An interpretable operating condition partitioning approach based on global spatial structure compensation-local temporal information aggregation self-organizing map for complex industrial processes

Abstract

Effective and interpretable partitioning of operating conditions in complex industrial processes is crucial for ensuring the efficiency and reliability of emerging integrated production processes. However, contemporary operation–condition partitioning models fall short in this domain for two main reasons. First, the parameters related to purely data-driven methods fail to involve physically meaningful semantics, resulting in suboptimal accuracy and interpretability. Second, there exists a spatial–temporal coupling interaction among process variables that reflects physically meaningful spatial–temporal correlations, further impeding effective interpretation. To address these challenges, an interpretable operation–condition partitioning method based on a Global Spatial Structure Compensation-Local Temporal Information Aggregation Self-Organizing Map is proposed. This approach formulates the operation–condition partitioning problem as a spatio-temporal feature extraction based on the operation–condition partitioning problem. First, a local separable aggregation (LSA) model based on the informer backbone is introduced to partition industrial variables into interpretable time-local groups. This ensures that the model can learn the underlying factors of the condition’s occurrence, thereby providing a reliable basis for interpretable root cause analysis. Next, a Global Spatial Feature Compensator (GSFC) is introduced that leverages the process mechanism knowledge and graph network to establish explicit spatial interrelationships among time-local groups in the LSA algorithm, thereby enhancing the model's inferential capabilities and security assurance. Subsequently, a spatiotemporal self-organizing map is designed to integrate the aforementioned spatial and temporal features, facilitating precise operation–condition partitioning decisions. Finally, the effectiveness and understandability of the proposed method are validated through comprehensive experiments in a real industrial application case study.