Engineering Structures 33 (2011) 31433154

Contents lists available at SciVerse ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Adaptive control of base-isolated structures against near-field earthquakes using variable friction dampers
Osman E. Ozbulut , Maryam Bitaraf, Stefan Hurlebaus
Zachry Department of Civil Engineering, Texas A&M University, 3136 TAMU College Station, TX 77843-3136, United States

article

info

abstract
This paper investigates the effectiveness of two adaptive control strategies for modulating the control force of variable friction dampers (VFDs) that are employed as semi-active devices in combination with laminated rubber bearings for the seismic protection of buildings. The first controller developed in this study is an adaptive fuzzy neural controller (AFNC). It consists of a direct fuzzy controller with self-tuning scaling factors based on neural networks. A simple neural network is implemented to adjust the input and output scaling factors such that the fuzzy controller effectively determines the command voltage of the damper according to current level of ground motion. A multi-objective genetic algorithm is used to learn the shape of the activation functions of the network. The second controller is based on the simple adaptive control (SAC) method, which is a type of direct adaptive control approach. The objective of the SAC method is to make the plant, the controlled system, track the behavior of the structure with the optimum performance. Here, SAC methodology is employed to obtain the required control force which results in the optimum performance of the structure. For comparison purposes, an optimal linear quadratic Gaussian (LQG) controller is also developed and considered in the simulations together with maximum passive operation of the friction damper. The results reveal that the developed adaptive controllers can successfully improve the seismic response of base-isolated buildings against various types of earthquake. 2011 Elsevier Ltd. All rights reserved.

Article history: Received 5 March 2011 Received in revised form 31 May 2011 Accepted 11 August 2011

Keywords: Variable friction damper Semi-active control Adaptive control Base isolation Near-field earthquake

1. Introduction One of the main considerations of civil engineers when designing a structure is to reduce the risks of damage and injuries caused by dynamic loads such as earthquakes and strong winds. Over past decades, base isolation has been found to be an effective means to protect structures and their contents from the destructive effects of dynamic excitations. However, recent studies have shown that near-field earthquakes characterized by longduration pulses with peak velocities result in significant relative displacements at the isolation level of a seismically isolated structure [13]. As a result of large isolator displacements, the size of the isolation device significantly increases. This may require very large seismic gaps between buildings or large bridge expansion joints. Therefore, these requirements, in return, increase the cost of the construction, which contradicts the primary goal of seismic isolation to design structures more efficiently and economically by reducing the earthquake forces transferred to the superstructure. In order to enhance the performance of base-isolated structures against near-field excitations, passive, active, and semi-active control devices have been proposed [410]. Passive systems

Corresponding author. Tel.: +1 979 458 1222. E-mail addresses: ozbulute@tamu.edu (O.E. Ozbulut), mbitaraf@tamu.edu (M. Bitaraf), shurlebaus@civil.tamu.edu (S. Hurlebaus). 0141-0296/$ see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.engstruct.2011.08.022

can reduce the deformations of the isolation bearings during strong ground motions; however, they can cause an increase in superstructure response due to large damper forces applied to the structure [1113]. Furthermore, in the case of a moderate or weak excitation, passive devices can have detrimental effects on an isolated structure since the desired isolation characteristics may be different for these ground motions and passive devices cannot be adapted online. On the other hand, active devices are generally able to control the seismic response of the isolation system for a wide range of loading conditions. However, active devices use an external energy source to produce the control forces imparted on the structure and they require a considerable size of power source, which makes them vulnerable to power failure. Also, an active control system has the potential to destabilize the structural system. Semi-active control devices have received considerable attention in recent years because of their great adaptability with low power requirements [1417]. These devices only absorb or store the vibratory energy and they do not input the energy to the system. Several researchers have studied the use of a semi-active device in a base isolation system in order to reduce the displacement response of an isolation system without an increase in superstructure response. Yoshioka et al. [18] performed experimental tests to demonstrate the effectiveness of a base isolation system that employs a magnetorheological (MR) damper. Symans and Kelly [19] investigated the performance of a variable viscous damper modulated by a fuzzy

3144

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

controller for the seismic protection of an isolated bridge. In another study, Nagarajaiah and Sahasrabudhe [20] proposed a variable stiffness device that is used in a sliding isolation system to reduce the seismic response of a base-isolated building, and the effectiveness of the proposed semi-active device was shown by performing analytical and experimental studies. Madhekar and Jangid [21] evaluated the dynamic response of a seismically isolated benchmark bridge equipped with viscous and variable dampers and assessed the performance of such dampers. One semi-active device that appears to be promising for structural control applications is the variable friction damper. Over the past few years, friction dampers have been widely investigated as an energy dissipation system to reduce the structural response due to dynamic loading [2225]. A typical passive friction damper consists of two solid bodies that slide over each other to develop friction, which absorbs energy. As is the case for other passive devices, passive friction dampers cannot adjust the slip force in real time according to the structural response. When the contact force is very large, the damper will not dissipate energy for moderate and weak ground motions since it may not slide. On the other hand, when the contact force is reduced, the damper will have small energy dissipation capacity under strong earthquakes due to its small sliding friction force. Therefore, a controllable normal force is favorable in order to ensure the required amount of energy dissipation for various levels of ground motion. A variable friction damper enables the adaptation of the contact force by an actuator, and it can effectively suppress the vibrations of a structure. Several variable friction dampers have been conceptually investigated and some have been implemented [2630]. In order to utilize the full capabilities of a variable friction damper employed in a smart isolation system, an effective control algorithm that is practically viable is needed. However, the task of developing an optimal controller is challenging due to uncertainties in the nature of ground motions and in the characteristics of isolation system. For instance, a controller designed for nearfield ground motions that cause significant deformations in the isolation system might develop large damper forces during a farfield earthquake of generally more moderate excitation. As a result, the isolation system may not perform as expected, and a significant increase in the acceleration response of superstructure can be observed. Alternatively, if the controller is designed for an earthquake with far-field characteristics, the damper force may not be large enough to effectively damp out the motion during a pulselike ground motion. In the past, several control algorithms have been proposed to modulate the contact force of variable friction dampers. Inaudi [31] developed a modulated homogenous friction (MHF) strategy that assumes that the slip force of the damper is proportional to the absolute value of the prior local peak of the damper deformation. He et al. [32] improved the MHF controller to enable slippage in the damper at different levels of seismic loading. Chen and Chen [29,33] suggested a viscous and Reid friction (VRF) control strategy in which the control force is computed as a function of the displacement and velocity of the damper. Xu and Chen [34] presented a modified VRF controller with a Kalman filter that uses the acceleration response as the feedback signal. Lu [35] proposed a control algorithm named predictive control that takes into account the features of friction dampers and predicts a slip force such that the damper remains in its slip state throughout the excitation. Ng and Xu [36] developed a non-sticking friction controller that uses the actuated velocity in the current state of the damper to ensure a continuous slipping and energy dissipation. In this study, two adaptive control strategies are developed in order to adjust the contact force of the variable friction dampers (VFDs) that are used in a smart isolation system. The first controller is based on an intelligent control strategy. In particular, fuzzy logic control theory is used to design an adaptive controller whose input

