You are on page 1of 11

Journal of Structural Engineering

Vol. 38, No. 4, October-November 2011 pp.327-337

No.38-29

Neurosemiactive control of 3-storey moment resistant frame with


magnetorheological dampers
K. Rama Raju*, A. Meher Prasad**, and Nagesh R. Iyer*
Email: prasadam@iitm.ac.in
*CSIR-Structural Engineering Research Centre, CSIR Campus, Taramani, Chennai-600113, India.
**Department of Structural Division, Indian Institute of Technology Madras, Chennai-600 036, India.
Received: 16 September 2010; Accepted: 13 October 2010

The paper presents a procedure for development of an optimal semi-active neurocontroller 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 the 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 further extended to train the
ANN corresponding to site-specific earthquakes based on the location of the building.
KEYWORDS: Seismic performance; magnetorheological dampers; toggle braces mechanism; dynamic characteristics; steel moment
resistant frame.

Magneto Rheological (MR) dampers are semiactive 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 dampers intrinsic
nonlinear behaviour. This nonlinearity can be captured
in a neural network and the same can be used to control
the structural systems.
One of the widely used control algorithm in civil
engineering is time domain control algorithm1. In this
algorithm, the control forces are obtained by minimizing
either objective or cost function. Command signals are
calculated based on these desirable control forces and
provided as input into an actuator so that the actuator
can produce force as close as possible to the desirable
one. Algorithms that belong to this category include the
instantaneous optimal prediction control method and
Linear Quadratic Regulator (LQR) method. Dyke et al
327

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

carried out analytical and experimental studies2-3 using


clipped-optimal control algorithm based on acceleration
feedback. In their approach, a linear optimal controller
combined with a force feedback loop was designed to
adjust the command voltage of the MR damper.
A new neuro-genetic control algorithm4 is presented
by Jiang and Adeli for finding optimum control forces.
The control algorithm does not need the pre-training
required in a neural network-based controller, which
improves the efficiency of general control methodology
significantly. Two 3D steel building structures, a 12Storey structure with vertical setbacks and an 8-Storey
structure with plan irregularity, are used to validate the
neuro-genetic control algorithm under three different
seismic excitations. Numerical validations demonstrate
the new control methodology significantly reduces the
displacements of buildings subjected to various seismic
excitations including structures with plan and elevation
irregularities.

Chang and Zhou1, proposed Neural Network (NN)


models to emulate inverse dynamics of MR damper which
calculates voltage signals based on a few previous time
steps of displacement, damper force, voltage signal, and
the desirable control force. These NN models were based
on the input-output generated using phenomenological
model proposed by Spencer et al5. These NN models
were used to calculate voltage signals as input into the
MR damper so that it can produce desirable optimal
control forces. Two control algorithms were selected
for validation. One was the optimal prediction control
for a Single Degree of System (SDOF) system and the
other was the LQR control for Multi Degree of System
(MDOF) systems and both were subjected to the first
20s of North South (NS) component of El Centro ground
acceleration time history. It is to be noted that it is
difficult to train NN for capturing the inverse dynamics
of MR dampers to desired levels of accuracy due to their
complex nonlinear behaviors. Hence, in present 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 modeling the inverse
dynamics of MR damper. A neurocontroller is developed
for controlling a 3-storey SMRF model for 5 types of site
specific earthquake loads.
NEURO-SEMIACTIVE CONTROL
In this paper, an analytical neuro-controller model for a
building model of Dyke et al2, is developed. The peak
ground acceleration of North South (NS) component
of El Centro ground acceleration time history, used for
this model. 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).
OPTIMAL SEMI-ACTIVE CONTROL WITH
MR DAMPERS
The linear quadratic regulator control of a multipledegree-of-freedom system which is subjected to NS
component of the 1940 El Centro ground acceleration
is studied. Program has been developed for control
algorithms by using MATLAB software package to
generate the control force for the next time step based
on the predicted state vector (displacement and velocity) and the predicted excitation force.

