A compressed tensor-based edge-deployable framework for multi-source thermal error compensation in face gear machining

Multi-source errors in precision face gear machining are nonlinear, multi-source, and time-varying, making real-time compensation a challenging task. To address the above challenges, an integrated framework, which combines physics-guided multi-source error modeling, tensor-based model compression, edge deployment, and bandwidth-aware signal scheduling, is proposed to achieve low-latency and high-accuracy error compensation. First, a three-layer system architecture is established, consisting of cloud-side model training, edge-side real-time inference, and sensing-side G-code adjustment. The physics-guided tensor modal decomposition model is developed to identify sensitive error sources by evaluating modal contributions with physical interpretability. To support real-time operation, a model compression strategy based on tucker decomposition and energy-guided truncation is employed, enabling lightweight deployment on edge devices with constrained computing resources. A bandwidth-aware encoding and scheduling mechanism is further introduced, incorporating dynamic sampling period and bit-depth modulation, as well as a priority-driven task queuing strategy formulated by a hybrid urgency-load-error function. Moreover, an improved truncation function method, which dynamically incorporates key nonlinear interactions and employs sensitivity-based dynamic truncation strategies, is proposed. A physics-guided tensor modal decomposition method (PG-TMDM) is proposed to extract directional sensitivity while preserving physical interpretability, outperforming conventional Sobol-based analysis in capturing nonlinear coupling and enabling real-time edge deployment. Then the high-efficiency multi-source error model is proposed based on the novel truncation function. Moreover, the physics-guided multi-source error modeling-based sensitivity analysis method is proposed. Experimental results on a five-axis gear grinding machine demonstrate that the proposed system reduces the maximum tooth surface error by 60.4 %, lowers the average error to 6.1 μm, and maintains end-to-end latency below 150 ms even under burst data traffic. The results also verify the system's robustness in high-load conditions and the effectiveness of the adaptive scheduling mechanism. This work provides a scalable and interpretable solution for high-accuracy, real-time multi-source error control in intelligent manufacturing environments.