Professional Documents
Culture Documents
CH 19
CH 19
Abstract
The chapter gives the procedure for development of an optimal semi-active neurocontrol
methodology for capturing the phenomenological model of a Magnetorheological (MR) damper
using Linear Quadratic Regulator (LQR) algorithm for controlling a 3-Storey Steel moment
resistant frame (SMRF) model. One of the important aspects of structural control is the time
delay associated with the control algorithm used to predict the control force. AI techniques
such as Artificial Neural Network (ANN) can be used to improve the efficiency/performance of
the control module. Keeping this in view, the possibility of application of feed forward neural
network, implementing LQR algorithm for semi-active control of MR damper in SMRF has been
explored. An explicit relation between control force and command signals (voltage) has been
developed for the given MR damper. The Neurocontroller is trained and tested with six types
of earthquake records scaled to Peak Ground Acceleration (PGA) of Design Basis Earthquake
(DBE). This methodology can be extended to train the ANN corresponding to site-specific
earthquakes based on the location of the building.
INTRODUCTION
Magneto Rheological (MR) dampers are semi-active control devices, that use MR
fluids to produce controllable dampers. To develop control algorithms that take
maximum advantage of the unique features of the MR damper, models must be
developed that can adequately characterize the damper’s intrinsic nonlinear behaviour.
This nonlinearity can be captured in a neural network and the same can be used to
control the structural systems.
NEURO-SEMIACTIVE CONTROL
In this chapter, an analytical neuro-controller model for a building model of Dyke et
al (1998) is developed. The peak ground acceleration of North South (NS) component
of El Centro [Dyke et al., 1998] Kobe, Northridge, Taft, Elcentro(SE) ground
acceleration time histories [ohtori et al., 2004] are reduced to 0.2g i.e. design base
earthquake (DBE). This excitation is used for emulation of the response characteristics
of building model with MR damper and the same is used for training the forward
neural network (FNN).
338 Neuro-semiactive control with MR Dampers
x Y
C1 C- Wen
Ku
f
f = c1 y + k1 ( x − x ) (1.1)
0
The evolutionary/hysteric variable z is given by
z = − γ | x − y | z | z | n−1 − β (x − y ) | z | n + A (x − y ) (1.2)
and this variable is governed by Wen (1976)
1 ⎡ α z + c0 x + k0 (x − y )⎤
y = (1.3)
(c0 + c1 ) ⎣ ⎦
Where x = the displacement of the structure at the attachment point of the MR fluid
damper, f = the force generated by the MR damper , y = an internal pseudo displacement
of MR damper, u = output of a first- order filter, v = the command voltage sent to
current driver, k1 = the accumulator stiffness, c0 = the viscous damping observed at
larger velocity, c1 = the dashpot to introduce nonlinear roll-off in the force velocity loops
Advanced course on Computational Structural Mechanics (ACCSM’09) 339
that was observed in the experimental data at low velocities, k0 is present to control
the stiffness at large velocities; x0 = the initial displacement of spring k1 associated
with normal damper force due to the accumulator, x = the calculated velocity of the
piston rod in determining the force generated in the damper model, γ, β and A, are the
hysteresis parameters; α = evolutionary coefficient; A total of 14 parameters (c0a, c0b, k0,
C1a, C1b, k1, x0, αa, αb, η, γ, β , A, n) are obtained to characterize the prototype MR damper
using experimental data and a constrained nonlinear optimization algorithm (Spencer,
Jr. et al 1997). By adjusting the hysteresis parameters for the yield element, γ, β and A,
one can control the shape of the hysteresis loops for the yielding element.
α = α(u) = αa + αbu (1.4)
c1 = c1 (u) = c1a + c1bu (1.5)
c0 = c0 (u) = c0a + c0bu (1.6)
where, ‘u’ is given as the output of a first- order filter given by
u = −η (u − v ) (1.7)
and m ‘v’ is the command voltage sent to the current driver.
