PARTICIPANT

■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 1 HIGH-TECH

SYSTEMS 2014

THE TRANSFORMATION
FROM TIME TO
FREQUENCY DOMAIN
A series of three articles will feature Frequency Response Function
(FRF) measurements. The FRF, describing the frequency-dependent
AUTHORS’ NOTE
behaviour of linear systems, is an indispensable tool in the engineering
Pieter Nuij, senior system
of dynamic systems. This series is not meant to be mathematically
architect, and David rigorous but aims at an understanding of the main steps to determine
Rijlaarsdam, group leader,
both work at NTS Systems
the FRF based on real measurements from an application point of view.
Development in Eindhoven, Part 1 deals with the transformation of time-domain signals to the
the Netherlands.
frequency domain and may serve as a stand-alone guide for basic signal
pieter.nuij@nts-group.nl
www.nts-group.nl analysis.
PIETER NUIJ AND DAVID RIJLAARSDAM

Why transform? Discrete spectrum

Although we ‘live’ in the time domain and naturally ask To determine the spectrum of a continuous time signal the
ourselves when things happen, asking the question how signal must be converted into a sequence of numbers that
often things happen often leads to revealing answers. Figure form the input for a numerical algorithm running in a
1 shows a time signal of the torque measured through the computer. Figure 3 shows the necessary steps for this
outgoing shaft of a gearbox. The spectrum of the same conversion.
signal is displayed in Figure 2 and clearly shows discrete
frequency components that can be related to the gear ratios The continuous time signal is conditioned and amplified/
of the gearbox. Although the time signal also contains this attenuated in an analogue electronics stage before it is low-
information, it is not visible in the time trace in Figure 1. pass-filtered to limit its frequency content. After filtering,
the signal is sampled and becomes a discrete time signal.
The individual samples are converted from an analogue
signal into a quantised signal with discrete values.
Three articles
Limiting the frequency content of the continuous time
This article covers the steps that are necessary to convert a time- signal before sampling is important, but it is sometimes
continuous signal into a discrete spectrum. Potential errors due to neglected, causing erroneous results. Figure 4a shows a
aliasing and leakage will be explained and solutions will be presented. continuous time signal with frequency fsignal that is sampled
An overview of several types of test signals will be presented. The with sampling frequency fsample. The resulting discrete time
second article will introduce the FRF, explaining the choice of test signal is a true representation of the continuous time signal.
signals in their relation to the coherence function and the In Figure 4b a different signal with frequency fsignal + fsample
measurement of the FRF in open-loop and in closed-loop systems. The is sampled with the same sampling frequency fsample. The
third article will focus on the extension of frequency-domain methods reconstructed discrete time signal seems to be the
towards nonlinear systems and the application of such methods to representation of a continuous time signal with frequency
define and optimise the performance of such systems. Each article will fsignal.
be illustrated with examples. This example clearly shows that sampling can introduce
errors called aliasing components. To prevent this from

8 MIKRONIEK nr 2 2014
 1 Time trace of torque
measurement.
2 Spectrum of torque
measurement.
3 Necessary steps to
convert a continuous
time signal into
numerical data points.

happening, the sampling frequency must be at least twice

the frequency of the highest frequency component present
in the continuous time signal. Since the frequency content 1

of the signal to be analysed is often unknown, low-pass

filtering to limit the frequency content is required. The
filter used for this purpose is called an anti-aliasing filter.

Like any filter the anti-aliasing filter will introduce

frequency-dependent phase shifts in the filtered signal. This
is not a problem when measuring FRFs as long as all
measurement channels used for the FRF calculations
experience equal phase shift. In the last block in Figure 3
the discrete time analogue signal is quantised, rounded to
the nearest discrete amplitude value of the 2n possible values
that can be represented by an n-bits digital word. Optimal
use of the amount of bits requires matching of the
amplitude of the signal to the voltage range that can be
accommodated by the AD converter (ADC). This is done
by choosing the correct amplification or attenuation factors
in the signal-conditioning front-end. Poor ranging will
cause harmonic distortion and a high noise floor but due to
the advent of high-quality 24-bits ADCs the ranging
requirements become less stringent. 2

Certified Precision
Engineer competencies
The content of this series of three articles is in
part covered in the “Experimental Techniques in
Mechatronics” course, provided by The High
Tech Institute (HTI). This course has been
selected for the DSPE Certification Program (see
page 58).
WWW.DSPEREGISTRATION.NL WWW.HIGHTECHINSTITUTE.NL
3

nr 2 2014 MIKRONIEK 9
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 1

4 Introduction of errors by
sampling.
(a) Signal (blue) with
frequency fsignal sampled
with frequency fsample
(red).
(b) Signal (blue) with
frequency fsignal + fsample
sampled with frequency
fsample (red). The
reconstructed signal is
the aliasing component
4a (black).
5 Analysis of time signal
x(t) in blue.
(a) Periodically
reconstructed time
signal based on
measurement signal
x1(t) in red.
(b) Some of the signal
components present in
x1(t).
4b

with frequencies being a multiple of 1/T Hz. So if T is 1 s,

x(t) will consist of a series of harmonics of 1 Hz.
A segment x1(t) of x(t) is captured and this measurement
record will be processed. If the length T1 of the segment
x1(t) is equal to T, then the result of the FFT will correctly
describe the frequency content of the time signal x(t). This
can be seen in Figure 5a, where the repetition of the
measurement record x1(t) results in a signal equal to the
5a
original signal x(t). Figure 5b shows some of the cosine
components present in the signal. The amplitude and
phase values of these cosine components are plotted in the
amplitude and phase spectra as displayed in Figure 6. The
units on the x axis in Figure 6 are cycles/T1, with T1 the
period time of the measurement record x1(t).

T1 can be calculated from the sampling frequency ( fs ) and

the size of the measurement record:
5b

T1 = # samples / fs

Knowing T1, the frequency axis can be scaled to cycles per

second [Hz]. The resulting frequency resolution Δf of the
Frequency analysis spectrum is equal to 1/T1. The highest frequency
The common way to describe the frequency content of a component that can be present in the spectrum has a value
time signal is to use a series of cosines with specific of fs /2 Hz. If a different segment of x(t) is captured, but the
frequencies, amplitudes and phases. The Fast Fourier length of the segment remains T1, the amplitude spectrum
Transform (FFT) is a very efficient algorithm for calculating will remain the same but the phase spectrum will be
these cosine-based signal components. However, as will be different. If the length of the segment is changed, the
demonstrated, the results must be interpreted carefully. amplitude spectrum will also change.

