A Novel Data Augmentation Method Based on Denoising Diffusion Probabilistic Model for Fault Diagnosis Under Imbalanced Data
Authors: Xiongyan Yang; Tianyi Ye; Xianfeng Yuan; Weijie Zhu; Xiaoxue Mei; Fengyu Zhou
Abstract:
Imbalanced data present a notable challenge in intelligent fault diagnosis, as the scarcity of fault samples often results in biased learning and reduced diagnostic accuracy. Conventional approaches, such as cost-sensitive learning, resampling, and generative adversarial networks (GAN), have achieved partial success, but remain limited by issues including overfitting, mode collapse, and unstable training. To address these limitations, this paper proposes a novel data augmentation method based on the denoising diffusion probabilistic model. Leveraging a diffusion mechanism that gradually adds and removes noise through a forward and reverse process, the proposed method enables stable and interpretable sample generation while enhancing both the authenticity and diversity of synthetic fault data.
In the proposed method, raw vibration signals are first converted into time-frequency images using the continuous wavelet transform and subsequently employed to train the denoising diffusion probabilistic model. After training, new fault samples are generated from Gaussian noise through the denoising process. To systematically evaluate performance of the proposed method, a comprehensive evaluation framework is developed, encompassing three dimensions: (1) the distributional similarity between generated and real data, quantified via maximum mean discrepancy; (2) the diversity and realism of generated samples measured by GAN-train and GAN-test metrics, and (3) the diagnostic performance under different imbalance ratios.
Extensive experiments on one public dataset and two real fault diagnosis testbeds demonstrate that the proposed method generates samples of higher quality and greater diversity compared to state-of-the-art data augmentation approaches, offering a novel solution to enhance fault diagnosis performance, particularly under challenging imbalanced data conditions.