Equation (1.7) is necessary to model the dynamics involved in reaching rheological
equilibrium and in driving the electromagnet in the MR damper.
MR Constraint Filter
The hysteric behavior of the damper is portrayed by the evolutionary variable in
Equation (1.2). By adjusting the parameters β, γ and A, the linearity in the unloading and
the smoothness of the transient from the pre-yielding to the post-yielding region etc., are
controlled. Spencer (1996) showed the upper limit of the evolutionary variable as
1
⎛ A ⎞⎟n
zu = ⎜⎜⎜ ⎟⎟ (1.8)
⎝ γ+ β ⎠⎟
The stiffness contribution in Equations (1.1) and (1.2) for MR damper can be
numerically shown to be small compared to the damping contribution and hence is
neglected in the following derivation (Chang and Zhou, 2002). Substituting Equation
(1.8) into Equation (1.3) and then Equation (1.1), the upper limit of the damper force
fu is approximated as
c1
fu = (α zu + c0 x ) (1.9)
c0 + c1
This represents a straight line in the f – x plane as shown in Fig.1.2. Under the rheological
equilibrium and a constant voltage condition, the steady state solution of the first order
filters Eq. (1.7) is u=v. This solution can be substituted into Equations (1.4)-(1.6) and
then subsequently into Eq. (1.9) which results in the following expression:
(c1a + c1b v) ⎡(α + αb v ) zu + (c0 a + cob v ) x ⎤
fu = (1.10)
⎡(c0 a + c1a )+ (c0b + c1b )v ⎤ ⎣ a ⎦
⎣ ⎦
For the prototype MR damper modelled by Spencer et al. (1997b) the parameters
are given in the Table 6.1. Using those values in equation 6.10 and substituting the
v = 0 (minimum) and v = 2.25V (maximum) separately produce the minimum and
340 Neuro-semiactive control with MR Dampers
the maximum force fmin and fmax, respectively, as shown in Fig.1.2. These damper
parameters are used for the MDOF example problem in thesis, and are different from
the damper parameters used by Dyke et al (1996) in their experimental and analytical
investigation of the same problem.
By substituting all parameters in Table 1.1 in above equation, it simplifies to a
quadratic form
v2(1320.16+10.33 x )+v(126911.26+1052.45 x -6.45fu)+
(25511.32+5943 x -304fu)=0 (1.11)
Where, v- voltage, x - velocity, fu –control force
Any intermediate line produces a straight line inside the range between two extremes.
The region represents an approximate range of forces realizable by the MR damper
under different voltage inputs. Another realizable region also exists in the third quadrant
and is anti-symmetrical to the one in the first quadrant. It should be emphasized that
this is only an approximate range and some errors exist especially in the pre-yielding
region where the velocity is small. The rule of the constraint filter is quite straight-
forward. For a given velocity value, the voltage is set at the max (min), when the
desirable control force is larger than fmax (smaller than fmin).
fmin (v =0V) v =1.125 f max (v =2.25V)
1200
1000
997
800
Force (N)
600
400
200
83.8
0
0 2 4 6 8 10
Velocity (mm/s)
Fig.1.2 Max and Min. Force vs. Velocity Curve
Advanced course on Computational Structural Mechanics (ACCSM’09) 341
Where, M, C and K are respectively, the n×n mass, damping and stiffness matrices
x(t) is the n-dimensional displacement vector, Fe (t ) is an r-dimensional vector
representing applied load or external excitation, Fe (t ) is the p-dimensional control
force vector, D defines the control force distribution matrix of size n × p, Ef defines
the excitation force location matrix of size n × r. In this equation, p denotes the total
number of control forces and r denotes the total number of excitation forces. Note that
one control force can affect several different degrees of freedom, and each ith DOF
that the jth control force affects is represented by Dij. Similarly one excitation force
can affect several degrees of freedom; each ith DOF that the jth excitation force affects
is represented by Ef.
