A wavelet-constrained band attention neural network for physics-informed interpretable tool wear prediction

Effective prediction of tool wear is crucial for enabling predictive maintenance in machining processes and averting catastrophic tool failures that lead to workpiece damage and production downtime. While deep learning techniques with temporal fusion are widely adopted for this task, they face three critical challenges: computationally inefficient and inflexible wavelet-based feature extraction, opaque multi-scale temporal fusion mechanisms that lack transparency, and insufficient robustness under noisy industrial conditions due to the absence of physical priors. To overcome these limitations, this paper presents the Wavelet-Constrained Band-Attention Network (WaveBAN). First, parametric wavelet kernels with learnable central frequencies are designed to derive interpretable band-wise features from raw machining monitoring signals. The sparse features across time steps are then processed through a Band-Attention Temporal Fusion (BATF) module based on a probabilistic mixture model, providing a clear mapping from multi-source temporal fusion to the final prediction results. The entire system is optimized using the Expectation-Maximization (EM) algorithm, allowing for optimization of model parameters as well as the dynamic learning of temporal and band-wise attention. Extensive experiments on two milling tool wear datasets, including a public benchmark and a factory-collected dataset, demonstrate that WaveBAN achieves competitive prediction accuracy while also providing transparent insights and enhanced robustness, thereby supporting more informed and effective machining process optimization.