반능동 MR 유체 감쇠기를 이용한 지진하중을 받는 구조물의 신경망제어 이헌재, 한국과학기술원 건설환경공학과 석사과정 한국전산구조공학회 추계 학술발표회 원광대학교, 익산 2002년 10월 19일 반능동 MR 유체 감쇠기를 이용한 지진하중을 받는 구조물의 신경망제어 이헌재, 한국과학기술원 건설환경공학과 석사과정 정형조, 한국과학기술원 건설환경공학과 연구조교수 오주원, 한남대학교 토목공학과 교수 이인원, 한국과학기술원 건설환경공학과 교수

 OUTLINE OF THE NEURAL NETWORK  ACTIVE CONTROL SYSTEM USING N. N. CONTENTS  INTRODUCTION  OUTLINE OF THE NEURAL NETWORK  ACTIVE CONTROL SYSTEM USING N. N.  SEMIACTIVE CONTROL SYSTEM USING N. N.  NUMERICAL EXAMPLE  CONCLUSIONS Structural Dynamics & Vibration Control Lab., KAIST, Korea

INTRODUCTION Problems Solutions Vibration control of seismically excited structure using artificial neural network was proposed by Ghaboussi et al. (1995) and Chen et al. (1995) Predetermining the Desired Response Need of Emulator Neural Network Problems New Training Algorithm using Cost Function Sensitivity Evaluation Algorithm Kim et al. (2000, 2001) Solutions Structural Dynamics & Vibration Control Lab., KAIST, Korea

Active control systems still have potential to destabilize the structural system. Semiactive control systems not only offer the reliability of passive control systems but also maintain the versatility and adaptability of fully active control systems. 기존의 신경망을 이용한 진동제어들은 능동제어시스템이었다. Structural Dynamics & Vibration Control Lab., KAIST, Korea

Outline of the Neural Network … … … : activation function : activation function Structural Dynamics & Vibration Control Lab., KAIST, Korea

Neural network using cost function (Kim et al., 2000) : state vector : control signal : weighting matrix : sampling number : total number of sampling times Structural Dynamics & Vibration Control Lab., KAIST, Korea

By applying the gradient decent rule to the cost at k-th step, the update for the weight can be expressed as : training rate Using the chain rule, the partial derivative of Eq. (6) can be rewritten as Structural Dynamics & Vibration Control Lab., KAIST, Korea

Let’s define the generalized error as Finally, the weight update can be simply expressed as In Eq. (9), the gain factor, , satisfies The bias is also updated by Structural Dynamics & Vibration Control Lab., KAIST, Korea

In the same manner, updates for the weight and bias between input layer and hidden layer can be obtained as Structural Dynamics & Vibration Control Lab., KAIST, Korea

Active Control System using N. N. Training rule (Sensitivity Evaluation Algorithm) CONTROLLER STRUCTURE (Neural network) Earthquake signal flow for control signal flow for training Structural Dynamics & Vibration Control Lab., KAIST, Korea

Semiactive Control System using N. N. Training rule (Sensitivity Evaluation Algorithm) Secondary (Clipped algorithm) : desired control force : control force of MR damper : velocity across the damper CONTROLLER STRUCTURE Primary Secondary (Neural Network) (Clipped algorithm) Earthquake signal flow for control signal flow for training Structural Dynamics & Vibration Control Lab., KAIST, Korea

NUMERICAL EXAMPLE Three-story building structure Dyke el al. (1996) Structural Dynamics & Vibration Control Lab., KAIST, Korea

Neural Network Control System displacement of 1st floor of 3rd floor velocity of 1st floor of 3rd floor control force ground acceleration Structural Dynamics & Vibration Control Lab., KAIST, Korea

Training and verification of neural network Time(sec) Acceleration (m/sec2) El Centro (0.348g) Time(sec) Acceleration (m/sec2) Northridge (0.334g) Time(sec) Acceleration (m/sec2) California (0.156g) Structural Dynamics & Vibration Control Lab., KAIST, Korea

Training neural network epoch cost function active control semiactive the change of the cost function Structural Dynamics & Vibration Control Lab., KAIST, Korea

