Adaptive Graph Construction and Feature-Enhanced Graph Neural Network for Few-Shot Bearing Fault Diagnosis under Variable Conditions

<title>Abstract</title> <p>To address the issue of insufficient cross-condition generalization capability caused by the scarcity of fault samples and the variability of operating conditions in real industrial scenarios, this paper proposes an Adaptive Graph Construction and Feature-Enhanced Graph Neural Network (AG-MS-GNN) for few-shot bearing fault diagnosis under variable operating conditions. The proposed method consists of three core modules. First, an adaptive physical graph construction mechanism is introduced, which integrates time-domain, frequency-domain, and demodulation-domain features of vibration signals to build a physically inter-pretable initial graph structure, and incorporates a learnable graph optimization layer to achieve end-to-end collaborative adaptation between the graph topology and the diagnostic task. Second, a negative-sample-free graph-level contrastive feature enhancement strategy is designed, which applies semantically preserved data augmentation to the original signals and performs view comparisons in the feature space, thereby avoiding the complexity of negative sample construction in traditional contrastive learning and effectively utilizing unlabeled data to enhance feature representation under few-shot conditions. Finally, a unified diagnostic model integrating multi-scale graph convolution and attention mechanism is constructed , which captures both local details and global contextual information while dynamically focusing on fault-sensitive feature channels. Experimental results on the CWRU bearing dataset demonstrate that the proposed method achieves a fault recognition accuracy of 95.5% under extreme few-shot and cross-condition scenarios, significantly outperforming baseline methods including Support Vector Machine, one-dimensional Convolutional Neural Network, Graph Convolutional Network, and Graph Attention Network. Ablation studies verify the effectiveness and synergistic necessity of the adaptive graph construction, multi-head graph attention, residual connections, and self-supervised enhancement modules. By integrating physically driven adaptive graph construction with negative-sample-free self-supervised feature enhancement, the proposed AG-MS-GNN method effectively resolves the generalization challenges arising from the coupling of sample scarcity and operating condition variations in bearing fault diagnosis, providing a high-accuracy and robust technical solution for intelligent fault diagnosis of rotating machinery.</p>