Selection and validation of predictive regression and neural network models based on designed experiments

Model selection and validation are critical in predicting the performance of manufacturing processes. The correct selection of variables minimizes the model mismatch error whereas the selection of suitable models reduces the model estimation error. Models are validated to minimize the model prediction error. In this paper, the relevant literature is reviewed and a procedure is proposed for the selection and cross-validation of predictive regression analysis and neural network models. Specifications on surface roughness and tolerances impact on manufacturing process plans, and differentiate product quality, and ultimately the product cost and lead times. Experimental data from a turning surface roughness study is used to demonstrate the developed concepts with regression and neural network techniques being used for the purpose of predictive rather than descriptive modeling.

Acknowledgements

The authors are grateful to the reviewers for their insights in revising the paper. This research has been partially funded by the Bradley University Heuser Research Awards grant 25-13-755 and No. 25-13-757 and Caterpillar Fellowship grant 25-11-154 awarded to Jack Feng. We are grateful to Professor Alan Miller for his valuable comments on the first draft and for verifying our models with his Fortran codes, some of which are located at http://users.bigpond.net.au/amiller.