Research On Fault Prognosis Methods For Critical Dynamic Mechanical Components Of Helicopter Transmission System

The Prognostics and Health Management(PHM)of helicopter powertrain is very important because it has no redundancy.As a key part of PHM,fault prognosis connects fault diagnosis and degradation state identification,and provides scientific decisionmaking suggestions for subsequent intelligent operation and maintenance of helicopter power train.Therefore,failure prognosis plays an indispensable role in the implementation of PHM work.Scientific fault prediction can effectively improve the safety performance and economic benefits of equipment,and related scholars have also carried out many method studies.However,there are still some problems,such as the limited method to determine the starting point of fault prediction,the performance of the characteristic indexes oriented to fault prediction is not ideal,the demand for expert experience is too large,it is difficult to build the fault prognosis model and the confidence of the fault prognosis results is not high.Therefore,there is still a certain gap between fault prediction and engineering practice.Therefore,in view of the above problems,aiming at the actual needs of helicopter critical component fault prognosis,this dissertation carries out the following specific research work:(1)An early fault diagnosis method based on Health Indicator(HI)amplitude abnormal change detection and Multipoint Optimal Minimum Entropy Deconvolution Adjusted(MOMEDA)fault enhancement is proposed.Aiming at the problem of determining the starting point of fault prediction,the related research work of early fault diagnosis is carried out.Firstly,the HI which can accurately characterize the degradation trend of the research object is obtained.Then,abnormal samples are obtained based on HI amplitude detection.Secondly,fault feature enhancement of abnormal samples is realized based on the MOMEDA method.Finally,early fault diagnosis is realized and the starting point of prediction is determined.(2)A Deep Convolutional Neural Network(DCNN)fault diagnosis method based on explainable training strategy is proposed.As the typical representative of deep learning methods,DCNN can adaptively excavate distributed sample features under big data and has strong universality of fault diagnosis.However,for the problem of fault diagnosis with limited sample tags,DCNN model training becomes a "stumbling block" restricting the wide application of this method.Aiming at the problem of how to train the DCNN model scientifically and improve its interpretability,the fault diagnosis method of DCNN under the guidance of interpretability training is studied.Firstly,the training samples are randomly dispersed and then input into the DCNN training model,and the theoretical basis of the random dispersion of DCNN training samples is mathematically derived to further improve the interpretability of the training process.Secondly,aiming at the problem of determining the number of training iterations of the DCNN model,a scientific method to quantify sample training batches is proposed.(3)A degradation state tracking and recognition method based on 1-Dimensional Deep Convolutional Neural Network(1DDCNN)and Principal Component Analysis(PCA)for information fusion is proposed.Appropriate HI can accurately characterize the degradation trend of fault prediction objects.The traditional HI extraction process relies too much on expert experience and has a large human interference factor.To solve the HI intelligent extraction problem,the HI extraction and degradation state tracking recognition method based on 1DDCNN and PCA information fusion is studied.Firstly,distributed fault features are extracted from degraded data adaptively based on1 DDCNN.Secondly,the distributed feature matrix dimension reduction is realized based on the PCA method,and HI with ideal monotonicity,robustness,and tendency is obtained by fusion to achieve accurate tracking of degradation trends.Finally,degradation state recognition of fault prediction object is realized based on HI amplitude change.(4)A degradation state tracking and recognition method based on Improved Phase Space Warping(IPSW)is proposed.Aiming at the problem of tracking and identifying the degradation state of fault prediction objects under variable working conditions,HI extraction under variable working conditions is studied.Firstly,phase space parameters of sample reconstruction are determined based on comprehensive information entropy.Secondly,based on the phase space warping theory,the phase space warping variables caused by fault changes are obtained.Thirdly,based on the condition correlation analysis method,the HI sensitive to fault change and immune to condition change is weighted.Finally,degradation state recognition of fault prediction object is realized based on HI amplitude change.(5)A fault prediction method based on HI and the dynamic model is proposed.In view of the problem of fault prediction model construction in different degradation stages,relevant research is carried out.Firstly,the comprehensive evaluation of degradation data of fault prediction object is optimized to obtain scientific HI.Secondly,the HI feature points after sparse are obtained based on Relevance Vector Machine(RVM).Thirdly,the parameter optimization of the variable order prediction model is carried out based on sparse samples,and the dynamic prediction model changed with the degradation stage is obtained.Finally,the fault prediction results are obtained by time extrapolation.(6)A fault prediction method based on Trajectory Enhanced Particle Filter(TE-PF)is proposed.Aiming at the problem of constructing a fault prognosis model and evaluating the uncertainty of prediction results,a fault prediction method named TE-PF is proposed.Firstly,from the perspective of degradation rate tracking and degradation trajectory strengthening,a general prediction model for the PF method is constructed.Secondly,the degradation trend information of historical samples and particle generated samples is effectively used to guide the parameter estimation of the general prediction model,and then,the multi-information fusion trajectory enhancement prediction model is obtained.Thirdly,the remaining service life prediction is extrapolated to obtain the fault prediction results.Finally,based on the statistical characteristics of PF particles,the uncertainty of prediction results is evaluated.In conclusion,aiming at the problem of failure prediction of helicopter powertrain components,the early fault diagnosis method based on HI amplitude detection and MOMEDA fault feature enhancement,fault diagnosis method based on interpretation training guidance DCNN under finite tag samples,HI intelligent extraction and degradation state tracking recognition method based on 1DDCNN and PCA information fusion under constant conditions,HI extraction and state tracking identification method based on IPSW under variable conditions,the fault prognosis method based on the HI with dynamic model and the fault prognosis methods based on TE-PF are proposed respectively,to provide methods to the problems of that the starting point of fault prognosis is not clear,the fault prognosis indicator performance is not ideal,expert experience is excess demanded,fault prognosis model building is difficult and failure prognosis results the confidence is not high.The effectiveness of the proposed methods is verified by experiments.