control
As with EMG signals, there is a linear relationship between the root mean square (RMS)
values of MMG signals and the force of the contraction within 20% to 80% of maximum
voluntary contraction (MVC) [25, 14]. This is one of the main characteristics that make
MMG especially suitable for prosthesis control. Barry et al. (1986) used a simple MMG-
based control strategy to allow below-elbow amputees to open and close a prosthetic
hand. However, basic sensor attachment issues precluded its practical use [13]. MMG
signals can be acquired using microphones [14, 17, 35] or accelerometers [24, 26, 35, 31].
Unlike their EMG counterparts, MMG sensors do not have to be in direct contact with
the skin. The low amplitude displacements (≈500 nm) evoked at the surface of the
skin can be amplified and transduced to acoustic pressure changes within a sealed air
chamber through the use of an elastic membrane. A microphone can then be used to
measure these pressure changes (acoustic pressure) within the air chamber [14]. Our
previous studies have evaluated the performance of silicon-embedded accelerometers for
MMG signal recording in terms of their signal-to-noise ratio (SNR) [31]. The main
purpose of this study is the evaluation and optimization of the SNR of silicon-embedded
microphones for MMG signal recording.
3.1.5 Materials & Methods
Sensor Design
Fig. 3.1 shows the schematic diagram of the transducer built for our experiments. An
electret microphone model EK3169 from EmKay Innovative Products was soldered to a
printed circuit board (PCB) with impedance coupling and amplification circuitry. The
initial embedding stage (casing) consisted of embedding the soldered microphone within a
silicon case. A plastic cylinder was attached to the microphone during the casing process
to prevent silicon from sweeping into the microphones case and to preserve the shape of
the air chamber during silicon curing. The second stage (sealing), consisted of adding a
thin silicon layer (<1 mm thick) at the bottom of the cured case after the plastic cylinder
had been removed. This constituted the driving membrane. A small bump was added to
the driving membrane to ensure good contact between the transducer and the vibrating
source.
Maximization of the SNR for the design shown in Fig. 3.1, allowed us to obtain opti-
mum values of height (H) and diameter (D) of the cylindrical air chamber, and optimum
silicon hardness for both the case and driving membrane of the transducer. A total of
20 combinations of different heights and diameters for the air chamber were tested main-
Figure 3.1: Schematic diagram of the MMG sensor design used in experiments. H stands
for height and D for diameter of the enclosed cylindrical air chamber.
taining constant silicon hardness (shore 20A for both case and driving membrane). Three
additional measurements were performed using the optimum air chamber dimensions (in
terms of its SNR) to test for the effect of different silicon hardness values for embedding
on the SNR of the transducer (i.e. shores 20A-65A, 65A-20A and 65A-65A for case and
driving membrane, respectively).
SNR Measurements
In order to standardize the tests, an artificially generated signal and white noise were
used to perform the SNR measurements using a custom-made mechanical stimulator. The
exact methodology used in the generation of the signal and construction of the mechanical
stimulator were detailed by Silva et al. (2002) [31]. The signal, a sum of in-phase scaled
sinusoidals, had a similar magnitude spectrum as an MMG signal from the wrist flexors
at 10% of MVC. White noise at 15% of the signal peak frequency amplitude (20Hz)
was used as additive noise. Ten repeated measurements of SNR were performed and
Kruskal-Wallis tests for equality of medians (p < 0.01) were used to test for significant
differences among the SNR values obtained for each combination of embedding properties
(air chamber geometry and silicon hardness).

Frequency Response
Additional frequency response tests were performed using sinusoidal signals from 1Hz to
4kHz to stimulate the transducers. Ten repeated measurements of the RMS value and
phase delay obtained for each stimulation frequency were acquired. These values were
used to determine the frequency response (magnitude and phase) of the transducers.
3.1.6 Results
Air Chamber Geometry
Table 3.1 shows the mean SNR values and standard deviations obtained for each combi-
nation of embedding properties. Note that, due to technical difficulties, the SNR values
for some of the combinations were extrapolated through the use of a weighted 2-D average
filter (the nearest values have the highest contributions). The combinations of diameter
and height which yield the highest SNR values are, in decreasing order, 7 & 8 mm, 10
& 6 mm, 10 & 4 mm and 13 & 2 mm, respectively. However, the Kruskal-Wallis tests
showed no significant differences among them. Table 3.2 shows the normalized sensitivity
values (with respect to the highest sensitivity value obtained) and standard deviations
calculated from the mean RMS values of the signals acquired with each sensor. It can be
seen from Table 3.2 that the 10 & 2 mm combination (diameter & height, respectively)
exhibited the highest sensitivity. It can also be noticed that sensitivity values do present
significant differences. The values in Table 3.2 were used to weight the SNR values in
Table 3.1. Therefore, the optimum air chamber geometry (maximum SNR value) results
from a combination of its SNR value and its sensitivity as a measure of performance. Fig.
3.2 depicts an interpolated surface (by cubic interpolation) representing the sensitivity-
weighted and normalized SNR values as a function of air chamber diameter and height.
The 13 & 2 mm combination (diameter and height, respectively) had the highest SNR
value and was therefore deemed the optimum for MMG signal recording.
Table 3.1: SNR values obtained for each combination of embedding properties tested
Diam. [mm] Height [mm]

2 4 6 8 10
7 21.0∗ 23.8±1.0 5.0±0.8 25.2±0.4 21.5∗
10 18.2±2.7 25.0±0.5 25.1±0.9 19.7±0.9 21.2∗
13 24.2±0.9 22.5±3.6 23.8±1.6 22.2±4.5 21.7±4.2
16 23.1±0.8 20.1±1.7 23.0±0.4 17.5±0.5 21.1∗
∗
These values were extrapolated
Table 3.2: Normalized sensitivity values obtained for each combination of embedding
properties tested
Diam. [mm] Height [mm]
2 4 6 8 10
7 0.54∗ 0.28±0.02 0.11±0.004 0.47±0.007 0.42∗
10 1.00±0.07 0.57±0.02 0.55±0.03 0.29±0.02 0.43∗
13 0.89±0.01 0.36±0.02 0.82±0.03 0.46±0.05 0.29±0.05
16 0.75±0.02 0.26±0.06 0.40±0.02 0.18±0.003 0.40∗
∗
These values were extrapolated
Figure 3.2: Normalized SNR surface obtained from interpolation of sensitivity masked
(Table 3.2) and normalized SNR values (Table 3.1) as a function of air chamber height
and diameter.
Silicon Hardness
Using the optimal air chamber dimensions from the previous analysis (13 mm diameter, 2
mm height), we tested for significant differences in the frequency response of the tranducer
due to changes in the silicon hardness of the embedding. Fig. 3.3 shows the comparison
of the magnitude spectrum for the three combinations of silicon hardness tested in terms
of their relative sensitivity. The reference (0 dB) is the highest mean RMS value obtained
over every sensor and every frequency. Note that the combination SH (i.e. soft silicone for
the case and hard silicone for the membrane), is the most stable and provides the greatest
difference between the region of interest (5-100 Hz) and the rest of the frequencies which
can be considered noise for the current application.
Microphone vs. Accelerometer
Data from previously published tests with silicon-embedded accelerometers done at our
laboratory were compared with the data obtained on the experiments reported here.
Figure 3.3: Magnitude response of silicon-embedded microphones with optimal air cham-
ber dimensions and different silicon hardness values. The legends of individual graphs
are labeled with two letters, the first being the hardness of the silicon case and the sec-
ond being the hardness of the driving membrane. S stands for “soft” (shore 20A) and H
stands for “hard” (shore 65A).
After some adjustments to compensate for differences in the amplitude of the stimuli
between microphones and accelerometers and the difference in the relative amplitude of
the background noise with respect to the driving signals, the mean overall SNR values
obtained were 22.69±2.97 and 16.53±3.83 dB for microphones and accelerometers, re-
spectively. This suggests that for MMG signal recording, silicon-embedded microphones
provide, on average, a 6dB enhancement in SNR over their accelerometric counterparts.
3.1.7 Conclusion
The results reported here demonstrate that silicon-embedded microphones can be used
effectively for MMG signal recording. Furthermore, when compared to silicon-embedded
accelerometers, microphones provide enhanced SNR performance. The results of this
study bring us one step closer to implementing a new and enhanced generation of elec-
trically powered prostheses using soft silicon sockets.

Chapter 4
Coupled MMG Sensor Pair
Having completed characterizations of silicon-embedded accelerometers and microphones,
we now turn our attention to the design of an optimum sensor for MMG signal recording
within the silicon-sleeve of an electrically powered prosthesis. Taking advantage of the

difference in the mechanical impedance in the coupling between the MMG signal source
and each of the single-sensor designs (i.e. silicon-embedded accelerometers and silicon-
embedded sensors), a dual sensor design for simple dynamic noise reduction is proposed.
This chapter introduces the coupled sensor, its basic properties, manufacturing processes
and mechanical model. The next chapter will deal with the complexities of real-time
algorithms for fusing the data obtained from these coupled sensors in order to generate
reliable control outputs. The first section of the present chapter is reproduced, with
permission, from a published paper.
Suggested sections to read with relevant new content are 4.1.1, 4.1.3, 4.1.4, 4.1.5, 4.1.6
and 4.2.
42
Chapter 4. Coupled MMG Sensor Pair 43
4.1 Coupled Microphone-Accelerometer Sensor Pair
for Dynamic Noise Reduction in MMG Signal
Recording
c
2003 IEE. Reprinted, with permission, from Silva J. and Chau T., “Coupled Microphone-
Accelerometer Sensor Pair for Dynamic Noise Reduction in MMG Signal Recording,”Electronics
Letters, Vol. 39, No. 21, pp. 1496-1498, 2003.
4.1.1 Abstract
External noise sources such as movement artefact may interfere with Mechanomyography
(MMG) signal recording, compromising signal detection for electrically-powered prosthe-
sis control. A novel coupled microphone-accelerometer sensor pair, intended to facilitate
noise reduction, was designed and tested on a below-elbow amputee. Sample recordings
demonstrate clear distinction between motion artefact and useful MMG signals.
4.1.2 Introduction
Mechanomyography (MMG) is the superficial measurement of the axial vibrations elicited
by contracting muscles [23]. Orizio et al. reported a linear relationship between the root-
mean-square (RMS) values of MMG signals recorded from the biceps brachii and the
force of the contraction between 20% to 80% of maximum voluntary contraction (MVC)
[25]. Therefore, MMG may substitute current electromyography (EMG) measurements
in electrically powered prosthesis control. Barry et al. have already used MMG sig-
nals acquired from the wrist flexors and extensors of below-elbow amputees to control a
free-standing prosthetic hand in a controlled laboratory setting [13]. However, external
mechanical noise sources such as movement artefact are known to cause considerable
interference, compromising signal detection and classification. Therefore, a reliable and

robust MMG sensor is a prerequisite to the practical use of MMG-driven electrically

powered prostheses. MMG signals have been measured with accelerometers [32, 35] and
microphones [35, 33, 14]. Considerably lower signal-to-noise ratios (SNR) have been
reported for accelerometers [32, 33]. This difference in noise sensitivity between mi-
crophones and accelerometers can be exploited in a coupled instrumentation set-up to
extricate information-bearing signal from noise.
4.1.3 Sensor Design
Fig. 4.1 shows a schematic diagram of the sensor design. It consisted of two micro-
machined, coupled vibration transducers (microphone1 - accelerometer2 pair) surface mounted
on a printed circuit board (PCB) and enclosed in a silicon case. Both sensors were aligned
to the direction of muscle vibration. Skin vibrations originating from contracting muscles
caused a proportional displacement of the silicon contact membrane. These vibrations
then modulated the acoustic pressure measured by the microphone inside the air cham-
ber [14]. The silicon acted as a passive low-pass filter that helped to increase the SNR
of the measurement [32]. The approximate surface area of the device was 1.9 × 1.9 cm
and 1 cm in height. Based on previous studies, a cylindrical air chamber (1.3 and 0.2 cm
in diameter and height, respectively) and silicon shores 20A and 35A for the silicon case
and membrane, respectively, were chosen for the design [33]. Note that the accelerometer
was placed behind the microphone. This caused a further decrease in the SNR of the
accelerometer since the silicon membrane and air chamber dampened any mechanical
vibration arriving from the microphone side. At the same time, the accelerometer was
capable of recording the direct effects of forces acting on the forearm as a whole, (e. g.
inertia caused by limb movement). In other words, there was a desirable mechanical
impedance mismatch between both transducers for signals arriving from the microphone
1
Omnidirectional electret condenser microphone MD6022ASC-0 from Emkay Innovative Products
2
Single axis accelerometer model BU7135 from Emkay Innovative Products
Figure 4.1: Schematic diagram of MMG sensor pair, placed against the skin.
side, while both transducers were sensitive to signals originating from external forces.
Therefore, the accelerometer can be seen as a dynamic reference sensor particularly sen-
sitive to noise.
4.1.4 Data Acquisition
Recordings were taken from a client with a traumatic below-elbow amputation. Three
sensor pairs (fig. 4.1) were placed around the distal end of the residual limb, separated
by equidistant angles (1200) at approximately 1.5 cm from the distal end of the stump.
An elastic fabric bracelet was used to hold the sensors in place. The client was asked to
perform one of three different activities namely: (i) extension; (ii) flexion; or (iii) random
limb movement during a two-second period defined as the activity period. The extension
and flexion activities corresponded to the typical contractions usually performed by the
client to open and close, respectively, her current EMG-based prosthetic hand. Signals
from the accelerometers were amplified (gain=240) using general-purpose operational
amplifiers. Audible alarms indicated the start and end times of the activity period.
Five recordings for each activity and sensor pair were acquired using the MATLAB
environment, and stored for further comparison.

