An integrated real-time tool wear monitoring and prediction framework with autonomous in-situ machine vision-based labeling

Tool wear in milling operations significantly impacts machining efficiency, surface quality and downtime, with flank wear (VBmax) serving as a critical failure indicator. Traditional tool condition monitoring (TCM) methods suffer from offline measurements, labor-intensive labeling and disjointed monitoring-prediction workflows. This study proposes an integrated real-time TCM framework with autonomous in-situ machine vision-based labeling to address these limitations. An automated data acquisition platform collects multi-sensor signals and high-resolution tool images during machining without interruptions, using CNC integration and FOCAS protocol for precise control. A lightweight U-Net model, enhanced with depthwise separable convolutions (DSConv) and Ghost modules, performs semantic segmentation of worn flank regions. Subsequent geometric contour analysis automatically measures precise VBmax values, thereby providing zero-human-intervention ground-truth labels for data-driven monitoring and prediction models. Multi-domain features from preprocessed signals are refined using Pearson correlation for input to a Kolmogorov-Arnold Network (KAN) for instantaneous wear regression. Accumulated KAN outputs feed a Reformer model for sequential wear prediction. Experimental results demonstrate automatic labeling with error below 5 μm, monitoring performance with RMSE and MAE consistently below 1.5 μm and 1 μm across public and experimental datasets, and superior prediction performance with RMSE and MAE below 0.6 μm and 0.5 μm compared to competitive benchmarks.