Few-Shot Learning Based on Meta-Learning and Graph Attention Networks for Bearing Fault Diagnosis

In industries such as chemicals, machinery, and tobacco, bearings are crucial components which operate under high load conditions. When they fail or experience irreversible wear, it can lead to accidents and significant economic losses. Therefore, effective health monitoring and fault diagnosis of bearings are vital for ensuring the safe and stable operation of industrial equipment. The complexity of working conditions and the scarcity of failure samples complicate the design and training of fault diagnosis models based on traditional deep learning. To address this, this paper proposes a few-shot fault diagnosis method utilizing meta-learning and graph attention networks. Data is processed into two-dimensional pixel values, and a convolutional neural network extracts sample features to create labeled graph data nodes. A multilayer perceptron calculates the similarity between nodes and generates connected edges automatically. The graph data is then input into the graph attention layer, where a multi-head attention mechanism updates node features, leading to classification. The use of metric-based methods enhances the model's generalization ability, yielding excellent accuracies of 92.89% under 5-way T-shot and 94.10% under 5-way 5-shot.