Research on Online Fault Diagnosis Data Enhancement Method for Portal Crane Based on Improved CycleGAN

Aiming at the characteristics such as the small number of fault samples and most of them are unlabeled of the gate drive motors, as well as the problems that lead to the difficulties in the application of traditional deep learning in fault diagnosis, this dissertation proposes a data enhancement method based on the improved CycleGAN for online fault diagnosis of portal crane. The article firstly describes the acquisition and preprocessing process of the portal crane dataset, including sensor arrangement, data collection, preprocessing and classification. Subsequently, an improved CycleGAN is proposed for data enhancement to address the problem of data imbalance and insufficient volume. CycleGAN achieves image transformation without pairwise data by introducing cyclic consistency loss and consists of two generators and two discriminators. The article elaborates the network structure, loss function and training process of CycleGAN, and verifies its effectiveness in one-dimensional vibration data enhancement through experiments. The experimental results show that after 2000 iterations, CycleGAN can effectively generate signals with better results, and the generated data mixed with the original data can significantly improve the performance of fault diagnosis.