A Method for Predicting Tool Wear Under Different Milling Conditions by Integrating Adaptive VMD with a Two-Stage Attention Mechanism

Reliable tool condition monitoring under high-noise and different milling conditions requires robust signal preprocessing and adaptive feature extraction. This study proposes a tool wear prediction framework that integrates Adaptive Variational Mode Decomposition (AVMD) with a two-stage Convolutional Block Attention Module (CBAM). AVMD uses signal spectral features to guide a directed search for decomposition parameters, reducing reliance on manually specified settings. The reference architecture processes reconstructed multi-sensor signals through multiscale convolution and a DenseNet–BiLSTM network, with two CBAM stages for feature recalibration. The first CBAM follows multiscale convolution and precedes DenseNet, while the second follows BiLSTM. Using the selected configurations, within-tool evaluations achieve R2 values of 0.9399–0.9573 on TC18 titanium-alloy milling data and 0.9907–0.9929 on the public PHM 2010 dataset. Exploratory cross-tool evaluations on PHM 2010 yield R2 values of 0.8954–0.9599 under fixed cutting conditions. t-distributed stochastic neighbor embedding and Shapley additive explanations are used to visualize feature representations and assess feature contributions to model predictions. The results demonstrate the potential of the framework for multi-sensor tool wear estimation in the evaluated machining scenarios.