To transfer the above equation in the first order form, let the state vector
{z}= ⎧⎨
x⎫
⎬ then
⎩ x ⎭
d ⎪⎧⎪ x⎪⎫⎪ ⎪⎧⎪ x ⎫⎪
⎪⎬
{z} = ⎨ ⎬=⎨ −1 −1 −1 (1.13)
dt ⎪⎩⎪ x⎪⎭⎪ ⎪⎩⎪−[M ] [C ] x −[M ] [K ] x + [M ] [Fc ] d ⎪⎪⎭
342 Neuro-semiactive control with MR Dampers
⎪⎧ x (t )⎪⎫⎪ ⎡ 0 I ⎤ ⎪⎧ x (t ) ⎪⎫ ⎪⎧ 0 ⎪⎫
z (t ) = ⎨⎪ ⎬ = ⎢⎢ ⎥ ⎨⎪ ⎬⎪ + ⎨⎪ −1 ⎬⎪ Fe (t )
⎪ ⎪ ⎥ ⎪ ⎪ ⎪ ⎪
⎩⎪ x (t )⎪⎭ ⎣−M K −M C ⎦ ⎩⎪ x (t )⎭⎪
− 1 − 1
⎩⎪M E f ⎭⎪
⎧⎪ 0 ⎫⎪ (1.14)
+ ⎪⎨ −1 ⎪⎬ Fc (t )
⎪⎩⎪M D⎪⎭⎪
Where, ‘0’ is a null matrix with all zero entries, and ‘I’ is the identity matrix. To
simplify Eq. (6.14), let
⎡0 I ⎤ ⎡ 0 ⎤ ⎡ ⎤
A=⎢ ⎥ ; B = ⎢ −1 ⎥; E=⎢ 0 ⎥ (1.15)
⎢−M −1 K −M −1C ⎥ ⎢M ⎥
E f ⎥⎦ ⎢ −1 ⎥
⎣ ⎦ ⎢⎣ ⎣ M D⎦
z (t ) = Az (t )+ BFc (t )+ E f Fe (t ) (1.16)
Here, the original second order differential equation is rewritten in the first order form
by introducing the state vector {z}. The state vector associated properties constitutes
the so-called State- Space. The size of the first order system, is increased by two
fold compared to the original second order system. However, the first order form
has certain advantages being adopted in the majority engineering applications. One
of the most commonly used methods of modern control theory is called optimal
control. Like optimal design method, optimal control centers on choosing a cost
function or performance index to minimize. Although this again raises the issue of
how to choose the cost function, optimal control is a powerful method of obtaining a
desirable vibration response. Optimal control formulations also allow a more natural
consideration of constraints on the state variables as well as consideration for reducing
the amount of time, or final time, required for control to bring the response to a desired
level
The optimum structural control of a MDOF modified for semi-active control of
structure subjected to ground motion is explained below.
The equations of motion for an n-degree-of-freedom structure with the mass matrix
M, damping matrix C, and stiffness matrix K equipped with m MR dampers and
subjected to ground acceleration are X g given by
.. .
M X (t )+ C X (t )+ KX (t ) = E f X g (t )+ DFc (t ) (1.17)
rate in order to consider the scaling effect. The time increment for this example is
set at 0.02s. The model is subjected to NS component of the 1940 El Centro ground
acceleration record (Fig.1.4) and it is to be controlled using prototype MR damper.
The time histories of relative floor displacement and velocity of uncontrolled and
controlled 3-Storey model when r=5 are as shown in Fig.1.5 and the peak displacement
and velocities in three storey are tabulated in Table 1.3. The reduction in displacement
and velocity is found to be about 72% and 77% respectively.
