Introduction of Regularization Effect to Structural Damage
Detection Scheme Using Neural Network
Young Gon Ko
Structural damage detection scheme is defined by inverse problem. In inverse problem, regularization techniques must be adopted to avoid ill-posedness which is characterized by sparseness of measurements and noise in measurements. In this paper, it is proposed how regularization effect is introduced to structural damage detection scheme using neural network.
Objective function is defined by least squared error between measured displacements and calculated displacements obtained by numerical model approximated to system parameter. The results, the outputs from optimization of objective function using Newton-Raphson and error back propagation, are compared. In optimization using error back propagation, regularization effect is introduced by changing slope of sigmoid function, which is one of activation functions, according to sensitivity of measurement displacement.
Similarly, regularization effect in structural damage detection scheme using neural network is introduced by changing slope of sigmoid function according to sensitivity of member. The validity of the proposed method is presented by examples.
neural network, error back propagation, damage detection, regularization, sensitivity, sigmoid function, truncated least squared error