Figure 5a shows a time signal x(t) of which frequency Figure 7 shows the situation that T2, the length of the
components are to be identified (= transformed from time measurement record x2(t), is not an integer amount of
to frequency domain). Since the signal is periodic with a periods T of x(t) any longer. The FFT algorithm will again
period of T, it consists of a series of cosine components determine the best fit of a series of cosines, but this time

10 MIKRONIEK nr 2 2014
 6 Spectra of measurement
signal x1(t) in case the
measurement record
length T1 is equal to
period T.
(a) Amplitude.
(b) Phase.
7 Results in case the
measurement record
length T2 is not an
integer amount of
6a
periods T.
(a) Time signal x(t) in
blue and periodically
reconstructed time
signal based on
measurement signal
x2(t) in red.
(b) Amplitude and
phase spectra of
measurement signal
x1(t) in blue and x2(t) in
red, both with starting
time t = 0 s.
6b

their frequencies will be multiples of 1/T2. This series,

however, describes a time signal with period T2, shown as
the red curve in Figure 7a. This signal clearly differs from
the original signal x(t), which has a period time T of 1 s.
The calculated amplitude and phase spectra in Figure 7b
in red will differ from the true spectra in blue. This
measurement error is called ‘leakage’ and will always be
present unless the length of the measurement record is an
exact multiple of the period T of the signal to be analysed. 7a

In the frequency analysis of stochastic signals (noise) this

error will always occur, since these signals are non-periodic.
The leakage error, however, can be reduced by the
application of weighting functions.

The Hanning weighting function is a commonly used

general purpose weighting function and is defined as:

w(t) = √(8/3) · sin2(2πt/T) 7b

Here, T is the record length.

In Figure 8a the weighting function is depicted in black. independent characteristics. This does not require the
Applying this weighting function to x2(t) results in a new signal to be periodic. A noise signal with time-independent
time trace x2hanning(t) = w(t) · x2(t), which represents a signal mean and variance also fulfils this criterion.
with reduced discontinuities, see Figure 8a in red. Figure 8b
shows the spectra of x1(t) in blue, x2(t) in red and x2hanning(t) In modal analysis a transient signal (hammer impact) is
in black in a logarithmic amplitude format. The leakage often used for system excitation. A transient weighting
error in x2(t) is clearly visible. This error even obscures the function is applied to improve the signal-to-noise ratio for
presence of a weak signal component at 8 Hz. the analysis of these short transient signals. The transient
weighting function is defined as:
Applying a Hanning weighting function significantly
reduces the leakage error, but it cannot be applied to all w(t) = 1 for t0 ≤ t < t0 + tw and 0 ≤ t0 < T – tw
signals. Since w(t) is not a constant but a function of time,
this type of weighting is only allowed for signals with time- w(t) = 0 elsewhere

nr 2 2014 MIKRONIEK 11
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 1

Here, t0 is the moment the window opens and tw the length

of the window. The system response signal is also a
transient and prone to leakage if the system has low
damping and must be weighted with an exponential decay
window. This window is defined as:

w(t) = e–(t – t )/τ for t0 ≤ t < T and 0 ≤ τ < T

0

w(t) = 0 elsewhere 8a

Here, t0 is the starting time of the weighting function, τ the

time constant and T the record length. Figure 9a shows the
transient window and Figure 9b the exponential decay
window, both dashed in black.

Signal units
So far the focus was on periodic signals that consist of a
series of discrete frequency components with constant 8b

amplitudes. Their power distribution over frequency is

discrete, so the power is concentrated in infinitesimally
narrow frequency lines. The ‘strength’ of these components
is independent of the frequency resolution of the spectrum
and can be calculated from the FFT lines like:

Ampl (k · Δf ) = | f (k)| with amplitude in [units]

or
Ampl (k · Δf ) = | f (k)| / √2 with amplitude in
[units RMS] 9a

Here, f (k) is the kth frequency line at (k – 1) · Δf [Hz].

Signals with a continuous power distribution cannot be

quantified with amplitude values. These signals are to be
quantified by their Power Spectral Density as function of
frequency, PSD( f ).
PSD( f ) can be calculated from the individual FFT lines by
averaging the power per line and scaling it with the 9b

equivalent noise bandwidth ENB of the FFT calculation:

PSD(k · Δf ) = |f (k)|2 / ENB [units2/Hz]

8 The aim of the Hanning averaging the power per line and scaling it with the
The ENB depends on the record length and the weighting weighting function. equivalent noise bandwidth of the FFT calculation and the
(a) The function (in
function used: black) smoothens out record length:
the discontinuities at the
beginning and end of
ENB = Δf · window_ factor the measurement record ESD(k · Δf ) = |f (k)|2 · T / ENB [units2 s/Hz]
and thus reduces
leakage.
For a rectangular weighting function (= no weighting (b) Leakage can obscure Test signals
function) window_ factor = 1, for Hanning signal components. Test signals play a decisive role in measuring the FRF of a
window_ factor = 1.5. 9 To be used for transient dynamic system. Important parameters are frequency
signals:
(a) transient weighting content, power and amplitude distribution as function of
For signals with a finite total energy, like transients, the function; time and the relation between peak value and RMS value
(b) exponential decay
Energy Spectral Density as function of frequency ESD( f ) function. (= crest factor).
can be calculated from the individual FFT lines by

12 MIKRONIEK nr 2 2014
The most basic test signal is the sinusoid. If the frequency is cannot be used for coherence measurements when only one
a multiple of the frequency resolution Δf = 1/T, then no phase realisation is used.
weighting function is required. The crest factor is only √2,
which makes it a well-suited signal for testing extremely Unlike multi-sine signals, random noise has a continuous
delicate systems. The sinusoid is ideal for measuring the spectrum and is not periodic. This results in leakage and so
Higher Order Sinusoidal Input Describing Functions the use of a Hanning weighting function is required. The
(HOSIDF) to determine nonlinear behaviour, as will be frequency range can be optimised by filtering. Within this
discussed in Part 3 of this series. Since all power is range the PSD can be flat (white noise) or inversely
contained in only one frequency line, many measurements proportional with frequency (pink noise). Noise is
are required to cover a frequency range. uncorrelated with all other signals. The cross-spectrum of
noise and any other signal will converge to 0 after
Measurement time can be reduced by constructing a test averaging, as will be explained in Part 2. This property is
signal that contains power at many frequency lines at the used to determine the frequency response function of a
same time. These multi-sine signals can be processed system under non-ideal measurement conditions.
without weighting function if all the frequency components
coincide with frequency lines of the FFT. The crest factor The pulse signal differs from the test signals mentioned
can be minimised by manipulating the phase of all earlier in the sense that its power distribution over
frequency components and can even become less than √2. frequency is time-dependent. For modal analysis
measurements the pulse is generated by a hammer impact.
In a special class of multi-sine signals distinct frequency By changing the hardness of the hammer tip, the frequency
lines do not contain energy. If a nonlinear system is excited range of the pulse spectrum can be influenced. A soft tip
by these special odd multi-sines, its response will contain will result in a wider pulse, which contains less high-
power in the non-excited frequency lines, which serve as frequency energy. As mentioned in the paragraph on
detection lines for nonlinear behaviour. It is important to weighting functions, an impact window can be used to
realise that these signals are periodic and thus consist of increase the signal-to-noise ratio of the measurement.
frequency components with a fixed amplitude and phase. In Triggering is important to align the window with the pulse.
Part 2 it will become clear that for this reason multi-sines To capture the total pulse, a pre-trigger delay is required. ◾

