To address multiple faults and fault severity detection under complex operating conditions such as strong noise, strong time variation and cross-working conditions, traditional CNNs, and wavelet-CNN hybrids often suffer from low accuracy and poor generalization, as they either adopt offline wavelet denoising or use fixed wavelet kernels, leading to lost adaptability. WGS-CNN is proposed with three synergistic mechanism-level innovations, forming a closed-loop design rather than a simple component combination. Multi-scale wavelet initialization with retained backpropagation anchors time-frequency prior knowledge, while enabling adaptive learning, laying a targeted foundation for feature extraction. Building on this prior injection, adaptive Gaussian windows with learnable scale factors dynamically constrain convolutional kernels to align with varying fault features, breaking static filtering limitations. Finally, square function activation fused with Gaussian denoising inherits the constrained features to integrate power spectrum enhancement, strengthening weak fault signals. Experiments show WGS-CNN achieves over 87% F 1 across complex scenarios, outperforming traditional CNNs and existing wavelet-CNN hybrids in accuracy and lightweight performance, providing an effective end-to-end solution with fundamental innovations in wavelet-deep learning fusion.
Pang et al. (Sun,) studied this question.