Semi-supervised multitask deep convolutional generative adversarial network for unbalanced fault diagnosis of rolling bearing

In the process of data collection of rolling bearing, it is inevitable to get unbalanced data, including unlabeled samples, relatively few fault samples and abundant normal samples. To further improve sample efficiency, this paper proposes semi-supervised multitask deep convolutional generative adversarial network (SM-DCGAN). One-dimensional raw vibration signals are transformed into two-dimensional grayscale images. Different from traditional DCGAN, the proposed SM-DCGAN fuses the tasks of discrimination and classification to form multitask discriminator, which can simultaneously train labeled and unlabeled samples. Subsequently, knowledge distillation method is used to achieve simple fault classifier from multitask discriminator. Then, unbalanced samples can be expanded by generator in SM-DCGAN. Unlabeled samples, labeled samples and expanded samples are input into the simple fault classifier to realize semi-supervised fault diagnosis. The results of rolling bearing fault diagnosis experiments fully prove that the proposed SM-DCGAN has high accuracy, great robustness and sufficient stability for unbalanced fault samples.