To overcome the difficulty of extracting steady-state intervals in complex cutting and the scarcity of plant sensors, we propose a train-time multimodal / run-time unimodal framework for CNC process-segment recognition. Multi-axis vibration and spindle-current signals are synchronously collected on workpieces of different materials to build a benchmark dataset. In the teacher network, a self-distillation spatial attention mechanism compresses intra-modal redundancy, while a vibration-guided cross attention module transfers high-frequency cues from vibration to the current stream, yielding a refined fused representation. Multi-stage knowledge distillation then equips a student network that requires only current RMS sequences at inference. Using current signals alone, the student attains 97 % accuracy while sharply reducing parameters and latency. Results confirm that high-precision, lightweight process-segment recognition is achievable with a single current sensor, paving the way for practical shop-floor monitoring.

CNC process-segment recognition; cross-modal knowledge distillation; attention mechanism; lightweight monitoring