and output scaling factors are tuned by a simple neural network according to the current level of ground motion. The adaptive fuzzy neural controller (AFNC) determines the command voltage by using isolation displacement and velocity as the two input variables. The second control method that is used to improve the performance of a base-isolated structure is the simple adaptive control (SAC) method. The goal of SAC is to force the controlled system to track the behavior of the model which exhibits the desired behavior. In other words, the error, or difference between the output of the controlled system and model, is reduced using the SAC method. This method does not require explicit system identification because in the direct method the identification and control functions are merged into one scheme in which the control gains are computed directly. SAC is suitable for control systems with uncertainty, disturbance, or any change in the characteristics of the system such as damage. Here, the required force to optimize the performance of a base-isolated structure against different earthquakes is obtained using the SAC method, and then, the command voltage of the variable friction damper is determined to generate the required force. Also, a linear quadratic Gaussian (LQG) controller is designed and considered in the simulations together with maximum passive operation of the variable friction damper for comparison purposes. A five-story building isolated by laminated rubber bearings is modeled together with variable friction dampers that are installed to the base of the building. Numerical simulations of the base-isolated building are performed and various response quantities are evaluated in order to assess the performance of the controllers. 2. Modeling of a base-isolated structure with variable friction dampers Consider an n degree of freedom linear base-isolated structure subject to earthquake ground acceleration xg . The equation governing the dynamic response of the structural system equipped with variable friction dampers located at certain levels of the structure is given by

Mx(t ) + Cx(t ) + Kx(t ) = MIxg (t ) + Du(t ),

(1)

where M, C, and K denote n n mass, damping, and stiffness matrices, respectively; D is an n 1 damper location vector; I is an n-dimensional identity matrix; x(t ), x(t ), and x(t ) are n 1 displacement, velocity, and acceleration vectors, respectively; and u(t ) is an n 1 control force vector. Rewriting Eq. (1) in state-space form gives

z(t ) = Az(t ) + Bu(t ) + Hxg (t ),

T

(2)

where z(t ) = [x(t ), x(t )] denotes the state vector of the system. Also, the system matrix A and force location matrices B and H are defined as follows:
A2n 2n = 0 M1 K I M1 C

B2n m =

(3)

M1 D

H2n r =

0 . I

In this study, the variable friction damper investigated experimentally and analytically by Lu and Lin [37] is adopted in the numerical simulations. This damper employs a piezoelectric actuator to provide a controllable contact force, as shown in Fig. 1. The friction force of the variable damper is proportional to the normal contact force and the friction coefficient between the friction pad and the friction bar of the damper. The contact force of the damper is given as N (t ) = Npre + Cpz V (t ), (4)

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

3145

Fig. 1. Schematic illustration of the components of the variable friction damper: (a) side view, (b) top view. Source: Lu and Lin [37].

Fig. 2. Acceleration and velocity response spectra for the selected earthquakes.

where N (t ) is the total normal contact force, Npre represents the constant preload, Cpz is the piezoelectric coefficient of the piezoelectric actuator, and V (t ) is the applied voltage on the stack actuator. Then, the friction force of the semi-active damper, f (t ), is given by

3. Adaptive control strategies Adaptive control is a method by which a controller is designed to have the capability of tuning its parameters automatically. The primary goal of using adaptive controllers is to improve the performance of the controller online in the face of changing/uncertain parameters. This section introduces two adaptive controllers that are designed to adjust the contact force of variable friction dampers according to the current level of external excitation. 3.1. Adaptive fuzzy neural control strategy One of the effective methods to deal with complex nonlinear systems is the fuzzy logic approach. Fuzzy logic enables one to describe relationships between the inputs and outputs of a controller using simple verbose statements instead of complicated mathematical terms. Due to its inherent robustness and simplicity, several researchers have used fuzzy logic theory to develop controllers for semi-active devices [3840]. The design of a fuzzy logic controller (FLC) involves four main steps: (i) fuzzification of variables, which includes the transformation of the crisp inputs to fuzzy variables by defining membership functions to each input, (ii) definition of a rule base that relates the inputs to output by means of ifthen rules, (iii) definition of an inference engine that evaluates the rules to produce the system output, and (iv) defuzzification, where the output variable that is a fuzzy quantity is transformed to a non-fuzzy discrete value. The performance of a conventional FLC depends on various controller parameters such as the scaling factors, the membership functions, and the rule base. In order to improve the effectiveness of the controller, an adaptive FLC can be designed by altering each of these parameters. Here, neural networks, trained with the Non-dominated Sorting Genetic Algorithm version II (NSGA-II), are employed to modify the input and output scaling factors of a fuzzy controller to ensure satisfactory controller performance for both near-field and far-field ground motions. By tuning the scaling