4.1.5 Results
Fig. 4.2 depicts sample recordings of rectified extension and random limb movement
signals for each transducer in a single sensor pair. During extension, the amplitude of
the microphone signal is directly proportional to the strength of the contraction. How-
ever, it can be readily observed that, without the additional information provided by
the accelerometer (bottom pair of graphs), the correct determination of the nature of
the signal, in this example, extension versus limb movement, would not be possible. In
other words, with the microphone alone, we would not be able to discriminate between
limb movement and useful MMG signals. Fig. 4.3 further emphasizes the importance
of the accelerometer with a graph of mean RMS values of the accelerometer signals over
all sensors for each activity. In contrast to the microphone, the accelerometer’s mean
RMS values during flexion (0.0567 ± 0.0123 V), extension (0.0584 ± 0.0101 V) and rest
(0.0568 ± 0.0107 V), were not significantly different as determined by the Kruskal-Wallis
test (p = 0.59). This verified the assumption that the accelerometer would have dimin-
ished sensitivity to signals arriving from the microphone side (i.e. muscle contraction).
Note however, the statistically significant increase (p = 3 × 10−6 ) in the mean RMS value
of the accelerometer signal (0.2184 ± 0.0720 V) during random limb movement.
Putting together the above results, a straightforward detection technique could simply
use the RMS value of the accelerometer signal as a dynamic threshold for the microphone
signal. Specifically, when the RMS value of the accelerometer signal is low, the micro-
phone signal is admissible for prosthesis control. On the other hand, a high RMS value
in the accelerometer signal indicates the presence of motion artefact and the microphone
signal should not be used directly for prosthesis control. We remark, however, that the
relationship between the microphone and accelerometer signals is not trivial. Further
studies are needed to determine the exact mathematical relationship between the trans-
ducers.
Figure 4.2: Sample measurements of extension (left) and random limb movement (right).
Note the considerable differences in the accelerometer signal (bottom pair of graphs).
Figure 4.3: Comparison of the mean RMS values and standard deviations among different
activities for the accelerometers. Note the increased RMS value in the accelerometer
during random limb movement.
4.1.6 Conclusions
The compact size of and complementary information provided by the coupled transduc-
ers may facilitate the practical use of MMG signals in prosthesis control. The sensor
pair may also be valuable in other electronics applications where the contamination of an
acoustic signal by ambient vibration confounds detection but cannot be easily removed
(e. g. similar power density spectrum and random phases for both signal and noise). The
straightforward detection technique suggested in this document exemplifies the powerful
potential of simple, hardware-based sensor fusion.

4.2 Coupled MMG Sensor Pair: Model Validation
and Optimization
4.2.1 Rationale: Model-Based Source Separation
Based on the information reported in the previous section, two interfering sources in the
microphone signal may be identified when using the proposed coupled sensor pair for
MMG signal recording: (1) the useful skin vibration due to muscle contraction, and, (2)
unpredictable external forces produced by limb movement, limb vibration (e.g. activation
of the terminal element motor) or sudden tapping, all of which induce relative movement
of the silicone membrane with respect to the sensor case, causing an undesirable pres-
sure change in the microphone’s air chamber [29]. It has already been established that
the information obtained from the transducers (microphone and accelerometer) signals
within a single sensor pair differs significantly for each of these sources. For instance, it
can be readily noticed that the accelerometer is only capable of detecting the external
undesirable forces. Nothing yet has been said about the nature of this detection by the
accelerometer or the relationship between the sources and the microphone signal. If the
exact mathematical relationship or an adequate approximation thereof between the in-
terfering sources and the signals acquired from the coupled MMG sensor pair are known,
transfer functions from the aforementioned sources (inputs) to the sensor signals (out-
puts) may be obtained. If validated, this model will automatically use the magnitude and
phase information simultaneously to provide the best possible estimate of the outputs.
Furthermore, if a linear time-invariant (LTI) system is assumed, the inversion of these
transfer functions will reveal the electronic circuit implementation that provides the most
efficient and reliable real-time source separation strategy. Other source characteristics
that support this model-based approach are:
1. The power spectra of both sources overlap in the bandwidth of interest (5-50Hz).
Therefore, no simple filtering techniques can be applied to track skin vibrations
under extreme noise conditions.
2. The onset of muscle contraction and external forces (source activation) is by nature
unpredictable (random phase), therefore, the phase information of the acquired sig-
nals will not be useful for source separation unless the relative phase shift observed
among sensors can be compared to the real phase shift, if any. A mechanical model
based on the physical interactions of the sensor components will provide the best
estimate for such a phase shift.
It is important to note however, that the performance of the proposed model-based source
separation will be extemely dependent on the reliability of the model itself. Therefore, if
such an approach is taken, special attention must be paid to the validation and precise
justification of the parameters involved. The ensuing sections present the proposed model
for the coupled microphone-accelerometer MMG sensor pair, together with its validation
and optimization. The main objective of developing a model is the generation of a reliable
source separation algorithm based on the fusion of data from individual transducers
within a single sensor pair.
4.2.2 The Model
Based on the assumption of a linear time-invariant (LTI) system, a one-dimensional model
for the coupled MMG sensor pair is proposed. This model accounts for the dynamic inter-
actions of the mechanical variables involved in practical MMG signal recording. Three-
dimensional and radial effects of the measured sources (i.e. muscle displacement lM (t)
and external force Fe (t)) are dismissed on the assumption that sensor design maximizes
sensitivity in a single axis (i.e. perpendicular to the sensor’s silicone contact membrane).
Furthermore, the small bump at the bottom of the silicone membrane restricts the loca-
tion of the skin vibration source to the center of the air chamber. Therefore, point source
localization is also assumed.
Air Chamber Mechanics
Figure 4.4 shows a graphical representation of the mechanical variables involved in the
determination of the mathematical relationship between a displacement l(t) of the sensor
silicone membrane and the acoustic pressure ∆Pa (t) inside the air chamber as measured
by the microphone. If the air chamber case is assumed stiff, the displacement l(t) can be
seen as a varying position source which changes the height of the cone-shaped membrane
surface generated during vibration. If we consider l(t) = 0 as the membrane position
at which the pressure of the air chamber equals the atmospheric pressure (i.e. when
the silicone membrane is flat), positive and negative displacements l(t) will then corre-
spondingly cause compression and rarefaction of the microphone’s air chamber. In this
sense, for every given position value l(t) = x there is a corresponding constant pressure
value Pa (x) inside the air chamber. This pressure, together with the elasticity of the
membrane (as a tendency to recover its original form), cause a spring effect which acts
locally exerting a restoring force FT (t) at the center of the membrane and opposite to
the vibration source, that is:
FT (t) = Fa (t) + Fm (t) (4.1)
where Fa (t) is the restoring force due to the air chamber compression/rarefaction and
Fm (t) is the restoring force due to the silicone membrane elasticity.
The magnitude of the force Fa due to air compression/rarefraction is given by the
relationship [1]:
Fa (x) = Am (x)∆Pa (x) (4.2)
where Am (x) is the membrane surface area and ∆Pa (x) is the pressure change (acoustic
Figure 4.4: Schematic representation of the mechanical effect produced by the silicone
membrane vibration in the microphone’s air chamber. An ideal linear spring model may
be assumed.
pressure) inside the air chamber with respect to the equilibrum position (l(t) = 0). The
area Am (x) and pressure change ∆Pa (x) are functions of the given position value, l(t) = x.
The cone-shaped membrane surface area for any given value l(t) = x is approximately:
√
Am (x) = πr r 2 + x2 (4.3)
where r is the air chamber radius (r = 6.5mm). The change in pressure ∆Pa (x) can be
calculated from Boyle’s law which for an isothermal compression gives:
∆Pa (x) = Pa (x) − P0 (4.4)

P0 V0
= − P0 (4.5)
Va (x)
P0 V0
= − P0 (4.6)
V0 − VCON E (x)
1atm(πr 2 h)
= − 1atm (4.7)
πr 2 h − 13 πr 2 x
where P0 and V0 are the respective pressure and volume values at equilibrum (l(t) = 0),
Va (x) is the air chamber volume for any given relative membrane position l(t) = x,
VCON E (x) is the cone-shaped volume formed by the vertical displacement at the centre of
the silicone membrane, and h is the air chamber height (h = 2mm). Finally, substituting
equations 4.3 and 4.7 into 4.2, the magnitude of the reaction force Fa (x) becomes:
!
√ 1atm(πr 2 h)
Fa (x) = πr r 2 + x2 − 1atm (4.8)
πr 2 h − 13 πr 2 x
Aditionally, as mentioned before, the elasticity of the silicone membrane will also
contribute to the spring effect. The restoring radial tension exerted on the cone-shaped
membrane surface can be easily calculated if the Young’s modulus of the silicone is
known. However, only the vertical component (i.e. in the direction of the displacement
l(t)) of this tension should be considered since the horizontal component will be opposed
by the stiff walls of the air chamber along the whole perimeter. The reaction force due
to membrane deformation Fm (x) then becomes:

!
√ x
Fm (x) = πTm Ym x2 + r2 −r √ (4.9)
x + r2
2
Figure 4.5: Reaction forces due to membrane displacement calculated for the air chamber
(Fa ) and silicone membrane (Fm ). The total force FT is also shown for comparison.
where Tm and Ym are the thickness (Tm = 0.5mm) and Young’s modulus (Y = 3 ×
106 N/m2 ) of the silicone membrane, respectively. Figure 4.5 depicts the forces caused by
air compression/rarefaction, and by the silicone membrane tension (Fa and Fm , respec-
tively) as functions of relative membrane position l(t) = x calculated using equations 4.8
and 4.9. This graph shows that the effect of the membrane on the total restoring force FT
(bold trace) is minimum and only considerable as the displacement l(t) = x approaches
extreme values (±2 mm). This demonstrates that we can still obtain a good approx-
imation to the total reaction force FT (t) by only considering the restoring force Fa (t)
due to air chamber compression/rarefaction. Furthermore, it is unlikely that l(t) will
ever acquire values lower than 0 mm since the small bump at the bottom of the silicone
membrane will prevent negative acoustic pressures inside the air chamber. Therefore,
in spite of the nonlinearity of the total reaction force FT , a local linearization is possible
in the range of interest (Figure 4.6). If the height of the bump is approximately 0.5 mm,
it is then possible to choose 0 to 1 mm as the range of interest for the linearization in

Figure 4.6: Local linearization of the reaction force FT exerted by the microphone’s air
chamber as a function of relative membrane displacement l(t) = x. The fitted equation
and the corresponding R2 value are shown.
which most of the l(t) values will fall. Using least squares linear regression, the ideal
linear function for the modeled spring becomes:
FT (x) = 2758.8x − 0.0937 (4.10)
which now has the form FT = ka x where ka = 2758.8N/m is the spring constant of the
linear “spring of air” described above. The small offset produced by the linearization can
be seen as an additional constant force applied to the “spring of air”. However, it can
be discarded from the model since any constant (static) force that does not cause limb
movement will be equally opposed by the elasticity of the silicone sleeve and no effect will
be detected with any of the transducers in the MMG sensor. If an electrical equivalent
is considered, the inductance La = 1/ka will represent the behaviour of the ideal linear
spring which couples every vibration l(t) to the acoustic pressure ∆Pa (t) measured by the
microphone. Three significant assumptions have been made for this section of the model:
(1) an ideal linear spring behaviour (R2 = 0.9966), (2) an isothermal compression of the
Figure 4.7: Electromechanical model of the coupled MMG sensor pair in Figure 4.1.
Corresponding electrical equivalents for mass, stiffness and damping coefficients of the
mechanical variables are shown.
air chamber and (3) a massless system. The assumption of an isothermal compression
was computationally validated. No significant differences (p > 0.9) were found when
restoring forces for isothermal and adiabatic compressions of the sensor’s air chamber
were calculated. On the other hand, the restoring force of the proposed “spring of air”
will rapidly exceed the inertia caused by the minute masses of the air column and silicone
membrane, thereby validating the third assumption.

