Steady-state segmentation in complex cutting is hindered by signal ambiguity and sparse shop-floor instrumentation. We introduce a train-time multimodal / run-time unimodal framework that couples multi-axis vibration with spindle-current data during training yet relies solely on current RMS sequences for inference. A benchmark dataset spanning multiple workpiece materials is first acquired. In the teacher network, self-distillation spatial attention prunes intra-modal redundancy, while vibration-guided cross-attention injects high-frequency cues into the current stream, yielding a compact fused representation. Multi-stage knowledge distillation transfers this representation to a lightweight student model. With only current signals, the student attains 97 % accuracy, outperforming the PCA-SVM and CNN-LSTM baselines by approximately 10 % and 3.5 %, respectively, while markedly reducing parameters and latency. The results confirm that high-precision, real-time CNC process-segment recognition is feasible with a single current sensor, enabling practical shop-floor monitoring.

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