In this study, an integrated framework of "perception-diagnosis-control" is proposed to solve the problems of difficult integration of multi-source heterogeneous data of mechanical equipment, slow adaptation to working condition drift and split diagnosis-control under the background of Industry 4.0. Core innovations include: 1) Building a dynamic feature fusion module based on graph attention network (GAT), modeling sensors as graph nodes, and adaptively calculating weights through attention mechanism to realize spatial association mining and key feature extraction of multimodal data, 2) Design an online incremental diagnosis framework based on elastic weight consolidation (EWC), use Fisher information matrix to quantify the importance of parameters, impose regular constraints on key parameters of old tasks, and control the performance loss caused by catastrophic forgetting within 2%, 3) Develop a multi-objective optimization control strategy integrating deep reinforcement learning (DRL) with digital twins. Employing the Proximal Policy Optimization (PPO) algorithm, construct a synergistic reward function balancing efficiency, energy consumption, and lifespan. Virtual debugging reduces physical testing risks by 70%. The validation on the IMS dataset of CNC machine tools shows that the proposed Dynamic Graph Attention Network (DGAT) fusion method has significantly better accuracy (95.3%) and F1 score (0.937) than traditional stitching methods, Compared with the fixed parameter method, DRL control strategy improves efficiency by 25%, reduces energy consumption by 8.6%, and reduces wear rate by 22.9%. This research realizes the leap from passive response to active optimization of equipment operation and maintenance, and provides an end-to-end intelligent decision-making scheme for intelligent manufacturing.
Lin et al. (Sun,) studied this question.