A smart structural health diagnosis system for few-shot bearing fault detection using Mamba and edge-attention GNN

Abstract

Bearings are core load-bearing components of rotating machinery, and local structural damage will easily cause severe faults and even catastrophic industrial equipment shutdowns. Current intelligent bearing diagnosis methods face three major bottlenecks in actual factory scenarios: severely limited labeled fault samples, weak global feature modeling ability of traditional convolution structures, and insufficient capacity to excavate implicit topological relations between limited training samples. To address these three practical difficulties, we develop a lightweight few-shot diagnostic network named PEIC-graph neural network (GNN)-Mamba.The proposed PEIC-GNN-Mamba framework addresses few-shot bearing fault diagnosis through a dual-stream collaborative architecture and graph-based relation reasoning. First, for complementary feature extraction, a dual-stream backbone is designed: the PEIC module captures fine-grained local vibration features using wide-domain strip convolutions and partial channel attention, while the patch-free 2D Mamba models long-range spatial-temporal dependencies via selective state space mechanisms to compensate for the global modeling limitations of traditiDonal convolutional neural networks. Second, to effectively model sparse sample relationships, support and query samples are constructed as graph nodes within a lightweight edge-attention graph reasoning module, which adaptively learns implicit topological associations and generates dynamic adjacency matrices. Finally, temperature-scaled cosine similarity is employed for metric classification, mitigating performance degradation under extreme data scarcity and enhancing overall cross-domain generalization. We conduct full experiments on the public Case Western Reserve University benchmark and a self-built industrial dataset with heavy background noise. Results demonstrate our method outperforms state-of-the-art Mamba, Transformer and graph-based few-shot models on all evaluation indicators. Ablation experiments confirm each structural improvement yields obvious diagnostic accuracy gains. Supported by linear computational complexity and low memory overhead, this framework is highly promising for real-time edge structural health monitoring.