Simultaneous identification of structural damage and nonlinear hysteresis parameters by an evolutionary algorithm-based artificial neural network

An evolutionary algorithm-based artificial neural network (ANN) to identify structural damage and nonlinear hysteresis parameters simultaneously is presented. To avoid the 'dimensional disaster', the principal component analysis technique is applied to eliminate the redundant dimensionality of the acceleration data. The acquired principal components, which covers over 95% of the variability in data, are then employed as the input to the ANN, while the system parameters to be identified are defined as the output. ANN is an effective tool to tackle complex problems in numerous fields. However, if using the gradient-descend algorithm to train the ANN model, with vanishing or exploding gradients, ANN may suffer the local minimal during the training process. To address this drawback, a new evolutionary algorithm, termed the K-means Jaya, is employed to train the ANN model to obtain optimal weights and biases by minimizing the discrepancies between real outputs and desired ones. The optimal weights and biases are then used to configure the ANN. The proposed method is applied to single and multiple degree-of-freedom nonlinear systems subjected to different external loadings. Results demonstrate the structural damage and nonlinear hysteresis parameters can be accurately identified.

Automated summary

An evolutionary algorithm-based artificial neural network is used to identify structural damage and nonlinear hysteresis parameters, with principal component analysis reducing redundant dimensionality of data. The approach successfully addresses the challenge of vanishing or exploding gradients in training the ANN model, achieving accurate identification of system parameters.