Model Integration
Figure 4.7 depicts an electromechanical equivalent of Figure 4.1 showing the remaining
mechanical variables involved in practical MMG signal recording when using the proposed
coupled MMG sensor pair3 . A variable velocity/voltage source dlM (t)/dt represents the
skin vibration due to muscle contraction. This source is coupled to the “spring of air” with
spring constant L−1 −1
a and the unknown mechanical damper Ra which prevents sustained
system oscillation. This spring and damper are in turn connected to a mass differential
Cs which represents the section of the soft silicone sleeve containing the coupled MMG
sensor. This mass is then connected to the forearm Cf by means of the silicone sleeve,
whose elasticity is represented by a second presumably linear spring with spring constant
L−1
l coupled to its corresponding mechanical damper Rl−1 . It is easy to note that the
stiffness coefficient L−1
l will be considerably higher than L−1
a since the sleeve is stiffer
than the air chamber (any force applied directly to the sensor will considerably deform
the air chamber before any local deformation in the sleeve may be noticed). However,
recall from Section 4.1 that any relative displacement of the silicone membrane l(t) can
be caused either by skin displacements lM (t) or by interfering forces Fe acting externally
on the coupled MMG sensor pair. This means that the coupling between the sensor and
the forearm is not completely stiff and some of the energy from external forces is still
being transmited to the air chamber. Therefore, although it may tend to infinity, the
stiffnes L−1
l of the sleeve cannot be discarded.
Figure 4.8a shows an electrical equivalent of the model proposed in Figure 4.7. A
variable current source substitutes the external force vector Fe (t). Figure 4.8b shows
the corresponding impedance circuit. Voltages at specific branches of the circuit now
represent the transducers’ signals within the coupled MMG sensor. These voltages, Vm (t)
for the microphone and Vs (t) for the accelerometer, represent the membrane relative
3
For a detailed explanation of the methods used for the determination of electrical equivalents please
refer to D’Azzo and Houpis (1995) and Oppenheim (1997) [4, 6]
Figure 4.8: Electrical equivalent circuit (a) and impedance circuit (b) of the model
proposed in figure 4.7.

velocity (with respect to sensor velocity) and sensor velocity, respectively. Note that the
physical transducer measurements are related to these voltages as follows:
dl(t)
Vm (t) = (4.11)
Z dt
Vs (t) = A(t)dt (4.12)
where l(t) is the relative membrane position with respect to the sensor case and A(t) is
the sensor acceleration. An additional non-linear transformation (equation 4.13) must
be applied to Vm (t) in order to obtain the value of acoustic pressure measured by the
microphone. From equations 4.7 and 4.11:

1atm(πr 2 h)
∆Pa (t) = R − 1atm (4.13)
πr 2 h − 13 πr 2 Vm (t)dt
Finally, the circuit in Figure 4.8b can be solved by obtaining functions for the outputs
(Vm (t) and Vs (t)) as functions of the sources (lM (t) and Fe (t)). In terms of impedance
(i.e. their Laplace transform) these functions are:

slM (s)Za (Zf + Zl + Zs ) − Fe (s)Za Zl Zs
Vm (s) = (4.14)
Zl (Zf + Zs ) + Za (Zf + Zl + Zs )
slM (s)Zl Zs + Fe (s)Zs [Za (Zf + Zl ) + Zf Zl ]
Vs (s) = (4.15)
Zl (Zf + Zs ) + Za (Zf + Zl + Zs )
where:
sLa Ra
Za = (4.16)
sLa + Ra
1
Zf = (4.17)
sCf
sLl Rl
Zl = (4.18)
sLl + Rl
1
Zs = (4.19)
sCs
Furthermore, inverse functions may be obtained by solving the system of equations 4.14
and 4.15 for the sources (lM (t) and Fe (t)) and by substituting the corresponding variables
from equations 4.11 and 4.12. In terms of impedance these functions are:
A(s)
s
Za Zl sl(s)[Za (Zf + Zl ) + Zf Zl ]
lM (s) = + (4.20)
Za (Zf + Zl ) Za (Zf + Zl )
A(s)
s
Za (Zf+ Zl + Zs ) sl(s)Zl Zs
Fe (s) = − (4.21)
Za Zs (Zf + Zl ) Za Zs (Zf + Zl )
where l(t) and A(t) can be calculated from the microphone and accelerometer signals,
respectively as:
3hM(t)
l(t) = (4.22)
M(t) + 1atmSM
A0 (t)
A(t) = (4.23)
SA
where M(t) and A0 (t) are the signals from the microphone and accelerometer, respec-
tively within a single coupled MMG sensor pair, and SM and SA are the corresponding
sensitivity values for each transducer. Equation 4.20 is then a model-based data fusion
algorithm, which uses the signals acquired from individual transducers within a single
coupled microphone-accelerometer MMG sensor pair to track the vibrating source (i.e.
skin vibration lM ) due to muscle contraction. This algorithm may be implemented us-
ing analog electronics and will dynamically eliminate any interference in the microphone
signal caused by external forces on the MMG sensor. Important assumptions and con-
siderations for the presented model are:
1. The measured value for the coupled MMG sensor’s mass Cs was 0.005Kg. This
parameter represents only a discrete aproximation of the sleeve’s mass differential
which contains the sensor. However, due to the significantly higher mass of the
sleeve and the forearm, the assumption of a discrete mass differential should not
affect the results considerably.
2. The forearm mass Cf must be individually measured for every user and must in-
clude the mass of the stump, the sleeve and the prosthesis. However, for model
validation and optimization, only the masses of the mechanical stimulator and the
silicone square section which contains the MMG sensor (i.e. the simulated fore-
arm) were considered Cf = 0.3Kg. The adjustment of Cf consequently changes
the model response. However, no decrease in model performance is expected for real
Figure 4.9: Schematic representation of the experimental setup for model validation
measurements.
MMG signal recording assuming that the proposed experimental setup sufficiently
characterizes the real conditions.
3. The stiffness coefficient for the silicone sleeve L−1

l and the value of the mechanical
dampers Ra−1 and Rl−1 are initially unknown. However, some constraints for the
possible values of this variables can be derived and used for model optimization.
4. The presented model considers exclusively the transduction of mechanical signals
measured by the coupled MMG sensor. Aditional electronic circuitry such as filters
and amplifiers conected to the transducers will modify these responses. Therefore,
a pre-processing stage which eliminates any effects from such electronic circuitry
must be added prior to the use of the proposed model.
4.2.3 Model Validation
In order to validate the proposed model, a controlled experiment involving real measure-
ments with the coupled MMG sensor was performed. Figure 4.9 shows the experimental
setup. A coupled MMG sensor pair embedded in a silicone square section (1200 × 1200 )
was fastened to a custom-made mechanical stimulator built with a flat type speaker4
mounted in a resonance box [32]. This mechanical stimulator was used to simulate the
useful vibrating source (i.e. skin displacement lM (t)) due to muscle contraction. In order
to facilitate suspension, the silicone soft socket must fit tightly on the patient’s stump.
This causes an initial tissue compression and increase of the apparent stiffness of the sil-
icone sleeve. Additional tissue compression may occur during practical MMG recording.
Further, some of the energy produced by external forces applied on the coupled MMG
sensor may be absorbed by the tissues as opposed to the microphone’s air chamber. In
order to recreate this possible soft tissue compression, and therefore, external force atten-
uation during MMG signal recording, a 2 mm silicone gel layer5 was placed between the
sensor and the mechanical stimulator. This gel possesses similar mechanical properties
as soft tissue and tissue simulants previously used in tissue mechanics research [21, 20].
The stimulator, together with the MMG sensor and silicone gel layer, were then secured
on a shaker platform6 . Note that the inertia caused by the vibration of the shaker can be
seen as a controled force Fe (t) acting externally (opposite to the direction of vibration)
on the MMG sensor. Therefore, limb movement, which causes significant interference in
the microphone signal, can be simulated. An accelerometer model BU-7135 from EmKay
Innovative Products, was used to measure the acceleration of the shaker platform. These
signals can then be used to approximate the value of the force of inertia Fe (t) acting
externally on the MMG sensor. The setup presented in Figure 4.9 provides us with an
environment similar to the real MMG signal recording conditions in which the parameters
of the skin vibration source lM (t) and the external force source Fe (t) can be controlled.
Furthermore, tests in which both interfering sources are simultaneously active as well
as single frequency measurements may also be performed and compared to the model.
4
Panasonic model WM-R57A
5
GEL-8150 from NuSil Technology
6
T2000 Series Electrodynamic Shaker System from Unholtz Dickie Corporation
Three sets of measurements for frequencies ranging from 5 to 100Hz were performed with
increasing values of sinusoidal forces applied by the shaker on the mechanical stimulator
(and therefore, the MMG sensor). The root mean square (RMS) magnitude of the cor-
responding sinusoidal accelerations caused by these forces on the shaker platform were
approximately 0.18, 0.71 and 1.41 g). Signals from each transducer within the coupled
MMG sensor pair and the accelerometer on the shaker platform were acquired simulta-
neously at a sampling rate of 1kHz during an average time window 0.5s in length. These
signals were stored for further analysis.
Model Optimization
The frequency response of each transducer in the coupled MMG sensor pair due to ex-
citation from the mechanical stimulator will not be as reliable as that due to shaker
excitation. This is because the vibration of the flat-type speaker in the mechanical stim-
ulator cannot be directly measured and may be biased by its own frequency response.
Furthermore, as reported in Section 4.1, the accelerometer does not present significant
sensitivity to skin vibrations. Thus, to provide reliable validation of the proposed model,
only frequency response measurements of signals acquired from the shaker were used
for model optimization. These empirical responses represent two thirds of the available
information for model optimization. It is conceivable that the optimal parameter values
may depend on whether the full set or chosen subset of frequency responses are con-
sidered. However, the chosen subset of frequency responses completely characterize the
undesired external force, suggesting that additional information may be redundant and
unlikely to significantly change the optimal parameter values. The empirical responses
were compared to the responses obtained from the model. In this case, the measured
external force Fe (t) is considered the input to the system and signals acquired with each
transducer within the coupled MMG sensor are the outputs.
A program was writen in the MATLAB environment to calculate the correspond-

ing frequency responses of the microphone and accelerometer within the coupled MMG
sensor due to an externally applied oscillatory force Fe (t). This program segmented
every recording (i.e. shaker acceleration, microphone signal and accelerometer signal)
into individual segments of one period in length. The RMS value, period and signal
delay with respect to the source for each cycle were calculated. Finally, the mean RMS,
frequency and phase over all the cycles in the recording window were obtained. These
values were then used to determine the empirical frequency responses of the microphone
and accelerometer due to shaker vibration. In order to obtain a better representation
of the empirical data, polynomial functions were fitted to the empirical magnitude and
phase response data for each system. These functions were then compared to the model
responses using a mean square error (MSE) criterion. A multiple criteria objective func-
tion was defined to optimize the model parameters. First consider the following single
criterion objective function:

N h i2
~ 1 X
MSE(θ) = ŷ(fi , ~θ) − y(fi) (4.24)
N i=1
where ~θ is a vector of model parameters, ŷ(fi , ~θ) is the model’s magnitude or phase re-
sponse, y(fi ) is the corresponding empirical magnitude or phase response, fi is a single
frequency value. To maximize the fit between empirical and model responses, four single
criterion functions, each in the form of equation 4.24 were formulated. In particular,
mic ~ mic ~
there were two objective functions. MSEmag (θ) and MSEphase (θ), for optimizing the
transfer function from the shaker to the microphone. Likewise, two objective functions,
acc ~
MSEmag acc
(θ) and MSEphase (~θ) were associated with optimizing the shaker to accelerom-
eter transfer function. The overall multiple criteria objective function, J(~θ), was then
formulated as the sum of the individual objective functions,
J(~θ) = MSEmag
mic ~ mic ~
(θ) + MSEphase acc ~
(θ) + MSEmag acc
(θ) + MSEphase (~θ) (4.25)
This multiple criteria function represents the total MSE between empirical and model
responses.
As mentioned earlier, the stiffness coefficient for the silicone sleeve L−1
l should be
higher than that of the air chamber L−1
a . This represents a parameter constraint. In
addition, none of the parameters can take negative values since they are all based on real
mechanical components. The optimization problem can now be formulated as follows,
Minimize J(~θ) (4.26)
such that
θ1 > L−1
a (4.27)
θ1 , θ2 , θ3 > 0 (4.28)
where ~θ = {θ1 , θ2 , θ3 } = {L−1 −1 −1 ~

l , Ra , Rl } and J(θ) is defined as in Equation 4.25.
Optimization was performed using the Nelder-Mead algorithm [22] in the MATLAB
environment and provided a set of values ~θ, which resulted in the best possible fit with
the empirical data.
4.2.4 Results
Figures 4.10 and 4.11 show the empirical magnitude and phase responses of microphone
and accelerometer, respectively, to the external force Fe (t). Polynomial regressions for
each graph are superimposed. As mentioned earlier, these regressions were used during
model optimization as representations of the corresponding empirical frequency responses
measured for each sensor.
Note that the slope m in each of the phase responses can be related to a time delay
∆t of the output signals given by the relationship:
m
∆t = (4.29)
360
Therefore, the real phase response can be approximated as a constant equal to the offset
of each linear relationship for the phase response (i.e. −22.8590 and 2.11280 for the
relative membrane displacement and sensor acceleration, respectively).

