.png){kind=icon}
.png){kind=icon}
Return
TFT-GCN: A Time-Frequency Based Model for Time Series Anomaly Detection
.png){kind=icon}
PRE
AI
Z
X
L
B
L
B
DOI：10.1109/tkde.2026.3707947.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Time series anomaly detection is essential for maintaining the stability and operational efficiency of industrial systems. However, most existing approaches emphasize temporal dependency modeling while paying limited attention to spectral characteristics embedded in the data. This limitation prevents models from effectively capturing both localized patterns and global trends, thereby reducing the accuracy of anomaly detection. To overcome these challenges, we introduce TFT-GCN, a novel anomaly detection model that integrates temporal and spectral analysis modules to simultaneously extract local and global information. Specifically, in the frequency-domain branch, we adapt the attention mechanism to extract critical information about frequency combinations, enabling the efficient extraction of global trends. In the time-domain branch, we introduce a cross-variable graph convolutional network, which not only models both homogeneous and heterogeneous interactions between series but also increases resilience to noise. Additionally, we incorporate a multi-scale attention mechanism that efficiently captures local dependencies across varying time scales to prevent the model from focusing too much on features at a single time scale. By combining the complementary strengths of temporal and spectral analyses, TFT-GCN offers a powerful framework for identifying anomalies. Experimental results on seven benchmark datasets verify that TFT-GCN surpasses most existing methods, highlighting its superior performance and robustness in time series anomaly detection.
Keywords:
Time series anomaly detection
time domain
frequency domain
Journal
IF:
10.4
Papers:
6.7K
Citations:
3.2W
Organization
Cited Papers