Robust Parameter Design Based On Nonparametric And Semi-parametric Methods

With the increasing improvement of people's living standard,customers have higher requirements on product quality.If any enterprise wants to stand out in the fierce competition,it must improve the quality and reduce the cost.The effect of improving product quality can be achieved through robust parameter design without adding extra cost.However,in many cases,the assumed response surface model structure does not conform to the reality,which will lead to the failure to obtain reliable research results,while the non-parametric method can deal with this problem well.However,nonparametric methods tend to rely on relatively large samples.In addition,most researchers mainly study normal response distribution data.For non-normal response data,the traditional treatment method is generalized Linear Models(GLM).However,for small sample test data,this method will lead to high uncertainty of model parameters in the modeling process.To solve these problems,in this paper,considering the model structure uncertainty problems,design of random error and the non-normal responses more robust parameter design problems,and combined with Bayesian sampling technology,hybrid genetic algorithm and LASSO variable screening technology,through the system modeling(response surface modeling,the generalized additive models-generalized additive models,GAM)and case study to verify the reliability of the proposed method.The specific research contents are as follows:(1)Multi-response optimization design combined with non-parametric method and Bayesian sampling technique.In this chapter,a non-parametric method is proposed to fit related multi-response problems.First,a multi-response surface model considering random fluctuations is constructed by combining non-parametric technology and Bayesian method.Then,the expected mass loss function is used as the objective function to optimize the parameters by hybrid genetic algorithm.Finally,Bayesian sampling trajectory diagram is used to demonstrate the reliability of model prediction.(2)Multi-response optimization design combining semi-parametric method and Bayesian sampling technique.The semi-parametric method can solve the problems of uncertain model structure and small sample size of test data.Half-parameter is composed of parameter part and non-parameter part.For the parameter part,classical least square OLS is usually used for parameter estimation,while for the non-parameter part,LLR is used for local linear regression estimation.Bayesian sampling is used to design the random error term,the multi-response surface model is constructed and the expected mass loss function is taken as the objective function.Consider the predictive response volatility,using the reliability problem of MCMC research results.(3)Multi-response optimization design combined with semi-parametric GAM model and Bayesian sampling technique.For the processing of non-normal data,the traditional methods have some errors,the traditional normal conversion methods will have errors in the transformation process,the generalized linear model method will have some errors,but for the small sample test data,the uncertainty degree of model parameters in the modeling process will be high.Semiparametric generalized additive model is put forward to deal with non-normal multi-response optimization problem,it integrated the semiparametric regression model based on the data of the flexibility of the characteristics and the generalized linear model to deal with the problem of non-normal responses to highlight the advantages,the research of this article the non-normal data robust parameter design problems,there are still consider to correct error.