Fault Detection of Air-spring Devices Based on GANomaly and Isolated Forest Algorithms

Aiming at detecting the crack-type and bulge-type faults of the high-speed train's air-spring devices, a computer vision-based image fault detection method is proposed. In this paper, we select the GANomaly network model, which is sensitive to bulge fault features, and histogram of oriented gradients (HOG) feature extraction combines with the isolated forest algorithm, which is sensitive to crack fault features to detection failures. Based on the means of sliding window segmentation, the positive and negative samples are divided into a large number of small pictures. They are easier to detect abnormal features. Then, these pictures are fed into the GANomaly network model. By comparing with the latent vector spaces obtained via encoding between positive and negative samples, bulge-type faults can be detected. HOG features are extracted from the small pictures, utilizing the isolated forest algorithm to detect crack type faults. Finally, marking a small picture with the highest anomaly score in the original image to complete precise location of fault object. or Math in Paper Title or Abstract.