Figure 4.10: Empirical magnitude and phase responses of microphone to an external force
Fe (t). Each point is calculated from the estimated force applied by the shaker platform
Fe (t) to the measured relative membrane position l(t) in the coupled MMG sensor. The
corresponding fitted polynomials are also shown as the solid lines.

Figure 4.11: Empirical magnitude and phase responses of accelerometer to an external
force Fe (t). Each point is calculated from the estimated force applied by the shaker
platform Fe (t) to the measured sensor acceleration A(t) in the coupled MMG sensor.
The corresponding fitted polynomials are also shown as the solid lines.
Figure 4.12: Empirical (solid line) and optimum model (dashed line) responses for the
transfer functions involved in the model optimization.

Finally, by minimizing the total MSE between the empirical and modeled responses,
the optimal values of 1/1.0682 × 10−6 N/m, 1/1 × 10−3 N/s and 1/1.6170 N/s were
obtained for the parameters L−1 −1 −1

l , Ra and Rl , respectively. Figure 4.12 compares the
empirical and optimum model responses for each of the transfer functions involved.
There is a maximum 2 dB error at 100 Hz in the optimized magnitude response of
the relative membrane position l(t). Correspondingly, the maximum error for the sensor
acceleration magnitude response is 1 dB. Maximum values of phase errors were 200 or
11% of 1800 (i.e. half a period) for the membrane displacement l(t), and 20 or 1.1% of
1800 for the sensor acceleration, A(t).
Once the model has been optimized, measurements in which both sources (i.e. shaker
platform and mechanical stimulator) are simultaneously active can be performed. These
measurements can be processed using equation 4.20 to track the original useful vibration
source (i.e. skin displacement lM (t) due to muscle contraction). Figure 4.13 depicts a
sample model-based source separation test performed with the Simulink toolbox in the
MATLAB environment. The real microphone and accelerometer recordings are shown
at the top. A sinusoidal external force with an amplitude of 1.7 N and frequency of 24
Hz was simultaneously applied with a sinusoidal skin vibration at 15 Hz and estimated
amplitude of 3 × 10−6 m. The tracked source lM (t) and its estimate lMd
(t) are shown
at the bottom of the figure. Note the complete elimination of the interfering external
force. The observed phase differences are mainly due to additional filtering applied to
the estimated skin vibration with the intention of eliminating the position offset.
4.2.5 Conclusions
The small errors obtained in the comparison of empirical and model responses validate
the proposed model. Other external variables such as the uncertainty in the precision of
the electronic components were the likely the cause of these small differences.
Figure 4.13: Sample source localization test for simultaneous active sources. M(t) and
A0 (t) are the microphone and accelerometer signals in Volts and lM (t) and lMd
(t) are the
measured and estimated skin vibration signals.

A model-based data-fusion algorithm for the discrimination of external forces in prac-

tical MMG signal recording has been proposed. This model provides the most direct
real-time source separation technique for the detection of skin vibrations under extreme
noise conditions.
Chapter 5
Multisensor Fusion
This chapter details the MMG-based data and sensor fusion algorithms devised and eval-
uated for the generation of control signals for externally powered prosthesis. Recall that
the proposed sensor for MMG signal recording [29] consists of two individual transducers
(see Chapter 4). Here, we refer to data fusion as the combination of information from
these transducers (i.e. accelerometers and microphones) within a single sensor pair. On
the other hand, we refer to sensor fusion, as the integration of information from the
outputs of different sensor pairs within a multisensor array. Section 5.1 (reprinted with
permission from a published paper), presents the first and simplest approach for the dy-
namic elimination of external noise sources and the detection of muscle contraction (i.e.
data and sensor fusion) based on the RMS value of the acquired signals. Section 5.2
discusses an integrated strategy where data and sensor fusion algorithms are combined
to provide MMG-based detection, estimation and clasification of muscle activity. The
proposed strategy represents the first ever attempt to generate multiple output signals
for prosthesis control using a MMG multisensor array embedded distally in a silicon soft
socket. This integrated strategy consists of three stages. The first is the pre-processing
stage, aimed at further increasing the SNR of individual sensor signals in the array. This
is followed by the detection stage which determines the prescence or abscence of muscle
72
Chapter 5. Multisensor Fusion 73
contractions in the acquired signals. Finally, upon detection of a contraction, the final
stage, that of clasification, specifies the nature of the contraction and the corresponding
control output. Sections with relevant new content are 5.1.1, 5.1.3, 5.1.4, 5.1.5, 5.1.6 and
5.2.
5.1 MMG-Based Multisensor Data Fusion for Pros-
thesis Control
c
2003 IEEE. Reprinted, with permission, from Silva J., Chau T., and Goldenberg A.,
“MMG-Based Multisensor Data Fusion for Prosthesis Control,” in Proceedings of the
25th Annual International Conference of the IEEE Engineering in Medicine and Biology
Society, pp. 2909-2912Cancun QR, Mexico, September 2003.
5.1.1 Abstract
Advantages in the functionality and comfort of soft silicon sockets or roll-on sleeves
over polyester laminate hard sockets for upper-limb prosthesis have been consistently
reported. However, attachment and wire breakage issues prevent the use of electromyog-
raphy (EMG) sensors with soft sockets in electrically powered prosthesis for below-elbow
amputees. Mechanomyography (MMG) is the measurement of the mechanical vibra-
tions elicited by contracting muscles. The use of MMG sensors embedded within the
soft silicon socket solves current attachment issues with EMG sensors and facilitates dis-
tal recording preventing wire breakage. In order to implement a practical MMG-based
detection system of muscle contractions for prosthesis control, three silicon-embedded

microphone-accelerometer sensor pairs were used to record MMG signals and movement
artifact around the distal end of the residual limb of a below-elbow amputee. A mul-
tisensor data fusion strategy for the generation of binary control signals based on the
root-mean-square (RMS) values of the segmented signals acquired with each transducer
was trained and used as a detector. A ninety five (95%) and eighty six percent (86%)
sensitivity were achieved in the detection of contraction signals from the wrist extensors
and flexors, respectively. The probability of error (false positive) in the discrimination
of movement artifact was thirteen percent (13%).
5.1.2 Introduction
Advantages in the functionality and comfort of soft silicon sockets or roll-on sleeves over
polyester laminate hard sockets for upper and lower-limb prosthesis have been consis-
tently reported [19, 27]. However, attachment and wire breakage issues prevent the use
of electromyography (EMG) sensors with soft sockets in electrically powered prosthesis
for below-elbow amputees [27, 16]. Therefore, alternative techniques for the measure-
ment of the muscular activity are necessary in order to facilitate the use of soft sockets.
When a muscle contracts, low frequency mechanical oscillations can be detected superfi-
cially [23, 25, 14]. The transduced signal is termed the surface mechanomyogram (MMG)
[orizio1993]. Most of the power of the MMG signal falls within the 5-50Hz bandwidth and
no significant frequency components beyond 100Hz have been reported for arm muscles
[23, 25, 14, 38, 17]. As with electromyographic (EMG) signals, there is a linear rela-
tionship between the root-mean-square (RMS) values of MMG signals and the force of
the contraction within 20% to 80% of maximum voluntary contraction (MVC) [23, 25].
This is one of the main characteristics that make MMG especially suitable for prosthesis
control. MMG signals recorded by a standard microphone have been used to operate
a free-standing prosthetic hand in a controlled laboratory setting [13]. Using a single-
channel, full-wave rectified MMG signal, Barry et al. demonstrated that MMG signals
could be intentionally used to open and close a prosthetic hand. The reported system
successfully discriminated between MMG signals arising from wrist flexion and exten-
sion and exhibited robustness to changes in skin impedance and microphone placement.
However, basic sensor attachment issues precluded its practical use [13]. Furthermore,
considerable noise interference from a number of sources (e. g. limb movement) re-
mained unresolved beyond the restricted conditions of a clinical environment. The use
of silicon-embedded MMG sensors for the measurement of muscular activity solves cur-
rent attachment issues with EMG sensors and facilitates distal recording, preventing wire
breakage [32, 33]. However, an efficient and robust signal processing strategy is still nec-
essary to generate a reliable control signal. The objective of this investigation was the
design of a MMG-based multisensor data fusion strategy for the detection of muscle con-
tractions under extreme noise conditions in order to facilitate practical implementation
for electrically powered prosthesis control.
5.1.3 Methodology
Sensor Design
MMG signals can be recorded using microphones [14, 17, 33, 35] or accelerometers
[32, 35, 24, 26]. Optimum embedding parameters have been reported for silicon- embed-
ded versions of both [32, 33]. These parameters were based on the optimization of the
signal-to-noise ratio (SNR) of MMG signals under controlled conditions where the noise is
assumed gaussian and the a priori SNR is known and positive (16 dB approximately). In
this case, the gain in SNR is about 3dB for silicon-embedded accelerometers and 9dB for
silicon-embedded microphones as compared to non- embedded accelerometers. However,
under extreme noise conditions the SNR can dramatically decrease to -18dB in silicon-
embedded accelerometers and -3dB in silicon-embedded microphones. This difference in
the sensitivity to noise of both sensors can be used in a coupled instrumentation set up
to determine the nature of the signal. Fig. 5.1 shows a schematic diagram and pictures
of the silicon-embedded microphone1 -accelerometer2 sensor pair used in the experiments.

The device had an approximate surface area of 1.9 x 1.9cm and a height of 1cm. The
1
Electret condenser microphone model MD6022ASC-0 from Emkay Innovative Products
2
Micro-machined accelerometer model BU-7135 from Emkay Innovative Products
sealed air chamber dimensions were 1.3 and 0.2cm in diameter and height, respectively.
Silicon shores 20A and 35A were used for the silicon case and membrane, respectively
[33]. Skin vibrations originating from contracting muscles and propagated through soft
tissue cause a proportional displacement of the silicon contact membrane. These vibra-
tions then modulate the acoustic pressure measured by the microphone inside the air
chamber. However, any force F acting on the sensor from the external side (away from
the skin) will also cause a relative displacement of the silicon membrane and, therefore,
considerable noise interference in the microphone.s signal. Forces associated with inertia
can cause such interference during limb movement. Thus, limb movement is probably
one of the main noise sources associated with practical MMG-based control of electrically
powered prostheses. The usefulness of the accelerometer arises then from the fact that it
can be used to measure the external forces acting on the sensor pair (the external force
signal can be calculated from the accelerometer signal scaled by the constant mass dif-
ferential on which the external force is acting). Note that the accelerometer is placed at
the back of the microphone. This causes a further decrease in the SNR of the accelerom-
eter because the microphone’s silicone membrane and air chamber filter the vibrations
from the muscles before they reach the accelerometer. The accelerometer acts then as a
dynamic reference sensor.
MMG Signal Acquisition
MMG signals were acquired from a client of the Musculoskeletal Program at the Bloorview
MacMillan Children’s Centre (female below-elbow traumatic amputee). Three sensor
pairs were placed around the forearm at approximately 1200 away from each other (Fig.
5.2). Sensors were placed about 1.5 cm from the distal end of the stump. After anti-alias
filtering, signals were sampled at 200 Hz and stored in a computer for further analysis.
All the recordings were 6s long with an activity period that extended from 2 to 4s. The
activity period was defined as the interval during which the client performed one of four
Figure 5.1: MMG sensor pair designed and used for the experiments. (a) Schematic
diagram of the design. (b) Accelerometer (posterior) view of MMG sensor pair before
embedding. (c) Microphone (anterior) view of MMG sensor pair after embedding; the
silicon contact membrane was added in the final embedding stage.

