Robust Parameter Design for Non-normal Response Based on Bayesian Ensemble Modeling
In complex products or advanced manufacturing processes,there are various uncertainties affecting the design of product quality that adversely affect the accuracy of the model,leading to poor parameter design.Therefore,a new Bayesian model averaging method for generalized linear models was proposed to solve the model uncertainty problem of non-normal responses for the robust parameter design problem with model uncertainty.Firstly,the variable indicators and model indicators were incorporated in the framework of the Bayesian generalized linear model to identify significant effects using the factorial effect principle.Secondly,the posterior probabilities of the variable indicators and model indicators were calculated by Bayesian sampling techniques to determine the model weights and model form,and simulated response values were obtained using a simulation procedure.Then,the quality loss function was constructed based on the simulated response values and the optimization scheme was determined in order to find the optimal solution.Then,the quality loss function was constructed based on the simulated response values and the optimization scheme was determined in order to find the optimal solution.Finally,the effectiveness of the method was demonstrated by practical micro-nano drilling and 3D printing cases.The results show that the proposed method in robust parameter design for solving model uncertainty under non-normal response has lower quality loss and is more reasonably robust than existing methods.
model uncertaintyrobust parameter designgeneralized linear modelsBayesian model averagingfactorial effect principle