Summary
Signal Units Weighting function Trigger condition
Sine f = n · Δf None
Ampl. [units] Free run
Sine f ≠ n · Δf Hanning
Multi-sine, linked to Δf Ampl. [units] None Free run
Random PSD [units2/Hz] Hanning Free run
Pulse ESD [units2 s/Hz] Transient window Pre-triggered on pulse

LITERATURE
• J.S. Bendat, A.G. Piersol, “Random data: analysis and
Coming up in Part 2 measurement procedures”, Chichester: Wiley-
Interscience, 2000, ISBN 0-471-31733-0.
• J.W. Cooley, J.W. Tukey, “An algorithm for the Machine
In Part 2 the Frequency Response Function will be introduced as the Calculation of Complex Fourier Series”, Math. Comp.
Vol. 19 (90), pp. 297-301,1965.
frequency-domain description of dynamic behaviour of linear dynamic • R.B. Randall, “Frequency analysis”, Naerum: Bruel &
systems. The article will focus on the practical aspects of FRF Kjaer, 1987, ISBN 87-87355-07-8.
measurements: choice of test signals, picket-fence effect,
measurement noise, and interpretation of the coherence function for
detection of nonlinearities. Both open-loop and closed-loop systems
will be considered.

nr 2 2014 MIKRONIEK 13
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 2

FROM SIGNAL ANALYSIS

TO SYSTEM ANALYSIS
Following the previous article on signal analysis, this article focuses on
AUTHORS’ NOTE
system analysis, specifically on the Frequency Response Function (FRF)
Pieter Nuij, senior system
as a tool to describe the frequency-dependent behaviour of linear time-
architect, and David invariant systems. This article will outline the move from signal analysis
Rijlaarsdam, group leader,
both work at NTS Systems
to system analysis and then explain the concept of the cross-spectrum,
Development in Eindhoven, the FRF and the coherence function. Finally, it will address the specific
the Netherlands.
problems that arise in the identification of closed-loop systems, paying
pieter.nuij@nts-group.nl
www.nts-group.nl special attention to the practical aspects of FRF measurements.

PIETER NUIJ AND DAVID RIJLAARSDAM

Input-output relationship 1 Mass-spring-damper

In the time domain, the input-output relationship of a system loaded with
excitation force f(t)
causal, linear, time-invariant system is described using a responds with
differential equation (DE). Figure 1 shows a mass-spring- displacement x(t).
damper system with input force f(t) and output
displacement x(t).

The DE describes the input-output relationship:

f (t) = mx(t)+ dx(t)+cx(t) 1

with m representing the mass, d the damping coefficient

and c the stiffness. Although this equation describes the
dynamics of the system completely, it requires solving for
x(t) to explicitly relate the displacement to the excitation
force. For more complex systems, solving the DE is often a Three articles
tedious job. An alternative to solving the DE in the time
domain is to transform the DE to the Laplace domain using In a series of three articles on Frequency Response Function (FRF)
the Laplace transform. The Laplace transform L{f(t)} is measurements, this second article introduces the FRF, explaining the
defined as: choice of test signals in relation to the coherence function and the
∞
measurement of the FRF in open- and closed-loop systems. The first
F ( s ) = ∫ e − st f ( t )dt article, which was published in the previous issue of Mikroniek,
0
covered the steps necessary to convert a time-continuous signal into a
with complex parameter s = σ + jω. As a result of this discrete spectrum. Potential errors caused by aliasing and leakage
d were explained and solutions were presented. An overview of several
transformation, f (t) = f (t) in the time domain becomes
dt types of test signals was presented. The third article will focus on the
a multiplication with ‘s’ in the Laplace domain: extension of frequency-domain methods towards nonlinear systems
and the application of such methods to define and optimise the
L{f '(t)} = sF(s) – f(0) performance of such systems. Each article is illustrated with examples.

32 MIKRONIEK nr 3 2014
 Bias errors in FRF estimates
As discussed in the previous section, the FRF can be
calculated from the Fourier transforms of the input and
output signals of the system to be described. In the ideal
situation as shown in Figure 2, the FRF calculation will
yield an unbiased estimate:
2
B( f ) Y ( f )
= = H( f )
A( f ) U ( f )

It is clear that the signal U(f ) must contain power in the

frequency lines of interest to generate proper results.
Without a proper excitation, no estimate is possible. In
practice, the measured signals will always include a certain
amount of noise. In the situation with noise V(f ) in the
3
output signal, as shown in Figure 3, the FRF estimate will
be biased:

B( f ) V( f )
= H( f ) +
A( f ) A( f )
Moreover, noise in the input signal, as shown in Figure 4,
will also result in a biased FRF estimate:

B( f )  W( f ) 
= H ( f ) 1 − 
4
A( f )  A( f ) 
Since noise is always present in realistic measurements, the
calculation of the FRF from the Fourier-transformed time
signal A(t) and B(t) will always be biased. A significant
improvement in the estimation of the FRF can be achieved
Applying this transformation to the DE results in: 2 Ideal, noise-free by using the cross-spectrum. This will be introduced in the
situation. next section.
3 Noise in the output
F ( s ) = ms 2 X ( s ) + dsX ( s ) + cX ( s ) signal.
assuming f '(0) = f (0) = 0 4 Noise in the input signal. The auto-spectrum and the cross-spectrum
The Fourier transform of a time signal a(t) is defined as:
From this expression, the input-output relationship
+∞
becomes a polynomial expression with variable s: A( f ) = F{ a( t )} = ∫ a ( t )e − j 2 πft dt
−∞

X(s) 1
= 2
F(s) ms + ds + c
which is called the transfer function. Next, substituting
s = jω = j2πf in the transfer function results in the FRF:
Certified Precision Engineer
X(f )
=
1
F(f ) ( j2πf ) m + j2πfd + c
2 competencies
The content of this series of three
X(f ) and F(f ) are the Fourier transforms of x(t) and f(t), so articles is in part covered in the
the FRF of a dynamic system can be determined from its “Experimental Techniques in
input excitation signal and the resulting output response Mechatronics” course, provided by The
time signal using a fast Fourier transform (FFT) analyser. High Tech Institute (HTI). This course
This is correct from a theoretical point of view, but the has been selected for the DSPE
estimate will be biased if the influence of measurement Certification Program (see page 26).
noise is not taken into account, as will be discussed in the WWW.DSPEREGISTRATION.NL WWW.HIGHTECHINSTITUTE.NL

next section.

nr 3 2014 MIKRONIEK 33
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 2

and results in a complex spectrum A(f ) with magnitude

and phase information describing the cosine components
that approximate the time signal a(t). The auto-spectrum of
a(t) is defined as:

