A reliable real-time tool wear monitoring framework based on temporal segmentation using domain-informed stacked AI models with physics-constrained predictions
摘要
Cutting-tool wear in CNC machining degrades part quality, increases scrap, and risks unexpected downtime, making real-time tool wear monitoring (TWM) essential. However, existing AI-based TWM approaches suffer from limited generalization and poor robustness under varying conditions. They also rely on static feature-importance scores, unable to capture evolving sensor–wear dynamics. To address these limitations, this paper introduces a hybrid data-driven TWM framework called the Domain-informed Stacked-Temporal LinearSVR model with Monotonicity enforcement and Physics-constrained Kalman smoothing (D-STL-MPK). In this framework, tool-wear progression over machining cycles is partitioned into fixed-length temporal segments. A domain-informed Linear Support Vector Regression (LinearSVR) model is trained on each segment to predict the degree of tool wear. Predictions are concatenated, enforced to be non-decreasing, and smoothed via a physics-constrained Kalman filter for realistic wear-progression estimate. Unlike traditional methods, the proposed framework applies cross-validation within each temporal segment to automatically identify the most informative sensor features for wear prediction in that phase. This enhances interpretability by revealing how the relevance of sensor modalities (such as force, vibration, and acoustic signals) evolves across wear phases. Experimental validation on the publicly available (PHM2010 and HMoTP) datasets shows that D-STL-MPK achieves root-mean-square errors (RMSEs) of 1.41