Graph contrastive learning of modeling global-local interactions under hierarchical strategy: Application in anomaly detection

Abstract

Lack of labeled samples and complexity of unit interactions pose significant challenges for effective anomaly detection in complex industrial processes. This work proposes an unsupervised anomaly detection framework, hierarchical strategy-based global-local graph contrastive learning (HS-GLGCL). First, a process topology graph is constructed based on prior knowledge and then enhanced by the technique of graph augmentation. The graph contrastive learning mechanism enables the model to learn intrinsic information from unlabeled samples. Local topological information is input into the model through structural coefficients to further accurately simulate information transmission between local variables amid complex unit interactions. When global topological information is found inadequately captured, isomorphic similarity is introduced to help the model obtain embeddings that can more accurately describe data distributions. Anomalies are detected and localized by setting a reconstruction error threshold. Additionally, reconstruction dissimilarity represented by Kullback-Leibler divergence is adopted to further confirm the model’s performance superiority and complete a thorough evaluation of the model’s reconstruction performance. The effectiveness of HS-GLGCL is validated in two case studies on data respectively from a sugar factory and a sour water treatment system.