Figure 5.2: Schematic diagram of sensor placement around the distal end of the residual
forearm.
possible activities namely: (i) remain in rest; (ii) opening contraction (extension); (iii)
closing contraction (flexion); and (iv) random limb movement. The terms “opening” and
“closing” in this case corresponded to the systematic reproduction of the contractions
usually performed by the client to open or close her current EMG-based prosthesis. Five
recordings were acquired for each task and each sensor (60 total). The main objective
was the discrimination among contraction signals and movement artifact.
Detection Algorithm
The RMS-based detection algorithm used in our experiments was defined as:
H1
.
RMSM ..>< γRMSA (5.1)
H0
where RMSM and RMSA are the RMS values of the microphone and accelerometer
signals, respectively. The right-hand side of 5.1 acts as a dynamic threshold applied to
the microphone signal for the detection of muscle contractions (H1 ). H0 in this case is the
null hypothesis of no contraction present in the recording. The detector’s performance
depends on γ, a constant parameter that may be associated with the relative response
between both sensors. The optimum value of γ was determined using a training algorithm
which will be detailed later. Note that a different value of γ may be defined for each
sensor pair. Signals were segmented into 0.2s non-overlapping decision windows and
band-pass filtered in the 5-50 Hz bandwidth. Given the sensor array in Fig. 2, it can
be easily demonstrated that a limited further decrease in the SNR of each accelerometer
signal A(t)i can be achieved by performing the following operation:
A(t)i − A(t)j − A(t)k (5.2)
where A(t)j and A(t)k are the remaining accelerometer signals. Equation 5.2 was then
used to further decrease the SNR of the accelerometer signals after filtering (i.e. the actual
contraction signal content was reduced as much as possible). The detection algorithm in
5.1 was then applied to each processed segment in a simulated real-time algorithm giving
us three decisions (one for each sensor pair). In order to reduce detection errors due to
local fluctuations, at least two out of the three sensors should have detected the signal
for it to be considered a true detection.
Performance Measures
It can be readily observed that every 0.2s segment is associated with a decision. Therefore,
the detection sensitivity of the algorithm was calculated as the ratio of the number of
contraction segments detected over the total number of contraction segments present
for all the signals with a contraction in their activity period. On the other hand, the
misdetection was calculated as the ratio of detected contractions from non-contraction
segments over the total number of non-contraction segments present. Finally, the overall
sensitivity was determined by calculating the ratio of detected contraction segments plus
undetected non-contraction segments over the total number of segments.

Training Algorithm
The training algorithm was based on the determination of the optimal value of γ in
5.1 which maximizes the detector’s performance. An initial value of γ was obtained
by calculating the mean ratio RMSM /RMSA from two of the resting signals (i.e. just
background noise). A number of values of γ around the value originally obtained were
then tested with two of the recordings for each activity. The detector’s overall sensitivity
(as defined in the previous section) for each value of γ was calculated and used to find
the function of a second order polynomial regression which related the values of γ with
the overall sensitivity measure. Finally, the maximization of this function yielded the
optimum value of γ.
5.1.4 Results
Table 5.1 shows the performance measures obtained for different sets of values of γ. Only
the value of γ for the sensor pair s2 is reported. Note however, that there was an optimal
value of γ for each sensor pair calculated during the training session which maximizes
the overall performance of the detection algorithm. Fig. 5.3 shows a comparison of the
output signals from the detection algorithm for an extension (“opening” contraction)
and movement artifact signals. A rectified and filtered version of the microphone’s sig-
nal (gray trace) is included as a reference. The detection sensitivity for this particular
contraction signal was 100% whereas the probability of error was 9.5% in the case of
the movement artifact signal. Note that, without the additional information provided
by the accelerometer, the correct detection of muscle contractions would be seriously
compromised.
Table 5.1: Performance measures for different sets of values of γ
Performance
Detection Detection
γ(s2 ) Sensitivity Sensitivity Probability Overall
(Extension) (Flexion) of Error Sensitivity
2.25 94.7% 86.7% 12.5% 91.4%
2.5 89.5% 71.4% 8.6% 91.6%
2.57∗ 89.5% 66.7% 7.3% 92.0%
2.75 89.5% 57.1% 6.5% 91.6%
∗
Optimal value calculated with the training algorithm for the reported sensor pair (s2 ).
Figure 5.3: Sample detection of an extension (“opening” contraction) signal from the
forearm and the discrimination of movement artifact.
5.1.5 Discussion
It can be easily seen from Table 5.1 that the detection performance for flexion (“closing”
contraction) signals is significantly lower than their extension counterparts. Furthermore,
the latter signals are less sensitive to changes in the value of γ. The reason is that RMS
values for flexion signals were also significantly lower than RMS values for extension sig-
nals. This is possibly due to a decreased muscular mass in the anterior side of the forearm
of the patient leading to a reduced vibration amplitude during contraction. Unlike their
EMG counterparts, the amplitude of MMG signals depends on the mechanical properties
of the muscle including its mass, shape and spatial distribution limiting the detector’s
performance. However, this behaviour could also be used for discrimination between
flexion and extension once a signal has been detected. Future work will be done on this
matter. Finally, note from Table 5.1 that a high sensitivity of detection for both flexion
and extension is achieved for the first set of values of γ. The training algorithm in turn,
provided a higher set of γ values for the maximization of the overall performance which
reduced the detection sensitivity; especially flexion detection. However, there is a high
probability that a majority of the presumably wrong decisions for the first set of values
(first row in table 5.1) were indeed unconscious contractions superimposed on movement
artifact, correctly detected by the algorithm. Therefore, an ideal training session must
include feedback from the patient in a real-time set-up possibly through the use of a
reward-punishment learning algorithm.
5.1.6 Conclusions
We have presented a simple and reliable multisensor data fusion strategy for the gen-
eration of binary control signals for an electrically powered prosthesis for below-elbow
amputees. Furthermore, the sensitivity of muscle contraction signal detection may facil-
itate strategies for the generation of multiple control signals (e.g. multiple contractions
over short periods of time may be used to switch among different motors) enhancing the
functionality of the whole system.
5.2 MMG-Based Multisensor Fusion Strategy for De-
tection and Classification of Muscular Activity
It has been demonstrated that the use of multiple sensors for the measurement of natu-
ral phenomena provides redundant information that, if properly combined, enhances the
SNR in most source localization and filtering applications [18, 34]. This approach also
facilitates the extraction of independent signal features for pattern classification [8]. As
mentioned at the beginning of this chapter, this thesis regards data fusion as the combi-
nation of information from transducers (i.e. accelerometers and microphones) within a

single MMG sensor pair, while sensor fusion is the integration of information from the
outputs of different sensor pairs within a multisensor array.
This section describes an integrated strategy where data and sensor fusion algorithms
are combined to provide MMG-based detection, estimation and clasification of muscle
activity. The proposed strategy represents the first ever attempt to generate multiple
output signals for prosthesis control using a MMG multisensor array embedded distally
in a silicon soft socket. This multisensor fusion strategy consists of three stages. The
first is the pre-processing stage, aimed at further increasing the SNR of individual sensor
signals in the array. This is followed by the detection stage which determines the presence
or absence of muscle contractions in the acquired signals. Finally, upon detection of a
contraction, the final stage, that of clasification, specifies the nature of the contraction
and the corresponding control output.

Figure 5.4: Schematic diagram of equation 4.20. The relative membrane position and
acceleration signals (l(t) and A(t), respectively) are individually filtered and added to
obtain an estimate of skin displacement lM (t).
5.2.1 Pre-processing
Model-Based Filtering (MBF)
The implementation of the proposed model-based source separation strategy in Section
4.2 for the estimation of skin displacement lM (t) (equation 4.20) represents one possible
approach to sensor fusion. This model-based source separation algorithm can not only
discriminate between muscle contractions and limb movement, but may also facilitate
detection of contractions under extreme noise conditions. Therefore, if proven reliable,
the model-based source separation algorithm will represent the most direct and efficient
data fusion algorithm for the enhancement of the SNR of signals from individual MMG
sensors. Recall that the proposed model was succesfully used to recover the simulated
muscle vibration signals (i.e. vibrations from the mechanical stimulator) in the experi-
mental setup of Section 4.2. However, considerable phase errors in the preliminary tests,
make it difficult to correctly separate interfering sources in real MMG recordings. This
implies that the proposed model is sufficient to describe the mechanical interactions of
the coupled MMG sensor on the shaker platform (see Figure 4.9), but is not sufficient to
describe these interactions in real MMG measurements. Possible routes for the generation
of an enhanced and more reliable model are described in Section 6.2.2. Figure 5.4 shows
a schematic diagram of equation 4.20, which describes the model-based source separation
algorithm. This equation may also be seen as the filtering and addition of two signals
(i.e. l(t) and A(t)). Both filters, Hm (s) and Ha (s) adjust the magnitude and phase of
these signals before addition. Filter Ha (s) focusses on calculating the contribution of
external forces on the relative membrane displacement signal l(t) using the acceleration
signal A(t). This contribution is eliminated from signal l(t) by addition, since the filtered
acceleration signal A0 (t) should be about 1800 out of phase with respect to the filtered
relative displacement signal l0 (t). This phase difference is expected since acceleration
is the second derivative of position. Therefore, if phase errors exist, the value of this
summation stage will be severely compromised. However, the RMS values of the signals
should remain unchanged since they are not dependent on phase. Therefore, the RMS-
based detection algorithm proposed in Section 5.1.3 may still be applied to signals l0 (t)
and A0 (t). These signals presumably provide better descriptions of muscle activity and
external force contributions, than the outputs M(t) and A0 (t) obtained directly from the
transducers in the coupled MMG sensor. This model-based filtering can be then defined
as a pre-processing stage which enhances the SNR of signals from individual transducers
within a coupled MMG sensor.
Accelerometer Sensor Fusion (ASF)
As mentioned in Section 5.1.3, the operation defined in equation 5.3 can be considered
a pre-processing stage for any array involving three MMG sensors distributed at a 1200
angle from each other around the patient’s stump (see Figure 5.2). This operation com-
bines information among accelerometer signals from different MMG sensor pairs in the
array to further decrease their SNR (i.e. increase sensitivity to noise). The algorithm is
based on the geometrical interactions that provide redundancy of the information about
Figure 5.5: Typical gain surface (Phase-based TFM) in the direction of sensor i as a
function of the phases, φk and φj , of the other two accelerometers in the array, at a single
input noise frequency.

external forces acquired by the accelerometers. It is defined as:
Ad
i (t) = Ai (t) − Aj (t) − Ak (t) (5.3)
where Ad
i (t) is the new estimate for acceleration in the direction of sensor i, while j and
k are sensor indices for the other two accelerometers in the array.
Any acceleration in the direction of sensor i will cause signals in the remaining sensors
j and k with half the magnitude and 1800 phase offsets with respect to signal Ai (t) (i.e.
Aj (t) = Ak (t) = Ai (t)cos(1200)). The assignment operation preserves signals among
sensors that are 1800 out of phase while completely eliminating in-phase signal compo-
nents. Therefore, the algorithm can be seen as a real-time implementation of a phase

error-based time-frequency mask (TFM) applied to each accelerometer signal. Figure 5.5
shows a typical normalized gain surface in the direction of sensor i as a function of the
phases φj and φk of the remaining accelerometer signals for a single frequency.
The resulting processed accelerometer signals can then be combined with the micro-
phone signals using any of the previously proposed data fusion algorithms. This pre-
processing algorithm may help to recover microphone signal components due to move-
ment inside the forearm that would otherwise be eliminated. An important example of
this situation is the onset of muscle contraction where the magnitude of muscle vibration
may cause readings in the accelerometers.
5.2.2 Detection
After enhancing SNR with the pre-processing algorithms, the subsequent task is to de-
termine whether the measured signals are due to muscle contraction (defined as the
hypothesis H1 ) or some other activity, unrelated to muscle activity, such as limb move-
ment or rest (defined as the null hypothesis H0 ). The resulting decision can then be
used to provide a suitable output signal for prosthesis control or as an input to a further
activity classification stage.

