Research On Parameter Estimation Of Reliability Models On Heavily Censored Data

A common method for predicting product reliability is to use parameter estimation methods to fit the life data of key product components to a specific distribution model(such as Weibull distribution),and then predict product reliability based on the feature quantities of the model.In this process,parameter estimation is one of the key links,and the accuracy of parameter estimation directly affects the accuracy of reliability prediction.In real life,the actual collected life data is often heavily censored due to the fact that many observation systems are required on site and the observation time is relatively short relative to the average life of the product.For this type of data,classical parameter estimation methods cannot provide robust estimates.Therefore,attempts need to be made to develop new parameter estimation methods.This paper mainly focuses on three novel parameter estimation methods.The specific research contents are as follows:(1)Under the condition of heavily censored data,a large-scale simulation numerical experiment is carried out by Monte Carlo method,and three novel parameter estimation methods are compared based on this numerical experiment.The comparison includes four aspects:accuracy,robustness,simplicity and applicability.The results show that the overall performance of the folded auxiliary-model-based method is better than the other two methods.(2)Because of the heavily censored data conditions,methods based on lognormal distribution and gamma distribution are limited.Attempts to apply folded auxiliary-modelbased method to lognormal and gamma distributions and validate with a large number of simulated datasets.The results show that the folded auxiliary-model-based method can provide accurate and robust parameter estimates for lognormal distribution and gamma distribution.(3)Considering the skewness characteristic of Weibull distribution,a parameter estimation method with Weibull distribution as an auxiliary model is developed,and a large number of simulation datasets are used to confirm that this method can provide accurate and robust estimation of gamma distribution.In addition,this method is compared with folded auxiliary-model-based method.The results show that the accuracy and robustness of the two are basically the same,but the newly developed method is simpler than the auxiliary modelbased method.