Gearbox fault diagnosis method based on dynamic temporal convolution under strong noise and limited sample operating conditions
TL;DR
The effectiveness of the proposed dynamic depthwise separable temporal convolution, hierarchical residual-connected, long short-term memory, hierarchical adaptive attention (DDSC-T-HRC-LSTM-LSTM-HAA), a lightweight single-branch framework integrating DDSC-T, HRC LSTM, and HAA is validated.
Abstract
Intelligent gearbox fault diagnosis is crucial for ensuring the safety and reliability of rotating machinery in industrial systems, yet limited labelled samples, strong noise, and edge deployment constraints remain challenging. To address these issues, this study proposes dynamic depthwise separable temporal convolution, hierarchical residual-connected, long short-term memory, hierarchical adaptive attention (DDSC-T-HRC-LSTM-HAA), a lightweight single-branch framework integrating DDSC-T, HRC LSTM, and HAA. DDSC-T adaptively reweights temporal kernels, HRC-LSTM captures long-range dependencies, and HAA emphasizes informative channels and positions. The effectiveness of the proposed method is validated on two gearbox datasets involving bearing defects and gear tooth-root cracks. The results show that, under the limited-sample setting with only 100 training samples per class on T1, the proposed model achieves mean accuracies of 99.7 ± 0.2% on T1 and 98.3 ± 0.3% on T2. It retains 71.6% accuracy at −5 dB and remains the best-performing method at −10 dB. Edge evaluation on an NVIDIA Jetson Xavier NX demonstrates 0.31 M parameters, 9.8 ± 0.4 ms inference latency, 11.2 ± 0.4 ms end-to-end latency, and 7.2 W average power, indicating practical deployability for online gearbox monitoring.