中国空间科学技术

中国空间科学技术 ›› 2026, Vol. 46 ›› Issue (5): 86-95.doi: 10.16708/j.cnki.1000-758X.2026.0079

• 数学与航天专题 • 上一篇    下一篇

不完整航天数据上异常检测算法性能退化预测

宁宝玲1,*，李文博2,*，刘成瑞2，刘文静2，王彤3   

  1. 1.黑龙江大学计算机与大数据学院，哈尔滨150000
    2.北京控制工程研究所，北京100094
    3.北京空间飞行器总体设计部，北京100094
  • 收稿日期:2025-12-31 修回日期:2026-05-22 录用日期:2026-05-25 发布日期:2026-09-23 出版日期:2026-09-30

A method for predicting the performance degradation of anomaly detection algorithm over incomplete monitoring data of spacecrafts

NING Baoling1,*, LI Bowen2,*, LIU Chengrui2, Liu Wenjing2, WANG Tong3   

  1. 1.School of Computer and Big Data, Heilongjiang University, Harbin 150000, China
    2.Beijing Institute of Control Engineering, Beijing 100094, China
    3.Beijing Institute of Spacecraft System Engineering, Beijing 100094, China
  • Received:2025-12-31 Revision received:2026-05-22 Accepted:2026-05-25 Online:2026-09-23 Published:2026-09-30

PDF (PC)

16

摘要： 异常检测是保障航天器健康在轨运行的重要手段。已有的异常检测算法依赖高质量的训练数据，用于通常是不完整的航天数据时会显著降低其检测性能。针对不完整航天数据，对如何评估其对异常检测算法性能的影响进行了研究。首先，基于缺失值对模型输入的影响，提出了三种粒度的缺失程度度量。然后，设计基于缺失度量的模型泛化性能度量方法，并理论证明检测模型性能退化的上界。最后，基于性能度量方法，设计异常检测算法性能退化预测算法ADForecast。该算法能够准确预测特点缺失比例下的异常检测准确率，并能够与各类异常检测算法集成。实验结果表明，该算法可以高效精准的预测大部分缺失比例下的多种异常检测算法的性能，使用相似航天器历史数据预训练参数的预测误差低于5%。

关键词: 不完整数据, 缺失数据, 异常检测, 机器学习, 数据质量, 性能退化预测

Abstract: Anomaly detection is vital for spacecraft health in orbit. Existing algorithms rely on high-quality training data and suffer marked performance loss on typically incomplete spacecraft data. This study examines the impact of incomplete spacecraft data on detection performance. First, three granular levels of missingness metrics are proposed based on missing values' effect on model inputs. Then, a generalization performance measure using these metrics is designed, and an upper bound for detection model degradation is theoretically proven. Finally, the prediction algorithm ADForecast is developed from this measure; it accurately forecasts detection accuracy at a given missing ratio and integrates with various detection algorithms. Experiments show ADForecast efficiently and accurately predicts the performance of multiple algorithms under most missing ratios; when pretrained with historical data from similar spacecraft, the prediction error stays below 5%.