Current
Driver
Control
Computer
Rheonetic SD-1000
Height: 158 cm
MR Damper
Mass: 304 kg
1
0
-1
-2
-3
0 2 4 6 Time (s) 8 10
Table 1.3 MDOF system response with and without control force
Responses Storey Uncontrolled With Control Reduction (%)
First 4.82 1.34 72.20
Displacement
Second 7.57 2.07 72.66
(mm)
Third 9.07 2.45 72.99
First 143.1 32.9 77.01
Velocity (mm/s) Second 225.4 52.0 76.93
Third 269.9 62.6 76.81
-20
40 Second Floor
Velocity
20
(cm/s)
0
-20
-40
40
Third Floor
20
Velocity
(cm/s)
0
-20
-40
6 F i r st F l o o r
3
Displacem
ent (mm)
0
-3
-6
10
Sec o nd F l o o r
Displaceme
5
nt (mm)
0
-5
-10
(mm)
ement
0
-10
4 5 6 Tim e (s)7 8 9 10
Fig 1.5 Displacement and Velocity time histories in first, second and third floors
of the building model
346 Neuro-semiactive control with MR Dampers
I I I
1 2 l
Input Layer
i=1,2,…l
W
hi
H Hidden Layer
h
h=1,2…m
W
oh
Output Layer
o=1,2…n
O O
2 n
( T1 - O1 ) ( T2 - O2 ) ( Tn - On )
Fig. 1.6 Typical three-layer back-propagation neural network
348 Neuro-semiactive control with MR Dampers
Generally, there is no direct and precise way of determining the most appropriate
number of neurons to include in each hidden layer and this problem becomes more
complicated as the number of hidden layers in the network increases. From the
understanding developed in graphical interpretation, it appears that increase in the
number of hidden neurons provides a greater potential for developing a solution
surface that fits closely to that implied by training patterns. In practice, however a large
number of hidden neurons can lead to a solution surface that while fitting the training
points, deviates dramatically from the trend of the surface at intermediate points to
that provides two literal of an interpretation of the training points. In addition, a large
number of hidden neurons slow down the operation of the neural network both during
learning and in views, if it is implemented using a software emulation and usually
the case conversely an accurate model of some or all features in the solution surface
may not be achieved with too few hidden neurons are included in the network. In
attempt to resolve this dilemma, a range of different configurations of hidden neurons
is normally considered and that with the best performance is accepted. Use could be
made however of a training system that evaluates automatically the utility of alternative
configurations of hidden neurons. One such technique starts by training a relatively
large network that is later reduced in size by removing the hidden neurons that don’t
significantly contributed to the solution. Yet another approach is the Radial-Gaussian
system, which adds hidden neurons to the network in a sequential manner training
each on the error left over from his predecessors. The number of hidden neurons
required to achieve a given level of desired accuracy is thus determined automatically
during training.
Properties of the Network used
The properties like transfer functions, performance functions, training functions,
learning functions are mainly problem dependant. Based on the nonlinearity and
complexity involved, following network is used which is found to be good for the
present problem after a number of trials.
generated from the LQR control algorithm for MDOF system using a program
developed in MATLAB. The force data for forward ANN is generated using the
equation (1.11). From the equation (1.12), the voltage corresponding to the forces and
velocities were calculated, and they were used for generating input data for ANN. The
equation (1.11) gives an explicit relation between control force and command signals
(voltage) and it is used to calculate next time step command signal (voltage) from the
next time step control force obtained from feed forward neural network.
MR Damper
fa
Building x , x
xg
FANN Control
Algorithm
.
v = 0 or max
F-V relation
Voltage No
filter for
MR d amper
Yes v
A single bay 3-storey building model and its properties are as shown in Fig 1.3,
considered by Dyke et al (1996) configured with a single MR damper and subjected
to El Centro NS component acceleration, is used for demonstration. As mentioned
earlier, the selection of network structure involves in the determination of number
of inputs, outputs, hidden layers and neurons in them and is usually done by trial
and error procedure. The properties like transfer functions, performance functions,
training functions, learning functions are mainly problem dependant. Based on the
nonlinearity and complexity involved, following network is used which is found to be
good for the present problem after a number of trials.
