A multi-source-calibrated DT-KG method for multidimensional tool-wear evolution analysis and intelligent prediction in machine tools

To address the insufficient adaptability and data-silo problems in the integration of digital twins with machine-tool condition monitoring, this study proposes an online-signal-driven and offline-calibrated method for the multidimensional evolution and intelligent prediction of machine-tool knowledge graphs. Using an LXK screw milling machine as the research object, the proposed method integrates machining parameters, sensor signals, and offline inspection data to support dynamic knowledge updating and state prediction. First, multi-source heterogeneous data are processed and represented in a unified form. Based on this, a macro-meso-micro coupled model is constructed to describe the heat-force-damage transfer relationships among cutting parameters, tool wear, observable tool microcracks, workpiece surface roughness and residual stress. Considering that meso- and micro-scale variables are difficult to measure fully online, a large amount of offline measured data is first used for model calibration and knowledge graph construction. After offline data-driven calibration, online machining parameters and sensor signals are used as inputs for dynamic state prediction and knowledge graph updating. Second, a hybrid knowledge graph construction strategy is proposed to improve the structured representation and reuse of machine-tool knowledge. Furthermore, a memory replay mechanism is introduced to suppress catastrophic forgetting during dynamic updating, and a 1:1 ratio of new to historical data is determined as the optimal mixing strategy. The results show that this strategy keeps the model forgetting rate below 5% and maintains old-knowledge retention above 95%, while preserving new-knowledge adaptability and knowledge coverage. Experimental results further verify the coupled trends among tool wear, microcrack evolution, surface roughness, and residual stress. This study provides a conservative and interpretable DT-KG method for machine-tool wear evolution analysis, and the constructed hybrid knowledge graph can support reverse fault tracing, graph evolution, and intelligent reasoning.