SAA ( f ) = A* ( f ) ⋅ A( f )

where * indicates the complex conjugate. Since the FFT

routine only uses a finite-length time record of length T of
the signal a(t), the result is an estimate Âi (f ) of the true
Fourier transform A(f ) at the frequencies k · Δf = k/T. From 5a 5b 5c

these individual estimates, the true auto-spectrum of a(t)

can be calculated by averaging over n measurement records:
5 Averaging of cross- The H1 and H2 frequency response function
1 n ˆ* ˆ spectrum estimates for
SAA ( f ) = lim ∑ A i ( f ) ⋅ Ai ( f ) estimators
n→∞ n i =1 increasing variation in
Δφ(fk ). The previous section explained that the averaged cross-
Unlike the complex spectrum A(f ), the auto-spectrum SAA(f ) (a) No phase variation spectrum of two uncorrelated signals is zero. This property
and fully correlated.
is real and describes the power distribution as a function of (b) Some phase will be used to determine a better estimate of the FRF in the
frequency. variation but still case of noise in the input and output measurements as
correlated.
(c) Fully uncorrelated. depicted in Figures 3 and 4. For noise in the output
The cross-spectrum SAB(f ) between the time signals a(t) and measurement (see Figure 3), the following relationships
b(t) is defined by: hold:

SAB ( f ) = A* ( f ) ⋅ B( f ) B = H ⋅ A +V
A* ⋅ B = H ⋅ A* ⋅ A + A* ⋅ V
and can be calculated using the FFT by averaging
B = Hthe
⋅ A +V A* ⋅ B A* ⋅ V
individual cross-spectra estimates: =H+ *
A* ⋅ B = H ⋅ A* ⋅ A + A* ⋅ V A ⋅A
*
A ⋅A
SAB ( f )
1 n ˆ* ˆ A* ⋅ B A* ⋅ V H1( f ) = if SAV ( f ) = 0 [v( t ) , a( t ) uncorrela
SAB ( f ) = lim ∑ A i ( f ) ⋅ Bi ( f ) = H + SAA ( f )
n→∞ n i =1 A* ⋅ A A* ⋅ A
SAB ( f )
H1( f ) = SAB(f )
Unlike the auto-spectrum SAA(f ), the cross-spectrum if SAV ( f ) = 0 [v(t), a(t) uncorrelated]
SAA ( f )
is complex valued. The magnitude of each individual cross-
spectrum estimate is the product of the magnitudes of the Likewise for the situation of noise in the input
two individual signal components with frequency fk. The measurement, as shown in Figure 4, the unbiased estimate
phase of the cross-spectrum is the phase difference Δφ(fk) for the FRF can be calculated using:
between the corresponding frequency components Â(fk)
SBB ( f )
and Bˆ( f ) . H2( f ) = if SBW ( f ) = 0 [w(t) , b(t) uncorr
k
SBA ( f )
SBB ( f )
H (f) if SBW ( f ) 0 [w( t ) , b( t ) uncorrelated]
Since the averaging operation is done on complex 2values,SBA ( f )
the magnitude of the averaged cross-spectrum will also In both cases, the excitation signal is expected to be
depend on the variation of Δφ(fk) over the measurement uncorrelated with the noise components in the measured
records as can be seen in Figure 5. If two signals are input and output signals. This is always the case if the
uncorrelated, the phase difference Δφ(fk) will be random excitation signal is a random signal.
and the magnitude of the averaged cross-spectrum will be
zero (see Figure 5c). The coherence function
In system analysis, it is important to quantify the degree of
The cross-spectrum emphasises common frequency linear relationship between the input and output signals. In
components in both signals, with a consistent phase statistics, this relationship is indicated by the correlation
relationship. This characteristic will be used in all practical coefficient, while in the frequency domain, it is indicated by
measurements where noise is always present. In the second the coherence function γ2(f ):
part of this paper, the variables in all derivations are 2
SAB ( f )
assumed to be a function of frequency and the notation ‘(f )’ γ2( f ) = 0 ≤ γ2( f ) ≤ 1
SAA ( f ) ⋅ SBB ( f )
is left out in the formulae for clarity.

34 MIKRONIEK nr 3 2014
 7a 7b

Figure 6 shows the relationship between the frequency

components at frequency fn of signal A(f ), which is the
measured input of a system, and B(f ), which is measured at
the output. Figure 6a clearly shows that the input-output
relationship is linear and the coherence is equal to one. The
input-output relationship in Figure 6b is no longer a
straight line through the origin, but varies due to noise in 8

either or both the measurement signals. Consequently, the

coherence is less than one. Figure 6c shows a perfect, noise-
free, non-linear relationship between the input and output contradicts the fact that the system is known to be 6 The coherence function
signals and again the coherence is less than one. Figure 6d nonlinear. Figure 7b shows the results for the same system, quantifies the amount of
linearity between input
shows the situation where input and output signals are not this time, however, measured with a random noise signal signal component A(fn )
related, resulting in a coherence equal to zero. with the same RMS value used in the chirp excitation. The and output signal
component B(fn ).
random noise excitation yields a coherence which is clearly 7 FRF (in blue) and
A coherence value less than one can be caused by: less than one over most of the frequency range, as is to be coherence (in red) of a
nonlinear system.
• Nonlinear system behaviour. expected for a nonlinear system. (a) Determined with a
• Noise in the measurements of a(t) and/or b(t). chirp, a deterministic
• Leakage in the FFT. FRF estimation in closed-loop systems signal.
(b) Measured with a
• Time-variant system behaviour. Unlike open-loop systems, closed-loop systems include a random noise excitation.
feedback path between output and input. This feedback can 8 Closed-loop system with
R(f) setpoint, T(f) test
It is a common misunderstanding that a measurement with cause unexpected problems in the FRF identification. signal and N(f) external
a coherence equal to one is proof of linear system Figure 8 is a model of a closed-loop system with controller noise source acting on
the system. A(f) is the
behaviour. If the measurement is based on only one data C(f ) and plant H(f ) with setpoint R(f ), test signal T(f ) and measured input signal,
block, i.e. without averaging across additional data blocks, external noise N(f ) acting on the system. B(f) is the measured
coherence will always be equal to one even in the case of output signal.

nonlinear system behaviour. Even if the coherence is Any attempt to determine H(f ) directly from the measured
determined using averaging, all deterministic test signals input signal A(f ) and the measured output signal B(f ) will
will yield a coherence value close to one if there is minimal fail because of the correlation between N(f ) and A(f ) and
measurement noise compared to the test signal. Examples R(f ) and A(f ):
of these types of test signals are the multi-sine and the
hammer impact if all impacts are equally strong. B = H ( N + T + CR − CB ) → B (1 + CH ) = H ( N + T + CR )
A = T + C ( R − H ( N + A )) → A (1 + CH ) = T + C ( R − HN )
Figure 7a shows the FRF and coherence of a measurement
B N + T + CR SAB SAN + SAT + CSAR
of a nonlinear system using chirp excitation. The coherence =H → =H⋅
A T + C ( R − HN ) SAA SAT + C ( SAR − HSAN )
is clearly one over most of the frequency range, which

nr 3 2014 MIKRONIEK 35
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 2

9 FRF of a resonance. The

horizontal lines indicate
the -3dB magnitude
levels of the respective
resonances.
(a) Measured with a
chirp.
(9) Measured with
9a 9b
white noise.