MODELING AND IDENTIFICATION OF


SYSTEM
The first step in the identification process is to develop
an input/output model for the MR damper. The model
developed can accurately reproduce the behaviour of the
MR damper. The magnetic field produced in the device
is generated by a small electromagnet in the piston rod.
Spencer et al5, proposed a phenomenological model to
portray the behaviour of a prototype MR damper. This
model is based on a Bouc-Wen hysteresis model and is
governed by the seven simultaneous Eqs. (1)-(7). This
simple mechanical idealization of the MR damper has
been used to accurately predict the behaviour of the
prototype MR damper over a broad range of inputs. The
applied force f predicted by this model is given by
f = C1 y& + k1 ( x x0 )

(1)

The evolutionary/hysteric variable z is given by


Z& = x& y& Z& Z

n1

n
( x& y& ) Z + A( x& y& ) (2)

and this variable is governed by


y& =

Z + C0 x& + k0 ( x y )

(C0 + C1 )

(3)

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 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 accu-mulator, x& = the
calculated velocity of the piston rod in determining the
force generated in the damper model, , and A, are
the hysteresis parameters; = evolutionary coefficient;
A total of 14 parameters (c0a, c0b, k0, C1a, C1b, k1,
x0, a, b, , , , A, n) are obtained to characterize
the prototype MR damper using experimental data
and a constrained nonlinear optimization algorithm5.
By adjusting the hysteresis parameters for the yield
element, , and A, one can control the shape of the
hysteresis loops for the yielding element.

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

328

= (u ) = a + b u

(4)

C1 = C1 (u ) = C1a + C1b u

(5)

(cla + clbv )
fu =

(C0 a + cla ) + (C0b + Clb )v

C0 = C0 (u ) = C0 a + C0bu

(6)

( a + b v) Z u + (C0 a + C0b v) x&

where, u is given as the output of a first- order filter


given by
u& = (u v)

For the prototype MR damper modeled5 by


Spencer et al. the parameters are given in the Table
1. Using those values in Eq. (10) and substituting v
= 0 (minimum) and v = 2.25V (maximum) separately
produce the minimum and the maximum force fmin and
fmax, respectively. These damper parameters are used
for the MDOF example problem in this study, and
they 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 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
(11)

(7)

and v is the command voltage sent to the current


driver. Equation 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 Eq. (2). By adjusting the
parameters , 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 et
al5, showed the upper limit of the evolutionary variable
as

Where, v- voltage, x& - velocity and 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 preyielding 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).

A n
Z u =

(8)

The stiffness contribution in Eqs. (1) and (3) for MR


damper can be numerically shown to be small compared
to the damping contribution and hence is neglected in
the following derivation1. Substituting Eq. (8) into Eq.
(3) and then Eq. (1), the upper limit of the damper force
fu is approximated as
fu =

Cl
( Zu + C0 x& )
C0 + Cl

(10)

(9)

This represents a straight line in the f x& plane.


Under the rheological equilibrium and a constant
voltage condition, the steady state solution of the
first order filters Eq. (7) is u=v. This solution can be
substituted into Eqs. (4)-(6) and then subsequently into
Eq. (9) which results in the following expression:

LQR CONTROL FOR MDOF SYSTEM


State Space Form Description

In general, equation of motion of dynamic system is


described by the second order differ-ential equation.

TABLE 1
MR DAMPER PARAMETERS FOR THE GENERALIZED MODEL
Parameter

c0a

c0b

k0

C1a

C1b

k1

x0

Value

2.1
kNs/m

.35
kNs/mV

4.69
kN/m

28.3
kNs/m

2.95
kNs/mV

.05
kN/m

.143 m

14
kN/m

69.5
kN/mV

3.63
mm-2

301

190
s-1

329

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

The analytical solution of second order differential


equation motion is essentially equivalent to solving
ordinary second order equation. On the other hand,
the first order state space form, description of dynamic
system has certain advantages over second order
form description. The second order equation can be
transformed into the first order equations and the first
order forms also can be transformed into second order
equations. Majority of existing computer software tools
are written for the first order systems.
This is due to the inherent nature of the first order
equations which are more convenient for numerical
computations. Another significant advantage of the first
order descriptions is that analysis of the equations can
be done in an explicit form. In recent years, calculating
the response of the system using an alternative method
called the state space method has become essential for
a broad range of dynamic problems. These structural
engineering situations include the use of manufactured
dampers in buildings, the placement of the base
isolators in buildings, and the use of structural control
methods to reduce earthquake and wind responses. The
state space method analyses the response of the system
using both the displacement and the velocity as the
independent variables, and these variables are called as
states. Consider a linear second order dynamic system,
M X&& (t ) + C X& (t ) + KX (t ) = E f Fe (t ) + DFc (t ) (12)
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, Fc(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 Eq. (12) in the first order form, let
the state vector, {Z } = xx& , then