f (t ) = N (t )sgn( ) if x = 0 x N (t ) f (t ) N (t ) if x = 0,

(5) (6)

where is the friction coefficient of the damper and sgn( ) denotes x the sign of the slip rate of the damper. The parameters Npre , Cpz , and are given as 1000 N, 1.10 N/V and 0.2, respectively. Also, the maximum driving voltage is 1000 V. During the motion of a structure, a friction damper has two possible states: sticking and slipping phases. A combination of stick and slip phases describes the complete state of the friction force. In the slip phase, the sliding velocity is non-zero and the friction force can be computed by Eq. (5). In the stick phase, two friction plates are stuck together, i.e., the sliding velocity is equal to zero. The absolute value of the friction force in the sticking phase, fs , can be approximated for numerical analyses as follows: fs = |fi + fr |

when |f | fs and x = 0,

(7)

where fi is the inertial force applied on the mass, and fr is the restoring force provided by structural stiffness. For a base-isolated building where variable friction dampers are installed between the first floor and ground, the approximate values of fi and fr are given as fi = mt ug (t ) and fr = kb xb (t ), where mt is the total mass of the structure, kb is the isolator stiffness, and xb is the relative displacement of the base isolators [32]. A total of six historical earthquake records that are commonly used in structural control research are employed as external excitation in the numerical simulations. The characteristics of selected ground motions are given in Table 1. Fig. 2 illustrates 5% damped elastic acceleration and velocity response spectra for the selected earthquakes.

3146

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

Table 1 Characteristics of the ground motions used in the analyses. Record (Station & Direction) El Centro Array #9 270 Pacoima Dam 164 Hollister-South & Pine 0 Sylmar-Olive View Med FF 360 H1170546.KOB E-W TCU068 N-S Earthquake 1940 El Centro 1971 San Fernando 1989 Loma Prieta 1994 Northridge 1995 Kobe 1999 Chi-Chi Magnitude (Mw ) 7.2 6.6 6.9 6.7 7.2 7.6 Peak ground acceleration (g) 0.21 1.22 0.37 0.84 0.63 0.36 Peak ground velocity (cm/s) 30.2 56.2 62.4 129.6 76.6 292.2 Source of dataa USGS CDMG CDMG DWP JMA CWB

a USGS = http://nsmp.wr.usgs.gov/ ; CDMG = http://docinet3.consrv.ca.gov/csmip/ ; DWP = http://www.ladwp.com ; JMA = http://www.jma.go.jp ; CWB = http:// www.cwb.gov.tw/.

Fig. 3. Block diagram of the adaptive fuzzy neural controller.

factors of input and output variables, the corresponding universe of discourse of the variable will enlarge or reduce, resulting in better specification of fuzzy parameters. The block diagram of the adaptive fuzzy neural controller is shown in Fig. 3. The fuzzy logic controller developed in this study employs isolation system displacement and velocity as two input variables and provides the command voltage of the damper. Seven triangular membership functions are defined for each input variable, as shown in Fig. 4. The fuzzy sets for the input variables are NL = negative large, NM = negative medium, NS = negative small, ZE = zero, PS = positive small, PM = positive medium, and PL = positive large. Note that universe of discourse for each input variable is defined from 1 to 1. Scaling factors are used to keep the input variables in the range of the universe of discourse. For each input, a reasonable scaling factor should be selected because, if the inputs are scaled to the values such that they become too small, then the innermost membership functions will be used frequently. On the other hand, if they become too big after being scaled, then the outermost membership function will be mostly employed, and this limits the effectiveness of the controller. Since the amplitudes of isolation deformation and velocity differ greatly for near-field and far-field ground motions, the decision on the scaling factors is done by a simple feedforward neural network, introduced later in this section. Five triangular membership functions are defined to cover the universe of discourse of the output variable voltage. The maximum driving voltage for the VFD is 1000 V. However, if the damper is operated at its full capacity during a far-field earthquake, the normal contact force of the damper will be too large and the damper will not dissipate much energy since it will not slide as expected. Therefore, a lower command voltage, i.e., a normal contact force, is desirable for moderate and weak earthquakes for effective energy

dissipation. On the other hand, when the contact force is reduced too much, there will not be enough energy dissipation for strong earthquakes due to the small sliding friction force of the damper. To overcome this difficulty, an output scaling factor that varies in the range of 0 to 1 is used to scale the command voltage of the damper. The output membership functions are equally spaced over the output domain, as shown in Fig. 5. The fuzzy sets for the output variables are VL = very large, L = large, M = medium, S = small, and ZE = zero. After the fuzzification of the input and output variables, a fuzzy rule base is defined for the FLC. The rule bases adopted for the developed fuzzy controller are given in Table 2. The control rules are in the form of ifthen statements, and they map the link between the input and output membership functions. Since the rules are words instead of mathematical equations, it is easy to interpret and modify the rules. For example, a rule in the Table 2 can be read as if the isolation displacement is negative large and isolation velocity is negative large, then the voltage is very large. The rationale used to form to rule bases is as follows: if the displacement and velocity of the isolation system is of opposite sign (i.e., the isolation system returns to its original position), then the output voltage is small, and if the isolation displacement and velocity have the same sign, then the output voltage is large. The magnitude of the output (the degree of being small or large) is linearly proportional with the magnitude of the input variables. When the displacement and velocity are almost zero or small, the command voltage is about zero; that means variable friction damper acts against motion as a passive Coulomb damper. The center of area method is used as a defuzzification method for the FLC to get a crisp output value. Neural networks have been used widely used for adaptive control of uncertain systems [41]. An artificial neural network

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

3147

Fig. 4. Input membership functions for the sublevel FLC.