Even if the setpoint signal R(f ) is equal to zero, the estimate difficult, but the torque signal can be calculated from a
of H(f ) will remain biased: simple motor current measurement if the motor constant is
known. Measuring small displacements at high frequencies
SAB ( f ) SAN ( f ) + SAT ( f )
= H( f ) ⋅ with a good signal-to-noise ratio is often difficult, while
SAA ( f ) SAT ( f ) − C( f ) H ( f )SAN ( f ) measuring accelerations at high frequencies is simple
Depending on the signal-to-noise ratios in the loop, the because of the increase in sensitivity with (j2πf )2. The
estimate of the FRF of the plant may vary between H(f ) and corresponding FRF can be calculated by dividing the FRF
–1/C(f ). by (j2πf )2. Time spent on optimising the signal-to-noise
A robust alternative to determine the plant FRF H(f ) in a ratio of the measurement signals is time well spent.
closed-loop system is the three-point method in which the
sensitivity S(f ) and the process sensitivity PS(f ) are Choice of frequency range and frequency resolution
determined: Take a broad frequency range for the first measurement to
get a good general overview of the region of interest. In the
STA 1 C  STR − HSTN  1
S( f ) = = +  = subsequent measurements, however, limit the frequency
STT 1 + CH 1 + CH  STT  1 + C( f ) H ( f ) range to the region of interest so as not to excite system
S H 1  STR − HSTN  H( f ) dynamics outside that range. The required frequency
PS( f ) = − TE = −  =
STT 1 + CH 1 + CH  STT  1 + C ( f )H( f ) resolution of the measurement depends on the damping at
PS( f ) the resonances or antiresonances. Too coarse a resolution
H( f ) =
S( f ) will underestimate the magnitude at the resonance
frequency because of the ‘picket fence effect’. A very
if R(f ), T(f ) and N(f) are uncorrelated, and averaged narrow resolution will not result in any additional quality
spectra and cross-spectra are used. improvement; it will only result in unnecessary
measurement time.
Tips for measuring an FRF in practice
Although a thorough understanding of the theory behind Figure 9a shows the measurements of a resonance at three
the FRF helps, it does not guarantee good measurement different frequency resolutions with a chirp test signal
results. Experience with experimentation is also required under leakage-free conditions. The estimate of the
for good results. This section focuses on some practical resonance magnitude depends on the frequency resolution
measurement aspects. used and shows a 2.2dB variation. As a rule, a resonance is
estimated correctly if the frequency interval between the
Choose the proper input and output signals -3dB magnitude frequencies is covered by at least five
The choice of input and output signal is often determined frequency lines. Figure 9b shows the measurement results
by the accessibility of the signals. Measuring torque is often of the same system but with a random noise test signal.

36 MIKRONIEK nr 3 2014
10

Apart from the ‘picket fence effect’, leakage errors are also 10 The –20dB rule to power in this frequency range is an option. Other options
evident. These errors show up as a difference in the determine the useful are zoom measurements in that limited frequency range.
frequency range in the
magnitude values for the same frequency line as a function case of impact This will allow the signal conditioning to be optimised to
of the resolution. Even using a Hanning weighting excitation. reduce the quantisation noise.
function, these differences amount to approximately 4 dB
at 4,100 Hz. Reduce unnecessary nonlinear behaviour as much as
possible
Hammer or shaker excitation for modal analysis? All real systems are in essence nonlinear, but for many
The choice of excitation is often dictated by practical systems, under normal operating conditions, the nonlinear
considerations like the accessibility of the test object, the behaviour can be ignored if the measurements are carried
ease of mounting a shaker and the required frequency range out carefully. Do not over-drive the system, check for
of excitation. In the case of impact excitation, the tip of the saturating power amplifiers and overloaded sensors. Listen
hammer determines the width of the excitation spectrum. for rattles in coverings, check for strong 50/60Hz
The useful frequency range starts at 0 Hz and is limited to harmonics. Optimise the sensor signal-to-noise ratio and
that frequency at which the magnitude is 20 dB less than choose proper ADC input ranging. Prevent aliasing and
the low-frequency magnitude, as can be seen in Figure 10. reduce leakage.

When testing very sensitive equipment, shaker excitation is Do the FRF measurements under the proper operating
often preferred over impact excitation. A band-limited conditions
white noise test signal in combination with a Hanning Friction in positioning systems often induces stick-slip
weighting function is a good starting point because the total behaviour. The stick phase system dynamics are very
signal power is evenly distributed over the frequency range different from the dynamics in the slip phase. For FRF
of interest. This reduces the risk of over-driving resonances, measurements in the stick phase, only very small excitations
which can happen with a sinusoidal excitation signal. are allowed in the 100 nm range. The slip phase dynamics
Moreover, because random noise is not a deterministic can be measured reliably by jogging the system.
signal, the information from the coherence function can be
used to assess the linearity of the measurement. Be critical about the measurement results
Figure 11 shows a typical evolution in the quality of an FRF
Optimise the signal-to-noise ratio of the signals measurement. The excitation signal is band-limited white
A dip in the coherence function often coincides with a low noise. The left column shows the magnitudes of the FRF
value in the excitation and/or response spectra. If the measurements, while the right column shows the
magnitude at the anti-resonance frequency is relevant, corresponding coherence plots. In both columns, the red
increasing the signal-to-noise ratio by increasing the input curves represent the final results. In the upper row in blue, a

nr 3 2014 MIKRONIEK 37
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 2

11

frequency resolution of 200 Hz was used with a rectangular With the same excitation signal level and weighting 11 FRF estimates for
weighting function. Both large leakage errors and picket function as in row two, the measurement in the third row is various weighting
functions, frequency
fence errors are present. done with a frequency resolution of 6.25 Hz. Once again, resolutions and
the leakage error is reduced as indicated by the increased excitation levels.
The second row shows the results for a measurement with value of the coherence function. The picket fence error is
equal excitation signal level and frequency resolution but also reduced. The remaining minima in the coherence
with a Hanning weighting function. The increase in function coincide with the antiresonances. The bottom row
coherence is caused by a significant reduction in leakage shows the final results. In this measurement, the RMS value
error. The picket fence error is not reduced. The coherence of the excitation signal is increased by a factor of 10, which
function still shows ranges with low values, however. These results in an increased signal-to-noise ratio of the output
ranges coincide with the resonances and antiresonances. signal. This results in a better coherence for the anti-
There are different reasons for the low coherence values of resonances. One should be careful about increasing the
the resonances and antiresonances. In the case of the excitation signal any further, as this may over-drive the
resonances the low coherence is due to leakage, while in the system resulting in a lower coherence value due to the
case of the antiresonances the low coherence is caused by induced nonlinearity. ◾
low signal-to-noise ratios, as will become evident from the
subsequent measurements.

LITERATURE
• P. Nuij, D. Rijlaarsdam, “Introduction to Frequency Response
Function Measurements (Part 1)”, Mikroniek, Vol. 54 (2), pp. 8-13,
Coming up in Part 3 2014.
• J.S. Bendat, A.G. Piersol, Engineering applications of Correlation and
Spectral Analysis, Chichester: Wiley-Interscience, 1993, ISBN: 0-471-
Part 3 will introduce frequency domain methods to use in practice to 57055-9.
model and optimise the performance of nonlinear systems. The article • H. Herlufsen, “Dual Channel FFT Analysis (Part 1)”, Brüel & Kjær
Technical Review, No. 1, 1984.
provides a brief overview and experimental examples of existing
modelling techniques and a novel method to assess and optimise the
performance of nonlinear systems using frequency domain-based
tooling.

