Harmful data enhanced anomaly detection for quasi-periodic multivariate time series

Abstract

Multivariate quasiperiodic time series (MQTS) anomaly detection has demonstrated significant potential across various practical applications, including health monitoring, intelligent maintenance, and quantitative trading. Recent research has introduced diverse methods based on autoencoders (AEs) and generative adversarial networks (GANs) that learn latent representations of normal data and subsequently detect anomalies through reconstruction errors. However, anomalous training set data can cause model pollution, which harms the ability to of the utilized model reconstruct normal data. The current data extreme imbalance creates an enormous challenge in terms of stripping out these anomalies. In this paper, we propose a GAN-based multivariate quasiperiodic time series anomaly detection method called IGANomaly (I represents isolation). This method isolates normal and harmful samples via pseudolabeling and then learns harmful data patterns to enhance the process of reconstructing of normal samples. First, the reconstruction error and potential feature distribution are jointly analyzed. Bimodal dynamic alignment is achieved through multiview clustering, thus overcoming the limitation of unidimensional determination. Second, dual reconstruction constraints for the generator and a gradient penalty mechanism for the discriminator are constructed. While maintaining the reconstruction quality achieved for normal samples, the propagation path of abnormal features is actively perturbed through a gradient inversion strategy. On three public datasets, IGANomaly achieves \(F1\ scores\) of 0.811, 0.846, and 0.619, demonstrating an average improvement of 18.9% over the best baseline methods.