Deep Learning-Driven Adaptive Machining Parameter Optimization for High-Precision CNC Milling

To address the limitations of single-modal data and low real-time performance in traditional anomaly diagnosis for CNC turning processes, this paper proposes a novel framework integrating adaptive multi-modal data fusion and lightweight graph neural network (GNN) for real-time anomaly diagnosis. First, multi-modal data (vibration, spindle current, and cutting force) are collected and preprocessed to extract time-frequency domain features. A mutual information-based graph construction method is designed to model the intrinsic correlations between multi-modal features, converting non-Euclidean feature data into structured graph data. Then, an event-driven lightweight GNN (EL-GNN) is proposed, which adopts a hierarchical propagation mechanism to reduce redundant computations and realizes millisecond-level inference. A cross-attention fusion module is embedded in the GNN to dynamically assign weights to different modal features, enhancing the robustness to noise. Experiments are conducted on a self-built CNC turning test platform and the public tool wear dataset. Results show that the proposed framework achieves an anomaly diagnosis accuracy of 98.73%, a recall rate of 98.51%, and a P99 inference latency of 28.3 ms , outperforming traditional machine learning methods and deep learning models by 3.2% − 8.9% in accuracy. This framework provides a reliable solution for intelligent predictive maintenance in CNC turning processes, balancing diagnostic accuracy and real-time performance.