As the core of Digital Twin (DT) systems, the analysis of historical and real-time time-series data in combination with anomaly detection methods can detect anomalies and predict the future state and behavior of industrial control systems faster and more accurately. However, the existing anomaly detection methods are difficult to model the complex nonlinear features inherent in the data, lack an effective expression of the correlation between the data variables, and the research is only limited to the time-domain modeling, ignoring the extraction and utilization of the frequency-domain information. In order to solve the above problems, this paper proposes a deep anomaly detection method named TF-DAD, In the time domain, the time convolution layer and the graph convolution layer jointly construct the spatio-temporal convolution block to realize the synchronous capture of time and space dependencies, In the frequency domain, the spectral information feature extraction is transformed into the magnitude and phase spectral image feature extraction task, the multi-branch attention mechanism in the Vision Transformer layer captures global context information of the image through multi-head attention, while concurrently using parallel lightweight dilated convolutions to extract local information with low computational complexity so as to realize the extraction of global information features of the image. The time-frequency features are jointly optimized to achieve more accurate anomaly detection by searching the optimal threshold through the grid. Experiments are conducted on SWAT, WADI and SMD datasets, and the experimental results show that the F1 metrics are improved by 6.14%, 7.96% and 9.02% over the suboptimal model, and 91.75% of the anomaly detection accuracy is achieved in the digital twin platform of the Key Laboratory of Information Security of Liaoning Petrochemical Industry, which verifies the effectiveness of the method for digital twin data.