Interpretable Multi-Resolution Mixing Neural Network for Tool Condition Monitoring

Drawing inspiration from wavelet analysis, we propose a wavelet packet decomposition-based layer designed for multi-resolution analysis and noise reduction of input signals. This approach accounts for both the gradual and abrupt features of tool wear while addressing the need for interpretability and noise robustness in tool wear monitoring tasks. The wavelet decomposition layer maintains the tree structure inherent in wavelet packet decomposition. It achieves this by stacking a deep 1D CNN with a shared filter and a stride of 2. Unlike traditional wavelet analysis, this approach moves away from reliance on fixed wavelets, and it learns end-to-end multi-channel filters instead. Additionally, the extra loss constructed by biorthogonal wavelets is used to constrain the filter training process, ensuring interpretability while enhancing the flexibility of filter optimization. This approach facilitates the extraction of multi-view and multi-resolution features from input signals. To further diminish noise based on signal characteristics and task-specific requirements, a gated linear soft-shrink function is employed to perform adaptive soft thresholding on the high-frequency detail coefficient obtained through wavelet decomposition according to the gating coefficient. Furthermore, the mix-res layer is introduced to enhance the interactivity between both intra-layer and inter-layer features. This is achieved by mixing threshold-processed wavelet coefficient matrices using dual multi-head self-attention, which leverages the contextual relationships inherent in these features. The proposed method demonstrates superior interpretability and noise robustness compared to existing approaches, such as LSTM, ResNet18, and Transformer, while also achieving advanced performance in tool wear monitoring tasks.