Anomaly Detection for Spacecrafts in the Presence of Incomplete Data

Data driven anomaly detection methods have been widely applied in spacecrafts. Most of the previous works assume training data are complete, but in practical settings, missing data usually appear. Therefore, the training data may miss some parts and become incomplete data. A natural strategy commonly used is to fill the missing parts using data imputation methods. However, there still lack of methods which can guarantee the imputation quality. When the training data are incomplete, it is hard to get rid of the “garbage in garbage out” issue, and the performance of anomaly detection models can not be guaranteed. In view of the above issues, this paper proposes a novel method for building anomaly detection models without relying on the assumption of perfect data imputation. The main idea is motivated by the observation that the time series data generated by spacecrafts have multiple dimensions, and the degrees of incompleteness on different dimensions are usually different. For the dimensions with large amounts of missing data, deleting them from the training data may let the model performance better. Utilizing the above idea, the proposed method first discovers the incompleteness patterns by mining the relationships between dimensions, builds multiple anomaly detection models on different training data obtained by deleting dimensions according to different patterns, and finally improves the detection performance by integrating all the models. Experimental results show that the proposed method can improve the performance of anomaly detection significantly when incomplete data exist.