Few-shot fault diagnosis of rotating machinery with two-branch prototypical networks

In the fault diagnosis of rotating machinery, vibration signal or spectrum is usually used. As a data-driven method, deep learning has been introduced into the field of fault diagnosis. But it often confronts with two critical difficulties: few fault cases and single data source. To this end, we employ the prototype network to solve the problem of few fault cases, and use the two-branch technique to combine data sources in time domain and frequency domain. We introduce the two-branch network structure into the framework of the prototype network and develop a two-branch prototype network (TBPN) for fault diagnosis. The TBPN model is constructed through three main steps. First, we extract information from vibration signals in time domain and frequency domain respectively, and feed them into the model as two branches. Second, the prototype representation of each fault in time domain and frequency domain can be learned through metric learners, and the distances between fault prototypes and query faults features are then calculated. Third, the distances in time domain and frequency domain are integrated and incorporated into the softmax function for multi-classification. The performance of TBNP is verified by a real-world application on fault diagnosis of rotating machinery with the case data accumulated by an industrial Internet enterprise in China. The results show that the TBPN model is suitable for fault diagnosis in the case of small data. Compared with using time domain signals or spectrum alone, their combination use can improve the effectiveness of fault diagnosis.