Ensemble time-frequency state-space model with representation disentanglement for tool wear recognition under heterogeneous sensors and variable machining conditions

Real-time tool wear recognition is pivotal for ensuring machining quality and operational safety. However, modern manufacturing systems are characterized by the dual complexities of multi-category machine tool collaboration and multiple machining conditions. Consequently, data-driven tool wear recognition models face significant adaptation challenges under coupled data distribution shifts induced by sensor configurations and machining conditions. To address this issue, this study proposes an ensemble time-frequency state-space model with representation disentanglement. For multi-sensor wear monitoring signals, independent time-frequency state-space modules are initially established to extract wear features from each signal channel individually. These modules adaptively focus on frequency bands associated with wear dynamics and model the time-frequency evolution trends across them. All independent modules are aligned to homogenize the cross-channel information, thereby complementarily enhancing the robust characterization of wear behavior within individual signal channels. Subsequently, a feature ensemble mechanism based on cycle-consistent negentropy actively assigns distinct fusion weights by quantifying the wear information content in each channel signal, thereby ensuring effective information fusion in heterogeneous sensor scenarios. Finally, an adversarial disentanglement strategy is designed to extract domain-invariant representations across varying machining conditions from the ensemble features, which are then mapped to wear values. The proposed model achieves precise extraction and generalization of wear representations under coupled data distribution shifts, with its performance validated through experiments on multiple cross-condition and cross-machine-type datasets.