Lightweight Cross-Condition Tool Wear Prediction via Transfer Learning for Resource-Constrained Edge Manufacturing

In resource-constrained manufacturing environments, accurate prediction of tool wear under variable cutting conditions remains challenging due to the trade-off between computational overhead and prognostic reliability. Conventional signal processing and deep learning methods suffer from high model complexity, cross-domain distribution discrepancies, and excessive computation cost, limiting their practical deployment on edge devices. This paper proposes an optimized cloud–edge collaborative transfer learning framework for lightweight remaining useful life (RUL) prediction of milling tools. On the cloud side, a low-complexity stacked auto-encoder built on a 1D-SqueezeNet is trained with cross-domain adaptation constraints to improve generalization across working conditions. Pretrained model parameters are then transferred to edge hardware for deep feature extraction. A kernel extreme learning machine (KELM) is further adopted for fast on-site tool wear regression prediction. When validated on full-lifecycle milling vibration data collected under real machining conditions, the proposed approach achieves nearly identical prediction performance on edge devices compared with cloud-side inference while also reducing edge inference latency to 0.01–0.021 s. These results demonstrate that the presented framework can deliver accurate prediction of milling tool RUL and provide rapid response under limited edge computational resources.