Multivariate Time Series Anomaly Detection with Fourier Time Series Transformer

Abstract

Anomaly detection in time series data plays a key role in automatic industrial operations. Due to the intricate temporal dependencies within time series data and the difficulty in obtaining labeled data, recent anomaly detection methods have primarily focused on the temporal domain features of time series data, neglecting the frequency domain features. However, spectral analysis can better utilize periodic information within time series data such as seasonal patterns, which helps capturing multi-scale and multiple frequency features. In this paper, we present a Fourier Time Series Transformer model (FTST for short), which combines the features of both time and frequency domains for time series anomaly detection. Specifically, the attention mechanism is utilized for modeling the temporal domain, while the Fourier Transform is employed to transform time series data into frequency domain data. The frequency domain features are then modeled using a Temporal Convolutional Network. By making full use of the temporal and frequency domains of time series data, FTST can significantly enhance the performance of time series anomaly detection. Experimental results on popular benchmark datasets demonstrate the anomaly detection performance of the proposed method.