Pang Jingyue, Liu Datong, Peng Yu, Peng Xiyuan
School of Artificial Intelligence, Chongqing Technology and Business University, China.
School of Electronics and Information Engineering, Harbin Institute of Technology, China.
ISA Trans. 2023 Sep;140:354-367. doi: 10.1016/j.isatra.2023.06.002. Epub 2023 Jun 6.
Spacecraft telemetry data are real-time data as the only basis for ground operation station and management system to judge the working performance and health status of spacecrafts in orbit. Telemetry data are high dimension, strong-dependent, and pseudo-periodic series, which bring great challenges to traditional anomaly detection methods of multivariate parameters. In this case, with the advantages of strong feature extraction and space injection ability, Mahalanobis distance (MD)-based approach has been a strong foundation for industrial system health monitoring. However, the typical MD-based method performs anomaly detection with a fixed threshold for MD series without capturing temporal evolution which cause high false alarms or missing alarms for complex abnormal modes. In this work, the temporal dependence Mahalanobis distance (TDMD) is realized based on multi-factors prediction which can effectively detect contextual and collective anomalies in multivariate telemetry series. Upper and lower limits with time series correlation and dynamic characteristics for the MD of each arriving multivariate point are constructed for online testing. Adequate experiments on simulated and real telemetry series verify the effectiveness and applicability of the proposed method.
航天器遥测数据作为地面操作站和管理系统判断在轨航天器工作性能和健康状态的唯一依据，属于实时数据。遥测数据是高维、强相关且伪周期的序列，这给传统的多变量参数异常检测方法带来了巨大挑战。在这种情况下，基于马氏距离（MD）的方法凭借其强大的特征提取和空间注入能力，为工业系统健康监测奠定了坚实基础。然而，典型的基于MD的方法对MD序列采用固定阈值进行异常检测，未捕捉到时间演变，这导致对复杂异常模式产生高误报或漏报。在这项工作中，基于多因素预测实现了时间相关马氏距离（TDMD），它能够有效检测多变量遥测序列中的上下文和集体异常。为每个到达的多变量点的MD构建具有时间序列相关性和动态特性的上下限，用于在线测试。在模拟和真实遥测序列上进行的充分实验验证了所提方法的有效性和适用性。
