| Flywheel is an important part of spacecraft attitude control.The failure of the flywheel system will affect the normal operation of the spacecraft and cause huge losses when it is serious.It is necessary to study the anomaly detection method of flywheel system.At the early stage,the data of flywheel system was less.At present,with the rapid development of aerospace industry,the data of flywheel system is greatly enriched.However,there are still big problems in the data structure.Most of the flywheel data are in the normal state,while only a small part of the data are in the abnormal state.In addition,a small amount of the fault data cover incomplete fault types,so all abnormal forms cannot be obtained,which is the prominent feature of the flywheel data.Aiming at the actual characteristics of flywheel data,this paper combines statistical analysis,clustering algorithm,neural network,relational degree measurement and frequent pattern mining algorithm to extract features from normal data and obtain invariant features contained in normal data,which can be used as the basis for anomaly detection.Based on the analysis of normal data,abnormal data can be detected.Firstly,time-frequency analysis combined with Ruili entropy was used to extract data features,and the rationality of the extracted data features was tested by cluster analysis.The verified data features can effectively distinguish the normal data from the abnormal data.According to the characteristics of the normal data,the threshold value of the normal data characteristics is obtained by combining the statistical theory.However,this method still has limitations.The method is only valid for certain data types.Secondly,the limitations of the threshold determination method are analyzed.In order to avoid matching the corresponding feature extraction method for each kind of data,an autoencoder is proposed to realize the intelligent data feature extraction.According to the characteristics of the extracted data,the region where the health data is located is determined by the algorithm constructed by the boundary,and the anomaly detection of the data is realized.According to the detection results,a density clustering algorithm is proposed to screen the health data,and the anomaly detection results of the algorithm are more ideal when combining the health data after screening with the boundary.Finally,since the research object of the algorithm in this paper is a single data,and the purpose of anomaly detection is achieved according to the characteristic changes of the research data,such research algorithm ignores the correlation between the data.The correlation degree measurement method is used to determine the strength of the correlation relationship between the data,and the selection of the associated data is realized.Combined with the knowledge of symbolic dynamics,the algorithm of frequent pattern mining of data is used to obtain the normal state of the health data.According to the set of health data states,the test data are judged.It is proved that the correlation relationship between data is feasible to realize anomaly detection. |
