Unsupervised anomaly detection for gearboxes based on the deep convolutional support generative adversarial network
Chengguang Zhang, Zhen Guo, Chuan Li
Abstract
The gearbox, as a key transmission device in the industrial field, may lead to severe vibrations or even failures when abnormalities occur. Therefore, with the increasing complexity of industrial automation, precise anomaly localization has become crucial. To address this issue, a gearbox condition monitoring method based on an unsupervised deep convolutional support generative adversarial network (DCSGAN) is proposed. First, high-dimensional data collected is used to train the generator, and samples are generated by the generator to calculate their reconstruction errors. Next, these reconstruction errors are utilized to train one-class support vector machine (OCSVM). During the testing phase, reconstruction errors are similarly calculated for the test data, and after being normalized using the same process as the training data, the errors are input into the trained OCSVM model for anomaly detection. The proposed method has been validated on a real gearbox dataset, and experimental results indicate that the DCSGAN outperforms other models in anomaly detection. The GitHub code for the proposed DCSGAN has been made public at: https://github.com/MR-ach/DCSGAN .