Physics-data fusion multi-modal deep learning for spindle current prediction in digital twin-driven machining state perception

Accurate perception of machining states is essential for intelligent Computer Numerical Control (CNC) systems, yet modeling key process variables remains challenging due to the nonlinear and multi-modal nature of cutting processes. Empirical models lack adaptability, while purely data-driven methods often neglect physical constraints. Direct measurement of cutting force is challenging in practice, as complex workpiece geometries, limited ranges of detection instruments, and high deployment costs make large-scale force data acquisition difficult. Therefore, spindle current is adopted as a more practical and accessible variable for state perception. In consideration of this problem, this study proposes a physics–data fusion multi-modal deep learning framework for spindle current prediction in a Digital Twin (DT)-driven environment. The framework integrates multi-modal inputs, including acceleration signals, STEP-NC semantic metadata, and cutter-workpiece engagement (CWE) geometric descriptors derived from the 3D engagement volume, into a unified representation that is processed by a data-driven branch for feature extraction. The architecture features a Sequence Head for residual regression and an attention-pooled Point Head for next-step prediction. Spindle dynamics constrain sequence outputs through a physics-informed module. Conditioned on process embeddings, a multi-objective loss adaptively balances dual heads and physical heuristics. The framework, deployed in a DT environment, supports synchronized multi-source data acquisition and real-time prediction. Experimental validation on a dedicated DT platform confirms framework feasibility, consistently maintaining R 2 > 0.95 and superior error metrics across diverse machining regimes.