Multichannel Anomaly Detection for Spacecraft Time Series Using MAP Estimation

• Published in: IEEE transactions on aerospace and electronic systems 2024-10, Vol.60 (5), p.5842-5855
• Main Authors: Li, Tianyu, Baireddy, Sriram, Comer, Mary, Delp, Edward, Desai, Sundip R., Foster, Richard H., Chan, Moses W.
• Format: Article
• Language: English
• Subjects: Anomalies; Anomaly detection; anomaly marked point process (Anomaly-MPP); Correlation; Data models; Datasets; Error analysis; Error detection; Machine learning; Object recognition; Predictive models; Sequences; Space vehicles; Spacecraft; Statistical models; System failures; Target detection; Time series; Time series analysis; transformer; Transformers; Water distribution; Water engineering; Water treatment
• ISSN: 0018-9251; 1557-9603
• DOI: 10.1109/TAES.2024.3400943

Automated anomaly detection in spacecraft telemetry systems is essential for analyzing abnormal events and system failures. A widely adopted strategy is to predict the target time sequences using a machine learning method first, then extract the anomalies from the residuals between the target time sequences and the predicted sequences by a thresholding method. Although thresholding-based anomaly extraction is simple and fast, it fails to take advantage of correlations between anomaly sequences over time and across channels. To make the process of anomaly extraction more flexible and more accurate, a statistical model referred as an anomaly marked point process (Anomaly-MPP) is proposed in this article. This model treats anomaly sequences as objects to be detected, making the anomaly detection a classical object detection problem. Formulating this as an optimization problem, we find the maximum a posteriori estimate of the set of anomaly objects in a multichannel time-series dataset, modeling the prediction error sequences generated from the output of a transformer with the proposed Anomaly-MPP for the posterior distribution. The prior distribution can incorporate domain knowledge and user-specified context into the problem formulation, thus providing additional detection “power.” By including a length prior energy term and a correlation prior energy term into the model, the anomaly extraction process not only considers the prediction error values, but also takes the length of detected anomaly sequences and the interchannel dependencies into account. A case study is given in the experimental section to illustrate the use of the model on a real dataset. Also, the effectiveness of our method is evaluated on an Mars Reconnaissance Orbiter dataset with inserted known anomalies and two public datasets: Secure Water Treatment and Water Distribution.