CBAM-enhanced diffusion model with minimal noise scheduling for data augmentation in fault diagnosis with imbalanced dataset

Abstract Acoustic signals, known for their non-contact nature, rapid response, and rich informational content, are widely applied in pipeline fault diagnosis. However, in real-world scenarios, fault events are rare and fault samples are difficult to obtain, resulting in significant class imbalance where normal samples dominate the dataset. To address this challenge, this paper proposes a convolutional block attention module (CBAM)-enhanced diffusion model with minimal noise scheduling for data augmentation in imbalanced fault diagnosis. A novel scheduling strategy is introduced by explicitly constructing a weighted term in the upper bound of the training loss to optimize the diffusion coefficient sequence. This strategy reduces the number of diffusion steps while preserving sample quality, thereby significantly improving generation efficiency. Furthermore, a CBAM-Enhanced U-net architecture is designed by integrating the CBAM and residual connections into the denoising network, enhancing the model’s ability to extract key acoustic features. Experimental results demonstrate that the proposed method achieves superior sample quality, improved distribution alignment with real data, enhanced generalization, and faster generation speed. The synthetic samples effectively enrich the minority class, balance the training dataset, and enhance the performance of downstream fault classification models.