SDC-LSTM: a lightweight and interpretable planetary gearbox fault diagnosis method

Planetary gearboxes are essential for ensuring the efficient and reliable operation of industrial equipment. However, fault diagnosis under complex operating conditions often encounters challenges such as high computational cost and limited noise robustness. To address these issues, a lightweight fault diagnosis model, designated SDC-LSTM, is proposed. The model incorporates a Separable Dilated Convolution (SDC) module designed to extract multi-scale local receptive field features from vibration signals, thereby substantially reducing parameter count and computational overhead. Furthermore, LSTM network is employed to capture temporal dependencies through its gated mechanism. Experimental evaluation on a wind turbine planetary gearbox dataset demonstrates that SDC-LSTM achieves a parameter count of only 0.18M and 46.93M FLOPs. Under noise-free conditions, it attains an average accuracy of 99.54% and a maximum accuracy of 99.72%, surpassing conventional models such as ResNet18, CNN-GRU, and MobileNet. Even under severe noise interference (σ = 0.6), SDC-LSTM maintains an accuracy of 80.95%, notably higher than the 61.74% achieved by ResNet18, indicating superior robustness.