Displacement of 3rd floor (El Centro) Time(sec) Displacement (cm) Time(sec) Displacement (cm) Structural Dynamics & Vibration Control Lab., KAIST, Korea

Acceleration of 3rd floor (El Centro) Time(sec) Acceleration (cm/sec2) Time(sec) Acceleration (cm/sec2) Structural Dynamics & Vibration Control Lab., KAIST, Korea

Acceleration (cm/sec2) Displacement of 1st and 2nd floor (El Centro) Time(sec) Displacement (cm) Acceleration of 1st and 2nd floor (El Centro) Time(sec) Acceleration (cm/sec2) Structural Dynamics & Vibration Control Lab., KAIST, Korea

Verification of neural network’s performance Displacement (cm) Time(sec) Acceleration (cm/sec2) Time(sec) 3rd floor, Northridge earthquake Structural Dynamics & Vibration Control Lab., KAIST, Korea

3rd floor, California earthquake Displacement (cm) Time(sec) Acceleration (cm/sec2) Time(sec) 3rd floor, California earthquake Structural Dynamics & Vibration Control Lab., KAIST, Korea

Results of control (El Centro earthquake) Active Semiactive Uncontrolled neuro-control neuro-control 0.549(1.00) 0.836(1.00) 0.973(1.00) 0.098(0.18) 0.133(0.16) 0.215(0.22) 0.138(0.25) 0.198(0.24) 0.216(0.22) (cm) 0.549(1.00) 0.317(1.00) 0.203(1.00) 0.098(0.18) 0.124(0.39) 0.082(0.40) 0.138(0.25) 0.136(0.43) 0.092(0.45) (cm) 880(1.00) 1065(1.00) 1414(1.00) 660(0.75) 569(0.53) 567(0.40) 777(0.88) 887(0.83) 639(0.45) (cm/sec2) (N) - 1000 803 Structural Dynamics & Vibration Control Lab., KAIST, Korea

Results of control (Northridge earthquake) Active Semiactive Uncontrolled neuro-control neuro-control 0.683(1.00) 1.102(1.00) 1.343(1.00) 0.093(0.14) 0.174(0.16) 0.210(0.16) 0.177(0.26) 0.254(0.23) 0.273(0.20) (cm) 0.683(1.00) 0.422(1.00) 0.248(1.00) 0.093(0.14) 0.108(0.26) 0.095(0.38) 0.177(0.26) 0.129(0.39) 0.084(0.34) (cm) 917(1.00) 1287(1.00) 1726(1.00) 661(0.72) 593(0.46) 659(0.38) 693(0.76) 837(0.65) 587(0.34) (cm/sec2) (N) - 1000 790 Structural Dynamics & Vibration Control Lab., KAIST, Korea

Results of control (California earthquake) Active Semiactive Uncontrolled neuro-control neuro-control 0.316(1.00) 0.476(1.00) 0.544(1.00) 0.043(0.14) 0.051(0.11) 0.081(0.20) 0.069(0.22) 0.084(0.18) 0.099(0.18) (cm) 0.316(1.00) 0.169(1.00) 0.104(1.00) 0.043(0.14) 0.046(0.27) 0.041(0.39) 0.069(0.22) 0.060(0.36) 0.042(0.40) (cm) 581(1.00) 662(1.00) 726(1.00) 358(0.62) 270(0.41) 286(0.39) 429(0.74) 515(0.78) 291(0.40) (cm/sec2) (N) - 522 537 Structural Dynamics & Vibration Control Lab., KAIST, Korea

CONCLUSIONS A semiactive neuro-control method using MR fluid damper is presented for seismic response reduction. The training algorithm based on a cost function and the sensitivity evaluation algorithm are used in neuro-control algorithm. The proposed semi-active neuro-control algorithm is quite effective to reduce seismic responses. And, the semi-active control system has many attractive features, such as the bounded-input, bounded-output stability and small energy requirements, ant so on. Structural Dynamics & Vibration Control Lab., KAIST, Korea

Thank you for your attention! The proposed semi-active neuro-control strategy using MR fluid dampers could be effectively used for control of seismically excited structures. Thank you for your attention! Structural Dynamics & Vibration Control Lab., KAIST, Korea