A neural network with two hidden layers H1 and H2 containing respectively 30 and
15 neurons is used for simulation. For neural network simulation, “Neural network
toolbox in MATLAB 6.1” is used. The transfer functions for the first and second layers
are logsig, and for output layer is purelin (Linear transfer function). The performance
function used for the neural network is Sum Squared Error (SSE).
In the present study SSE is set to 4x10-5. The neural network is trained with training
function, trainlm (Levenberg-Marquardt backpropagation algorithm) with 5000
epochs. The learning function of neural network is learngdm (Gradient descent with
weight/bias learning function). The input data generated in section 1.7.4 for FNN
consist of last three steps of displacements (Xn, Xn-1, Xn-2) for each floor, three steps of
forces (fn , fn-1, fn-2), three steps of voltages (Vn,Vn-1,Vn-2), and one output corresponding
to next time step control force, fn+1 is given as input for training forward neural
network as given in Fig.1.7 (c). The equation (1.11) gives an explicit relation between
control force and command signals (voltage) and it is used to calculate next time step
command signal (voltage), Vn+1 (also expressed as v ), from the next time step control
force, fn+1 (also expressed as f ) obtained from feed forward neural network (FNN).
v
M R D am p er
v
x
E rro r b a ck
P ro p ag atio n
In v erse N N
m o d el ∑
v
H 130 H 215
fn
fn -2 H 129 H 214
Vn
Vn -2
f̂
X 1n
F o rce an d
V oltag e rela tion
X 1n fn +1
X 2n
Vn +1
X 2n
X 3n H 12 H 22
X 3n -2
H 11 H 21
It is found from this study that training the network for inverse dynamics of MR
damper is very difficult because of its complex nonlinear behaviour which results in
inaccurate predictions of command voltages. Hence, this command voltage is input
into the MR damper which then produces force, fa required to act on the building.
forces/voltages. The results for both trained and tested output values are found to be
matching well.
The same forward neural network is trained with 7500 time steps of forces/voltages
from five earthquake time histories, i.e., NS component of El Centro used by Dyke et
al (1998), Northridge, Kobe, Taft and SE component of El Centro (0-15secs of each)
given by Ohtori et al (2004) which are reduced to design basis earthquake acceleration
0.2g for this problem. Actual training control force data from LQR algorithm and
testing data from dynamic NN model are without cut-off. Then the explicit relationship
between forces and voltages given in Equation 1.11 is used to convert forces without
cut-off to voltages with cut-off by passing through the constraint filter shown in Fig.1.7
(lower and upper limits of 0 V and 2.25V respectively). The trained and tested control
force data used for Forward ANN without cut-off for 5 types of earthquakes (0-15s
each) are as shown in Figs. 1.9 and 1.10. From figures it is noted that trained and
predicted output values of forces from forward Neural Network are in good agreement
with actual values. The training and predicted command voltage data obtained from
forward dynamics NN with cut-off for 5 earthquake time histories (0-15s of each) are
shown in Figs. 1.11 and 1.12 and are found to be in good agreement. The validity of
the NN predictions is checked for random sample time histories of control forces and
voltages with trained data (Figs. 1.10 and 1.12). The good agreement between the
control forces and voltages predicted with those from simulation over random time
segments for five earthquakes considered suggest that the neuro-controller developed
can be used to control the vibration of structures.
