A load-adaptive multi-level feature fusion method for predicting metal fatigue crack length based on lamb wave and strain signals

Accurate identification of metal fatigue cracks can promptly warn about the structural service condition. Based on piezoelectric lamb wave and strain signals, this paper proposes a multi-level feature fusion method for heterogeneous signals, combined with the self-attention mechanism, to accurately predict crack length under multiple load conditions. Firstly, the long-sequence piezoelectric lamb wave data is segmented, and the segmented piezoelectric lamb wave data features are extracted through long short term memory (LSTM). The output results of multiple LSTMs are fused through a one-dimensional convolutional neural network (1DCNN). At the same time, the strain signals under various load conditions are pre-processed for features, and then the features of the strain signals are fused through the 1DCNN module. Through the self-attention mechanism for adaptive feature fusion of the two sensor signals, the weights of piezoelectric lamb wave features and strain signal features can be adaptively optimized and adjusted. Combined with the strain feature preprocessing, the model can better adapt to different load conditions. Finally, the experiment on the compression and tensile (CT) specimens is conducted to verify the fusion model and the prediction results of the single sensor. Multiple evaluation metrics are used to compare and analyze the prediction results of the fusion model and the single sensor, verifying the effectiveness of the proposed fusion method for crack length prediction.