GCS-EN-SCNs-wd: a generalized class-specific cost-sensitive mechanism regulate stochastic configuration networks with direct link for transformer fault diagnosis

Abstract Data-driven intelligent fault diagnosis breaks conventional rules and achieves advancements in transformer fault diagnosis. However, intelligent models exhibit diagnostic preferences for the normal status due to the scarcity of fault class samples in the field. Thus, this paper proposes a novel transformer fault diagnosis method, which enhances learning efforts for minority classes while accounting for recognition accuracy of the majority, to address the imbalanced sample sizes between normal and fault statuses. Specifically, prioritizing minority classes in imbalanced learning may lead to a decline in diagnostic performance, particularly for the majority. Hence, this paper constructs a generalized class-specific cost-sensitive mechanism based on class-effective sample size (GCS-EN) to adjust learning efforts across classes. Additionally, stochastic configuration networks with direct links (SCNs-wd) is introduced. This model enables structural adaptation and incorporates the proposed cost-sensitive mechanism to improve performance for classes requiring varying levels of learning attention. Experimental results demonstrate that the proposed model achieves superior diagnostic performance for both fault and normal statuses compared to conventional imbalanced learning models, validated on published datasets and field data. The model attains an accuracy (Acc) of 95.34%, average accuracy (Avg-acc) of 93.63%, G-mean of 0.93, and MAUC of 0.96. This work provides a promising solution for power transformer fault diagnosis in practical field applications.