Generative adversarial networks for data augmentation in machine fault diagnosis

• Published in: Computers in industry Vol. 106; pp. 85 - 93
• Main Authors: Shao, Siyu, Wang, Pu, Yan, Ruqiang
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.04.2019
• Subjects: Auxiliary classifier generative adversarial networks; Data augmentation; Fault diagnosis; Induction motor; Signal generation; Fault diagnosis; Induction motor; Signal generation; Data augmentation; Auxiliary classifier generative adversarial networks
• DOI: 10.1016/j.compind.2019.01.001
• ISSN: 0166-3615

•Generative adversarial network is able to generate realistic samples.•Model with one dimensional convolution operation achieves best performance.•Quality of generated samples is evaluated by statistical features and experiments.•Generated samples serve as training data in machine fault diagnosis tasks.•Generated samples enrich dataset and improve fault classification performance. Generative adversarial networks (GANs) have been proved to be able to produce artificial data that are alike the real data, and have been successfully applied to various image generation tasks as a useful tool for data augmentation. In this paper, we develop an auxiliary classifier GAN(ACGAN)-based framework to learn from mechanical sensor signals and generate realistic one-dimensional raw data. The proposed architecture contains two parts, generator and discriminator, and both of them are built by stacking one-dimensional convolution layers to learn local features from the original input. Such stacked structure is able to learn hierarchical representations through convolution operation and easy to train. Batch normalization is performed within generator to avoid the problem of gradient vanishing during training, and category labels are used as the auxiliary information in this framework to help train the model. The proposed approach is designed to produce realistic synthesized signals with labels and the generated signals can be used as augmented data for further applications in machine fault diagnosis. In order to evaluate the performance of the generative model, we introduce a set of assessment to evaluate the quality of generated samples, including statistical characteristics and experimental verification. Finally, induction motor vibration signal datasets are utilized to investigate the effectiveness of the proposed framework.