RMS-Based Detection of Muscle Contractions (RBD)
The reprinted paper in Section 5.1 [30] presents the first approach to data and sensor
fusion in which the RMS value of accelerometer signal is used as a dynamic threshold
for the RMS value of the microphone signal (equation 5.1). The algorithm not only en-
hances the SNR but also refutes or verifies the hypothesis of muscle contraction within
a given processing window, for each MMG sensor pair. Subsequently, an “overall” deci-
sion can be obtained by counting the outcomes across individual sensors and chosing the
most frequent decision (sensor fusion). In this case, data and sensor fusion strategies are
merged into one processing stage. The main advantage of this algorithm is the simplicity
of its implementation, foregoing the need for complex electronic signal processors. How-
ever, a training stage is still needed to adjust the detection parameter γ for each sensor.
This training strategy can be defined as an optimization problem in which the outcome
measure is the number of incorrect decisions made by the algorithm. Thus, an optimal
set of γ values can be obtained. If model-based filtering is applied for the purpose of
RMS-based detection (RBD), the afore-mentioned microphone and accelerometer signals
will become the filtered relative membrane vibration l0 (t) and acceleration A0 (t) signals,
respectively.
5.2.3 Classification
As mentioned earlier, MMG signals propagate through fluid and tissue surrounding the
contracting muscles. This property facilitates signal acquisition from virtually any loca-
tion around the patient’s stump. However, the nature of this propagation depends on the
mechanics of the materials involved. In principle, one of the main effects of these mate-
rials is that of a mechanical damper whose damping coefficient increases with increased
propagation distance from the contracting muscle. Consequently, the signal amplitudes
detected by sensors located distal to the source would be diminished as compared to

Figure 5.6: Feature space for muscle activity classification. Every point P is defined by
three RMS values derived from the microphone signals. A linear classifier can be defined
for the differentiation between wrist flexions and extensions.
those signals from sensors placed proximal to the source. Therefore, a consistent differ-
ence in the amplitude of signals acquired at different points around the forearm should
be observed for different muscle activity patterns (i.e. when different muscle groups are
active). This propagation property, together with other particular characteristics of mus-
cle vibration, such as the acoustic dipole behaviour (further explained in Section 6.2.2),
could be used to perform a spatial estimation of muscular activity in the forearm cross-
section monitored by the MMG sensor array. A proposed algorithm for such estimation
is presented in Section 6.2.2. However, a simpler approach is also possible.
RMS-Based Activity Classifier (RAC)
Once a contraction has been detected, MMG signal propagation properties may facilitate
the definition of a muscle activity classifier based on significant differences among the
RMS values acquired by different sensors in the array. This would allow us, for instance,
to differentiate between known activities such as wrist flexion and extension, paving the
way for the generation of multiple control outputs.

Using the three available RMS values of the microphone signals M1 , M2 and M3
or the filtered relative membrane vibration signals l10 , l20 and l30 (if model-based filtering
was applied) a three-dimensional feature space can be defined (see Figure 5.6). For every
g , RMS
decision window, we can define a point P = (RMS g , RMS
g ) in this 3-dimensional
1 2 3
g , i = 1, 2, 3, denotes a normalized RMS value for sensor i.

feature space, where RMS i
One way to define this normalization for the ith sensor is as follows.
g = RMSi
RMS i , i = 1, 2, 3 (5.4)
RMS1 + RMS2 + RMS3
g ≤ 1. It is important to normalize the RMS values to reduce the effect

where 0 ≤ RMS i
of the variability introduced by contractions of different strengths.
A two-category linear classifier (a plane in the 3D feature space) can then be defined,
together with the associated classification rule. In this case, the optimal plane will divide
the feature space so that a majority of points from wrist flexions are located on one
side of the plane while a majority of points from wrist extensions are on the other. The
classification rule can then be defined as:
Hf
.
d..>< 0 (5.5)
He
where d is the distance from the point P to the plane, and Hf and He are the hypotheses
that the contraction is a wrist flexion or extension, respectively. In this case, d will be
positive for flexions and negative for extensions. The distance d can be calculated as:
g + θ RMS
θ1 RMS g + θ RMS
g +θ
1 2 2 3 3 4
d= q (5.6)
θ12 + θ22 + θ32
where θ1 , θ2 , θ3 and θ4 are the parameters defining the plane equation. The vector
θ̂ = {θ1 , θ2 , θ3 , θ4 } is initially unknown, but it can be optimized in a training session if
estimates of the true target distances dT are known.

Figure 5.7: Flow chart for the proposed integrated data/sensor fusion strategies.
5.2.4 Algorithm Integration
Figure 5.7 shows a flow chart where the proposed data/sensor fusion algorithms are iden-
tified within an integrated framework. This framework uses the six signals available from
the transducers in the array (i.e. three microphone signals M(t) and three accelerometer
signals A0 (t)) to provide a final decision on the nature of the contraction, if at all present.
Table 5.2 shows the four possible pre-processing combinations or strategies for the
proposed framework. In the subsequent subsections, we will evaluate the performance
and utility of each pre-processing combination.

Table 5.2: Possible pre-processing combinations for MMG-based detection and classifi-
cation of muscle contractions.
Strategy MBF ASF
No. Applied? Applied?
1 No Yes
2 No No
3 Yes Yes
4 Yes No
5.2.5 Methodology
Strategy Evaluation
The most important task in the generation of control outputs for prosthesis control is
the reliable classification of the user’s intention. Current EMG-driven externally powered
prosthesis measure the electrical activity of superficial muscles and use that information
to generate an estimate of this intention. Therefore, it makes sense to compare the
proposed MMG-based detection and classification strategies against their EMG counter-
parts. This comparison would demonstrate whether it is possible to obtain information
comparable to that provided by EMG sensors using a MMG multisensor array. The
chosen methodology for evaluating the proposed strategies was thus a systematic com-
parison between EMG and MMG. Both sets of signals were simultaneously acquired from
below-elbow amputees wearing a soft silicon socket and performing different activities.
The information acquired from the EMG signals was considered as the target (i.e. true
intention) for every test. The objective was to determine how closely the information
afforded by EMG sensors could be recovered using the MMG-based strategies proposed.
Data Aquisition
MMG and EMG signals were simultaneously acquired from two clients of the Muscu-
loskeletal Program at the Bloorview MacMillan Children’s Centre using custom-made
soft silicone sockets. An array of three coupled MMG sensors (described in Section 4.1)
was embedded distally (i.e. approximately 1.5 cm from the distal end of the client’s
stump) around the silicone sleeve. Figure 5.8 shows one of the constructed sockets. Note
the equiangular sensor distribution around the socket (i.e. equally spaced at angles of
1200 ). The chosen number of sensors in the array was governed by the maximum number
of connection wires that can be currently implemented in the interface between the soft
silicon socket and the additional processing circuitry in the hard shell of an externally
powered prosthesis (see Figure 2.3). The subjects were below-elbow amputees and cur-
rent EMG-driven prosthesis users. Subject 1 was a traumatic amputee and subject 2 was
a congenital amputee. Recordings were performed for three different activities namely,
wrist flexion, wrist extension and random limb movement. The activities named “wrist
flexion” and “wrist extension” were defined as the contractions currently performed by
the EMG-driven prosthesis users to respectively close and open their prosthetic hand.
Feedback was provided to subjects by allowing simultaneous EMG-based control of a
free-standing prosthetic hand. The necessary EMG electrodes3 for such control were
placed underneath the silicon sleeve at the muscle sites used in their current prosthesis.
For the first set of recordings the subjects were asked to perform three “normal”
strength contractions to alternately open and close the free-standing prosthetic hand. For
the second set, the instructions were the same except that this time “strong” contractions
were solicited from the patients. A third set of recordings consisted of short repetitive
contractions to either open or close the free-standing prosthetic hand. With the fourth
set, clients were asked to randomly move their residual limbs. Ten repeated recordings
3
13E125 myoelectric electrodes from Otto Bock Group
Figure 5.8: Sample soft silicone socket built for MMG signal recording. Note the em-
bedded multisensor array containing three coupled MMG sensors at equi-distant angles
around the distal end of the stump.
with an average duration of 7s were performed for each set. Signals from each of the three
accelerometers and microphones in the array as well as from two EMG electrodes (i.e.
one for flexion and one for extension) were acquired using the MATLAB environment for
a total of 8 signals sampled at 1kHz and stored for further analysis. Figure 5.9 portrays
the experimental setup for one of the EMG-driven prosthesis users.
Target Signals
As previously mentioned, the information obtained from the EMG sensors was assumed
to be sufficient for the reliable estimation of the true user intention in every test. The
EMG sensor outputs were already rectified and filtered (Figure 5.10 top). Therefore,
together they provided an analog estimate of the energy of the measured signals. For the
generation of the target detection signal, a program was written in MATLAB to threshold
both EMG signals (i.e. from wrist flexors and extensors) at 0.05V. This threshold is
sufficient to recover all the electrical activity due to muscle contraction in the EMG
sensors. This produced a binary output indicating the presence of a muscle contraction
whenever either of the EMG signals exceeded the threshold (middle graph of Figure 5.10).
Figure 5.9: Experimental setup for real MMG signal recordings from an EMG-driven
prosthesis user.
To generate the target classification signal, the electrical signal from the wrist ex-
tensors was subtracted from the electrical signal from the wrist flexors. In this way, a
target estimate of the true distance dT for the algorithm presented in Equation 5.6 can
be obtained. Note that in this case, co-contractions will tend to zero while well differenti-
ated contractions will be positive for flexion and negative for extension. Therefore, once
contractions have been detected, the sign of the true distance estimate will determine
the type of contraction (i.e. hypothesis Hf or He ).
Figure 5.10 shows an example of acquired EMG signals for one of the subjects. The
top graph shows the original EMG signals acquired, and the middle and bottom graphs
show the target binary detection and distance dT obtained with the above procedure.
Training
As mentioned earlier, both the RMS-based detection of muscle contraction and the RMS-
based classification algorithms require the optimization of certain parameters to maximize
detection and classification accuracy (i.e. γ values for detection and vector θ̂ for classi-
fication). Therefore, 60% of the available recordings (i.e. 60% of signals with “normal”,
“strong” and “short” contractions) were used for training or parameter optimization.
Figure 5.10: Original EMG signals (top) and derived targets (middle and bottom) for
the optimization of the proposed MMG-based strategy.

Processing windows of 0.2s duration were used in all cases. This window size allows
us to recover information from the lowest frequency considered for MMG signals (5Hz)
while still allowing enough time to provide a natural real-time response. To avoid patient
frustration and confusion, the maximum acceptable delay between the beginning of the
contraction and the prosthesis response is reported to be 300ms [15].
For the optimization of the detection γ values, the target detection signal obtained
from the EMG sensors was compared to the detection signal obtained from the proposed
RBD algorithm. In this case, the outcome measure is the ratio of incorrect decisions
to the total number of possible decisions (i.e. total number of 0.2s windows). The
optimization problem is then to find the set of γ values that minimize this ratio.
For the optimization of the classification vector θ̂, the target distances obtained from
the EMG signals were compared to the calculated distances using Equation 5.6. The
outcome measure in this case is the MSE between these distances. The optimization
problem is thus to find the vector θ̂ which minimizes the calculated MSE. Note that the
vector θ̂ represents the parameters of the plane in the 3D feature space which maximally
separates wrist flexion and extension.
Both optimizations were performed using the Nelder-Mead algorithm for function
minimization [22] within the MATLAB environment.
Strategy Testing and Cross-Validation
Once the optimal values of γ and θ̂ were obtained, the remaining 40% of available signals
were used to test the performance of the proposed strategies on untrained data. For both
cases (i.e. detection and classification), performance is reported as the ratio between
correct decisions (as compared to the target EMG-derived signals) and the total number
of possible decisions or 0.2s windows available.
In order to validate the proposed strategy, a ten-fold cross-validation was performed.
Cross-validation provides a statistical distribution of performance for repeated train and

Figure 5.11: Mean and standard deviation values of the cross-validated accuracy of de-
tection for each subject and each proposed strategy of Table 5.2.
test procedures in which subsets for training and testing are randomly selected each time.
The mean across trials then represents a more robust indicator of the performance of the
proposed strategies [10].
5.2.6 Results
Detection
Figure 5.11 shows the mean and standard deviation values of the cross-validated accuracy
of detection for subjects one (top) and two (bottom). These evaluations were performed
using the MMG-based strategies proposed in Table 5.2, using the EMG information
(target EMG-derived signals) as the gold standard.
Note that the accuracy for subject one is highest (0.8213±0.0145) when the second
Figure 5.12: Sample MMG-based detection (top) and EMG-based detection (bottom) of
muscle contraction.
strategy is used while the fourth strategy provides the highest accuracy (0.8251±0.0077)
for subject two.
Figure 5.12 shows a sample MMG-based detection signal (top) and the correspond-
ing EMG-based detection signal (bottom). For reference, the original signals from the
microphones in the coupled MMG sensors and EMG electrodes are superimposed. It is
important to note that, as shown in Figure 5.12 the error reported in MMG-based detec-
tion is largely due to additional muscle activity detected exclusively by MMG sensors.
This phenomenon is somewhat expected and perhaps desirable, given that the activity
of deep muscles can be detected by MMG sensors, but not surface EMG electrodes. The
additionally detected activity is considered erroneous when compared to EMG detection.
Hence, the accuracies reported above must be interpreted with care since they are only
relative to the performance or current EMG sensors.