38 MIKRONIEK nr 3 2014
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 3

FREQUENCY DOMAIN
MODELLING AND
PERFORMANCE
OPTIMISATION OF
NONLINEAR SYSTEMS
AUTHORS’ NOTE
Concluding a series of three, this article deals with frequency domain
David Rijlaarsdam, group
leader, and Pieter Nuij, senior methods when applied to systems subject to nonlinear dynamical
system architect, both work at effects. To meet increasing system requirements, techniques have to
NTS Systems Development in
Eindhoven, the Netherlands. deal with the performance-degrading effects of nonlinearities. When
Maarten Steinbuch is full applied with care, frequency domain methods provide practically
professor in the department of
Control Systems Technology at applicable tools to model and optimise the performance of nonlinear
Eindhoven University of
Technology. Johan Schoukens
systems. This article provides a brief overview and experimental
is full professor in the examples of existing modelling techniques, as well as a novel method
department of Fundamental
Electricity and Instrumentation
for performance assessment and optimisation using frequency domain
at the Free University of based tooling.
Brussels (VUB), Belgium.
david.rijlaarsdam@ DAVID RIJLAARSDAM, PIETER NUIJ, MAARTEN STEINBUCH AND JOHAN SCHOUKENS
nts-group.nl
www.nts-group.nl

Introduction
Three articles Increasing performance requirements on high-performance
(motion) systems require novel techniques to deal with
In a series of three articles on Frequency Response Function (FRF) performance-degrading effects of nonlinearities. For linear
measurements, this third article focusses on the extension of and time-invariant (LTI) systems, frequency domain
frequency-domain methods towards nonlinear systems and the methods are widely accepted in the engineering community
application of such methods to define and optimise the performance for modelling as well as performance optimisation
of such systems. The first article [1] covered the steps necessary to purposes. Although systems may be specifically designed to
convert a time-continuous signal into a discrete spectrum. Potential minimise nonlinear effects, nonlinearities such as magnetic
errors caused by aliasing and leakage were explained and solutions fields may be inherently present in the design. In particular,
were presented. An overview of several types of test signals was some applications require the presence of nonlinear effects,
presented. The second article [2] introduced the FRF, explaining the such as friction in the motion stage of an electron
choice of test signals in relation to the coherence function and the microscope. Whether performance is measured by speed,
measurement of the FRF in open- and closed-loop systems. Each accuracy, reproducibility, smoothness, or other
article is illustrated with examples. performance measures, the effects of nonlinearities become
increasingly important in high-precision applications.

22 MIKRONIEK nr 4 2014
1 2a

To accurately detect, analyse and compensate performance- 1 Schematic depiction of

degrading effects in controlled nonlinear dynamical the (motion module) of
the TEM system. For
systems, practically applicable tools are required. Therefore, simplicity, only one
following up on [1] and [2], this article discusses the pitfalls direction of motion is
depicted. All transfer
and potential gains when using frequency domain functions in this article
techniques to model and optimise the performance of are computed from the
motor voltage to the
nonlinear systems. Note that this article aims to provide an measured displacement.
accessible, but very brief overview of existing methods to 2 Low-speed jogging
address nonlinear systems in the frequency domain. For a response of the TEM
system and the effect of
more complete and formal overview and comparison, see, optimised friction
for example, [3]. compensation.
(a) No Coulomb friction
feed forward.
The main carrier throughout this paper is an industrial case (b) Optimal Coulomb
friction feed forward.
study of optimal friction compensation in a Transmission
Electron Microscope (TEM) system as depicted in Figure 1. 2b

Based on this example the effects of nonlinearities in the

frequency domain are introduced as well as state-of-the-art
Certified Precision modelling techniques. Moreover, a practically applicable
Engineer competencies method is introduced to analyse, visualise and optimally
compensate performance-degrading nonlinear effects in
The content of this series of three dynamical systems. Application of this technique has
articles is in part covered in the resulted in a reduction of the achievable smooth jogging
“Experimental Techniques in motion to a level of 20 nm/s in an industrial TEM
Mechatronics” course from (Figure 2) [4].
Mechatronics Academy, offered to the
market by The High Tech Institute. This Frequency domain modelling of nonlinear
course has been selected for the DSPE systems
Certification Program (see page 46).
WWW.DSPEREGISTRATION.NL WWW.HIGHTECHINSTITUTE.NL Overview of modelling techniques
WWW.MECHATRONICS-ACADEMY.NL Frequency domain methods are widely accepted and have
been an impetus for the development of modelling and

nr 4 2014 MIKRONIEK 23
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 3

The success and popularity of frequency domain methods

for LTI systems is largely due to the combination of Fourier
analysis and the fact that a sine wave is an eigenfunction of
LTI systems, which satisfy the properties of superposition
and homogeneity. For nonlinear systems, this is generally
not the case, which implies that a conventional frequency
domain model captures only a part of the system’s
dynamics, as illustrated by Figure 3.

Figure 3 depicts three Frequency Response Functions

(FRFs) of the TEM system depicted in Figure 1. These FRFs
are measured using white-noise inputs with three different
rms values. It becomes clear that the FRF captures only a
subset of the system’s dynamics as the dependence of the
dynamics on the input amplitude is not captured by a single
3
FRF. This effect, called ‘gain compression / expansion’, is
one of several effects of nonlinear behavior in the frequency
domain, such as desensitisation, intermodulation and the
generation of harmonics (see Table 1). Moreover, although
3 Frequency response control design techniques for LTI systems. Although the the coherence of a single FRF (not shown) may indicate
function of the TEM for
varying input levels.
linearity assumption can rarely be satisfied in practice, nonlinear behavior if the excitation signal is random, it
linear models and analysis often suffice when systems are provides no indication of nonlinearity if the system is
operated around a given working point. Given the subject to a deterministic excitation signal such as a
widespread acceptance and success of frequency domain multi-/swept sine [8].
methods for LTI systems, several approaches exist that
extend frequency domain methods towards nonlinear Hence, traditional frequency domain tools and models fail
systems, e.g. the Generalised Frequency Response Function in the presence of nonlinearities, requiring an extension of
(GFRF) [5], nonlinear FRF [6], describing functions [7] and conventional frequency domain methodologies if
linear approximations [8]. Table 1 provides an overview of nonlinearities are present. Two modelling approaches will
such methods and the nonlinear effects / information be discussed that allow detection and modelling of a subset
captured by each model. of nonlinear effects in the frequency domain. Moreover, it

Table 1 Overview of frequency domain modelling approaches for nonlinear systems.