Fig. 5. Output membership functions for the sublevel FLC. Table 2 Fuzzy rule base for the sublevel FLC. Isolation displacement Voltage NL NM NS ZE PS PM PL NL VL VL L L M S ZE NM VL L L M S ZE S NS L L M S ZE S M ZE L M S ZE S M L PS M S ZE S M L L PM S ZE S M L L VL PL ZE S M L L VL VL

Isolation velocity

A simple neural network that consists of three artificial neurons as shown in Fig. 7 is constructed to specify the input and output scaling factors of the fuzzy controller. As discussed earlier, nearfield earthquakes usually possess long duration pulses with peak velocities. Therefore, the ground velocity xg is selected as the input of each neuron in order to determine the characteristics of the ground motion. The range of the input xg is defined to be [100, 100] cm/s, where the upper limit of |100| cm/s is set as a saturation point and seismic records with ground velocities beyond this value are directly classified as near-field earthquakes. The outputs of the network are scaling factors for the isolation displacement Sd , isolation velocity Sv , and command voltage Sv olt . The activation functions for artificial neurons that produce the scaling factors Sd and Sv are selected to be triangular basis functions, while a bipolar sigmoid function is chosen for Sv olt . These functions are defined as F1 (x) = c1 (c2 c1 ) g x F2 (x) = c3 (c4 c3 ) g x F3 (x) = 1 exp( g ) x 1 + exp( g ) x (8) (9)

(10)

Fig. 6. Schematic representation of an artificial neuron.

consists of a number of simple artificial neurons that are usually organized into three layers, namely, an input layer, a hidden layer, and an output layer, with random connections between the layers. Fig. 6 shows the schematic representation of an artificial neuron. The input signals, represented by x1 , x2 , . . . , xn , are modified by synaptic weights. The output of a neuron is specified by an activation function whose input is the sum of weighted inputs.

where c1 , c2 , c3 , c4 , and are the constants that define the shape of the activation functions. The shapes of these functions are illustrated in Fig. 8. Note that the activation functions F1 and F2 produce a maximum value of c1 and c3 for scaling factors Sd and Sv , respectively, and the magnitude of the scaling factors decreases when the absolute value of the ground velocity increases. On the other hand, the activation function F3 yields an output scaling factor Sv olt that increases with the magnitude of ground velocity and changes in the range [1, 1]. However, the absolute value of Sv olt is used as the scaling factor and, therefore, the output of fuzzy controller, i.e., the command voltage of the damper, is scaled by

3148

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

Fig. 9. Chromosome representation of the neural network.

near-field earthquakes, the 1940 El Centro and 1999 Chi-Chi earthquakes are used to represent far-field and near-field excitations, respectively, during the GA simulations. In order to evaluate each chromosome, four objective functions are defined and computed as follows:
Fig. 7. Neural network architecture.

max xb (t ) J1 =
t (FF )

a number between 0 and 1. The variation of the output activation function F3 with the constant is shown in Fig. 8. Note that this function is highly nonlinear, and the value of significantly affects the output of the function. A genetic algorithm (GA) is a search technique that emulates biological evolutionary theories for optimization and search. The method has been successfully applied to solving a wide range of optimization problems. In recent years, the use of genetic algorithms for training neural networks has been explored in several studies [42,43]. During training of a neural network, a genetic algorithm can be used to determine the network topology and/or weights and/or transfer functions. In this study, a multi-objective genetic algorithm optimizer, namely NSGA-II, is implemented to search the optimum activation function parameters of the neural network that generate the input and output scaling factor of the fuzzy controller. NSGA-II is a computationally fast and elitist evolutionary algorithm based on a non-dominated sorting approach. Among a pool of initial random candidate values that reside within a user-defined range, NSGA-II generates a set of Pareto-optimal solutions through an iterative process. In particular, it compares each solution with every other solution in the population to determine if it is dominated, and then evaluates the solutions in accordance with given performance objectives. A detailed description of the NSGA-II algorithm can be found in [44]. In order to determine optimal activation functions using NSGAII, five parameters, namely, c1 , c2 , c3 , c4 , and , are encoded into the GA chromosome, as shown in Fig. 9. From preliminary simulations, it is found that the parameters c1 (maximum value) and c2 (minimum value) of input scaling factor Sd are within the ranges [7.0, 12.0] and [1.5, 3.0], respectively. Similarly, the ranges for the parameters c3 and c4 of input scaling factor Sv are within the ranges [1.5, 4.0] and [0.5, 1.5], respectively, and the parameter for output scaling factor Sv olt is varied in the range [3.0, 20.0]. For evaluation of candidate solutions during NSGA-II optimization, a seismic excitation is required. Since here the goal is to have a single controller that is effective against both far-field and

max xb (t )
t (FF )

J2 =

t (FF ),f

max xf (t ) ,
(11)

t (FF ),f

max xf (t ) max xf (t ) max xf (t )

max xb (t ) J3 =
t (NF )

max xb (t )
t (NF )

J4 =

t (NF ),f

t (NF ),f

where xb and xb denote the uncontrolled and controlled base displacement, respectively; xf and xf are the uncontrolled and controlled absolute floor acceleration, respectively; t is time; and f represents the story that is considered. The objective functions J1 and J3 evaluate the peak base displacement in the controlled structure normalized by the corresponding displacement in the uncontrolled structure during far-field (FF) and near-field (NF) earthquakes, respectively; the objective functions J2 and J4 compute the peak absolute floor acceleration in the controlled structure normalized by the corresponding acceleration in the uncontrolled structure during far-field and near-field earthquakes, respectively. Here, the term uncontrolled refers to an isolated building without any semi-active dampers. With the abovedescribed settings, a population of 50 chromosomes is initialized for a total of 100 GA runs.
3.2. Simple adaptive control strategy The simple adaptive control (SAC) technique was introduced by Sobel et al. [45] and developed further by Wen and Balas [46] and Barkana and Kaufman [47]. In the simple adaptive control method, the desired behavior is defined by the behavior of the model. This method evaluates the system behavior by comparing the system output and the model output, and its goal is to match the response of the controlled system (plant) with the model that exhibits the desired behavior. The governing equations [40] for the plant (the controlled system) and the model (the system with the desired performance) are

xp (t ) = Ap xp (t ) + Bp up (t ) + di (t )

(12)

Fig. 8. Shapes of activation functions.

