Research on multi-feature fusion pulsed eddy current testing method of pipe wall thickness based on CNN-LSTM-Attention algorithm

Pulse eddy current (PEC) testing is a key non-destructive methods for pipeline defect detection. Unknown lift-off from uneven pipeline coating degrades defect assessment and thickness inversion accuracy in testing.Conventional PEC uses linear mapping, requiringprecise lift-off for accurate thickness prediction. To address the interference caused by unknown lift-off in thickness prediction, a method is proposed that fuses multi-segment temporal features from the eddy current signals. Specifically, the CNN-LSTM-Attention decouples nonlinear relationships in eddy current voltage signals across time, thickness, and lift-off, eliminating lift-off interference for accurate defect evaluation and thickness assessment. In order to enhance the selected voltage features, the Squeeze-and-Excitation (SE) channel attention mechanism is employed, combining CNN and LSTM to strengthen the relationship between temporal voltage features and thickness. The tube wall thickness can be directly inferred from the eddy current voltage excluding unknown lift-off influence. Results show that the CNN-LSTM-Attention outperforms traditional linear mapping method as well as CNN-LSTM, PCA-SVR, CNN-GRU-Attention, and PCA-CNN-LSTM-Attention in evaluation capability, with the thickness assessment error being controlled within 0.2 mm, effectively eliminating the influence of unknown lift-off values in practical tests.