Access Type

Open Access Dissertation

Date of Award

January 2015

Degree Type


Degree Name



Industrial and Manufacturing Engineering

First Advisor

Kyoung-Yun Kim


Numerous and various research projects have been conducted to utilize historical manufacturing process data in product design. These manufacturing process data often contain data inconsistencies, and it causes challenges in extracting useful information from the data. In resistance spot welding (RSW), data inconsistency is a well-known issue. In general, such inconsistent data are treated as noise data and removed from the original dataset before conducting analyses or constructing prediction models. This may not be desirable for every design and manufacturing applications since every data can contain important information to further explain the process. In this research, we propose a prediction modeling framework, which employs bootstrap aggregating (bagging) with support vector regression (SVR) as the base learning algorithm to improve the prediction accuracy on such noisy data. Optimal hyper-parameters for SVR are selected by particle swarm optimization (PSO) with meta-modeling. Constructing bagging models require


more computational costs than a single model. Also, evolutionary computation algorithms, such as PSO, generally require a large number of candidate solution evaluations to achieve quality solutions. These two requirements greatly increase the overall computational cost in constructing effective bagging SVR models. Meta-modeling can be employed to reduce the computational cost when the fitness or constraints functions are associated with computationally expensive tasks or analyses. In our case, the objective function is associated with constructing bagging SVR models with candidate sets of hyper-parameters. Therefore, in regards to PSO, a large number of bagging SVR models have to be constructed and evaluated, which is computationally expensive. The meta-modeling approach, called MUGPSO, developed in this research assists PSO in evaluating these candidate solutions (i.e., sets of hyper-parameters). MUGPSO approximates the fitness function of candidate solutions. Through this method, the numbers of real fitness function evaluations (i.e., constructing bagging SVR models) are reduced, which also reduces the overall computational costs. Using the Meta2 framework, one can expect an improvement in the prediction accuracy with reduced computational time. Experiments are conducted on three artificially generated noisy datasets and a real RSW quality dataset. The results indicate that Meta2 is capable of providing promising solutions with noticeably reduced computational costs.