Multilayer structured wavelet kernel network enabling interpretable feature extraction for industrial intelligent monitoring

In this study, a novel interpretable deep learning framework named Multilayer Structured Wavelet Kernel Network (MSWKN) is proposed for intelligent industrial monitoring. The model introduces trainable wavelet-inspired convolutional kernels that embed the mathematical principles of wavelet decomposition into a hierarchical network architecture. Unlike conventional convolutional neural networks, MSWKN enables end-to-end multi-scale representation learning through fully differentiable wavelet bases, allowing each convolutional layer to adapt its center frequency, bandwidth, and receptive field in response to the intrinsic spectral distribution of the input signal. This design bridges the gap between physical interpretability and data-driven adaptability, offering transparent insight into the underlying time–frequency structure of monitored events. Experimental validation across three industrial scenarios, pipeline leakage monitoring, laser machining monitoring, and planetary gearbox fault diagnosis, demonstrates that MSWKN achieves state-of-the-art accuracy while maintaining clear interpretability. Layer-wise analyses reveal that each wavelet block autonomously focuses on distinct physical attributes, progressively capturing global trends, rhythmic textures, transient modulations, and fine impulsive details in a hierarchical manner consistent with classical wavelet theory. Compared with existing wavelet-based or attention-driven architectures, MSWKN delivers superior performance, stability, and noise robustness without sacrificing physical transparency, providing a promising path toward trustworthy and physics-consistent AI for industrial signal understanding.