Research Of Remaining Useful Life Prediction For Products Based On Machine Learning

With the fusion of information technology and industrial systems,the data acquired by distributed sensor networks,along with other informational data,like data for internet of things and cross-border data,emerge as the so-called industrial big data deemed as the core power for new generation of industry evolution.For modern engineered systems,it is necessary to monitor the key equipment and execute predictive control.The collected data is analyzed intelligently to assess and predict the system health states,reducing the operational risk of equipment and system.Besides the predictive maintenance can be realized to optimize the usage,maintenance and support strategies,known as the prognostics and health management(PHM).Remaining useful life(RUL)estimation is the core technique of PHM system,however,it is of great difficulty to output a credible estimation with high accuracy.This might be caused by the following factors: complexity of failure mechanisms,uncertainty of sensory data,unpredictability of future operational conditions,and etc.Especially as a lack of “Run-to-Failure(RtF)” data for practical system,it is hard to assess the RUL estimation methods.Confronted with the challenges of PHM,the data fusion and ensembles of methods/algorithms are two parallel approaches.As a pivotal of artificial intelligence(AI),machine learning(ML)contains abundant data modeling algorithms.ML is suitable for health diagnosis and RUL prediction for its high capacity of selflearning and fusion analysis.This dissertation researched the methodology of RUL estimation with ML systematically,the applied classification and regression algorithms are Hierachical Clusetring(HC)and Relevance Vector Machine(RVM)respectively.The issuses involve the health index(HI)synthetization by multi-dimensional sensory data,health state classification and diagnosis,performance and RUL estimation.The research focused on the improvement of RUL estimation accuracy and uncertainty analysis of the predictions.The main contents are as follows:(1)For solving the problem that how to transform multi-dimensional sensory data into onedimensional health index,a health index synthetization method based on HC and similarity analysis was proposed.The original sensory data was firstly preprocessed for further analysis by unsupervised clustering learning algorithms.Then the entire RtF data can be divided into four main clusters,corresponding to four health states,i.e.healthy,sub-healthy,degraded and failure states.Taking the advantage of the HC,the main cluster can be further divided into smaller clusters according to data feature and practical demand.Based on similarity analysis between data points and the clusters,we can diagnose the health state of the system.Finally the distances between the data points with failure class center(“failure baseline”)can be calculated to generate the onedimensional HI.The case studies show that the synthesized health indices(SHI)were highly correlated with the input sensory data.It means that SHI can integrate the information of each dimension of data,meanwhile preserve the inherent degradation feature of the system.(2)For solving the problem of RUL estimation when there exist enough RtF samples for training,based on model-matching concept,we proposed an improved model matching(IMM)method with the fusion of multi degradation models.Based on SHI,the degradation model was built with the relevance vector regression(RVR)algorithm,multi similar degradation models were selected for RUL estimation so as to improve the estimation accuracy.Furthermore,the “tails” of degradation were cut which exceed the center of failure cluster so that the predicted RULs are safer for maintenance decision.Finally,the uncertainty anlysis can be accomplished as we can acquire the samples of RUL estimation for one testing sample.Through the case study and comparisons,the proposed method was demonstrated to have a great advantage to output more accurate RUL estimation results.(3)In most of the practical scenarios,the RUL is estimated on basis of partial degradation data,so that the sole approach for RUL estimation is building the partial degradation model and extrapolating the model to the failure threshold.If the basic RVR model is extrapolated with multisteps,the predicted results will deviated from the true degradation trajectory severely,which can not satisfy the requirement of long-term prediction for health status and RUL.To overcome this shortcoming and preserving the advantages of RVM method,a modified RVR model was created by adding one or more column vectors representing the overall degradation pattern into the design matrix.This is denoted as RVR-NDM method.Besides how to adjust the model parameters in real scenario was also researched to obtain an optimized prediction model.For demonstration,two cases with different overall degradation patterns were studied.The results showed that the RVRNDM has an obvious advantage over basic RVR and Generalized Linear Regression(GLR).(4)For a type of Sterling cryocooler practically used in space,applying the proposed methods to predict the performace and RUL.Firstly,the health states for cryocooler were classified into four classes.So that the prognostic time can be determined for long-term performance and RUL predicitons by three different methods,i.e.RVR,GLR and RVR-NDM.Then the basic RVR model was extrapolated with one-step to accomplish the short-term performance predcitons.For the ground testing datasets of the rack thermal control drawer onboard space station,experiments of working condition classification and health index synthesization were executed.Through practical engineering cases,it is demonstrated that the effectiveness of the proposed methods.