FreqWave-TranDuD: A Multivariate Time Series Anomaly Detection Method Based on Wavelet and Fourier Transforms
Yashi Huang, Peishun Liu, Rongjia Han, Tian Yin, Quanjie Dou, Yibao Song, Mengqi Luo · IEEE Access · 2025
Time series anomaly detection (AD) plays a crucial role in network systems. It enables the timely detection of anomalies and root cause analysis, helping to prevent unnecessary losses. Existing methods have not fully exploited the frequency information embedded in time series data, limiting their ability to effectively capture global and periodic patterns. In this paper, we propose a novel time series anomaly detection method based on frequency domain feature extraction, referred to as FreqWave-TranDuD. The method adopts an encoder-decoder deep learning architecture, integrating Fourier Transform and Wavelet Transform to extract comprehensive time-frequency information. Fast Fourier Transform (FFT) is used to capture global periodic patterns, while Wavelet Transform captures local patterns. Additionally, LSTM is employed to capture temporal dependencies in the time series. The extracted features are then preliminarily concatenated and fed into the encoder-decoder architecture for accurate anomaly detection. Extensive experiments on six public datasets from domains such as aerospace detection and water treatment demonstrate that our method, FreqWave-TranDuD, outperforms other advanced baseline models.