A small sample piezoelectric impedance-based structural damage identification using Signal Reshaping-based Enhance Attention Transformer

Abstract

The electromechanical impedance-based (EMI) method is a non-destructive damage identification approach that can effectively reveal the health condition of a structure. Conventional deep learning-based EMI methods require massive amounts of data. However, the number of structural EMI samples that can be obtained in applications is limited. In this research, a new Signal Reshaping-based Enhanced Attention Transformer network is proposed for small sample structural health monitoring. The proposed network constructs a Signal Reshaping-based Enhance Attention module in the Transformer-based network, which can effectively enhance the overall feature extraction by reshaping each of the input signals. Optimal parameters of the Signal Reshaping-based Enhance Attention module were analyzed. The experiment shows that the proposed network can detect the state and location of bolts and mass variation in a cantilever beam with an overall accuracy of 98.61 % with small samples, which is 23 % better than a conventional Transformer. In addition, the performance of the proposed network was compared to one-dimensional convolutional neural network (1DCNN), very deep convolutional network (VGG), recurrent neural network (RNN), and long short-term memory (LSTM). Besides, the advantage of the proposed system is evaluated in a case study of glass plate damage identification.