200
ANN Simulation Training Instances
150
Control Force (N)
200
150 ANN Simulation Tes ting Ins tances
Actual Tes ting Ins tances
100
50
0
-50
-100
-150 El Centro (NS) Kobe Northridge Taft El Centro (SE)
-200
0 4 8 12 16 20
Time (sec)
Fig.1.10 Testing and Prediction of control force data for FANN without cutoff
Advanced course on Computational Structural Mechanics (ACCSM’09) 353
1.6
1.2
0.8
0.4
0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75
Time (s)
Fig.1.11 Training and testing of command voltage data obtained from control
force from Forward ANN with cut-off for 5 EQ time histories,15s of each
ANN Simulation Random Testing Instances
3
Voltage (v)
2.4
1.8
1.2
0.6
0
0 4 8 Time (s) 12 16 20
Fig.1.12 Test voltage and command voltage data obtained from ANN control
force simulation (Random sample from 5 EQ time histories each 4s)
SUMMARY
Development of Neuro controllers for controlling the dynamic response of MDOF has
been taken up in this chapter. In order to achieve the objective, a forward NN (which
finds the force required to produce the desirable control force under current response
condition) is trained for solving forward dynamics of MR damper. It is noted that it
is difficult to train NN for capturing the inverse dynamics of MR dampers (which
finds the voltage required to produce desirable control voltage under current response
condition) to desired levels of accuracy due to their complex nonlinear behaviour.
Hence, in this study an explicit relation between control force and command signals
(voltage) has been developed for the given MR damper and the same is used for
modelling the inverse dynamics of MR damper. In the present investigation the
complex nonlinear behaviour of MR damper is represented by the phenomenological
model [Spencer, 1997] based on Bouc-Wen model.
A single bay 3-storey building model considered by Dyke et al (1996) configured with
a single MR damper and subjected to five types of earthquake excitations representing
the design basis earthquake loads is considered for training the NN. The LQR algorithm
has been implemented for finding the control force for r = 5 (Equation 1.22). The
inputs to the NN model include the last three time steps of structural displacements
354 Neuro-semiactive control with MR Dampers
at the location where the damper is attached, damper forces and command voltages.
The output is the command force to be supplied to the MR damper. Neural network
architecture is developed for a semi-active control of a MR damper using trial and
error procedure. The neurocontroller is further modified by implementing the cut-off
voltages relevant to the MR damper considered. The good agreement of the between
the control forces and command voltages predicted with those from simulation over
random time segments for 5 earthquakes considered suggest that the neurocontroller
developed can be used to control the vibration of structures.
While the methodology presented in this chapter is general in nature, similar neuro-
controllers can be developed and used depending upon the location of the building,
which in turn decides the DBEs to be considered for training the NN.
REFERENCES
Chang, C. C. and Zhou, L. (2002). “Neural Network Emulation of Inverse Dynamics
for Magnetorheological Damper” Journal of Structural Engineering, ASCE, 128,
pp. 231-239
Dyke, S.J., (1996), “Acceleration Feedback Control Strategies for Active and Semi-
Active Control System: Modeling, Algorithm Development, and Experimental
Verification”
Dyke, S. J., Spencer, Jr., B. F., Sain, M. K., and Carlson, J.D. (1998). “Experimental
Study of MR Dampers for Seismic Protection” Smart Materials and Structures,
7, pp.693-703.
Ohtori, Y., Christenson, R. E., and Spencer Jr., B. F (2004), “Benchmark Control
Problems for Seismically Excited Nonlinear Buildings”, Journal of Engineering
Mechanics, ASCE, 130, pp 366-385.
Rama Raju, K., Meher Prasad. A and Muthumani. K (2003), “An overview of
present and future directions in active and passive control analysis and design
methodologies used in civil engineering structures”, National seminar on Smart
Materials, Intelligent Structures & Systems in Civil Engineering, Bangalore.
Rama Raju. K., Raghu.M., and Prasanth.S.,(2007), “Neural network based technique
for Preliminary design of reinforced concrete columns”, Journal of structural
Engineering, 34, October-November 2007, pp297-305.
Rama Raju. K (2009), “Studies on seismic performance enhancement of buildings with
viscous fluid dampers and magnetorheological dampers”, PhD thesis submitted to
Indian Institute of Technology, Madras.
Spencer. Jr., B. F., Dyke, S. J., Sain, M. K., and Carlson, J.D. (1997). “Phenomenological
Model for Magnetorheological Dampers” Journal of Engineering Mechanics,
ASCE, 123, pp.230-238