A Semi-supervised Deep Learning Model with Consistency Regularization of Augmented Samples for Imbalanced Fault Detection

With increasing requirements on reliability, maintainability and safety in modern ICT systems, fault detection, as an indispensable part of AIOps, has become essential in cloud computing or communication network environments. However, due to the lack of effective labels and class imbalance on faulty samples, fault detection performance based on the common classification model can't meet the system's operational requirements. Some recent approaches of SSL propose a consistency regularization loss to solve the problem of insufficient labels. However, these approaches are mainly for images based on artificial data augmentations but not feasible for all data types, and class-imbalance problem is not considered simultaneously. So, we propose a semi-supervised method for imbalanced fault detection with few labels, called SSLCR-IFD. In the method, we use a semi-supervised deep classifier based on consistency loss to solve the lack of labels, in which two sample augmentation methods based on clustering and GAN are used. Furthermore, a selective pseudo-labeling self-training strategy is proposed to solve the class-imbalance problem. Compared with the standard data augmentation, our methods alleviates the need for domain knowledge and can be used on multiple types of tasks. Finally, experiment results show that our method outperforms the baseline methods on two different AIOps tasks.