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

3149

of the semi-active friction dampers. In LQG control design, the control force is computed by minimizing a given performance index using the isolation velocity and displacement as feedback signals. Here, ground acceleration is considered as stationary white noise, and a quadratic performance index with infinite horizon is selected as J = lim 1

E
0

z(t )T Qz(t ) + u(t )T Ru(t ) dt ,

(20)

Fig. 10. Block diagram of the adaptive control system.

yp (t ) = Cp xp (t ) + d0 (t )

(13) (14) (15)

where the matrices R11 and Q2n2n are referred to as weighting parameters, whose magnitudes are assigned according to the relative importance attached to the state variables and to the control force in the minimization procedure. In this study, the weighting matrix Q is selected as follows:

xm (t ) = Am xm (t ) + Bm um (t )
ym (t ) = Cm xm (t ),

Q=

M 0

0 , K

(21)

where Ap and Am are state matrices, Bp and Bm are the input matrices, and Cp and Cm are the output matrices for the plant and model, respectively. The vectors xp and xm are the n 1 plant state vector and nm 1 model state vector. The plant output vector is yp , ym is the model output vector, up is the m 1 input control vector, and um is the m 1 input command vector. The order of yp is m. The dimension of the model state may be less than the dimension of the plant state, but it should be large enough to create the desired command for the plant. However, the order of model outputs is equal to the order of the plant outputs because yp is supposed to track ym . The variables di (t ) and d0 (t ) represent input and output disturbances. Fig. 10 shows the block diagram of the adaptive control system. The control command can be calculated by up = K(t ) [ym yp xT (t ) uT (t ) ]T = K(t ) rT (t ), m m where K(t ) = [Ke (t ), Kx (t ), Ku (t )] = KI (t ) + KP (t ). (17) (16)

where M and K are the mass and stiffness matrices, respectively. The above form of the Q matrix represents the sum of the kinetic and strain energies, and it is commonly adopted in structural control applications. The selection of the parameter R is based on the goal of balancing the isolation deformation and superstructure acceleration responses, and, through a trial and error process, R = 0.1 is chosen for the simulations. The control force obtained from optimal control theory is termed the active control force, and it cannot always be achieved by semi-active devices. A variable friction damper can only provide a resistant force whose direction is always opposite to that of the slip velocity. Also, the desired control force can be realized when the damper is in the slipping state. Hence, a modified clipped control strategy is used to achieve the desired force as closely as possible [36]. In this approach, the following criteria are used to adjust the desired contact force:

ACTIVE u N (t ) =
0

if uACTIVE ub < 0 if uACTIVE ub 0,

(22)

The overall gain K(t ) is the sum of the integral gain KI (t ) and the proportional gain Kp (t ), which can be calculated from

where uACTIVE is the control force obtained from the LQG control method, and is the Coulomb friction coefficient. 4. Numerical study A five-story base-isolated building, as studied by Johnson et al. [49], was selected to investigate the performance of the developed adaptive controllers. A lumped-mass structure model with one degree of freedom per floor was used in the numerical simulations. The model of the smart base-isolated structure with variable friction dampers is shown in Fig. 11. The fundamental period of the building is 0.3 s, with a damping of 2% in the first mode. The mass of each floor is equal to 5897 kg. The damping coefficients for the five floors are 67, 58, 57, 50, and 38 kN s/m, and the stiffnesses of the five floors are 33,732, 29,093, 28,621, 24,954, and 19,059 kN/m, respectively. The isolation system consists of low-damping rubber bearings with a base mass of 6800 kg, assumed to have linear force deformation behavior with viscous damping. The total damping coefficient and the stiffness of the rubber bearings are 7.45 kN s/m and 232 kN, respectively. For the given isolation parameters, the base-isolated building has a period of 2.5 s and a damping ratio of 4% for the fundamental mode. In order to improve the performance of the base-isolated building against different earthquakes, variable friction dampers with a total force capacity of 16.8 kN are installed to the base of the structure. Time history analyses of the base-isolated building were performed in MATLAB/Simulink [50] for the six historical earthquakes.

KI (t ) = (ym (t ) yp (t )) rT (t ) T KI (t ) Kp (t ) = (ym (t ) yp (t )) r (t ) T.
T

(18) (19)

The rate of adaptation is controlled by the positive definite scale matrices T and T. The proportional gain is used to increase the rate of convergence of the plant performance to the desired performance, and the integral gain is required for the stability and tracking of the system. The coefficient in Eq. (18) can be very small, and it is used to prevent the integral gain from reaching very high values or diverging in the presence of the disturbance [48]. In order to define the model with the desired behavior to control the performance of the structure with VFDs, the behavior of the model should always be better than the behavior of the controlled system and should not depend on the external loads applied to the structure. Therefore, here, the model used for SAC is designed such that it always exhibits better performance than the plant. In particular, it is defined as a structure having the output equal to zero. In this study, the output is the velocity of each floor of the structure, and a velocity equal to zero results in the minimum displacement and drift for the controlled structure. 3.3. Optimal controller For comparison purposes, a linear quadratic Gaussian (LQG) control algorithm is also adopted to tune the contact normal force

3150

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

Fig. 11. Model of the smart base-isolated building as a six degree of freedom system.

