類神經網路在結構主動控制上之應用

Abstract

Currently, most control algorithms useds for the active control of civil-engineering structures are developed based on optimal control. The standardoptimal-control technique requires the entire time history of structural response to compute the necessary control force, so it has to minimize thethe objective function in the global sense, while the instantaneous optimal- control technique minimizes the objective function within each small timeinterval to compute the necessary control force. Because of the difficultyto obtain the system variables of a real structure, it's hard for the activecontrol algorithms of structures according to energy in design. In this paper,a new active control algorithm, Adaptive Neural Controller (ANC) is developedby artifitial neural network. The ANC consists of two components : (1) aNeural Emunator Network to represent the structure to be controlled, and (2)a Neural Action Network to determine the control action on the structure.Because of the adaptability and error tolerance of the artifitial neural network, the above difficulty can be solved by the ANC. Besides, the error backpropagation learning algorithm (BP) and the conjugate gradient method are used in this paper to update the weights of the ANC. BP learning algorithm based on the gradient descent method has a non-system convergent rate for a constant learning ratio. The conjugate gradient method improvesthe calculation of gradient direction of BP learning algorithm. Instead ofconstant learning ratio, the step length in the line search of the conjugate gradient method is adapted during the learning process. By the analysis ofcases, the ANC has a good effect on the control of structures both by the BP algorithm and by the conjugate gradient method, and the conjugate gradientmethod has a better convergent rate than the BP algorithm.