Improved generative adversarial network for vibration-based fault diagnosis with imbalanced data

Effective fault diagnosis is essential for maintaining the safe running of machine systems. Recently, the data-driven methods have shown great potential in intelligent fault diagnosis. However, the data collected in actual situations may be imbalanced which brings difficulties for the diagnosis. In this paper, an improved Generative Adversarial Network (GAN) is proposed to enhance the fault diagnosis performance with imbalanced data. Compared with the traditional GANs, the improved GAN introduces an auxiliary classifier to boost the training process and an Auto-Encoder based method for similarity estimation of generated samples. Meanwhile, an online sample filter is designed to ensure the selected samples meet the requirements of both accuracy and variety simultaneously. Experiments are implemented on the benchmark data from the Case Western Reserve University and the XJTU-SY datasets, and the results are compared with those by other intelligent methods, which proves the advantage of our proposed method in imbalanced fault diagnosis.

Automated summary

This study proposes an improved Generative Adversarial Network (GAN) to enhance fault diagnosis performance with imbalanced data, introducing an auxiliary classifier, an Auto-Encoder based method, and an online sample filter to ensure accuracy and variety of selected samples. Experimental results demonstrate the advantage of this method in imbalanced fault diagnosis.