Research On Key Technologies Of In-orbit Spacecraft State Prediction And Diagnosis Based On Data Driven Method

On-orbit spacecraft operate in harsh space for a long time.Due to its complicated structure and difficult maintenance,once a failure occurs,it will cause huge losses.Therefore,while continuously improving the reliability of the spacecraft itself,more and more attention is being paid to the health management of the orbital spacecraft.At present,the health management of on-orbit spacecraft is mainly focused on the research of fault diagnosis methods.Fault diagnosis is an afterthought maintenance method,and there are serious deficiencies in the processing effect.When an abnormal change occurs in a spacecraft,usually the corresponding telemetry parameter also changes,so it can be predicted by the telemetry parameter,and it can be intervened in time before the fault occurs.This paper takes the telemetry data returned by the orbiting spacecraft as the research object.The main contribution and findings of this paper include:(1)The pre-processing method of telemetry data for on-orbit spacecraft is studied.By analyzing the actual telemetry data,the basic process of telemetry data preprocessing is studied,including data filling,data filtering and periodic data extraction.Aiming at the filtering problem of telemetry data,a filtering method based on the sliding window fastest tracking differentiator is proposed.The data is divided through the sliding window,and the parameters of the local fastest tracking differentiator are designed according to the stability of the data in each window to realize the telemetry data.Ideal filtering.In the subsequent analysis of telemetry data,the data is usually divided in a fixed period given by expert experience,and this method may cause truncation or redundancy of information when dividing non-linear and non-stationary telemetry data.Since the on-orbit spacecraft operates according to the orbital period,a large amount of telemetry data returned by it also presents an approximate periodic cyclic change law.A pseudo-period extraction method for spacecraft telemetry data is proposed,divided by the similar form of the telemetry data in the time domain.The basic analysis unit of telemetry data,and on this basis,the pseudo-period of the data is extracted as a unit for dividing data.(2)Research on spacecraft telemetry data prediction method based on decomposition and integration.There are a lot of non-linear and non-stationary data in the spacecraft telemetry data,so the prediction effect directly using the classical model prediction method is limited.In order to improve the applicability of prediction methods,a decomposition and integration architecture prediction model is adopted,and a fractional order DGOM(1,1)power model based on morphological variational mode decomposition is proposed.Firstly,the data decomposition method is analyzed,and the superiority of the variational modal decomposition is pointed out.For the parameter setting of the variational modal decomposition,the morphological filtering method is used to determine the initial value of the variational modal decomposition parameter by lifting the data spectrum envelope,And gives the parameter optimization method.Because the gray system prediction model has strong applicability and simple modeling process,the gray prediction model is selected as the prediction method of the telemetry data.According to the characteristics of the telemetry data,a score is established based on the GM(1,1)model The order DGOM(1,1)power model improves the scope of application of the GM(1,1)model.It combines the variational modal decomposition with the fractional order DGOM(1,1)power model to form a decomposition and integrated prediction model.It can handle different types of telemetry data prediction problems and has good prediction accuracy.(3)Studied the prediction effect of deep neural network of gate recurrent unit(GRU)in deep learning model on telemetry data.For the massive telemetry data transmitted by the orbiting spacecraft,the deep learning neural network is used for prediction.Due to the gradient dispersion and explosion problems in the recurrent neural network(RNN),the long-term and short-term memory network(LSTM)is mainly used to predict the data.Compared with LSTM,GRU has a simple network structure.According to the fast requirements of spacecraft state prediction,on the basis of minimizing GRU(MGU),a faster multi-resolution minimum-gated recurrent unit neural network model(MMGU),the network has a more concise network structure and good prediction accuracy.(4)Research on data-driven abnormal diagnosis technology.According to the requirements of the ground measurement and control station for the rapidity and accuracy of the on-orbit spacecraft anomaly diagnosis,the spacecraft anomaly diagnosis is performed from a single feature angle and a multiple feature angle respectively.Research on data-driven anomaly diagnosis technology of on-orbit spacecraft.According to the requirements of rapidity and accuracy of on-orbit spacecraft anomaly diagnosis,spacecraft anomaly diagnosis is carried out from single feature angle and multi-feature angle respectively.A single-feature anomaly recognition method based on the phase plane trajectory feature and Rényi entropy feature of the telemetry data is proposed.For a large number of unlabeled spacecraft telemetry data,multiple features of the telemetry data are extracted and processed using semi-supervised ELM.Realize multi-feature semi-supervised ELM identification,and realize accurate and rapid diagnosis of telemetry data faults.