Innovative data-driven algorithm for defect parameter identification in large-scale structures

This study proposes an innovative data-driven algorithm that combines the scaled boundary finite element method with an autoencoder and a causal dilated convolutional neural network for defect identification in large-scale structures. The scaled boundary finite element method simulates the propagation of waves in large-scale structures containing various types of defects. Conveniently, the scaled boundary finite element method can simulate different types of defects within structures and, by discretizing only the boundaries of structures, efficiently generate sufficient training data. To simulate wave propagation in large-scale structures, an absorbing boundary model based on Rayleigh damping is established, avoiding computations across the entire structural domain. The affinity propagation clustering algorithm is employed to optimize the number and layout of sensors, and the optimized multi-sensor data serve as the original training samples for autoencoder feature extraction. Autoencoder exhibits strong nonlinear feature extraction capabilities, mapping the high-dimensional original input feature vector space to a low-dimensional latent feature vector space to obtain low-dimensional latent features for network model training. This effectively improves the learning efficiency of the network model. The constructed causal dilated convolutional neural network model ensures orderliness of temporal data and achieves a larger receptive field without increasing neural network complexity, thereby capturing more historical information. Numerical examples demonstrate that the proposed model can accurately identify the quantified information of defects in large-scale structures. Compared with the previous model, the proposed model exhibits improved robustness.