Content at frequencies

Content at frequencies
Legend:

not present in input

Gain compression /

Intermodulation

present in input
Desensitisation

green:
effect captured / info available
Harmonics
expansion

red:
effect not captured / info not
available
dependence of dependence of combination of generation of
orange: gain on input response at one input frequencies frequency
effect partly captured / info available, excitation level frequency on to produce new components in
requires additional processing input at another frequencies in the output at
phase

phase

frequency output multiples of input

gain

gain

frequency
Conventional FRF
Generalised FRF [5]
Nonlinear FRF [6]
sinusoidal
Describing function generalised
HOSIDF [7]
Best linear approximation [8]

24 MIKRONIEK nr 4 2014
4a 4b

will be shown later on that analysis of nonlinear effects in 4 Transformation of a 5a

nonlinear system to a
the frequency domain provides a useful way to define and best linear
optimise the performance of nonlinear systems. approximation (BLA)
and corresponding
nonlinear distortion.
Experimental results (a) Nonlinear system.
In the following, two approaches to modelling the non- (b) BLA and nonlinear
distortion.
linear TEM system are illustrated based on experimental 5 Analysis of the TEM
results. The best linear approximation (BLA) and Higher system.
(a) Best linear
Order Sinusoidal Input Describing Function (HOSIDF) are approximation.
selected from Table 1 as they provide models that are easily (b) Output spectrum:
computed from measured data and provide insightful average (black), variance
due to nonlinearities
representations, which are usable in practice. First, a best (red) and stochastic
linear approximation of the TEM system is derived, disturbances (orange),
also energy appearing at
yielding separate quality measures relating to random odd (blue) and even
disturbances (noise) and nonlinear influences. Second, the (green) non-excited
frequencies.
HOSIDF is used to explicitly investigate nonlinear effects in
the TEM system. 5b

Linearised modelling: best linear approximation

In this section, the estimation of a linear model in the In this example, the BLA is derived by exciting the TEM
presence of nonlinearities is addressed. Consider a system with a set of so-called ‘special random, odd
nonlinear system as depicted in Figure 4a and assume the multisines’, i.e. different realisations of multisine signals
system can be transformed to a structure as in Figure 4b with equal amplitude spectra, but randomised phase
(for existence and invariance of this transformation, see realisations [8]. The adjectives ‘special’ and ‘odd’ refer to
[3]). In Figure 4b, the Fourier transform of the output Y( f ) the fact that the multisine signal is constructed only from
then depends on the input U( f ) and consists of a sinusoids which are an odd multiple of some base frequency
component that is generated by an LTI system B( f ), a and some of these odd multiples are set to zero in the
distortion S( f ) due to nonlinearities and an output excitation spectrum (‘special’). Such specially designed
disturbance N( f ), i.e. signals allow to:
1. compute the BLA on spectral lines which are present in
Y( f ) = B( f )U( f ) + S( f ) + N( f ) the excitation signal;
2. detect nonlinearities by energy appearing at non-excited
Then, the pair (B( f ), S( f )) is defined as the best linear spectral lines (intermodulation, harmonics);
approximation (BLA) [8], if it satisfies: 3. assess model quality by separately quantifying variance
due to nonlinearities and output disturbances.
B( f ) = arg min E{(y(t) – B( f )u(t))2}
B(f )
For details on the analysis required to attain the
Here, E{ } denotes the expected value. Several studies information mentioned above, see [3]. Figure 5a shows the
concerning the existence and uniqueness of the BLA can be BLA of the TEM system for different excitation levels (rms
found [3] [8]. In this section, however, attention is focussed values of the multisine signal). Furthermore, Figure 5b
on practical identification of the BLA and the information shows the output spectrum generated by the BLA and the
available therein. separation between the variance on this average due to
nonlinear effects and due to output disturbances. Moreover,

nr 4 2014 MIKRONIEK 25
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 3

6a 6b

it shows the energy level at non-excited odd and even 6 Higher-order sinusoidal Furthermore, Figure 6b shows the first HOSIDF as a
describing functions
multiples of the base frequency of the multisine which are (HOSIDFs) of the TEM
function of frequency and excitation amplitude. The fact
not present in the excitation signal. system. that the even HOSIDFs are close to zero and the behaviour
This shows that the variance caused by nonlinearities is (a) First, third and fifth of the gain characteristics of the odd HOSIDFs, both point
HOSIDF for f0 = 20 Hz
mainly due to odd nonlinear effects, which points towards a (grey: variance on towards a point-symmetric nonlinearity, the effect of which
point-symmetric nonlinearity such as (Coulomb) friction average). reduces with increasing excitation amplitude, e.g. friction.
(b) First HOSIDF.
(a point-symmetric nonlinear function of the system’s
velocity). Hence, using the BLA when working with Apart from providing insight into and quantification of the
nonlinear systems not only provides a best approximation nonlinear effects present, the HOSIDFs have been shown to
of the FRF, but yields both quantitative measures of the allow computation of the parameters defining the non-
type and level of nonlinearities present and the quality of linearity. Moreover, the HOSIDFs can be used to optimally
the measurement (signal-to-noise ratio). design nonlinear compensators to reduce performance-
degrading effects in nonlinear systems. This is addressed in
Nonlinear modelling: Higher Order Sinusoidal Input the next section.
Describing Functions
The second modelling approach specifically models the Frequency domain performance optimisation of
generation of harmonics due to nonlinearities. Due to its nonlinear systems
similarity to the classical sinusoidal describing function, the
model is referred to as the Higher Order Sinusoidal Input Time versus frequency domain performance
Describing Function (HOSIDF) [7]. The HOSIDFs are While the previous section dealt with modelling nonlinear
defined for sinusoidal inputs with frequency f0. The k th effects in the frequency domain, this section focusses on
HOSIDF Hk ( f0, a) is a function of both frequency f0 and using frequency domain analysis to optimise the
amplitude a and is defined as follows: performance of nonlinear systems [3] [4]. Defining and
optimising performance of nonlinear systems in a
Hk ( f0, a) = Ys (kf0, a) / Usk ( f0, a) practically applicable manner is nontrivial. For example,
performance indicators for LTI systems such as bandwidth,
Here, the subscript s refers to the single-sided spectrum of sensitivity and gain/phase/modulus margin cannot be used
the input U( f ) and output Y( f ). For k = 1, the above and global optimisation of a given performance indicator is
definition equals the definition of a conventional FRF not straightforward.
under sinusoidal excitation and that of a conventional
sinusoidal describing function. The added value of the The method introduced in the following is based on the
HOSIDFs comes from explicitly – both in gain and phase concept behind the HOSIDFs, i.e. nonlinear effects generate
shift – modelling the generation of harmonic frequency harmonics when the system is subject to a sinusoidal input
components by the nonlinear system. (often harmonics are indeed a necessary and sufficient
condition for the existence of nonlinear behaviour [3]). The
Figure 6a shows the first three odd HOSIDFs for the TEM method also assumes that the nonlinear effects are
system at f0 = 20 Hz for varying input amplitude. performance-degrading and should therefore be

