Haoran Liang, Lei Song, zhongsong ma, and jiangyong duan
real-time anomaly detection, hierarchical clustering, unsupervised learning, concept drafting
Anomaly detection is one of the key technologies to ensure its safety and reliability of space imaging payload in orbit. Under the influences of complex operating environment, uplink instructions, equipment performance and so on, concept drifting phenomenon is obvious and it is difficult to define the boundary between normal and abnormal data, thus traditional anomaly detection method based on threshold value will be invalid. Besides, the lack of abnormal samples which makes the serious imbalance of positive and negative sample will increase the difficulty of anomaly detection. As a result, a kind of unsupervised learning method based on hierarchical clustering was proposed. The method includes offline training and online detecting parts. In offline training process, selecting multivariate training data to cover most working modes, and acquiring the clusters of training data to establish the normal model. Then in online detecting process, detecting real-time data based on the normal model above. Finally, we utilize the real data of some space imaging payload to verify the validity of the proposed method. The result shows that this method could detect the abnormal state effectively in real-time without physical model and abnormal sample and easy to implement.
Important Links:
Go Back
