Cross-Alloy Transferability of Spindle-Synchronous Vibration Signatures in CNC Milling Using Physics-Informed Autospectrum Features

Data-driven recognition of machining conditions in CNC milling processes represents a significant research focus within process monitoring applications. This work assesses whether physics-informed, spindle-synchronous vibration representations remain transferable between Al 6061 and Al 7075 using multi-axis and multi-location autospectrum data. Spindle speed is used as a controlled and physically traceable process-state label rather than as an unknown CNC variable to be reconstructed. Vibration signals were collected under various combinations of feed rate, spindle speed, and depth of cut. Subsequently, autospectrum descriptors associated with spindle rotation, harmonic components, frequency-band energy, and sensor location were extracted. Preliminary cross-alloy analysis demonstrated that spindle speed recognition exhibited more stable transfer performance than feed rate or depth of cut recognition. Multiple machine learning models were evaluated, with the combination of spindle-related autospectrum descriptors + LightGBM model selected as the final configuration. The resulting model achieved 0.8708 accuracy, 0.8701 Macro-F1, 0.7500 MCC, and 0.9615 ROC-AUC in the cross-alloy spindle speed recognition task. In two-way transfer analyses, Macro-F1 scores of 0.8748 and 0.8661 were obtained for the Al 6061 to Al 7075 and Al 7075 to Al 6061 scenarios, respectively. Ablation studies indicated that spindle-related frequency descriptors offer advantages over comprehensive feature representations. SHAP-based explainability analysis revealed that harmonic band, tooth-passing band, and sensor-location-dependent spectral features were most influential in model decisions. These findings indicate that physics-informed autospectrum descriptors preserve discriminative spindle-synchronous vibration information across the investigated Al 6061 and Al 7075 alloy domains.