Research On Methods For Data-driven Life Prediction Of Electrical Product

As products' performance will degrade throughout their life span until unit failure happens at which point the product will have to be replaced.Condition-based monitoring is used to develop product residual life predictions which are based on the running performance of the product.By analyzing a product's lifespan and performance with condition based monitoring,the residual life prediction of the product can be calculated.Condition based monitoring is essential to the development of corresponding product replacement and timely system operational maintenance schedules.This thesis develops and extends two main data driven methods for residual life prediction through theoretical investigation and instance analysis by focusing on lifespan prediction of electrical components.A study of similarity measurement between reference sample and test sample using an extended similarity measure algorithm was the first of the two proposed methods.The algorithm was based on slope distance and variable weighted Euclidean distance.It combined the advantages of slope distance and variable weighted Euclidean distance,and more weight was assigned to a product's most recent status than to its former status.A comparative analysis of electromagnetic relays data indicated the superiority of the extended similarity measure algorithm for residual life prediction.Proportional hazards model(PHM)was the second method of residual life prediction.The implementation of PHM for electrical product life prediction can be based on a time independent covariate or a time dependent covariate.The time independent covariate PHM of electrical appliances was established by focusing on the effect of different environmental stress' on characteristic parameter and mean life of electrical appliances.Time dependent covariate PHM described the product life prediction model by recognizing chaos characteristic in the covariate time series.Parameter estimation strategy was described in detail in the process of maximum likelihood estimation method.The two methods of residual life prediction were used to study the lifespans of electromagnetic relays.The comparison of the results indicated that both the characteristic life and mean life decrease significantly with the increase in ambient temperature.One of the benefits of similarity based modelling is that a recession model is not necessary which helps to remove a level of difficulty from dealing with the data as well as the large prediction errors that common with recession models.A product's running state can be combined with covariate monitoring information in the PHM which can produce a small error relatively.Both methods possess their own merits,which improve upon and supplement traditional reliability approaches.Even though electromagnetic relays were the only component used for this study,the methods used to study their product residual life can be applied to other electronic componets for reliability study.