Relationship-aware multichannel diffusion augmentation and graph convolutional fusion for small-sample rotating machinery fault diagnosis

Annals of Forest Science · Springer Nature · 2026-11-16 · DOI: 10.1007/s10845-026-03025-9

Аннотация

Limited labeled data in multi-sensor rotating machinery fault diagnosis can lead to inadequate coverage of the training distribution and insufficient exploitation of cross-sensor statistical relationships. To address these issues, this study proposes a Relationship-Aware Graph Convolutional Network with Diffusion Augmentation (RAGCN). First, a conditional diffusion model, termed the Residual-Attention U-Net-based Diffusion Model (RAUnet-DM), is developed to generate complete synchronized multi-sensor windows and to model the class-conditional multichannel joint distribution, thereby reducing the distortion of cross-sensor statistical relationships that can arise when channels are generated independently and recombined. Second, sensor channels are treated as graph nodes, and a sample-adaptive functional correlation graph is constructed from the absolute Pearson correlation coefficients within each window. A shared one-dimensional encoder and a graph convolutional network are then used for cross-sensor feature fusion. Synthetic-data quality is evaluated using a comprehensive consistency index (CI), correlation preservation degree (CPD), and frequency-domain coherence relation similarity (FCRS), which quantify overall distributional consistency, preservation of zero-lag correlation structure, and similarity of frequency-domain coherence patterns, respectively. Experiments on the SEU gearbox dataset and an independently acquired rolling-bearing dataset show that, with 10% real training data, RAGCN achieves diagnostic accuracies of 95.66 and 92.36%, respectively. Controlled experiments further show that synchronized multichannel joint generation better preserves class-conditional cross-sensor statistical relationships, while the sample-adaptive functional graph can exploit this relational information to improve multi-sensor feature fusion. These results demonstrate the effectiveness of RAGCN for small-sample multi-sensor fault diagnosis of rotating machinery.

Читать у издателя →

Похожие статьи