For the adaptive fuzzy neural controller, the optimum values of the parameters c1 , c2 , c3 , c4 , and are found to be 9.3, 2.1, 3.3, 1.1, and 3.4, respectively. For the SAC method, the values of , T, and T are chosen to be 0.1, 100 I44 , and 10 I44 , respectively. The results for the applied ground motions are presented in Table 3. The evaluated response quantities are maximum isolator displacement xb,max , maximum interstory drift ds,max , maximum floor ac celeration xs,max , and maximum damper force Fd,max . The results for the base-isolated building without any damper (uncontrolled structure), for the maximum passive operation of the variable friction dampers, and for the modified clipped-optimal controller described above, are also given in Table 3 for comparison purposes. The minimum value of each response quantity for various controller cases is indicated in bold in the table. It can be seen that passive operation of VFDs with maximum voltage significantly reduces the peak isolation deformation for all considered excitations without an increase in interstory drifts, except in the case of the San Fernando earthquake. Specifically, the reduction of the maximum base displacement is of the order of 32% to 76%. However, there is a significant amplification in the maximum floor acceleration for most of the cases. In particular, the maximum floor acceleration increases by 306%, 156%, 131%, and 23% for the San Fernando, Kobe, El Centro, and Loma Prieta earthquakes, respectively. The semi-active control of VFDs notably improves the performance of base-isolated building concerning the peak acceleration response at the cost of slight deterioration in the peak isolation deformation. When the performances of three control strategies are compared, it can be seen that developed adaptive control methods are more effective than a clipped-optimal controller, especially in controlling the acceleration response of the base-isolated building. For example, there are 80%, 95%, and 109% increases in the peak floor accelerations compared to the uncontrolled cases for the El Centro, Kobe, and San Fernando earthquakes, respectively, when the VFDs are modulated by a clipped-optimal controller, while the same increases are only 30%, 47%, and 58% for the AFNC and 30%, 28%, and 78% for the SAC method. Also, when the performances of the adaptive controllers are compared, it is observed that the results for the AFNC and for the SAC method are mostly close to each other; however, the SAC method produces slightly better results

for the structural response of the base-isolated building while the AFNC commands smaller control forces for most of the excitation cases. Figs. 12 and 13 shows the time histories of the isolator displacement and top floor acceleration for the semi-active control of the base-isolated building with the AFNC and with the SAC method subjected to the El Centro and Chi-Chi earthquakes, respectively. The results of the uncontrolled structure are also provided as a measure of performance evaluation of semi-active controllers. Also, a forcedisplacement diagram of the VFDs and the time history of the command voltage for the AFNC and for the SAC method are presented in Figs. 14 and 15 for the same excitation cases. As shown in Fig. 14, in case of a far-field earthquake such as the El Centro excitation, both adaptive controllers operate the VFDs as passive Coulomb dampers with nearly zero voltage. Therefore, the structural responses for both controllers are very close, as illustrated in Fig. 12. When the responses of the base-isolated building subjected to the Chi-Chi earthquake (which has near-field characteristics) are compared for developed adaptive controllers, it can be seen that the peak isolator displacement and top floor acceleration are close to each other. However, the structural response ceases earlier when the dampers are controlled with the SAC method. Also, as is shown in Fig. 15, the controller developed using the SAC method is more aggressive and commands higher voltage values. Overall, both controllers reduce both the displacement and acceleration response of the base-isolated building when it is subjected to the Chi-Chi earthquake. In order to further assess the performance of the developed controller, the energy balance equations of the base-isolated structure are now established. Energy concepts can be used to evaluate the performance of seismically excited structures since the vibrations of structures due to an earthquake can be described as an energy transferring process. For a linear base-isolated structure with an installed variable friction damper, the seismic input energy is the sum of the kinetic energy, the strain energy, the energy dissipated by structural damping, and the energy dissipated by the VFD. The energy balance equations of the base-isolated structure can be expressed as EK + E + ES + EH = EI . (23)

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154 Table 3 Maximum responses of the base-isolated building to various seismic excitations. Earthquake El Centro Response xb,max (cm) ds,max (cm) xs,max (m/s2 ) Fd,max (kN) xb,max (cm) ds,max (cm) xs,max (m/s2 ) Fd,max (kN) xb,max (cm) ds,max (cm) xs,max (m/s2 ) Fd,max (kN) xb,max (cm) ds,max (cm) xs,max (m/s2 ) Fd,max (kN) xb,max (cm) ds,max (cm) xs,max (m/s2 ) Fd,max (kN) xb,max (cm) ds,max (cm) xs,max (m/s2 ) Fd,max (kN) Uncontrolled 20.37 0.12 1.32 20.57 0.12 1.47 35.75 0.20 2.36 82.17 0.46 5.34 27.94 0.16 1.89 60.19 0.34 3.88 Passive-max 4.95 0.12 3.05 16.80 10.28 0.18 5.97 16.80 24.24 0.19 2.91 16.80 49.46 0.33 4.11 16.80 13.68 0.15 4.84 16.80 32.96 0.23 3.14 16.80 Optimal 10.45 0.09 2.37 16.80 11.50 0.10 3.07 16.80 30.14 0.20 2.91 16.80 60.20 0.36 4.76 16.80 18.14 0.18 3.70 16.80 39.27 0.24 3.57 16.80 AFNC 10.40 0.08 1.71 9.28 11.36 0.09 2.32 8.97 29.71 0.20 2.81 12.43 56.37 0.35 4.17 13.60 17.16 0.15 2.79 10.70 35.05 0.24 2.97 16.79

3151

SAC 10.69 0.08 1.71 16.80 10.49 0.12 2.62 16.80 26.59 0.19 2.49 16.80 51.18 0.33 4.24 16.80 15.60 0.12 2.42 16.80 34.72 0.22 2.67 16.80

San Fernando

Loma Prieta

Northridge

Kobe

Chi-Chi

Fig. 12. Time histories of (a) isolator displacement and (b) top floor acceleration of uncontrolled and semi-active systems subjected to the El Centro earthquake.