{}

{z&} =

d x
= [ M ]l [C ]x&
dt x&

{} {

[ M ]l [ K ]x +[ M ]l [ Fc ]d

(13)

{ }

x& (t )
Z& (t ) = &&
x(t )

I
=

M l K M l C

{ xx&((tt)) }+M

(14)

l Fc (t )
+
F
(
t
)
e
l
Ef
M D

Where, 0 is a null matrix with all zero entries, and


I is the identity matrix. To simplify Eq. (14), let
O

I
A =
;
l
l
M K M C
O
O
; E = l
B = l

M D
M E f

Z& (t ) = Az (t ) + BFc (t ) + E f Fe (t )

(15)

(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 of engineering applications. One of the most
commonly used methods of modern control theory is
called optimal control. Like optimal design method,
optimal control centres 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.

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

330

The equations of motion for an n-degree-of-freedom


structure with the mass matrix M, damp-ing matrix C,
and stiffness matrix K equipped with m MR dampers
and subjected to ground acceleration X&& g are given by
M X&& (t ) + CX& (t ) + KX (t ) = E f X&& g (t ) + DFc (t ) (17)
Where the n-dimensional vector x = the relative
displacement; the m-dimension vector, FC = the control
force vector generated by the dampers; and D and I =
the nm and the n1 location matrices for the control
forces and the excitation respectively. Above Eq. (17)
can be rewritten in the state-space form as
z& = Az + Bu + E &&
xg

(18)

Where, z = [x x& ]T a 2n-dimensional state vector.


The matrices A, B, E are defined in Eq. (15).
The classical LQR algorithm has been studied for
active control and for semi-active control of structures.
Let the control force fc in Eq. (16) be obtained by
minimizing the following infinite horizon performance
index:

T
T
J =
Z Qz + f c R f c
dt

Solution for state space form equation (18) is given


Z k +1 = Fs Z k + H d Fk + Guk
Where, for a constant function,
Fs = e At , H d = e At H t ,
G = A1 (e At I ) B, FK = Hg
Numerical Example for MDOF System

The performance of the proposed control strategy is


now evaluated through numerical simulation. A three
story building model2 is as shown in Fig. 1, configured
with a single MR damper proposed is adopted for
demonstration. The MR damper is rigidly connected
between the ground and the first floor of the structure.
The structural matrices are
3-Story Scale-Model Building
xa3

xa2

(19)

Current
Driver

Where, Q = a positive semi-definite and R = a


positive definite weighting matrix, respectively. For
Closed-loop control, minimizing Eq. (19) subject to the
constraint of Eq. (18) results in a control force vector fc
regulated only by the state vector z
1
f c = Rl BT Pz = Gz
(20)
2
Where, G represents the gain matrix; and P is
the solution of the steady-state Riccati Equation. In
the current study, the weighting matrix R and Q are
assumed as
K
Q =
O
T

(21)

(22)

R = rD K D

Where, D = the n m location matrix, r = the


weight parameter reflecting the relative importance
of the reduction in the state vector z as compared to
the required control force vector, fc . In present study r
value is taken as 5.

331

(23)

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

xa1
f, xd
xg
Control
Computer

Rheonetic SD-1000
MR Damper
Fig. 1

Height: 158 cm
Mass: 304 kg

Three storey model (Dyke, 1996)

12
6.84
0
N
K s =10 6.84 13.7 6.84 ;
m

6.84 6.84
0

175 6.84
0

Ns
Cs =50 100 50
m

50
50
0

1
1
98.3 0
0



M s = 0
98.3 0 Kg ; D = 0 ; E f =1

0
98.3
0

Acceleration (m/s2)

The first three natural frequencies are calculated to be


4.46, 15.81, and 23.64 Hz. Their corresponding modal
damping ratios are 0.3%, 0.6% and 0.6% respectively.
The first 20 s of the NS component of the 1940 El
Centro ground acceleration record is again used as the
excitation. The acceleration is to be reproduced at five
times the recorded 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. 2)
and it is to be controlled using prototype MR damper.
The time histories of relative floor displacements of
controlled and uncontrolled 3-storey frame model when
r=5 are as shown in Fig. 3. The peak displacements in 3Storey frame model with LQR algorithm are compared
with Clipped Optimum Control (COC) results by
Dyke et al3, and are given in Table 2. The reduction
in displacements with LQR and COC are found to be
about75% and 78% respectively.
E l C e ntro NS

