Branch Voting and Hierarchical Activation for Anomaly Inference in Refinery Units

Anomaly inference in refinery units is challenged by hierarchical propagation, branch interactions, and limited prior knowledge of node activation. This study proposes a branch voting and hierarchical activation method for refinery anomaly inference. Anomaly phenomenon nodes, anomaly cause nodes, and their connections are constructed from process mechanisms, field investigations, technical documents, and expert knowledge. A full test set is expanded from highly credible cases under process constraints, and a graded loss rule library is established according to the minimum evidence required for cause activation. During online inference, upstream nodes vote within each branch, and anomaly cause nodes are activated based on accumulated support and minimum activation values before propagating to deeper layers. A PSO Greedy algorithm is further developed to optimize connection weights and minimum activation values through global search and local refinement. Case studies show that the method preserves expert-specified activation logic and supports hierarchical inference from anomaly phenomena to deeper causes. Compared with conventional PSO, PSO Greedy achieves lower mean loss and a higher zero loss success rate.