Few-Shot Rotating Machinery Bearing Fault Diagnosis With Denoising Diffusion-Based Data Augmentation

Fault detection and diagnosis of bearings are crucial for maintaining the operational integrity of rotating machinery. However, the scarcity of fault samples, particularly for rare faults, leads to severe data imbalance, which consequently degrades model performance and increases false alarm rates. Although deep learning methods have improved diagnostic accuracy, models based on one-dimensional data often struggle to effectively capture the complex spatiotemporal correlation characteristics inherent in bearing faults. To address this challenge, this study proposes an innovative data augmentation framework that integrates a Denoising Diffusion Probabilistic Model (DDPM) with Gramian Angular Field (GAF) transformation, and presents an intelligent fault diagnosis method for bearings under small-sample conditions. The proposed method first processes vibration signals acquired from a three-channel sensor using the Piecewise Aggregate Approximation (PAA) algorithm to generate RGB-formatted Gramian Angular Difference Field (GADF) images. The size of the resulting images is then reduced by employing bicubic interpolation-based downsampling. Subsequently, building upon these prepared samples, we introduce an improved U-Net architecture within the DDPM, designated the IGM-DDPM, which achieves the fusion of deep holistic and salient features to enhance the sample generation process. The quality of the generated samples is evaluated using the Fréchet Inception Distance (FID) as a metric. Finally, an improved fault diagnosis model based on EfficientNet-B0 integrated with Convolutional Block Attention Module (CBAM) is proposed. By refining the number of input channels at each stage, this model mitigates overfitting and improves diagnostic effectiveness. Experimental results from multiple designed validation tests demonstrate that the proposed image enhancement method can effectively improve fault features within the images and maintain satisfactory diagnostic performance.