4
3
2
1
0
-1
-2
-3
0

Displacement (mm)

Displacement (mm)

Fig. 2

Time (s)

10

El Centro NS acceleration time history (Dyke, 1996)

6
First Floor

3
0
-3
-6

8
6
4
2
0
-2
-4
-6
-8
-1 0

Second Floor

Displacement
(mm)

10

W ith o u t c o n tro l

w ith c o n tro l
Third Floor

5
0
-5

-1 0
4

Fig. 3

6 Time (s) 7

10

Displacement time histories in first, second and third floor of the


building model

ARTIFICIAL NEURAL NETWORKS OPTIMUM


CONTROL WITH MAGNETOR-HEOLOGICAL
DAMPER
ANN can be massively parallel and therefore is said
to exhibit parallel distributing processing. Here, only
basic concepts about the selection of neural network
topology for feed forward neural networks are
described.
Number of Hidden Layers and Hidden Neurons in
Network
Most of the proposed neural network architectures
rely largely upon experience, intuition and insights
about the physical system, which is being modeled.
The most difficult aspects of using the artificial neural
networks for capturing the design space are the proper
choice of the size and types of neural network. As a
rule of thumb of Hajela and Berke6 suggested that
the number of nodes on the hidden layer is somewhere
between average and sum of input-output nodes,
is a rule for effectively chosen as an approximation
somewhere between that of a first second order
polynomial approximation. A single hidden layer with
nodes numbering between the average and the sum
of input-output nodes in the literature is a good first
start.
The number of hidden layers and the number of nodes
in each layer is a problem dependent7-8. However numerical experience with several representative problems
suggests that its number be minimized for computational
efficiency. It can be shown that, in the absence of nonlinear activation functions, a multi layer network can be
reduced to a equivalent single layer network.
In the neural network, the weights can be thought
of as unknowns in an equation and the number of
equations is also equal to the number of training pairs
the number of nodes on the output layer. For example,
having 11 training pairs is analogous to having 451
equations to solve. The number of unknowns should
be equal to the number of weights (number of nodes
on the input layer times the number of nodes on the
hidden layer plus the number of nodes on the hidden
layer times the number of nodes on the output layer)
plus the number of nodes on the hidden layer plus the
number of nodes on the output layer (last two terms
represent a bias term on these nodes). A neural network
must be determined (equations greater than or equal
to unknowns) to obtain an adequate approximation.
JOURNAL OF STRUCTURAL ENGINEERING
Vol. 38, No.4, OCTOBER-NOVEMBER 2011

332

TABLE 2
MDOF SYSTEM RESPONSE WITH AND WITHOUT CONTROL FORCE
Responses

Storey

Uncontrolled

With Control LQR

With Control COC

Reduction (%)
LQR

COC

Displacement (mm)

First

5.38

1.34

1.14

75.09

78.81

Second

8.2

2.07

1.85

74.76

77.44

Third

9.6

2.45

2.12

74.48

77.92

Notation: LQR: Linear Quadratic Regulator, COC :Clipped Optimum Control

In general, an accurate method of determining the


most appropriate number of neurons to include in each
hidden layer does not exist. 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 an 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 do not significantly contributed to
the solution. Yet another approach is the Radial-Gaussian
system, which adds hidden neurons to the network in a
se-quential 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.
DATA FOR NEURAL NETWORK
The emulation of the inverse dynamics of the MR
damper can be regarded as an identification problem
333

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

for a complex and unknown nonlinear system voltage


as output. The input-output relationship of this system
can be replicated by the collection of input and output
data, selection of a network structure, training of the
network, and validation of the trained network.
The quality of the trained network is directly
related to the quality of the training data. In order to
make the verified model fully represent the underlying
system, the data must contain information in the
entire operating range of the system. The selection
of a network structure involves the determination of
number of inputs and outputs, hidden layers and nodes
in the hidden layers, and is usually done by the trial
and error. A commonly used method of validation is
to investigate prediction of errors using a few sets of
test data. The data of the control force for the next time
step based on both predicted state vector (displacement
and velocity) and the predicted excitation force (as
given by Eq. 1) has been 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 Eq. (11). From Eq. (12), the
voltage corresponding to the forces and velocities were
calculated, and they were used for generating input
data for ANN. The Eq. (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.
FORWARD NEURAL NETWORK WITH MR
DAMPER
The model given in Fig. 4(a) shows that the forces
generated by the MR damper fu depend on the voltage
v and the displacement response of the structure x
at the location where the damper is attached. This
dynamic behavior is termed as the forward dynamic

