Detecting Anomalies in Satellite Telemetry Data based on Causal Multivariate Temporal Convolutional Network
Zefan Zeng, Jiugang Lei, Guang Jin, Chi Xu, Lu Zhang · 2022
Prediction-based methods are widely used in satellite telemetry data anomaly detection. However, pursuit to detect as many anomalies as possible often leads to unsatisfactory false positives. Moreover, the interdependence (causality) between telemetry parameters is often overlooked as an important tool for data analysis, resulting in poor model interpretability. We propose a Causal Multivariate Temporal Convolutional Network (CMTCN) model, which first discovers anomaly propagation causality among parameters by constructing a causal network based on Anomalous Transfer Entropy (ATE), then applies a multi-feature input Temporal Convolutional Network (TCN) to predict telemetry data. We also propose a Peaks Over Thresholds (POT) -based anomaly score calculation method and anomaly identification criterion for the residuals. CMTCN can effectively capture and exploit the interdependence between features, which not only increases the interpretability of the model, but also is more sensitive to anomalies. Case studies on two public telemetry datasets and a multivariate time series dataset show that our method outperforms all baselines in F1-score.