A Self-Supervised Contrastive Learning Anomaly Detection Method with Fuzzy Rough Sets

Anomaly detection has garnered significant attention as a critical method for maintaining system reliability. However, real-world datasets typically demonstrate nonlinear correlations when containing significant noise and missing data points. Conventional anomaly detection approaches typically examine anomalies from a single granularity level, they exhibit inherent limitations when handling such data. This paper proposes a self-supervised contrastive learning-based fuzzy rough anomaly detection method. Data augmentation incorporated into self-supervised contrastive learning, aiming to emulate the intricacies of real-world data scenarios. It further addresses data uncertainty through fuzzy rough set theory. Specifically, the proposed approach utilizes an embedding network trained via contrastive learning to extract low-dimensional features. A fuzzy relation matrix is constructed from these compressed representations. Upper and lower approximations are applied to compute the precision of fuzzy rough approximations. The final output quantifies the anomaly level of fuzzy information granules. For each object, the anomaly factor is calculated by aggregating the anomaly levels of all associated fuzzy information granules, which are derived from multiple fuzzy granules of the object. Experimental results across six datasets demonstrate that our method outperforms five comparative approaches in anomaly detection, consistently achieving superior AUC metrics.