behaviour of the MR model damper. The feed forward


model with one hidden layer can satisfactorily emulate the forward dynamic behavior of the MR damper.
Fig.4(a) illustrates schematically the training of
forward dynamic behavior of MR damper with forward
NN model proposed here. Similarly, inverse dynamics
of the MR damper demonstrated by Chang and Zhou,
2002 with forward NN model is as shown in Fig. 4(b).
A feed-forward neural network model is proposed for
implementation of forward NN model as shown in Fig.
4(c).
f
MR Damper
v
Force and
Voltage relation

x
Forward
NN model

Error back
Propagation

(a) Forward NN Model


v

MR Damper
v
x
Inverse NN
model

Error back
Propagation

(b) Inverse NN Model


H130

H215

fn
fn-2

H129

Vn
Vn-2

H214

Force and
Voltage
relation

X1n
X1n-2

f n+1

X2n
X2n-2
X3n

Vn+1
H12

H2 2

X3n-2
H11

H21

(c) Neural Network proposed for Forward NN


Fig. 4

Training of Forward NN models for MR Damper model

A single bay 3-storey building model and its


properties are as shown in Fig.1, 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 4 105. 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 func-tion). The input data generated 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. 4(c). The Eq. (11) gives an explicit relation
be-tween 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)
obtained from feed forward neural network (FNN).
NEURO CONTROL METHODOLOGY USING
MR DAMPER
The way flexible building is controlled under earthquake
ground acceleration using a MR damper is illustrated
in Fig. 5 with the proposed neurocontroller. As shown
in Fig. 5, the next step force from FNN is converted to
next step voltage, using the force-voltage relationship
developed (Eq. 11) for the MR damper and passed
on to Voltage filter for MR damper. If the desirable
control voltage is not realizable, the voltage filter is set
at either zero or the maximum level. If the desirable
JOURNAL OF STRUCTURAL ENGINEERING
Vol. 38, No.4, OCTOBER-NOVEMBER 2011

334

control force is realizable by the MR damper, next step


voltage, v , and the building responses are passed on to
MR damper.
It is found from this study that training the network
for inverse dynamics of MR damper is very difficult
because of its complex nonlinear behavior 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.
RESULTS AND DISCUSSION
After evolving a methodology for training the ANN, an
exercise is taken up to check the working of Forward
neural network and for arriving at the initial connection
weights. The forward neural network is trained with
2950 time steps of forces/voltages from El Centro NS
component time history and tested for the next 50 time
steps of forces/voltages. The results for both trained and
tested output values are found to be matching well.
MR Damper

fa
Building

x , x&

&x&
g

SUMMARY

FANN Control
Algorithm

f
F-V relation

v
Voltage filter
for MR
damper

Yes

No

v = 0 or max
v

Notation : F-V=Force-Voltage; FANN=Feed Forward ANN

Fig. 5 Control Strategy Using MR Damper

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

335

by Ohtori et al9, 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 cutoff. Then the explicit relationship between forces and
voltages given in Eq. (11) is used to convert forces
without cut-off to voltages with cut-off by passing
through the constraint filter shown in Fig. 5 (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 (015 seach) are as shown in Figs. 6 and 7. 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. 8 and 9 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. 7 and 9).
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.

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

Development of Neuro-controllers for controlling


the dynamic response of MDOF has been taken up in
the present study. 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.

Control Force (N)

200
150
100
50

ANN Simulation Training Instances


Actual Training Instances

0
-50
-100
-150
-200

El Centro (NS)
0

Kobe

10

15

20

25

30

35

40

El Centro (SE)

Taft

Northridge
45

50

55

60

65

70

75

Time (sec)
Fig. 6

Training and Prediction of control force data for FANN without cut-off, 5 Earthquake time histories together, 15s of each

200
150

Control Force (N)

ANN Simulation Testing Instances


Actual Testing Instances

100
50
0
-50
-100

El Centro (NS)

-150
-200
0

Kobe

Northridge

El Centro (SE)

Taft
12

16

20

Time (sec)
Fig. 7

Testing and Prediction of control force data for FANN without cutoff, 5 Earthquake time histories each 4s (random samples)