Fig. 13. Time histories of (a) isolator displacement and (b) top floor acceleration of uncontrolled and semi-active systems subjected to the Chi-Chi earthquake.

3152

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

Fig. 14. Forcedisplacement diagram of VFDs and time history of the command voltage for (a) the AFNC and (b) the SAC method for a base-isolated building subjected to the El Centro earthquake.

Fig. 15. Forcedisplacement diagram of VFDs and time history of the command voltage for (a) the AFNC and (b) the SAC method for a base-isolated building subjected to the Chi-Chi earthquake.

Each term in Eq. (23) can be obtained by integrating the individual force terms in Eq. (1) over the entire relative displacement history. The terms are given as follows: EK = E =
0

1 2

t xT Mxt
t

(24)

ES =
0

xT Cdx =
t

xT Cxdt

(25)

the base-isolated building with the AFNC and with the SAC method for the El Centro earthquake. It can be seen that although the input energy somewhat increases when the VFD is installed in the structure, the VFD successfully dissipates most of the input energy. Also, Fig. 17 illustrates the sum of the kinetic energy and strain energy (EK + ES ), which is known as the total energy of the system or the damage energy. It should also be noted that the maximum damage energy of the base-isolated building with the VFD is 60% smaller than that without any damper. 5. Conclusions

xT Kdx

(26)

EH =
0

uT DT dx
t

(27)

EI =
0

xg MIdx.

(28)

In the above equations, EK is the absolute kinetic energy, E is the damping energy, ES is the elastic strain energy, EH is the hysteretic energy provided by the VFD, and EI represents the absolute input energy. Note that EH is zero for a base-isolated building without any damper. Fig. 16 shows the time histories of the input energy EI , the viscous damped energy of the building and rubber bearings E , and the energy dissipated by the VFD EH , if it is present, for the uncontrolled base-isolated building and the semi-active control of

The goal of this study was to investigate the effectiveness of two adaptive control strategies that were employed to modulate the normal force of variable friction dampers installed in a baseisolated building. The developed adaptive controllers aim to reduce the isolation system deformations without increasing the superstructure accelerations significantly during various types of seismic excitation. The first controller is an adaptive fuzzy neural controller that has an online tunable inputoutput universe of discourse in order to determine the command voltage of the damper according to current level of ground motion. For the design of the second controller, simple adaptive control strategy is employed to obtain the damper force for the optimum performance of the baseisolated structure. Maximum passive operation of the VFDs and

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154

3153

Fig. 16. Energy time histories for uncontrolled and controlled base-isolated buildings subjected to the El Centro earthquake. [7] Bitaraf M, Barroso LR, Hurlebaus S. Adaptive control to mitigate damage impact on structural response. J Intell Mater Syst Struct 2010;21:60719. doi:10.1177/1045389X10361993. [8] Nagarajaiah S, Narasimhan S. Seismic control of smart base isolated building with new semi-active variable damper. Earthq Eng Struct Dyn 2007;36: 729749. [9] Shook DA, Roschke PN, Ozbulut OE. Superelastic semi-active damping of a base-isolated structure. Struct Control Health Monitoring 2008;15:74668. [10] Hurlebaus S, Gaul L. Smart structure dynamics. Mech Syst Signal Process 2006;20:25581. [11] Kelly JM. Role of damping in seismic isolation. Earthq Eng Struct Dyn 1999; 28(1):320. [12] Mazza F, Vulcano A. Nonlinear response of RC framed buildings with isolation and supplemental damping at the base subjected to near-fault earthquakes. J Earthq Eng 2009;13:690715. [13] Alhan C, Gavin H. A parametric study of linear and non-linear passively damped seismic isolation systems for buildings. Eng Struct 2004;26(4): 48597. [14] Gavin H, Alhan C, Oka N. Fault tolerance of semiactive seismic isolation. J Struct Eng 2003;129:92232. [15] Symans MD, Constantinou MC. Semi-active control systems for seismic protection of structures: a state-of-the-art review. Eng Struct 1999;21(6):46987. [16] Alhan C, Gavin HP, Aldemir U. Optimal control basis for performance comparison of passive and semiactive isolation systems. J Eng Mech 2006;132(7): 70513. [17] Sahasrabudhe S, Nagarajaiah S. Experimental study of sliding base-isolated buildings with magnetorheological dampers in near-fault earthquakes. J Struct Eng 2005;131(7):102534. [18] Yoshioka H, Ramallo JC, Spencer BF. Smart base isolation strategies employing magnetorheological dampers. J Eng Mech 2002;128(5):54051. [19] Symans MD, Kelly SW. Fuzzy logic control of bridge structures using intelligent semi-active seismic isolation systems. Earthq Eng Struct Dyn 1999;28:3760. [20] Nagarajaiah S, Sahasrabudhe S. Seismic response control of smart sliding isolated buildings using variable stiffness systems: an experimental and numerical study. Earthq Eng Struct Dyn 2006;35(2):17797. [21] Madhekar SN, Jangid RS. Variable dampers for earthquake protection of benchmark highway bridges. Smart Mater Struct 2009;18:118. [22] Cheryy S, Filiatrault A. Seismic response control of buildings using friction dampers. Earthq Spectr 1993;9(3):44766. [23] Mualla IH, Belev B. Performance of steel frames with a new friction damper device under earthquake excitation. Eng Struct 2002;24(3):36571. [24] Gaul L, Hurlebaus S, Wirnitzer H, Albrecht H. Enhanced damping of lightweight structures by semi-active joints. Acta Mech 2008;195:24961. [25] Gaul L, Nitsche R. The role of friction in mechanical joints. Appl Mech Rev 2001;54(2):93106. [26] Kannan S, Uras HM. Active control of building seismic response by energy dissipation. Earthq Eng Struct Dyn 1995;24(5):74759. [27] Hirai J, Naruse M, Abiru H. Structural control with variable friction damper for seismic response. In: Proceedings of 11th world conf earthquake eng. 1996. p. 88693. [28] Morita K, Fujita T, Ise S, Kawaguchi K, Kamada T, Fujitani H. Development and application of induced strain actuators for building structures. Proc SPIE 2001;4330:42637. [29] Chen C, Chen G. Shake table tests of a quarter-scale three storey building model with piezoelectric friction dampers. Struct Control Health Monitoring 2004;11:23957.

