Abstract:
To address the susceptibility of planetary gearbox fault feature extraction to noise interference and the relatively high complexity of diagnostic models for wind power applications, a lightweight diagnostic model integrating wide-kernel multiscale convolution and one-dimensional anti-aliased downsampling was constructed. The model combined parallel depthwise convolutions with short and wide kernels and pointwise convolution to extract and fuse multiscale features. Fixed low-pass filtering was applied before downsampling along the temporal dimension, while compression of the hidden channels and learnable weighted residual fusion were employed for feature reorganization. Comparative and ablation experiments were conducted using a planetary gearbox dataset. The results showed that, under the primary operating condition, the proposed model achieved an average diagnostic accuracy of 93.61% across six signal-to-noise ratio levels from −6 dB to 4 dB, exceeding that of the multiscale kernel-based residual convolutional neural network by 5.76 percentage points. The model had 0.149 M parameters and a computational cost of 11.618 MFLOPs. In the ablation experiments under low signal-to-noise ratio conditions, removing either the wide-kernel multiscale convolution module or the anti-aliased downsampling module led to maximum accuracy decreases of 8.90 and 9.62 percentage points, respectively, indicating that both modules contributed to improved diagnostic accuracy under low signal-to-noise ratio conditions.