Feature Modeling And State Evaluation Of Power System Data For Space TT&C Mission

In the rocket mission,the working state of the power system is very important to the success or failure of the flight,which needs to be evaluated in real time.In the existing evaluation methods of launch site,the red line method,which can be used for real-time evaluation,is sensitive to data noise;and its missing alarm and false alarm is often high;the expert system method,is suitable for mature and stable engine models after flight mission,because its maintenance cost is too high.Researchers used to apply machine learning methods to data of engine test run and simulation,but the actual data of previous space missions is still put on the shelf,and compared with test run and simulation data,the actual data has more real and complex advantages;machine learning method is often difficult to train accurate models in the face of data sets with fewer samples,when tranfer learning methods can solve the problem of small amount of data in the newly developed engine.Aiming at the goal of real-time evaluation and post analysis of the launch site,firstly,the mission data of various types of engines in the launch site in the past five years were collected,the outliers were removed,the learning database structure was designed,and the data samples were stored in the database according to the engine type and different operation stages,so as to carry out parameter anomaly detection and system fault diagnosis modeling for each stage of each engine.Through DTW-AGNES clustering,seven kinds of abnormal conditions of param-eter change trend were obtained.Then,according to various parameters of different engine working stages,the appropriate feature vectors were obtained by constructing feature engineering.Different machine learning classification models were used to train and test the parameters in each stage of each engine,and compared with the envelope method in the traditional method.Finally,obtained the suitable model for each param-eter.The GSP algorithm of sequential pattern mining was improved,and the starting time sequences of all parameters in each stage of the engine were obtained.In order to solve the sequential pattern analysis with time interval,two methods,TEAX and TEAF,were proposed for clustering mining and classification diagnosis for the first kind of fault.Based on the obtained event sequence patterns,used methods of association analysis to mine the strong correlation between parameters from the data,and analyze the fault pattern to diagnose the second kind of fault.In view of the poor performance of machine learning in small sample areas,tried to use the transfer learning strategies to solve the constraints of the model training caused by the insufficient sample size in the target area.The Parameter Anomaly Detection System(PADS)and the Engine Fault Diagnosis System(EFDS)were designed,and parameters anomalies and engines faults were de-tected in two tasks of CZ-7A and CZ-3B in 2020,as real-time performance of models and systems were well.