面向多极值质量特性的全局式序贯性实验设计方法

针对多极值质量特性的全局性建模和参数优化问题, 为降低实验设计样本量, 提高模型预测性能, 提出一种全局式序贯性设计方法. 首先在一定的初始实验设计方式下, 建立过程粗略的SVR模型; 而后根据模型中支持向量的分布, 在支持向量样本各维度的45度角或轴向方向同步地增加实验点, 再拟合过程新的SVR模型; 如此迭代进行, 当模型精度达到要求或样本量达到上限时终止序贯性设计. 仿真与实证研究表明, 该方法能够在可行域全局的范围内将实验点合理地分配在质量特性的多个极值附近, 避免了传统的单路径式序贯性设计只能发现单个极值的不足, 充分提高了实验效率; 与均匀空间网格设计、 拉丁超立方设计和均匀设计相比, 在样本量接近的情况下, 基于全局式序贯性设计的SVR 模型的预测均方误差至少降低了30%; 而在预测误差较为接近的情况下, 全局式序贯性设计的样本量至少降低了12%.

Global sequential design for multi-extreme quality characteristics process

For the global modeling and parameter optimization of manufacturing process featured with multi-extreme quality characteristics,a global sequential design(GSD) approach is proposed to reduce the experimental sample size and improve the predictive performance as well.Firstly,the proposed approach sets up an elementary model of the process using support vector regression(SVR) under a certain kind of initial experiment manner.Secondly,after identifying support vectors of the initial experiment sample, new experiments are added simultaneously at the 45 degree or axial direction of support vectors in each dimension of process parameters.Thereafter a sequential model of the process is set up.Finally,the sequential steps mentioned above are iterated until the lower prediction mean square error(MSE) or the upper sample size acceptable is reached.The simulation and empirical study show that,GSD approach can rationally allocate the experimental runs around the multi-extreme quality characteristics within the whole feasible region of process parameters,which overcome the shortcoming of only finding single extreme using traditional line steepest ascent sequential design and therefore improve the experimental efficiency. Furthermore,compared with uniform grid design,Latin hypercube sampling and uniform design manner, under the same sample size,the prediction MSE of GSD based SVR model decreases 30%at least;while under the same prediction MSE of SVR model,the sample size of GSD decreased 12.0%at least.