MSR-GAN: multi-scales decomposition representations for unsupervised anomaly detection

Abstract

Time series anomaly detection is crucial in many fields due to the unique combinations and complex multi-scale time-varying features of time series data, which require accurate analysis. However, previous research has failed to adequately address these complexities, lacking effective decomposition and multi-scale modeling to comprehensively capture differences between normal and abnormal time points at various scales. To address this, our study aims to propose an innovative approach, the Multi-Scale Reconstruction Network (MSR-GAN). It features a Multi-Scale Decoupling Module (MTD) to separate input time series into different-scale components and models the reconstruction as parallel full-scale time series recovery. Furthermore, a Reconstructed Residual Collaborative Learning Module (RRCL) is constructed to perform inter-scale interactions by adaptively calculating importance scores for generator weight control. Extensive experiments demonstrate MSR-GAN’s state-of-the-art performance on multiple benchmark datasets for time series anomaly detection, thus providing a more effective solution, enhancing monitoring and handling of abnormal situations in related fields, and promoting the further development of time series analysis techniques.