Prior knowledge enhanced graph attention network for real-time anomaly detection in tobacco drying processes

The tobacco drying process in tobacco processing production lines exhibits the characteristics of high-dimensional time series, scarcity of labeled data, and dynamic process coupling, which pose significant challenges for anomaly detection. This article proposes an unsupervised anomaly detection framework based on graph neural networks. The main innovations include: 1) a prior knowledge-driven graph representation method that models the production line as a directed graph, where nodes represent process units and edges encode material flow relationships, enabling spatio-temporal feature learning based on GNN; 2) a hybrid detection architecture that combines the advantages of predictive and reconstruction models, integrating historical patterns with short-term trends to enhance detection sensitivity for sudden faults and gradual deviations; and 3) introduction of a dual graph attention network to adapt to the bidirectional influence between process parameters and quality indicators caused by the PID feedback control mechanism. Experiments conducted on an actual production line of a tobacco manufacturing company demonstrate that this method outperforms existing approaches in terms of accuracy, recall, and F1-score, achieving 0.958, 0.979, and 0.976, respectively. The research results provide a robust solution for real-time anomaly detection in complex industrial production environments.