An Improved Conditional Generative Adversarial Method for Wind Turbine Gearbox Fault Diagnosis With Imbalance Data

Abstract

The problem of data imbalance is widely present in wind turbine (WT) gearbox fault diagnosis, resulting in significant differences in diagnostic accuracy for majority and minority classes. To address this challenge, an improved conditional generative adversarial network (ICGAN) was proposed to mitigate the impact of imbalanced data on fault diagnosis. The ICGAN uses a 1-D convolutional neural network (CNN) to extract feature vectors rich in detailed information from real samples as conditional signals. This enhances the learning ability of the generator for intrinsic patterns and structures in the data, ensuring that the generated data not only contain the intrinsic features of real samples but also meet specific condition requirements. In addition, considering that Gaussian mixture models (GMMs) can effectively capture the multimodal characteristics of signals, ICGAN uses the signals synthesized by GMM as the generator inputs, enabling the generator to cover more sample space and increase the diversity of generated samples. Moreover, the GMM can adaptively adjust parameters according to the feedback of the discriminator to improve the adaptability of the generator to the data, thereby enhancing the generalization performance of the ICGAN. The Wasserstein distance and the Bayesian optimization algorithm are also used to improve the stability and efficiency of model training. The experiments show that the data generated by ICGAN are more similar to real samples compared to other generative adversarial network (GAN)-based models. Using the vibration data generated by ICGAN for multiclass fault diagnosis of real WT gearbox, the average diagnostic accuracy can reach 94%, with the accuracy for minority classes also reaching 92%.

Topics & Concepts