Frequency-Domain Dual Masking for Enhanced Time Series Anomaly Detection

In the era of big data, anomaly detection is pivotal for maintaining system integrity and reducing economic losses within data mining research. While unsupervised deep learning methods excel at capturing nonlinear temporal dependencies, existing approaches exhibit limitations, including susceptibility to training biases from anomalous intervals and insufficient modeling of inter-channel correlations. To address these challenges, we propose a novel approach integrating dual masking in the frequency domain (FDM-AD). This method converts the time series into the frequency domain and segments it into patches. It employs an amplitude masking strategy combined with a binary mask matrix derived from a linear probability matrix. This masked representation is then processed in a latent space to optimize features. Subsequently, time-domain reconstruction is performed to obtain a filtered signal. An anomaly score is calculated based on the discrepancy between the original and the reconstructed signals. This approach is capable of effectively detecting both point anomalies and subsequent anomalies. Comprehensive experiments on benchmark public datasets validate the superiority of the proposed method.