Semi-supervised variational bi-directional sampling on multi-class imbalanced electric power data for fault diagnosis

Abstract

Imbalanced data is one of the inevitable problems in electric power industrial fault diagnosis, which will affect the performance of classification models. Therefore, a semi-supervised variational bi-directional sampling (SVBDS) method is proposed to balance the dataset with data-level and feature-level combined. First, a novel bi-directional sampling framework suitable for the dataset with various imbalance degrees between multi classes is built, which corresponds with the real dataset structure. This framework can separate normal data and balance the rest fault sets. Second, a two-step synthesis sampling method is proposed for minority fault categories based on generative model to obtain homo-distributed and diverse data. This method can quantitatively generate specific categories and ensure generative quality. The experimental results on the actual industrial power transformer dataset and the wind turbine dataset demonstrate the effectiveness and application prospect of the proposed method.