Fault diagnosis of rolling bearings with imbalanced samples using conditional temporal denoising diffusion generative adversarial network

Accurate fault diagnosis of rolling bearings is essential for ensuring the safe and reliable operation of rotating mechanical systems. However, samples imbalance, especially the scarcity of real fault samples is widely present in engineering applications which weakens the diagnosis performance of traditional generative models due to the mode collapse and low sampling efficiency. To overcome these challenges, this study presents a fault diagnosis framework for few-sample bearing fault diagnosis, which is built upon an enhanced conditional temporal network derived from Denoising Diffusion Generative Adversarial Networks (DDGAN), designated as Conditional Temporal Denoising Diffusion Generative Adversarial Network (C-TeDDGAN). This framework enables the accurate synthesis of diverse fault samples through the integration of a class-conditional generation mechanism. Concurrently, a temporal correlation loss function is developed to enforce temporal continuity and correlation constraints on vibration signals. Lastly, a bidirectional cross-validation scheme integrating “Training on Generated data and Testing on Real data (TSTR)” and “Training on Real data and Testing on Generated data (TRTS)” is established to filter high-fidelity synthetic samples. Experimental evaluations on both a high-speed aero-engine bearing dataset and a laboratory test rig dataset demonstrate that the diagnostic accuracy across varying data imbalance ratios surpasses that of comparative benchmark models. Ablation studies confirm the synergistic benefits of the temporal correlation loss and cross-validation module, thereby verifying the efficacy of the proposed framework in enhancing diagnostic accuracy for imbalanced fault diagnosis and effectively alleviating the few-sample imbalance issue.