FS-TFD-MEN: an asynchronous motor few-shot fault diagnosis framework based on integrated metrics

Abstract As the ‘heart’ of industrial power, induction motors are widely used in engineering machinery that requires converting electrical energy into mechanical energy, and accurate monitoring of their operating conditions is crucial for ensuring the safety and efficiency of industrial production. Compared with traditional signal processing and machine-learning approaches, emerging deep learning techniques have shown great potential for motor fault diagnosis amid the rapid development of big data. However, under real-world operating conditions, fault samples are often scarce and difficult to collect, resulting in only a few labeled samples being available for each health condition. This limitation restricts the performance of many existing diagnostic models in few-shot fault diagnosis tasks. Therefore, this study proposes a novel few-shot time–frequency dual-branch metric ensemble network (FS-TFD-MEN) capable of performing non-destructive diagnostics on motors using monitoring data. First, a dual-branch feature extractor is designed for time-domain and time–frequency-domain signals to preserve local temporal variations and effectively reveal spectral components and time-varying characteristics, thereby yielding more robust feature representations. Then, an ensemble metric is constructed by integrating a multi-scale attention prototype metric with a Mahalanobis-distance-based metric, which captures global correlations between the query set and the support set while disentangling local features from the support set. Finally, a metric gating network is introduced to adaptively fuse the weights of the two metric branches and emphasize the branch with higher relevance, thereby improving recognition performance under few-shot conditions. Experimental results demonstrate that FS-TFD-MEN can effectively diagnose various faults of induction motors in few-shot scenarios, achieving a maximum accuracy of 96.13% when only six samples are available for each health condition.