ANN Simulation
El Centro (NS)

2.4

Kobe

Actual Training Instances


Northridge

Taft

El Centro (SE)

Voltage (v)

2
1.6
1.2
0.8
0.4
0
0

10

15

20

25

30

35

40

45

50

55

60

65

70

75

Time (s)
Fig. 8

Training and testing of command voltage data obtained from control force from Forward ANN with cut-off for 5 Earthquake time histories,15s of
each

ANN Simulation Testing Instances

Actual Testing Instances

Voltage (V)

2.5
2
1.5
1
0.5
0
0
Fig. 9

Time (s)

12

16

20

Test voltage and command voltage data obtained from ANN control force simulation, 5 Earthquake time histories each 4s (random samples)

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

336

A single bay 3-Storey building model configured


with a single MR damper and subjected to 5 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 optimum control force. The inputs to the NN model
include the last 3 time steps of structural displacements 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.
The architecture of the NN is developed by a trial and
error procedure.
The neurocontroller is further modified by
implementing the cut-off voltages relevant to the MR
damper considered. The good agreement 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 paper
is general in nature and the similar neurocontrollers can
be developed and used depending upon the location of
the building which decides the DBEs to be considered
for training the NN.
In the present study, testing instants are taken from
the given training instants for ANN. This is because to
train ANN with number of time histories to simulate
any random time history is not possible. Further
investigations are required to find the possibility of
classifying acceleration time histories based on their
frequency distribution, bandwidth and include this
information also as input training instance for ANN.
This may help to train ANN within the required
bandwidth of frequencies, so that the trained ANN
can simulate any random time history coming in the
bandwidth of frequencies.

REFERENCES
1.

2.

3.

4.

5.

6.

7.

8.

ACKNOWLEDGEMENT
9.
This paper is being published with the kind permission
of Director, CSIR-Structural Engineering Research
Centre, Chennai-600113. The authors acknowledge
the inputs given for this work by Prof. D. Janakiram
of Computer Science and Engineering Department,
Indian Institute of Technology, Madras.

337

JOURNAL OF STRUCTURAL ENGINEERING


Vol. 38, No.4, OCTOBER-NOVEMBER 2011

Chang, C. C. and Zhou, L., Neural Network


Emulation of Inverse Dynamics for Magnetorheological Damper Jl. of Struct.Engg., ASCE, Vol.
128, No. 2, 2002, pp 231239.
Dyke, S.J., Spencer Jr. B.F., Sain, M.K. and
Carlson, J.D., Experimental Verifi-cation of SemiActive Structural Control Strategies Using Acceleration Feedback, 3rd International Conference
on Motion and Vibration Control, September 1-6,
Chiba, 1996, Japan.
Dyke, S. J., Spencer, Jr., B. F., Sain, M. K.,
and Carlson, J.D., Experimental Study of MR
Dampers for Seismic Protection Smart Mat. and
Structs., Vol. 7, No. 5, 1998, pp 693703.
Jiang, X. and Adeli, H., Neuro-genetic algorithm
for non-linear active control of structures, Intl.
Jl. for Numerical Methods in Engg., Vol. 75, No.
7, 2008, pp 770786.
Spencer. Jr., B. F., Dyke, S. J., Sain, M. K., and
Carlson, J.D., Phenomenological Model for
Magnetorheological Dampers Jl. of Engg. Mech.,
ASCE, Vol.123, No.3, 1997, pp 230238.
Hajela, P., and Berke, L., Neural Networks in
Structural Analysis and Design: An Overview,
Computer Systems in Engineering, No.1-4, 1998,
pp 525538.
Rama Raju. K., Raghu.M., and Prasanth.S.,
Neural network based technique for Preliminary
design of reinforced concrete columns, Jl of
Struct. Engg., SERC, Vol. 34, No. 4, OctoberNovember 2007, pp 297305.
Rama Raju, K., Studies on seismic performance
enhancement of buildings with viscous fluid
dampers and magnetorheological dampers, PhD
Thesis, Department of Civil Engineering, Indian
Institute of Technology, 2008, Chennai.
Ohtori, Y., Christenson, R. E., and Spencer Jr., B.
F., Benchmark Control Problems for Seismically
Excited Nonlinear Buildings, Jl. of Engg. Mech.,
ASCE, Vol. 130, No. 4, 2004, pp 366385.
(Discussion on this article must reach the editor before
January 31, 2012)

You might also like