Sensor Embedding and Variant Transformer Graph Networks for Multi-source Data Anomaly Detection

With the rapid development of sensor technology and the proliferation of multi-source data, anomaly detection of multi-source time series data has become more and more important. In the past, anomaly detection methods often deal with the temporal information and spatial information contained in the data separately, which makes the spatio-temporal information in the data unable to be fully utilized by the model. To this end, this paper proposes a fusion of sensor embedding and temporal representation networks to solve this problem. In addition, we adopt graph neural network to better model multi-source heterogeneous data, and enhance the accuracy of anomaly detection by combining the double loss function of reconstruction loss and prediction loss. This approach not only facilitates the learning of normal behavior patterns from historical data but also enhances the model's predictive capabilities, allowing for more accurate anomaly detection. Experimental results on four multi-source sensor datasets show the superiority of the proposed method compared with the existing models. Further analysis show that the model enhances the interpretability of anomaly detection through the analysis of anomaly associated sensors.