Figure 5.13: Sample MMG-based estimation of muscle activity (top) and its correspond-
ing EMG-based counterpart (bottom).
Classification
Once the most accurate strategies from Table 5.2 have been chosen, it is possible then to
train the classification algorithm, yielding MMG-based estimates of muscle activity (i.e.
the distance d in Equation 5.6) that are optimally fitted to the EMG-derived estimates
of true activity.
Figure 5.13 depicts a sample normalized MMG-based estimate of muscle activity (top)
and its corresponding normalized EMG-derived estimate of true activity (bottom) for
subject two. The MMG-based estimate was calculated using the optimal values for vector
θ̂ obtained during training. As previously explained, additional activity unbeknownst to
the EMG sensors was detected by the MMG sensors. It is also important to note that
the amplitudes of the MMG estimates may not correlate with those of the EMG-derived
estimates since MMG and EMG signals arise from different physiological phenomena.
Finally, there is a small temporal delay in the MMG-based estimates. This delay may
be due to the fact that mechanical effects always succeed electrical signals (commands)
sent by the neuromotor control system to generate muscle contractions. However, despite
these minor differences between MMG and EMG estimates of muscle activity, our results
demonstrate that muscle activity can be tracked with MMG sensors in a way similar to
that of EMG sensors. In other words, most of the information provided by EMG sensors is
also contained in the signals acquired with MMG sensors. Furthermore, from the cross-
validation tests, we obtained classification accuracies of 0.7090 ± 0.0394 and 0.7001 ±
0.0519 for the first and second subjects, respectively. These accuracy measures represent
the ratio between correct classifications and the total number of decisions available. In
this case, for every available point in the recorded signals, a correct classification is defined
as an agreement between the signs of the EMG-derived estimate and the MMG-based
estimate of muscle activity. As with detection tests, it is also important to note that the
reported accuracies only describe the agreement between MMG and EMG estimations
of muscle activity. The proper interpretation of the reported accuracies will be discussed
below.
5.2.7 Discussion
Through a comparison to a simple EMG-based method, we have demonstrated that the
proposed MMG-based strategies provide sufficient information for the detection, estima-
tion and classification of muscle activity. However, the potential of MMG signals far
exceeds the simple emulation of EMG-based muscle activity detection. The following
physiological, usability, robustness and functional considerations must factor into the
careful interpretation of the reported MMG classification accuracies.
1. The ubiquitous detection capability of the MMG sensors is not accounted for in
the accuracy measure. The accuracies reported above specify the amount of in-
formation gathered by EMG sensors that can be recovered using MMG sensors.
Although EMG-derived target signals were assumed correct, they do not necessar-
ily reflect the users’ true intentions. In other words, the comparisons conducted in
the preceeding sections attempted to fit the available information from the MMG
sensor array into current EMG-based standards, an excercise which may actually
limit the potential of MMG-based systems. Recall that MMG sensors are capable
of detecting contraction signals from virtually every muscle in the residual limb of
the client, due to MMG signal propagation properties, while surface EMG is lim-
ited to the contraction of superficial muscles. This ubiquitous detection capability
of MMG sensors may naturally favour the generation of multiple (> 2) control
outputs.
2. The relatively high MMG accuracy implies that MMG-driven prostheses may be
highly usable by current EMG system users. We can also discuss the reported
performance measures from the clients’ perspective. If EMG-derived target signals
are assumed sufficient for the estimation of the users’ true intention, the values
reported for accuracy would indicate the percentage of this intention that would be
correctly detected and classified with the proposed MMG-based strategy without
the need for additional training. This means that, if a client switches from EMG- to
MMG-driven prostheses, about a 60% detection and classification accuracy would
be retained prior to any additional training.
3. Client training is likely to further enhance MMG classification accuracy. The previ-
ous observation also implies that client training in the use of MMG-driven prosthesis
may considerably overcome the initial errors reported here. In the same way a ded-
icated musician masters the use of any musical instrument, MMG-driven prosthesis
users may gradually learn to target every contraction to the correct location in the
defined 3D feature space.
4. The robustness of the MMG system to limb movement is not considered in the
accuracy measure. In practical applications, EMG-driven systems will always have

a lower performance than the one assumed in this research. Recall that every opti-
mization was performed using only “normal”, “strong” and short contractions and
thresholding the EMG signals at 0.05V. In practice, this threshold is significantly
higher due to the considerable amplitude of EMG signals acquired during random
limb movement. Figure 5.14 shows a comparison between a sample MMG-based
detection (top) and a sample EMG-based detection during random limb movement
(bottom). Original sensor signals are included in the background as a reference.
The accelerometer signals are also included in the middle graph to emphazise their
importance in these circumstances. Note that, at a threshold of 0.05V most of
the movement is incorrectly detected as intentional muscle contraction when using
EMG electrodes. On the other hand, the MMG-based strategy proposed is capa-
ble of eliminating most of this noise (top graph in Figure 5.14). The dashed line
in the bottom graph indicates a more realistic threshold (0.3V) for practical use
in EMG-driven prosthesis, which would avoid most of the observed misdetections.
With the proposed MMG-based strategies, there is no need for ad-hoc parameter
adjustment to account for limb movement. Furthermore, the proposed MMG-based
strategy has the potential to detect and make use of weaker contractions than those
employed by EMG systems.
5. The ideal training target is user intention, which may more truly reflect multifunc-
tion MMG classification performance. The optimization of the proposed MMG-
based strategy is ultimately not dependent on the availability of EMG signals. The
algorithms may use the expressed client intention as a better target for training.
Furthermore, a more complex classifier which optimally differentiates among more
than two activities may also be designed and applied to the generation of mul-
tiple outputs. As mentioned earlier, MMG signal propagation properties endow
MMG-based control strategies with potential that surpasses that of their EMG
Figure 5.14: Sample MMG-based detection of muscle activity (top) and its corresponding
EMG-based counterpart (bottom) during random limb movement. The middle graph
shows the accelerometer signals in the three coupled MMG sensors in the array.
counterparts.
5.2.8 Conclusions
The potential of MMG signals for prosthesis control has been demonstrated. Further-
more, as compared to EMG-based systems, MMG sensors can provide enough information
for the detection, estimation and classification of muscle activity.
The results lay the groundwork for the development of the first MMG-driven proto-
type prosthesis. The succesful evaluation of such a prototype by prosthesis users would
provide further support and justification for the application of MMG in prosthesis control.
Chapter 6
Conclusions
6.1 Contributions
The work presented in this thesis has solidly laid the groundwork for the development
of a new generation of MMG-driven externally powered prosthesis. Contributions to the
field of Rehabilitation Engineering, particularly in the area of mechanomyography and
externally powered upper extremity prosthetics, are listed below.
1. A procedure for embedding micromachined accelerometers into silicon for the pur-
pose of MMG measurement [31, 32].
2. Characterization of silicon embedded acclerometers and the optimization of SNR

through the adjustment of silicon embedding properties [31, 32].
3. A procedure for embedding micromachined microphones into silicon for MMG
recording, including a unique design and practical implementation of a silicon air
chamber and contact membrane [33].
4. Characterization of silicon-embedded microphones and SNR optimization through
the adjustment of silicon-embedding properties and air chamber geometry [33]
105
Chapter 6. Conclusions 106
5. Systematic comparison of silicon-embedded accelerometers and mircophones in

terms of SNR in MMG measurement [31, 32, 33].
6. The novel design and multistep manufacturing process for a coupled microhpone-
accelerometer MMG sensor pair for dynamic noise reduction in practical MMG
signal recording [29]. The implemented coupled MMG sensor not only facilitates
noise discrimination in prosthesis control but may also be used in many other
clinical applications such as force measurements, fatigue monitoring and evaluation

of muscle mechanics [36].
7. The development and theoretical and empirical validation of a mechanical model of
the novel sensor pair for MMG measurement. This model provided insights on the
dynamic interactions of the sensor components and the mathematical relationships
between individual transducers in a coupled MMG sensor.
8. The first ever reported design and application of a sensor array for distal MMG
signal recording [30].
9. The demonstration that activity from multiple forearm muscles can be measured
reliably and distally from contracting muscles, without specific sensor-to-muscle

assignment. This contribution is a complete departure from the concepts of con-
ventional muscle activity monitoring, where sensors are anatomically located at
and functionally mapped to single muscle sites in a rigid one-to-one fashion. In
the conventional case, the location of each sensor involves a painstaking search for
the optimal recording site. With the proposed MMG sensor array, this search is no
longer necessary.
10. The design of a simple MMG-based multisensor fusion strategy for the detection
and discrimination of muscle activity. This strategy demonstrates the usefulness
of MMG signals in prosthesis control and represents a starting point for more
complex multisensor algorithms. Such algorithms could in turn further enhance

the performance of current MMG systems.
11. Inaugral demonstration that the information provided by an MMG sensor array is a
superset of the information contained in a conventional two-site EMG configuration

for powered upper limb control in two amputees.
6.2 Future Work
6.2.1 The Prototype
As mentioned at the end of the previous chapter, the work presented in this thesis sup-
ports the implementation of a prototype MMG-driven externally powered prosthesis for
below-elbow amputees. An evaluation of such a prototype by current and new below-
elbow prosthesis users would help to assess the potential enhancement of functionality and
comfort provided by the combination of MMG sensor arrays and silicon soft sockets. Fur-
thermore, the prototype implementation and evaluation will reveal practical challenges
and the necessary steps towards the establishment of standards in the development of a
new generation of externally powered prostheses for upper-limb amputees.
Some important considerations for an initial evaluation are:
1. Functionality. The development of the prototype prosthesis will facilitate functional
performance comparisons between EMG and MMG-driven systems. This compar-
isons will help to identify the relative merits and shortcomings of both types of
systems, and will reveal specific conditions in which current EMG-driven prosthe-
ses be functionally inadequate. Section 5.2 already showed a clear comparison in
which the performance of MMG-based systems is superior to that of EMG under
extreme noise conditions (e.g. limb movement).

2. Training. It is expected that a new prosthesis user will feel more comfortable learn-
ing to control an MMG-driven prosthesis rather than an EMG-driven prosthesis.
This is due to MMG signal propagation which facilitates detection of muscle activ-
ity from the entire residual limb, which together with the classification algorithm
proposed, provides more natural control. On the other hand, EMG sensors can only
measure muscle activity from a single site. This requires the accurate activation
of individual muscles. Therefore, current issues with muscle activity classification

caused by co-contraction may be overcome with simple MMG detection strategies.
In turn, these strategies would significantly reduce the user training period.
3. Adaptability. Although it is not essential, it is desirable that current EMG-driven
prosthesis users do not have any difficulty switching to MMG-driven prosthesis.
In other words, the adaptation period should ideally be minimized. A prototype

MMG-driven prosthesis would provide insights into the proper methods necessary
to facilitate this transition.
4. Cost. Hardware and software implementation of MMG-driven systems are initially
less costly than their EMG counterparts. The construction of a prototype prosthesis
would provide clear cost estimations.
MMG-driven prostheses will soon become a real alternative at the Bloorview MacMil-
lan Centre. However, they not only have potential as an alternative to current technolo-
gies but as means to further develop the field of Rehabilitation Engineering by advancing
the frontier of externally powered prosthetics research.
6.2.2 The Model
The model-based source separation strategy proposed in Section 4.2 was not sufficiently
validated with real MMG measurements. This means that all or part of the assumptions
for the development of the model for data fusion within the coupled MMG sensor pair
may be further improved upon. There is no doubt that the development of a reliable
model will enhance the performance of MMG-driven systems, simplifying hardware and
software for prosthesis control. Some suggestions for the enhancement of the proposed
model are as follows.
1. The proposed model assumes point sources where radial effects of the measurement
system represented by the residual limb and silicon sleeve were ignored. This might
have been a sufficient approximation for the experimental tests but not for real
MMG measurements. Therefore, the integration of radial effects into the model
should be investigated.
2. Ideal linear components were also assumed in the proposed model. This may have
considerably affected model performance especially in modeling the silicon sleeve.
Therefore, numerical models for continous mechanics should also be investigated,
for example, through the application of finite element analysis (FEA).
Spatial Force Estimation
The proposed model in Section 4.2 deals with the mechanical interactions within a single
coupled MMG sensor pair. However, it is also possible to define a model-based sensor fu-
sion strategy which tracks the mechanical interactions among potential sources (muscles)
and surrounding media within the forearm.
During a sustained contraction the muscle can be successfully modeled as a vibrating
sphere moving along a single dimension in space [14]. This model also corresponds to
an acoustic dipole [5]. However, if we are to measure MMG signals at multiple locations
around a cross-section of the muscle, an additional acoustic dipole (perpendicular to the
original, and vibrating 900 out of phase) is necessary to recreate previously reported
observations by Ouamer et al. (1999) [26]. Therefore, the displacement of the skin in the
Figure 6.1: Schematic diagram of a cross-sectional MMG measurement and the associated
parameters.
direction of the transducer can be approximated for a single frequency ω as:
3|F |cos(φ)
lM (t, ω) = − sin(ωt − kr − β) (6.1)
2πρr 3 ω 2
where k is the wave number, ρ is the density of the media surrounding the muscle,
r 6 β are the polar coordinates of the distance vector between the source (muscle) and
the transducer and cos(φ) is the attenuation factor due to the relative orientation of
the muscle and the transducer (see Figure 6.1). Note that the force F, which can be
associated to the strength of the contraction, modulates the amplitude of the measured
sinusoidal vibration. This model of muscle contraction can be used to generate a spatial
force estimation algorithm which tracks the muscle activity on the forearm cross-section
monitored by the MMG sensor array. It can be seen from equation 6.1 that, if we consider
the phase delay kr produced by the time delay of arrival (TDOF) as negligible (kr β),
microphone pairs should have a non-frequency dependent difference in phase equal to the
difference between β angles for every potential source. Then, a phase-error mask such as
the one depicted in Equation 6.2 can be defined,
Φij (ω) = 0.5[1 + cos(6 li (ω) − 6 lj (ω) − βij )] (6.2)
where 6 li (ω) and 6 lj (ω) are the phases of the discrete fourier transform (DTF) of each
corresponding relative membrane position signals from the two coupled MMG sensors
in the array used to generate the mask and βij is the difference in β angles among the
microphones from which the mentioned displacement signals are derived. This mask
will have no effect on frequency components whose phase is equal to the difference βij
while attenuating frequency components whose phase differs from it. For a combination
of three sensors, two masks will be obtained for each skin vibration signal. These can
be averaged or multiplied to obtain the overall masking effect. A cross-section of the

