Intelligent Fault Diagnosis of Bearings With Both Working Condition Variation and Target Data Scarcity

Intelligent fault diagnosis (IFD) methods allow effective feature extraction of mechanical big data and perform well in fault diagnosis tasks. Numerous domain adaptation based IFD methods have also been applied to achieve domain invariant fault diagnosis under working condition variation. Nevertheless, a sufficient amount of unlabeled testing data need to be available in training, which is not well suited to real-world applications. In real-world applications, only data in machine normal condition can be easily accessed. Targeting at this more hash real-world data scenario with both domain shift and target domain fault data scarcity, a two-stage method is proposed. First, a cycle-consistency based one-dimensional generative adversarial network is designed for fault data generation to provide data closer to the target domain distribution which relives the target data scarcity problem. Secondly, based on the self-challenge mechanism, a parallel inverse attention (PIA) module is proposed, and inserted into a diagnosis network to dig the learning potential of all features, which alleviates the problem of imbalance optimization among features. Massive experiments on two bearing datasets reveal that the proposed method has better adaptability under the more hash real-world fault data scenario.