CrossFuN: Multiview Joint Cross-Fusion Network for Time-Series Anomaly Detection

Time Series are often used to record the various states (i.e., metrics) of a system. Detecting the anomaly state is challenging because temporal dynamics and inter-metric dependencies need to be learned simultaneously, and anomaly types are diverse due to the complexity of the time series. Many anomaly detection models still leave some challenges unresolved. They mainly ignore the importance of information from frequency domain while concentrating on time domain modeling and further neglect the cross-domain effects of time and frequency domains. In this paper, we proposed a novel Multi-View Joint Cross Fusion Network (CrossFuN) to detect diverse types of anomaly, which has wide applicability to both univariate and multivariate time series. Particularly, based on the assumptions of Time-Frequency Heterogeneity and Time Frequency Coordination, a time-frequency joint cross fusion block is designed to simultaneously model the information of both the time and frequency domains, and captures the relationship between the time domain and the frequency domain. Moreover, taking advantage of the attention mechanism, CrossFuN can capture temporal dynamics and inter-metric dependencies. We conduct extensive experiments on seven real-world datasets to demonstrate the effectiveness of CrossFuN.