Robust Anomaly Detection via Neighborhood-Adversarial Signal-Noise Disentanglement

The rapid expansion of IT systems has intensified the need for robust unsupervised anomaly detection in multivariate time series. While existing methods leverage adversarial autoencoders to amplify reconstruction errors for anomaly discrimination, challenges persist in handling complex noise interference and distinguishing subtle anomalies in high-dimensional temporal data. In this paper, we propose a novel framework named Neighborhood Adversarial Signal-Noise Disentanglement (NASND). First, we introduce a staged modeling strategy with coarse-to-fine granularity to construct a standard set of normal samples, which utilizes DBSCAN clustering and neighborhood generation algorithms. Second, a neighborhood adversarial mechanism is designed based on a generative adversarial network (GAN) architecture, enforcing the model to learn shared normal patterns across different domains. Third, we propose a signal-noise disentanglement method to separate environmental noise from intrinsic features. Compared to vanilla GAN-based methods, NASND demonstrates superior performance in capturing temporal dependencies. The performance of NASND is validated by experimental results in a large number of experiments conducted on real industrial datasets, which show significant improvements in model accuracy (11.32%), recall (12.12%), and F1 score (13.71%) compared to the state-of-the-art baseline.