A knowledge transfer-based intelligent decision support method for fault management

Abstract

In practical operations, fault management often depends on the expertise of onsite operators, yet manual judgments are limited in timeliness and consistency. To support onsite operators, this paper proposes a decision support approach to recommend the optimal intervention action for fault by comparing risk-reward of candidate actions. A significant challenge is quantifying action rewards, due to the unavailability of data on action consequences during the decision stage. In response, we introduce a symptom description-based knowledge transfer to evaluate action rewards without such data. First, risk prototypes are introduced, which are trained by historical fault data to transform risk magnitude into quantifiable distances between the prototypes. Then, fault symptom descriptions are incorporated as risk knowledge, upon which a generalized mapping function between risk prototypes and symptoms is established. This mapping function is realized through a zero-shot learning paradigm, enabling the knowledge transfer from observed symptoms to those not yet seen. Finally, an online recommendation strategy is developed, which identifies residual symptoms post-action and maps these to the risk prototypes in the feature space. By analyzing the distances between post-action risk prototypes, the risk-reward of actions is assessed, allowing for action recommendations based on their risk-reward rankings. The proposed method is validated by the benchmark Tennessee Eastman process. The results show that with a well-designed symptom matrix, it is possible to identify the optimal intervention action for fault management under zero-sample conditions.