An enhanced generative adversarial network for longer vibration time data generation under variable operating conditions for imbalanced bearing fault diagnosis

As a typical data augmentation method, a generative adversarial network is widely applied to solve data scarcity problems for imbalanced bearing fault diagnosis. However, these methods still face challenges in generating longer data due to the risk of mode collapse and instability during training. To address this issue, an enhanced generative adversarial network is proposed for generating longer vibration time data to improve imbalanced bearing fault diagnosis under variable operating conditions. Firstly, a dual cross-frequency attention block is integrated into the discriminator to adaptively extract intra-component and inter-component features across low and high frequency components decomposed using Wavelet, thereby facilitating generator to generate longer synthetic time data with higher frequency resolution. Furthermore, the sequence information block is introduced to generate synthetic time data under variable operating conditions by incorporating specific operating condition information into the generator. To expedite the synthetic data generation process under variable operating conditions, healthy data corresponding to these conditions are used as input for the generator, replacing random noise. Finally, the superiority of the proposed method is validated through experiments on two bearing datasets for imbalanced bearing fault diagnosis. Experimental results on these two datasets demonstrate that the proposed method with 2048-length synthetic data achieves the highest accuracy of 98.75 % and 96.88 %, respectively, outperforming state-of-the-art methods. Therefore, the proposed method can effectively address the challenge of generating longer vibration time data, improving diagnostic accuracy in imbalanced bearing fault diagnosis under variable operating conditions.