Bidirectional heterogeneous synergistic fault detection using multiple local data in large-scale industrial systems

Abstract

Technological advances have increased the complexity of industrial processes, such as semiconductor and manufacturing systems, leading to large-scale system integration. Consequently, the operational states of such systems rely heavily on complex and high-dimensional data for an effective representation. Existing strategies, such as local and local–global methods, focus on capturing local features and their interactions with global characteristics but often overlook the heterogeneity among local input units and the interconnections between subsystems within the same larger-scale system, resulting in flawed assumptions and information loss during modeling. To tackle these challenges, this paper proposes a bidirectional heterogeneous synergistic model (BHS) based on multiple local groups. Specifically, a heterogeneity-constrained agglomerative hierarchical clustering method is developed to capture and optimize the heterogeneity between local groups. Next, multiple feature extractors are constructed to capture fine-grained local features, enhancing the capability of large-scale systems to represent critical information. Subsequently, a bidirectional attention mechanism based on mutual information is proposed to synergistically uncover subsystem correlations within the same system, compensating for the loss of multiscale synergy during local modeling. Finally, feature fusion is employed to integrate information across subsystems, enabling unsupervised modeling for large-scale industrial systems. Experimental results from a simulation process, a benchmark process, and a practical semiconductor measurement task demonstrate the superiority of the proposed approach in fault detection tasks for large-scale industrial systems.