Research On Bearing Remaining Useful Life Prediction Method Based On Improved Gaussian Mix Model And Optimal Relevance Vector Machine

The prognostics and health management(PHM)of mechanical equipment is the key technology to guarantee the self-service support and operation cost reduction for the equipment,its usage can manage key machine components and machines so as to ensure the safe and reliable potation of the whole equipment.Rolling element bearing is one of key mechanical components in the complex equipment and its degradation cycle is long and failure rate is random,which leads its safety operation always main concern in industry.The failure of the bearing in service may leads to the stop of the machine,even causes serious production accidents.Therefore,accurate assessment for the performance degradation of the bearing will provide scientific guidance for service strategy and the management of spare parts.Recently,the development of computer technology provides a technical basis for the PHM of bearings,in which the data-driven method is favored by researchers and engineers because of its low cost and easy realization.On the basis of literature review and trend analysis,this thesis starts from the data-driven methods and focuses on three contents,including clustering analysis,health indicator construction,and prediction model for the remaining useful life(RUL).Main aims of this research are to improve the performance of the health indicator and prediction accuracy of the prediction model.Main research methods and contributions are summarized as follows:First,for the Gaussian mixture model,the true distribution of the data is not considered.An improved Gaussian mixture model(GMM)is proposed to process a large scale of bearing data and accurately describe the bearing degradation process.It is well known that the use of single feature is not enough to accurately follow the bearing degradation trend.In the improved GMM,the multi-domain features are extracted from the bearing vibration data and can describe the bearing degradation comprehensively and accurately.Using these features,an improved Gaussian mixture model is constructed to improve its clustering performance.In fact,real data are usually located or near the subflow of the ring space and the corresponding classic GMM has the advantage of nonlinear multi-modal modeling,but this model does not consider the actual distribution of the data.In order to solve this problem,the local geometry of the near-neighbor map approximate data is selected in the improved GMM,and then the Hellinger distance is introduced to measure the distribution of data.After that,the regularization term is defined to optimize parameters of the model.The results of the simulation and experiments verified that the proposed GMM has a better discriminative ability for unlabeled data.Second,a new health indicator is constructed on the basis of the improved GMM.An ideal health indicator not only describes the monotony and consistent degradation trend of bearing,but also accurately identifies the initial failure that can be used as early warning and further is useful for accurately assessing the bearing RUL.In order to facilitate subsequent maintenance management,a new HI is defined to easily track different degradation stages.The Jensen-Rényi divergence(JRD)is firstly introduced to the improved GMM and then fuses the probability output of the improved GMM.Then,the tanh function is used to normalize the solution result to obtain the health indicator.The results of simulated and experimental data sets demonstrate that the HI not only identifies the bearing initial failure accurately,but also can provide a reference for the prediction of the remaining useful life of the bearing..Finally,a new RUL prediction model is constructed.The RUL of bearings is to determine the time interval from the current inspection moment to the bearing failure threshold,which is an important support for the transition from after-the-fact maintenance to preventive maintenance.Most of prediction models are used for short-term prediction and cannot satisfy the application requirement for the long-term prediction.Therefore,a RUL prediction model based on the optimal relevance vector machine(RVM)is proposed for long-term prediction.The constructed health index first obtains the characteristic matrix by using the phase spatial reconstruction to reduce the influence of errors caused by the long-term prediction.Then,a multi-exponential degradation model is built to improve the dynamic tracking ability of the model,and the optimal relevance vectors are extrapolated to the failure threshold so as to realize the long-term prediction of the bearing.The simulation and experimental results verify the effectiveness of the proposed methods in this thesis,and the comparison with other methods also indicates that these methods are applicable and accurate.The prediction results of some experiments demonstrate the proposed methods has better performance for the long-term prediction.Furthermore,these methods and results can be used for the research of bearing RUL prediction,meanwhile,are provide valuable reference and appliable technology for datadriven prediction and PHM technology of other key machine components and machines.