Multi-response Robust Parameter Design Based On Bayesian Approach

As an important technology of continuous quality improvement, robust design optimization is a way to deal with the problems of uncertainty optimization, and it can reduce and control the generation of volatility from the source. In the early design stage of product design or craft process, application of robust parameter design can improve the quality of the product or process effectively. The traditional robust design is mainly used in the single response system, but in the current society, people's demands tend to be diversified, so the products which have multiple quality characteristics gradually become the mainstream in the competition market. Accordingly, in the field of quality research, the multi-response robust design is gradually become the main trend of the quality design.In the general multi-response robust optimization, usually involves a series of problems like robust measurement of multi-variant process, the tradeoff among multiple responses, model uncertainty, the measurement of optimal results in reliability and so on. But the traditional methods usually just consider one aspect or several aspects among these problems.For above problems, in this paper, taking the multi-response system as the research background, we systematically study the problems of model building which consider both robust and reliability based on desirability function, multivariate quality loss function, Bayesian posterior probability method, multi-objective optimization techniques, heuristic optimization algorithm and the empirical analysis, etc. Some main results are summarized as follows:(1) Bayesian modeling and optimization of multi-response based on improved desirability function. Aiming to the problems of importance degree on multiple responses and reliability evaluation, under the frame of the Bayesian statistical analysis, this paper combined the desirability function which considering objective weight with the Bayesian posterior possibility, then build a desirable and robust constrained model. Firstly, this method based on the experimental data to determine the objective weight of each response by entropy weight theory, then according the Bayesian posterior sample to build overall desirability function which considering objective weight. Secondly, through Bayesian posterior probability approach, this method can gain the probability of the posterior sample meet the specification limit, and the probability of not less than a desired target value as a constraint condition. Thirdly, due to the nonlinear constraints, the hybrid genetic algorithm is used to optimize the objective function. At last, through a case to show that the new model not only measures responses objectively, but also ensures the possibility of responses fall within the specification.(2) Multi-response optimization design considering quality loss and conformance probability. Based on the above research, we found that a constraint model needs higher performance computer and longer optimal time in the process of optimization, so it will brings troubles and inconvenience to the researcher. Also in the past study, some researchers point out that the desirability model can't consider the quality loss that responses deviate from the target value, and the traditional desirability model can't resolve the correlation problem among each response. For these problems, based on the above research, this paper combines the multivariate loss function with the Bayesian posteriori possibility, and through the desirability function to normalize the two parts, then, it constructs a new non-desirability function which is a dimensionless robust reliability model with no constraint. At last, through cases study to demonstrate that the proposed method can get a group of robust and reliable parameter solution.In the end, the thesis summarizes the above research results, and also discusses some challenging topics which deserve further research in the future.