Meta-Learning With Intraclass and Interclass Optimization for Few-Shot Fault Diagnosis

Recent years have witnessed a booming interest in the data-driven paradigm for fault diagnosis. However, it is usually difficult to collect sufficient faulty data for model training in practical applications, thus limiting the application of these intelligent diagnosis methods. In this article, we develop a novel method named meta-learning with intraclass and interclass optimization (MLIIO), which targets training an effective metric-based fault classifier using limited data. On the one hand, an intraclass aggregation loss function is proposed to enable sample features from the same class to gather together. This yields a compact representation manifesting the central tendency for the same categories. On the other hand, an interclass discriminative loss function is proposed to enforce sample features from the different classes to maintain a large margin, which further ensures that the metric space has a clearer discriminative boundary. By applying the episodic training mechanism to optimize the proposed losses, general, and discriminative feature representations can be learned to more efficiently identify new failure scenarios with scarce data. Experimental results on a public rolling bearing dataset and a real-world railway turnout dataset showcase that the proposed MLIIO approach outperforms several state-of-the-art methodologies.