Investigation On Fault Prediction Methods Based On Process Neural Network For Liquid-Propellant Rocket Engines

Fault prediction for liquid-propellant rocket engines(LRE)plays an important role in the LRE health monitoring.With big data technology and artificial intelligence moving on such a pace,data-driven fault prediction methods for LRE already become current research hotspot.Taking a large-scale liquid oxygen/hydrogen rocket engine as the object investigated,the research on process neural network(PNN)fault prediction methods for LRE of this dissertation is governed by the ideas of ‘the general framework to specific methods,single model to combined model,single algorithm to multi-algorithms,theoretical analysis to application integration'.The main research achievements in the dissertation are summarized as follows.The fault prediction methods for LRE based on PNN are proposed.The relevant concepts of fault prediction for LRE is described by mathematical language,and the work procedure and steps of fault prediction for LRE are summarized.Further the hierarchical structure and function module of LRE fault prediction are analyzed,and on this basis a general framework and strategies of fault prediction for LRE is constructed.Finally,based on the general framework and strategies,a PNN fault prediction method is proposed for the stable process and startup process of LRE.The general framework and strategies and PNN fault prediction method are verified by the engine ground test data.The results show that the process of fault prediction for LRE can be regulated efficiently by the model,and the fault prediction method based on PNN can obtain the accurate results of engine fault prediction and component level fault isolation.However,there are some drawbacks of the method,such as complex modeling process,difficult optimizing,low precision of prediction and poor ability to predict the development trend of fault.The PNN fault prediction methods for LRE based on increment learning is proposed.In order to predict the LRE stable process data timely,a fault threshold adaptive update method(FTAU)is proposed.And the method solve the problem of that the ATA can not deal with the soft fault.In order to solve the problem of low precision of prediction and poor ability to predict the fault development trend of off-line PNN,the weights update PNN(WU-PNN)and the output accommodation coefficient update PNN(OACU-PNN)are proposed respectively according to the difference ways of network output.The FTAU,WU-PNN and OACU-PNN are verified by the engine ground test data respectively.The results show that the timeliness and validity of fault prediction are improved effectively by the FTAU.The fault development trend is predicted more accurately by the WU-PNN and OACU-PNN than the off-line PNN.A kind of combined prediction PNN is proposed.In order to improve the modeling efficiency and generalization ability of PNN prediction method,the combined prediction PNN method(CP-PNN)is proposed based on the analysis of single network prediction performance and combined network generalization error.A dynamic weights combining method is also proposed based on the prediction error of PNN submodel,and it is very effective for the CP-PNN to predict the variation trend of LRE characteristic parameter.Finally,in order to improve the prediction efficacy of combined PNN,an error predicting correlation prediction PNN method(EPCP-PNN)is proposed.The CP-PNN and EPCPPNN are verified by the engine ground test data respectively.The results show that the process of prediction model establishing is simplified by the above methods and generalization ability of prediction method is also improved by the above methods.The PNN prediction method based on LRE data sample reconstruction is proposed.How the quantity,quality and representativeness of LRE data sample influence on the PNN prediction model optimizing and PNN prediction ability are analyzed.The methods of eliminating gorss errors,data filtering and phase-space reconstructing are also given in this dissertation.On the basis of the periodicity and multiple tendency analysis for LRE data,a multiscale parallel PNN(MP-PNN)is proposed.In order to deal with lack of LRE data sample,a segmental prediction PNN(SP-PNN)is proposed.The MP-PNN and SPPNN are verified by the engine ground test data respectively.The results show that sample length can be reduced and the model can be established more easily for the MP-PNN and SP-PNN.The prediction performance of PNN for the LRE characteristic parameters is improved effectively.A kind of multi-algorithms ensemble prediction method is proposed.The theory of multi-algorithms ensemble and segmental prediction are analyzed firstly.In order to deal with the limitation of single prediction method for LER characteristic parameters,a multialgorithms dynamic weights combining ensemble prediction Method(MDWCE)is proposed based on AdaBoost.RT framework secondly,and the training method of MDWCE is given.Finally,a online multi-algorithms ensemble prediction method(Online-MEP)is proposed to deal wit the online prediction problem.The MDWCE and Online-MEP are verified by the engine ground test data respectively.The results show that the difficulty of establishing the prediction model can be reduced by MDWCE,and it has strong operability and satisfactory accuracy for LER characteristic parameters prediction.The results also show that Online-MEP can obtain the online prediction model satisfying the practical situation efficiently and fastly.Due to the lack of LRE data sample in some situations,the single online prediction method can not predict the LER characteristic parameters satisfactorily.However,the above problem can be solved by Online-MEP effectively,and Online-MEP has a fast speed to establish the prediction model and predict the variation trend of LER characteristic parameters.Based on the engineering requirements of LRE health monitoring,this dissertation has studied the fault prediction methods based on PNN.The results obtained will be helpful for the fault prediction of China's current large-scale LRE and next generation reusable LRE.It can be sure that the achievements obtained have important engineering application values.This dissertation summarizes the fault prediction problem essence of PNN from the LRE operational engineering practice,abstracts characteristic parameters prediction models,and proposes efficient predicting approaches.These research works and results have some theoretical significance.