MFRM: Masked Frequency-Refined Modeling for Multivariate Time Series Anomaly Detection

Xiangchao Wang, Jiaxin Yin, Peizhuo Lv, Jie Yang, Yuanyuan Qiao

Keywords: anomaly detection, time series, frequency domain, learnable frequency masking, reconstruction model

TL;DR: MFRM introduces a novel MTSAD method with learnable frequency masking and a two-stage architecture to refine normal series modeling, addressing inadequate frequency information utilization and over-generalization in temporal reconstruction models.

Abstract: Frequency-domain information can reveal complex characteristics such as periodicity and seasonality in time series, playing a crucial role in multivariate time series anomaly detection. Since the frequency domain features a long-tailed distribution, existing temporal reconstruction models exhibit a fundamental bias toward the information-concentrated low-frequency bands, while severely underutilizing the discriminative power of fine-grained frequency details, making the detection of complex anomalies particularly challenging. In this paper, we introduce MFRM, a novel reconstruction model that strategically leverages frequency-domain information for enhanced anomaly detection. Our key innovation lies in a learnable frequency masking module that adaptively identifies and extracts frequency components most correlated with normal behavioral patterns, enabling fine-grained frequency details utilization. Furthermore, by disrupting the original spectrum of anomalous series through its frequency masking mechanism, the MFRM exacerbates reconstruction difficulties for anomalies in the time domain and offers a novel perspective to mitigate the over-generalization issue. Extensive experiments on seven benchmark datasets demonstrate MFRM's state-of-the-art performance.

Primary Area: learning on time series and dynamical systems

Submission Number: 6691