Research On CMG Anomaly Detection Method Based On Data Driven

As one of the most critical components of modern spacecraft,Control moment gyro(CMG)is generally used in spacecraft attitude adjustment for the ability to achieve stable large moment output with low energy consumption.CMG can directly determine the service life of spacecraft due to it’s effectiveness and reliability.Data-driven spacecraft anomaly detection is becoming prevalent,with the massive historical operation data of spacecraft and the explosive development of machine learning technology.The loss of space mission can be significantly reduced through detecting and even predicting equipment faults in time.Based on the actual operation data of a CMG,a anomaly detection method for CMG based on ARMA(Auto-Regressive and Moving Average)was designed and implemented,moreover,a anomaly prediction method for CMG based on Informer model was proposed,and then a CMG anomaly detection system was designed and constructed.The research is organized as follows:The characteristics of time series data generated by CMG test platform were analyzed,and the overall data preprocessing procedure was designed.Discarding noise points,missing values,abnormal communication values and error values without practical significance was adopted as cleaning method.In the data dimension specification,the original data sampled in 8Hz frequency was aggregated twice in different time scale in order to get four kinds of time series data i.e.minutes,half hours,hours and days series.For the data conversion,four kinds of time series data were normalized by z-score method.After the preprocessing stage,the high-quality data that meet the requirements of model processing were obtained.A method for fault extraction before anomaly happening and a anomaly detection method based on ARMA were proposed.Based on the feature extraction principles of trend and monotonicity of sequence,some common statistics of each sequence were calculated,including the quadratic statistics of the new sequence formed by common statistics,the correlation of multiple time series variables,etc.And the statistics with great difference between normal data and abnormal data were selected as fault features.So that they can satisfy the monotonicity and trend,moreover,have the distinguishing property.In addition,in order to obtain the concrete manifestation of the new fault,the ARMA model was used for anomaly detection.The optimal ARMA model was generated by feeding normal data,and the abnormal data was tested with using residual of model as detection standard.A method of applying the informer model to the long-term prediction of CMG time series data was proposed in order to detect faults in advance.A training method on CMG abnormal data set was designed.Compared to ARMA model,the superiority of Informer model in multi-step prediction was verified by experiments.The CMG anomaly detection system with graphical user interface was designed by using Tkinter framework.The requirements of the project specification were analyzed and then,the specific functions of the system were put forward,the overall architecture and sub-modules of the system were designed.The research results of this paper were integrated in the system,and an anomaly detection platform with friendly interface,simple operation and stable performance was achieved.