A novel class imbalance-robust network for bearing fault diagnosis utilizing raw vibration signals

Recently, although vast intelligent fault diagnosis methods are proposed, their validities are mostly confirmed via balanced datasets, which cannot always hold for the class-imbalance problem prevails among datasets in real-world applications. Hence, a class imbalance-robust network is proposed for bearing fault diagnosis, which tackles class imbalance both in the feature extraction and classification stages. For feature extraction, balanced sparse filtering (BSF) is proposed, which innovatively introduces kurtosis into balancing the discriminative feature extraction capabilities of different classes. Meanwhile, the balancing matrix is also proposed in BSF to remedy the parameter updating imbalance caused by class imbalance. For feature classification, the balancing matrix is also embedded into softmax regression to enhance the balancing capability. Furthermore, extensive experiments on bearing vibration signal datasets are conducted in validity confirmation. Additionally, an interesting property of BSF is investigated, and the phenomenon that class imbalance is actually a two-edge sword is interpreted.