26 MIKRONIEK nr 4 2014
compensated. However, before moving to performance harmonics and at the excitation frequency (Table 2, dark
optimisation, consider the essential difference between green) provides a suitable and practically measurable
time and frequency domain based performance assessment; performance measure. Note that optimising nonlinear
see Table 2. compensators by removing harmonics in the output of a
system provides a globally optimised controller for a large
Table 2 illustrates the difference between optimisation class of systems, including magnetic fields, actuator/sensor
based on time and frequency domain analysis with respect nonlinearities and several forms of friction. Hence, a static
to performance in terms of tracking and compensation of nonlinear compensator optimised in this way, provides
nonlinear effects. If a control system is optimised based on optimal compensation for any input/reference signal,
tracking error, the output of such a system will approximate although it was tuned using a sinusoidal input only [3].
the sinusoidal reference in both amplitude and phase (light
red). The controller will, however, sub-optimally The method described above has, so far, been applied to
compensate the nonlinearity as the input-dependent term optimise the performance of a military surveillance camera,
in the performance measure clouds the influence of the linearise amplifier characteristics and compensate friction
harmonics generated by the nonlinearity (dark red). The in a TEM motion module [4]. Moreover, simulations are
resulting controller only provides optimal tracking for the successful in linearisation of the influences of magnetic
amplitude and frequency of the signal applied during fields in maglev applications and fully automated, adaptive
tuning. implementations have been successfully applied in practice.
For the purpose of this article, attention is focussed on the
If instead the harmonics are optimally compensated by application of friction compensation in a TEM.
minimising the corresponding frequency domain
performance measure only (dark green), the corresponding The TEM system depicted in Figure 1 is subject to
output will match the input in frequency, but possibly differ conventional Coulomb friction feed forward and the
in amplitude and phase (light green) resulting in a sub- corresponding feed-forward gain is to be optimally tuned.
optimal tracking error. The solution comes from separating In this application the feed-forward gain is incrementally
the problem and using the appropriate performance increased while the system is subject to a sinusoid with a
measure for each job [4]. Hence, first optimise a nonlinear frequency significantly below the bandwidth of the closed-
controller to optimally compensate nonlinear effects based loop system. The results are depicted in Figure 7.
on a frequency domain performance measure (e.g. dark
green) and then optimise a conventional LTI controller to Figure 7a shows the energy present at the individual
achieve tracking performance and disturbance attenuation, harmonic lines in the spectrum as a function of the feed-
e.g. using loop shaping. forward gain. Moreover, the cost function combining all
statistically relevant harmonics is depicted. From Figure 7a
Next, an example of frequency domain based friction it becomes clear that around 0.2 V the overall energy at
compensation in a TEM is presented to illustrate the above. harmonics in the output is minimised (individual
harmonics have slightly different optima due to the nature
Frequency domain based friction compensation of the nonlinearity and compensator). As observed in
As indicated in Table 2, nonlinear effects can be assessed in Figure 7b, the harmonics in the output spectrum have
the frequency domain by considering harmonics present in indeed been reduced significantly when comparing the
the output when the system is subject to a sinusoidal input. uncompensated (red) and optimally compensated (green)
For example, the ratio between the energy present at situation.

Table 2 Time and frequency domain performance measures for optimal tracking and nonlinear compensator design.
Time domain Frequency domain
minimise harmonic disturbance
A cos(2πf0t + φ0)
Optimal nonlinear K Y ( kf 0 )
↓ ∑
compensator
B cos(2πf0t + ψ) k =0 Y ( f0 )
k ≠1

minimise tracking error and harmonic disturbance

A cos(2πf0t + φ0)
2 K 2
Optimal tracking ↓ Y ( f 0 ) − U ( f 0 ) + ∑ Y ( kf 0 )
A cos(2πf0t + φ0) k =0
k ≠1

nr 4 2014 MIKRONIEK 27
 ■ INTRODUCTION TO FREQUENCY RESPONSE FUNCTION MEASUREMENTS – PART 3

7a 7b

The practical implications and global nature of the Ministry of Economic Affairs under the BSIK programme. 7 Frequency domain
based optimisation of
optimised tuning become clear from Figure 2. Comparing This work was supported in part by the Fund for Scientific
Coulomb friction feed
Figure 2a and 2b yields that selecting the optimised Research (FWO-Vlaanderen), by the Flemish Government forward in the TEM
parameter setting based on the frequency domain based (Methusalem), and by the Belgian Government through the system.
(a) First five harmonics
performance measure depicted in Figure 7a yields a Interuniversity Poles of Attraction (IAP VI/4) and performance
significant performance improvement in terms of Programme. ◾ measure (Table 2, dark
green) as a function of
smoothness during low-speed jogging motion at 20 nm/s. controller parameter.
(b) Difference between
the output spectrum
Conclusions without (red) and with
Although frequency domain methods are widely applied (green) optimised
Coulomb friction feed
and accepted for modelling and performance optimisation
forward. Grey lines
of linear and time-invariant system, their applicability to indicate statistically non-
systems containing nonlinearities is nontrivial. However, significant data.

increasing performance requirements on high-performance

systems require novel techniques to model and deal with
REFERENCES
the performance-degrading effects of nonlinearities. When
[1] P. Nuij and D. Rijlaarsdam, “Introduction to frequency response
applied with care, frequency domain methods provide function measurements – Part 1”, Mikroniek, 54(2):8-13, 2014.
practically applicable tools to model and optimise the [2] P. Nuij and D. Rijlaarsdam, “Introduction to frequency response
function measurements – Part 2”, Mikroniek, 54(3):14-22, 2014.
performance of nonlinear systems. This article provides a [3] D. Rijlaarsdam, Frequency domain based performance optimization of
brief overview of existing frequency domain based systems with static nonlinearities, Ph.D. thesis, Eindhoven University of
Technology, ISBN: 978-90-386-3149-3, 2012 (available at www.
modelling techniques for nonlinear systems and a novel davidrijlaarsdam.nl).
method to assess and optimise the performance of such [4] D. Rijlaarsdam, P. Nuij, J. Schoukens and M. Steinbuch, “Frequency
domain based nonlinear feed forward control design for friction
systems. Both the modelling approach and performance compensation”, Mechanical Systems and Signal Processing, 27(2):551-
optimisation are illustrated using a TEM (Transmission 562, 2012.
Electron Microscope) system suffering from the [5] R. Yue, S. Billings and Z.-Q. Lang, “An investigation into the
characteristics of nonlinear frequency response functions. Part 1:
performance-degrading effects of friction during low-speed Understanding the higher dimensional frequency spaces”,
jogging motion. International Journal of Control, 78(13):1031-1044, 2005.
[6] A. Pavlov, N. van de Wouw and H. Nijmeijer, “Frequency response
functions for nonlinear convergent systems”, IEEE Transactions on
Acknowledgement Automatic Control, 52(6):1159-1165, 2007.
[7] P. Nuij, O. Bosgra and M. Steinbuch, “Higher-order sinusoidal input
The work reported in this article was carried out as part of describing functions for the analysis of non-linear systems with
Condor, a project under the supervision of the Embedded harmonic responses”, Mechanical Systems and Signal Processing,
20(8):1883-1904, 2006.
Systems Institute (ESI) with FEI Company as the industrial [8] R. Pintelon and J. Schoukens, System identification: a frequency
partner. This project was partially supported by the Dutch domain approach, Wiley-Blackwell, ISBN: 978-04-706-4037-1, 2012.

28 MIKRONIEK nr 4 2014