Abstract Although graph neural networks (GNNs) have achieved remarkable success for fault diagnosis, the mainstream node classification paradigm suffers from low efficiency. Furthermore, intrinsic correlations within raw signals are rarely exploited to address the insufficient fault knowledge in model training. To address these issues, a time–frequency constraint-guided graph-level feature representation learning method for few-shot fault diagnosis is proposed. It overcomes the inevitable challenges of graph reconstruction and model retraining associated with node-level diagnosis models, thereby enhancing the real-time responsiveness. Specifically, a feature-enhanced chain graph (FECG) with only a few edges is introduced, which improves the interpretability and efficiency of edge construction in input graphs. Further, a time–frequency constrained graph convolutional network (TFCGCN) is developed, which can guide the gradient descent direction during model training based on the designed time–frequency constraint loss, reducing the model's reliance on labeled faulty samples. To mitigate the attenuation of time–frequency constraints during cross-layer propagation, a node feature transfer technique is proposed, and it also enhances the feature extraction capabilities of graph convolution layers. Through ablation experiments and comparisons with various existing models, the effectiveness and superiority of the proposed FECG-TFCGCN were validated, with several extremely unbalanced training sets on axial flow pumps and machine tool spindles.
Zhang et al. (Sat,) studied this question.