forearm may then be divided in a matrix of potential sources applying the mask 6.2 to
each. This would provide a map of muscle activity (i.e. spatial force estimation) on the
monitored forearm cross-section.
This strategy, together with an enhanced model of the coupled MMG sensor, could
eventually facilitate the generation of multiple real-time quasi-analog signals for the con-
trol of complex terminal mechanisms. The optimization of such algorithms could con-
ceivably allow control of extremely complex tasks such as digit manipulation.
6.2.3 Pattern Recognition
The classification strategy proposed only uses one feature (i.e. normalized RMS) per
sensor in the array and applies a simple linear discriminant to differentiate between
flexions and extensions. However, it is also possible to increase the dimensions of the
feature space by including additional signal features such as peak frequency and relative
phase differences among sensors. This may initially enhance signal classification and
ultimately provide additional control signals.
As suggested in the foregoing discussion, the full potential of MMG-driven systems is
yet to be exploited. This thesis has laid the groundwork for such exploration, particularly
in the utility of MMG signals, and has provided a solid foundation for the development
of MMG-driven prosthesis for upper limb amputees.

Bibliography
[1] Vibrations and Waves. Norton & Company, Inc., 1971.
[2] Electromyography for Experimentalists. The University of Chicago Press, 1986.
[3] Log-Linear Models. Springer-Verlag, 1990.
[4] Linear Control System Analysis and Design. Mc-Graw Hill, 1995.
[5] Engineering Noise Control. E & FN Spon, 1996.
[6] Signals & Systems. Prentice Hall, 1997.
[7] Introduction to Surface Electromyography. Aspen Publishers Inc., 1998.
[8] Multi-Sensor Fusion: Fundamentals and Applications with Software. Prentice Hall,
1998.
[9] Practical Nonparametric Statistics. John Wiley & Sons, 3rd edition, 1999.
[10] Pattern Classification. John Wiley & Sons, 2000.
[11] A Anstey. Pediatric prosthetics: A demographic study of pediatric amputee inci-
dences at the bloorview macmillan centre, May 2003.
[12] DT Barry and NM Cole. Muscle sounds are emitted at the resonant frequencies
of skeletal muscle. IEEE Transactions on Biomedical Engineering, 37(5):525–531,
1990.
112
Bibliography 113
[13] DT Barry, JA Leonard, AJ Gitter, and RD Ball. Acoustic myography as a con-

trol signal for an externally powered prosthesis. Archives of Physical and Medical
Rehabilitation, 67(4):267–269, 1986.
[14] A Courteville, T Gharbi, and JY Cornu. Mmg measurement: A high-sensitivity
microphone-based sensor for clinical use. IEEE Transactions on Biomedical Engi-

neering, 45(2):145–150, 1998.
[15] K Englehart, B Hudgins, P Parker, and RN Scott. A microprocessor-based multi-
function myoelectric control system. In Proceedings of the 23rd Canadian Medical

and Biological Engineering Conference, May 1999.
[16] TAZK Gaber, CM Gardner, and SGB Kirker. Silicone roll-on suspension for upper
limb prosthesis: User’s views. Prosthetics and Orthotics International, 25(2):113–
118, 2001.
[17] MS Goldenberg, HJ Yack, FJ Cerny, and HW Burton. Acoustic myography as an
indicator of force during sustained contractions of a small hand muscle. Journal of
Applied Physiology, 70(1):87–91, 1991.
[18] DL Hall and J Llinas. An introduction to multisensor data fusion. Proceedings of
the IEEE, 85(1):6–23, 1997.
[19] M Heim, M Wershavski, ST Zwas, I Siev-ner, H Nadvorna, and M Azaria. Silicone
suspension of external prosthesis: a new era in artificial limb usage. Journal of Bone
Joint Surgery, 79-B:638–640, 1997.
[20] D Kalanovic, MP Ottensmeyer, J Gross, G Buess, and SL Dawson. Independent
testing of soft tissue visco-elasticity using indentation and rotary shear deformations.
In Proceedings of the 11th Annual Medicine Meets Virtual Reality Conference, pages
137–143, January 2003.

Bibliography 114
[21] AE Kerdok, SM Cotin, MP Ottensmeyer, A Galea, RD Howe, and SL Dawson.

Truth cube: Establishing physical standards for real time soft tissue simulation. In
Proceedings of the International Workshop on Deformable Modelling and Soft Tissue
Simulation, November 2001.
[22] JA Nelder and R Mead. A simplex method for function minimization. Computer
Journal, 7:308–313, 1965.
[23] C Orizio. Muscle sound: Bases for the introduction of a mechanomyographic signal
in muscle studies. Critical Reviews in Biomedical Engineering, 21(3):201–243, 1993.
[24] C Orizio, D Liberati, C Locatelli, D De Grandis, and A Veicsteinas. Surface
mechanomyogram reflects muscle fibres twitches summation. Journal of Biome-
chanics, 29(4):475–481, 1996.
[25] C Orizio, R Perini, B Diemont, M Maranzana Figini, and A Veicsteinas. Spectral

analysis of muscular sound during isometric contraction of biceps brachii. Journal
of Applied Physiology, 68(2):508–512, 1990.
[26] M Ouamer, M Boiteux, M Petitjean, L Travens, and A Sal‘es. Acoustic myography
during voluntary isometric contraction reveals non-propagative lateral vibration.

Journal of Biomechanics, 32(12):1279–1285, 1999.
[27] Y Salam. The use of silicone suspension sleeves with myoelectric fittings. Journal
of Prosthetics and Orthotics, 6(4):119–120, 1994.
[28] Information & Statistics Division NHS Scotland. The amputee statistical database
for the united kingdom 2001/02. Technical report, National Amputee Statistical
Database, 2003.
Bibliography 115
[29] J Silva and T Chau. Coupled microphone-accelerometer sensor pair for dynamic
noise reduction in mmg signal recording. Electronics Letters, 39(21):1496–1498,
2003.
[30] J Silva, T Chau, and AA Goldenberg. Mmg-based multisensor data fusion for pros-
thesis control. In Proceedings of the 25th Anual International Conference of the
EMBS, pages 2909–2912, september 2003.
[31] J Silva, T Chau, and S Naumann. Quantifying the signal-to-noise ratio of silicon-
embedded sensors for mechanomyography. In Proceedings of the 27th Canadian
Medical and Biological Engineering Conference. CMBES, November 2002.
[32] J Silva, T Chau, S Naumann, and W Heim. Systematic characterisation of silicon-

embedded accelerometers for mechanomyography. Medical & Biological Engineering
& Computing, 41(3):290–295, 2003.
[33] J Silva, T Chau, S Naumann, W Heim, and AA Goldenberg. Optimization of the
signal-to-noise ratio of silicon embedded microphones for mechanomyography. In
Proceedings of the Canadian Conference on Electrical and Computer Engineering,
volume 3, pages 1493–1496. IEEE, May 2003.
[34] PK Varshney. Multisensor data fusion. Electronics & Communication Engineering
Journal, pages 245–253, 1997.
[35] M Watakabe, K Mita, K Akataki, and Y Itoh. Mechanical bahaviour of condenser

microphone in mechanomyography. Medical & Biological Engineering & Computing,
39(2):195–201, 2001.
[36] JC Wood and DT Barry. Time-frequency analysis of skeletal muscle and cardiac
vibrations. Proceedings of the IEEE, 84(9):1281–1294, 1996.

Bibliography 116
[37] YT Zhang, C Basil, R Mandayam, and G Douglas. A comparative study of simulta-

neous vibromyography and electromyography with active human quadriceps. IEEE
Transactions on Biomedical Engineering, 39(10):1045–1052, 1992.
[38] LQ Zhao, YT Zhang, W Herzog, and DF Yuan. A theoretical study of muscle
vibrations: The influence of transducer mass on vibromyographic signals. In 20th
Canadian Medical and Biological Engineering Conference. CMBES, 1994.
View publication stats

Mechanomyography Sensor Design and Multisensor Fusion For Upper-Limb Prosthesis Control

Uploaded by

Copyright:

Available Formats

Mechanomyography Sensor Design and Multisensor Fusion For Upper-Limb Prosthesis Control

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mechanomyography Sensor Design and Multisensor Fusion For Upper-Limb Prosthesis Control

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Mechanomyography Sensor Design and Multisensor Fusion for Upper-Limb

Activity Detection View project

Brain-Computer Interfaces View project

The user has requested enhancement of the downloaded file.

A thesis submitted in conformity with the requirements

c 2004 by Jorge Silva

Mechanomyography Sensor Design and Multisensor Fusion for Upper-Limb Prosthesis

Master of Applied Sciences

Graduate Department of Mechanical & Industrial Engineering

of remnant muscles (mechanomyography or MMG). This document details the technical

development of a new generation of MMG-driven upper-limb externally powered prosthe-

sis. Fundamental contributions include a systematic characterization and optimization

powered prostheses research.

A mis padres, mi amor y respeto incondicionales.

To my supervisor, my endless admiration and gratitude.

To my partner, my most valuable possession: my life.

ception Laboratory; and to Dr. Z. W. Zu in the Vibration Laboratory at the University

resources for the development of this work.

1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2.1 Quantifying the Signal-to-Noise Ratio of Silicon-Embedded Sensors for

2.1.3 Materials & Methods . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2 Systematic Characterisation of Silicon-Embedded Accelerometers for Mechano-

2.2.3 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . 18

3.1 Optimization of the Signal-to-Noise Ratio of Silicon-Embedded Micro-

phones for Mechanomyography . . . . . . . . . . . . . . . . . . . . . . . 33

3.1.4 MMG Sensors & Prosthesis Control . . . . . . . . . . . . . . . . . 35

3.1.5 Materials & Methods . . . . . . . . . . . . . . . . . . . . . . . . . 36

4 Coupled MMG Sensor Pair 42

4.1 Coupled Microphone-Accelerometer Sensor Pair for Dynamic Noise Re-

duction in MMG Signal Recording . . . . . . . . . . . . . . . . . . . . . 43

4.1.3 Sensor Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.1.4 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4.2 Coupled MMG Sensor Pair: Model Validation and Optimization . . . . . 49

4.2.1 Rationale: Model-Based Source Separation . . . . . . . . . . . . . 49

4.2.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4.2.3 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.2 MMG-Based Multisensor Fusion Strategy for Detection and Classification

5.2.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

5.2.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

6.2.1 The Prototype . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

6.2.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

2.1 Embedding properties of the prepared samples. . . . . . . . . . . . . . . 9

2.5 SNR values (dB) measured in embedded and non-embedded transducers.

Increasing applied forces are labelled F1 -F4 . . . . . . . . . . . . . . . . . 25

3.2 Normalized sensitivity values obtained for each combination of embedding

5.1 Performance measures for different sets of values of γ . . . . . . . . . . . 81

5.2 Possible pre-processing combinations for MMG-based detection and clas-

sification of muscle contractions. . . . . . . . . . . . . . . . . . . . . . . . 92

2.1 Schematic diagram of the silicon-embedded accelerometers prepared. “TT L ”

2.2 Schematic diagram of the mechanical stimulator. Samples were placed

2.3 Schematic diagram of externally powered below-elbow prosthesis. Cur-

that holds terminal element is adjusted over the hard socket . . . . . . . 16

sise its flexibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.5 Schematic diagram of silicon-embedded accelerometer samples. TT L =top

layer thickness; TBL =bottom layer thickness . . . . . . . . . . . . . . . . 19

sponse from sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20