Proceedings of the 1997 ASCE Structures Congress,

Portland, Oregon, April 13-16, 1997.

H1 Static and Dynamic Output Feedback Control

of the AMD Benchmark Problem
H. Allison Smith  (member), Scott E. Breneman y (student member) and
Olivier Sureau y
Abstract
This study develops an active control methodology for the AMD benchmark prob-
lem of Spencer, et al. (1997) based on dynamic output feedback controllers designed
using an H1 based approach. Kalman lter estimators of the states of a reduced or-
der model of the benchmark structure are coupled to static state feedback controller
gains to develop the dynamic feedback controllers. A method is outlined for designing
H1 feedback controller gains, and a comparison is made between the eectiveness of
H1 static output feedback and the dynamic acceleration feedback controllers. The
results quantify the performance increase obtained with the additional complexity of
the dynamic output feedback controllers compared to the static acceleration feedback
controllers.
Introduction
For the AMD benchmark problem of Spencer, et al. (1997), controllers are de-
signed using two H1 based approaches: one approach uses direct static output feed-
back of sensor measurements, and the second approach uses a dynamic output feed-
back controller that consists of static state feedback controller gains with a Kalman
lter state estimator. The controllers considered in this study are designed by a
continuous-time H1 controller approach, then discretized for simulation with the
benchmark model.
The controllers are developed from a state space design model of the form:
x_ (t) = Ax(t) + B u(t) + B w(t)
u w (1)
z(t) = C x(t) + D u(t) + D w(t)
z zu zw (2)
y(t) = C x(t) + D u(t) + D w(t)
y yu yw (3)
where x(t) is the state vector, u(t) is the vector of control inputs, w(t) is the vector
of disturbance inputs, y(t) is the vector of sensor measurements, z(t) is the vector of
regulated outputs, and A, B , B , C , D , D , C , D , and D are matrices of
u w z zu zw y yu yw


Assistant Professor, Stanford University, Dept. of Civil Engineering, Stanford, CA 94305.
y
Research Assistant, Stanford University, Dept. of Civil Engineering, Stanford, CA 94305.

1 Smith et al.
the appropriate dimension developed from the evaluation model. The static feedback
controller gains dene the controller command signal as u(t) = ,KG y(t), and the
dynamic output feedback controllers are of the form:
x_ (t) = A x (t) + B y(t)
c c c c (4)
u(t) = C x (t) + D y(t)
c c c (5)
where xc (t) is the state vector of the dynamic controller and Ac , Bc , Cc and Dc are
the state description of the regulator. The idealized continuous closed loop system
can then be described in the standard form by:
x^_ (t) = Aclx^ (t) + Bcl w(t) (6)
z(t) = C x^ (t) + D w(t)
cl cl (7)
where x^ (t) is the augmented state vector representing the dynamics of the system
and controller. For this system, the H1 norm of the transfer function Tzw , from the
excitation w(t) to the output z(t), is dened as:
k Tzw k1 =: sup kk wz(t)
(t)
k2
k2 (8)
w(t)

