Fault Diagnosis on Imbalanced Data Using an Adaptive Cost-sensitive Multiscale Attention Network

Abstract

Mechanical fault diagnosis using vibration signals is a difficult problem in real-world applications, since few equipments in an abnormal state making abnormal data difficult to obtain, so that original samples are usually highly imbalanced. Imbalanced data make fault diagnostic models' classifier be biased toward normal samples, resulting in misdiagnosis of abnormal samples. To decrease the influence of imbalanced data on fault diagnosis, most methods based on oversampling have high computational burden due to the generation of many redundant samples, and the generated samples can also bring new noise to the diagnostic model leading to the decrease of diagnostic accuracy. In this paper, an adaptive cost-sensitive multiscale attention network, termed CS-MAN, is proposed, which consists of a Multiscale Convolutional Attention Network (MAN) for 1-D vibration signals and an Imbalanced Cross-Entropy (ICE) loss function. Specifically, MAN, as a feature extractor, constructs a broader feature space to obtain more discriminative features. The ICE considers the contribution of the imbalanced proportion of each sample and the characteristics of Difficult Samples Mining (DSM) to the overall misclassification cost. In this way, ICE can adaptively assign more suitable misclassification cost to decrease the influence of imbalanced data. Compared with oversampling-based methods, our method has lower computational burden and higher diagnostic accuracy. Experimental results on an imbalanced bearing dataset verify that our method outperforms the state-of-the-art methods based on over-sampling.