基于频域通道交互增强的多维时间序列异常检测方法OA

To address the insufficient characterization of local frequency variations and channel representations in exist-ing multivariate time series anomaly detection methods,a frequency-domain channel interaction enhancement method for multivariate time series anomaly detection(TFDG-AD)was proposed.First,the fast Fourier transform was applied to the input sequence and partitions the spectrum into sub-bands,where the real and imaginary components were combined to construct sub-band representations for describing local anomaly features in different frequency ranges.Then,channel relationships were adaptively learned from the current sub-band representations to build sparse normalized channel graphs,on which residual graph propagation was performed to aggregate information from related channels and obtain enhanced sub-band representations.Finally,the enhanced representations were reconstructed back to the time domain,and anomaly scores were calculated by combining time-domain reconstruction errors with frequency-domain errors.Ex-periments on multiple public datasets show that TFDG-AD effectively improves detection performance and outperforms several baselines.