Controller Design
For the design of static output and state feedback controller gains, an iterative
gradient search was implemented to nd a locally optimal solution that met a pre-
scribed attentuation of the H1 norm of the closed loop system. Dening the H1
norm of a closed loop system as:
CL =: k Tzw (cl) k1 = k Ccl (sI , Acl ),1 Bcl + Dcl k1 (9)
and the open loop H1 norm as:
OL =: k Tzw (ol) k1 = k Cz (sI , A),1 Bw + Dzw k1 (10)
then the increase in performance attributed to the controller can be characterized
by the H1 norm ratio  =: CL = OL . To characterize the control eort cost of this
performance gain, we dene the controller eort H1 norm as:
u =: k Tuw k1 = k Cclu (sI , Acl ),1 Bcl + Dclu k1 (11)
where Cclu and Dclu are matrices mapping the augmented system states, x^ (t), and
excitation, w(t), to a measure of the controller eort. For this study, the measure
of controller eort is taken as the controller command signal, u(t); however, other
measures can be used, such as the actuator displacement or acceleration. To nd
a candidate controller, the optimization routine minimizes the control eort norm
u subject to a performance criteria constraint of  < des , where des is a specied
desired attenuation ratio. This procedure is used to nd a sequence of controllers
with varying performance and control eort characteristics by solving for controllers
with sequentially lower values of des , while using a previous controller solution as the
2 Smith et al.
 des J1 J2 J3 J4 J5 u xm xam
0.90 0.5507 0.9295 0.1087 0.1126 1.0105 0.0156 0.1424 1.8087
0.75 0.4823 0.7995 0.2310 0.2354 0.9469 0.0469 0.3026 1.6950
0.65 0.4302 0.7071 0.3341 0.3373 0.9302 0.0737 0.4377 1.6651
des J6 J7 J8 J9 J10 max(u) max(xm ) max(xam )
0.90 0.6012 1.1048 0.1109 0.1384 1.2048 0.0337 0.3571 5.6114
0.75 0.5547 1.0363 0.2468 0.2589 1.2281 0.1143 0.7339 5.8319
0.65 0.5300 1.0038 0.3727 0.3848 1.2647 0.1978 1.1139 5.9995
Table 1: Static Feedback Performance Indices

starting point for the optimization routine. The advantages of this solution method are
apparent in the ease at which more complex H1 design modelling can be included in
the solution process. The denition of the H1 norms can be extended to include non-
linear, time-varying, or uncertain model descriptions without signicant modication
of the algorithms developed. Examples of this controller design approach including
bounded uncertainty models and actuator saturation can be found in Breneman et
al. (1997) and Chase et al. (1996).
The design of a dynamic output feedback controllers can be accomplished by
formulating the optimization over the entire dynamic regulator, Ac , Bc , Cc , Dc , and
including sensor noise in the excitation, w(t). This approach leads to optimization
problems that can require extensive computational eort. A simpler approach is to
couple an H1 state feedback controller gain with a state estimator designed by an
H1 or alternate approach, such as a Kalman lter.
Example Controllers
The static and dynamic output controllers presented herein are designed for the
AMD benchmark problem using a conguration where the sensor measurements y(t)
are the absolute oor accelerations of the structure, the acceleration of the actuator
piston, and the direct measurement of the ground acceleration. The external excita-
tion, w(t), is the ground acceleration, and the regulated output is the full regulated
output vector, z(t), dened by Spencer et al. (1997) which includes the structural
displacements, velocities, and accelerations, and the actuator displacement, velocity
and acceleration. A 12th order reduced dynamic model used for design of the exam-
ple controllers was obtained by standard balanced model truncation techniques with
tools available in the MATLAB controls toolbox. A frequency domain comparison of
this reduced order model and the evaluation model shows that the dominant dynamic
characteristics of the system are retained in the design model for the frequency ranges
of most interest.
Static output feedback controller gains were generated for this conguration by
the described optimization method for des values of 0.95 to 0.10 in 0.05 increments and
the performance measures J1 though J10 obtained from a SIMULINK analysis with
the evaluation model. Below an attenuation ratio of des = 0:65, the controllers fail to
meet the constraint that the maximum acceleration of the actuator piston is to be less
than 6g's. Above des = 0:65 all the actuator eort contraints are met. The stochastic
response performance indices, J1 through J5, are approximated by simulating the full

3 Smith et al.
 des J1 J2 J3 J4 J5 u xm xam
0.70 0.4931 0.8105 0.1215 0.1268 0.8209 0.0310 0.1592 1.4694
0.40 0.3700 0.5840 0.3068 0.3106 0.6008 0.0834 0.4019 1.0755
0.20 0.2562 0.4032 0.5543 0.5544 0.6374 0.1671 0.7261 1.1409
des J6 J7 J8 J9 J10 max(u) max(xm ) max(xam )
0.70 0.5493 1.0376 0.1347 0.1463 1.0986 0.0806 0.3920 5.1221
0.40 0.4952 0.9536 0.3586 0.3661 0.9836 0.2725 1.0881 4.4155
0.20 0.4055 0.7491 0.8353 0.8782 0.9668 0.6929 2.4975 4.8822
Table 2: Dynamic Feedback Performance Indices

