Fault severity classification based on information flow and support vector machines

Abstract

Fault severity describes the level of impact caused by a fault. In the event of a malfunction in a nuclear power plant, a vast amount of data and alarms are generated, making it challenging for operators to identify and determine the state of the power plant in a short time. If the fault severity can be effectively identified during a malfunction, it would allow the operators to respond quickly and prevent further losses. Compared to other machine learning methods, the Support Vector Machine (SVM) has superior classification capability in scenarios with limited nuclear power samples. However, if the feature representation of the SVM training sample is inadequate, resulting in similar, indistinguishable training data, it will affect the SVM's classification of fault severity. To address this issue, using the Liang-Kleeman Information Flow (IF) can characterize information transmission from a causal perspective. By introducing IF to combine features of the training data, insufficient sample features can be replaced with the IF that describes information from a causal viewpoint, thereby enhancing the SVM's accuracy in fault severity classification. This paper takes the Main Steam Line Break accident in the primary loop of the nuclear reactor as an example and conducts simulation experiments on the PCTRAN simulation platform. The experimental results show that under 5 different fault degrees, the accuracy of fault degree evaluation of SVM model combined with IF is improved. This method can effectively improve the classification accuracy of different fault severity degrees in MSLB steam flow data with similar changes.