Adaptive Wavelet Threshold Attention and Explainable Lightweight CNN for Strongly Time-Varying Fault Diagnosis and Feature Extraction Mechanism of Rotating Machinery

Abstract To address the problems of fault feature drift, weak information submerged by noise in rotating machinery under strongly time-varying conditions, as well as frequency domain rigidity, low-frequency distortion, and high deployment cost of traditional CNNs, an explainable lightweight CNN based on adaptive wavelet threshold attention is proposed. The model captures full-frequency features through an adaptive wavelet convolution layer, strengthens fault features and suppresses noise via an adaptive threshold temporal attention enhancement module, and achieves lightweight design with a single-layer architecture. On bearing variable-speed and gearbox variable-speed-variable-load datasets, the average diagnostic accuracy reaches 93.75%, and the F1-score exceeds 79% under low signal-to-noise ratio conditions. Its performance stems from the synergistic closed-loop of wavelet convolution, adaptive threshold temporal attention enhancement, and gradient feedback. Adaptable to industrial edge devices, the model provides a reliable technical solution for strongly time-varying fault diagnosis.