Unsupervised bearing raceway surface defect detection based on improved f-AnoGAN

Abstract For addressing bearing raceway surface defects, supervised deep learning methods face challenges such as insufficient defect samples and imbalance between defect and non-defect samples. To overcome these issues, we propose an enhanced fast anomaly detection with generative adversarial networks (f-AnoGAN) unsupervised detection algorithm. Firstly, SE-Block modules based on SE attention mechanism and residual structure are integrated into the encoder, aiding the network in focusing on channel information while alleviating gradient vanishing problems. Secondly, transfer learning is introduced to effectively enhance the algorithm’s detection performance and generalization ability. Finally, utilizing bearing images collected from industrial sites, a self-built bearing raceway surface defect dataset is constructed, and extensive experiments are conducted. Experimental results demonstrate that the improved algorithm achieves an area under curve score of 99.96% on the self-built bearing raceway surface defect dataset, representing a 7.07% improvement over the f-AnoGAN algorithm, meeting the requirements for online detection in bearing industry applications.