Spatio-Temporal Deep Feature Fusion Network-Based Distributed Quality-Related Anomaly Monitoring for Large-Scale Manufacturing Processes

Large-scale manufacturing processes are characterized by complex process, dynamic coupling between variables, and quality heritability between sub-processes, which poses great challenges to traditional process monitoring. In this article, a novel distributed quality anomaly monitoring framework based on a designed spatio-temporal deep feature fusion network (STDFFN) is proposed. First, sub-block division is performed based on the process knowledge, and minimal redundancy maximal relevance is introduced to select quality-related coupling variables between sub-blocks to achieve reconstruction and precise partitioning. Second, a new STDFFN is designed, combining the advantages of convolutional neural network and long short-term memory network in extracting spatial and temporal features, respectively. Canonical correlation analysis (CCA) is employed to quantify the correlation between the extracted features and quality variables, constraining the extraction of deep spatio-temporal quality-related features. A two-stage training strategy is proposed to ensure the completeness of quality-related information. The convolutional block attention module is introduced to dynamically focus on important quality-related information while accelerating network convergence. Then, based on the deep spatio-temporal features, we design the local monitoring statistic. Bayesian inference is used to obtain the global quality anomaly monitoring results, and combined with variable forgetting contributions to achieve visualization of fault propagation in large-scale sequential manufacturing processes. Finally, the effectiveness of the proposed method is validated by the actual hot strip mill process (HSMP) data.