Fig. 17. Time histories of damage energy for uncontrolled and controlled baseisolated buildings subjected to the El Centro earthquake.

an optimal controller developed using LQG theory, along with a modified clipped control strategy, was also considered in order to evaluate the performance of the adaptive controllers. The results from numerical simulations with several ground motions demonstrate that variable friction dampers that operate as semi-active devices by employing adaptive control strategies developed in this study can effectively improve the response of base-isolated buildings against both far-field and near-field ground motions. References
[1] Jangid RS, Kelly JM. Base isolation for near-fault motions. Earthq Eng Struct Dyn 2001;30:691707. [2] Shen J, Tsai MH, Chang KC, Lee GC. Performance of a seismically isolated bridge under near-fault earthquake ground motions. J Struct Eng 2004;130(6):8618. [3] Heaton TH, Hall JF, Wald DJ, Halling MV. Response of high-rise and base isolated buildings to a hypothetical Mw 7 blind thrust earthquake. Science 1995;267:20611. [4] Dicleli M. Supplemental elastic stiffness to reduce isolator displacements for seismic-isolated bridges in near-fault zones. Eng Struct 2007;29:76375. [5] Barbat AH, Rodellar J, Ryan EP, Molinares N. Active control of nonlinear base-isolated buildings. J Eng Mech 1995;121(6):67685. [6] Ozbulut OE, Hurlebaus S. Optimal design of superelastic-friction base isolators for seismic protection of highway bridges against near-field earthquakes. Earthq Eng Struct Dyn 2011;40:27391.

3154

O.E. Ozbulut et al. / Engineering Structures 33 (2011) 31433154 [40] Kim HS, Roschke PN, Lin PY, Loh CH. Neuro-fuzzy model of hybrid semi-active base isolation system with FPS bearings and an MR damper. Eng Struct 2006; 28:94758. [41] Zbikowski R, Hunt KJ. Neural adaptive control technology. World scientific series in robotics and intelligent systems, Vol. 15. Singapore: World Scientific; 1996. [42] Kitano H. Neurogenetic learning: an integrated method of designing and training neural networks using genetic algorithms. Physica D: Nonlinear Phenom 1994;75:22538. [43] Whiteley D, Starkweather T, Bogart C. Genetic algorithms and neural networks: optimizing connections and connectivity. Parallel Comput 1990;14: 34761. [44] Deb K, Pratap A, Agrawal S, Meyarivan T. A fast elitist non-dominated sorting genetic algorithm for multi-objective optimization: NSGA-II. IEEE Trans Evol Comput 2002;6(2):18297. [45] Sobel K, Kaufman H, Mabius L. Implicit adaptive control for a class of MIMO systems. IEEE Trans Aerosp Electron Syst 1982;18(5):57690. [46] Wen J, Balas M. Finite-dimensional direct adaptive control for discrete-time infinite-dimensional Hilbert space. J Math Anal Appl 1989;143:126. [47] Barkana I, Kaufman H. Simple adaptive control of large flexible space structures. IEEE Trans Aerosp Electron Syst 1993;29(4):113749. [48] Barkana I. Gain conditions and convergence of simple adaptive control. Int J Adapt Control Signal Process 2005;19(1):1340. [49] Johnson EA, Ramallo J, Spencer BF, Sain MK. Intelligent base isolation systems. In: Proceedings of 2nd world conference on structural control. New York: Wiley; 1998. [50] The Math Works Inc. MATLAB 7.3, Natick, MA, 2006.

[30] Narasimhan S, Nagarajaiah S. Smart base isolated buildings with variable friction systems: H-infinity controller and SAIVF device. Earthq Eng Struct Dyn 2006;35(8):92142. [31] Inaudi JA. Modulated homogenous friction: a semi-active damping strategy. Earthq Eng Struct Dyn 1997;26:36176. [32] He WL, Agrawal AK, Yang JN. Novel semiactive friction controller for linear structures against earthquakes. J Struct Eng 2003;129(7):94150. [33] Chen GD, Chen CC. Semi-active control of the 20-storey benchmark building with piezoelectric friction dampers. J Eng Mech 2004;130(4): 393400. [34] Xu YL, Chen B. Integrated vibration control and health monitoring of building structures using semi-active friction dampers: part I-methodology. Eng Struct 2008;30:1789801. [35] Lu LY. Predictive control of seismic structures with semi-active friction dampers. Earthq Eng Struct Dyn 2004;33:64768. [36] Ng CL, Xu YL. Semi-active control of a building complex with variable friction dampers. Eng Struct 2007;29:120925. [37] Lu LY, Lin GL. A theoretical study on piezoelectric smart isolation system for seismic protection of equipment in near-fault areas. J Intell Mater Syst Struct 2009;20:21732. [38] Symans MD, Kelly SW. Fuzzy logic control of bridge structures using intelligent semi-active seismic isolation systems. Earthq Eng Struct Dyn 1999;28: 3760. [39] Choi KM, Cho SW, Jung HJ, Lees IW. Semi-active fuzzy control for seismic response reduction using magnetorheological dampers. Earthq Eng Struct Dyn 2004;33(6):72336.