evaluation model for 300 seconds using a dominant excitation frequency, !g , equal
to the rst dominant frequency of the idealized continuous closed loop system. The
performance metrics are presented for the static output feedback controllers with des
values of 0.9, 0.75 and 0.65 in Table 1.
For the dynamic output feedback controllers, the state feedback controller gains
are obtained using the same method as used for the static output feedback controller
gains, assuming the controller has complete access to the states of the 12th-order
reduced order model. A series of controller gains were again obtained, using des values
of 0.95 to 0.10 in increments of 0.05. A Kalman lter sensor error gain matrix, L, is
obtained from the standard LQE method as implemented in the MATLAB Controls
Toolbox command LQEW. For the Kalman lter design, the weighting matrices are
equivalent to an RMS value of the ground excitation of 0.1g and the RMS values of
the sensor noise at 0.01 volts, with the additional assumption that the noise terms are
independent. Other weighting matrices can be used in the estimator implementation;
however, simulation results did not show a signicant improvement in the composite
regulator's performance. The resulting continuous regulators are discretized via the
Tustin method, and the performance indices determined with MATLAB Simulink
analysis. The performance indices for the dynamic output feedback controllers for
the 12th order model designed for des of 0.7, 0.4, and 0.2 are presented in Table 2.
The dynamic output controllers designed for des <0.15 do not meet the maximum
absolute actuator acceleration constraint of 6g's.
To compare the eectiveness of the two controller design methods described,
Figure 1 plots the sum of J6 and J7 versus the sum of J8 and J9. In this comparison,
the structural response is characterized by the normalized maximum structural drift,
J6, and the normalized maximum structural acceleration, J7. The actuator eort
is characterized by the normalized actuator displacement, J8, and the normalized
actuator velocity, J9. Figure 1 shows data points for the controllers from the static
output feedback (SOF) and the dynamic output feedback (DOF) methods that meet
all the actuator constraints. As is apparent in Figure 1, the dynamic output controllers
perform consistently more eectively than the static output controllers considered in
this comparison. Additionally, much greater controller eort can be applied before the
actuator contraints are not met. Although more ecient dynamic and static output
controllers may be designed by more complex modelling, this study quanties the
improvements of controller performance gained by the additional complexity of using
a H1 dynamic output feedback controller with the AMD benchmark problem.

4 Smith et al.
 1.8

DOF Controllers
1.7 SOF Controllers

1.6

1.5
Response J6+J7

1.4

1.3

1.2

1.1

1
0 0.5 1 1.5 2 2.5
Controller Effort J8+J9

Figure 1: Static Versus Dynamic Output Feedback Eciency

Conclusions
In summary, this research presented H1 based controller design approaches for
the AMD benchmark problem, where both static and dynamic acceleration feedback
controllers were studied and compared. The performance of example controllers, which
met the contraints posed by the benchmark problem, were compared at various levels
of actuator eort. Simulation results indicate that the H1 dynamic output feedback
controllers were signicantly more ecient than the static output feedback controllers
and capable of attenuating the structural response to lower levels of dynamic response.
Support for this research was provided in part by the National Science Foundation
Presidential Young Investigator Award No. BCS-9058316. In addition, the authors
wish to recognize the collaborative research eorts of Professor Rahmat Shoureshi at
the Colorado School of Mines.
References
1. Breneman, S. E., J. G. Chase, and H. A. Smith (1997). \Robust And LTI H1
Static Output Feedback Design For Systems with Limited Actuator Authority",
Proceedings of the 1997 American Control Conference, Albuquerque, NM.
2. Chase, J. G., S. E. Breneman, and H. A. Smith (1996). \Robust H1 Static
Output Feedback Control With Actuator Saturation", submitted to the ASCE
Journal of Engineering Mechanics
3. Spencer Jr., B. F., S. J. Dyke, and H. S. Deoskar (1997). \Benchmark Prob-
lems in Structural Control", Proceedings of the 1997 ASCE Structures Congress,
Portland, OR.

5 Smith et al.