Many existing fault diagnosis methods based on deep learning (DL) require numerous fault samples to train the diagnosis model. However, in industrial applications, rotating machines (RMs) operate in normal states for most of their service life with fault events being rare and thus failure samples are very limited. To solve the problem above, a novel wide residual relation network (WRRN) is proposed for intelligent fault diagnosis of the RMs. Specifically, the WRRN is trained by performing a series of learning tasks in RMs with sufficient samples to obtain knowledge about how to diagnose, and then it is directly transferred to realize fault task of the RM with small samples. In this method, a wide residual network-based feature extraction module is used to generate representative fault features from input samples, and a relation module is designed to calculate the relation score between the sample pairs so as to determine their categories. Extensive experiments are conducted on two RMs to validate the WRRN method. The results demonstrate that the WRRN can accurately identify the fault types of the RMs with only small samples or even one sample. The WRRN significantly outperforms the existing popular methods in diagnostic performance.
Keywords: fault diagnosis; few-shot learning; relational network; rotating machines; wide residual network.