Multi-response non-parametric profiling using Taguchi's qualimetric engineering and neurocomputing methods: Screening a foaming process in a solar collector assembly

• Profiling of seven controlling factors for a solar-panel foaming operation.
• Three foaming characteristics are considered.
• Multi-factorial multi-responses are ranked and fused to a single master response with non-parametric transformations.
• Concurrent distribution-free optimization is effected through ANN linear and perceptron methods.
• Foam polyol temperature is found to be the predominant active effect.

Taguchi's data programming techniques in synergy with data analysis tactics based on artificial neural networks have been fruitful in illuminating intricate manufacturing phenomena. We present a non-parametric approach to treat multi-response multi-factorial datasets created with Taguchi's orthogonal-array samplers. Replicated response datasets are compressed utilizing the signal to noise ratio and then they are homogenized by simple rank-ordering. The multiple response layout is reduced to the more tractable uni-response arrangement by using the super-ranking concept to enact the fusing of the individual responses. The ‘ranked-and-fused’ dataset is subjected to conversion by linear and three-layer perceptrons. The performance of a group of examined effects is assessed according to the perceptrons’ sensitivity analysis output. Using Wilcoxon's one-sample test, statistical significance is assigned to the accumulated ranking scores obtainable from a series of independent perceptron runs. We discuss the efficiency status for each of the two engaged perceptron options on affecting prediction accuracy as well as the influence of data fusion on the SNR-compressed datasets. Our robust neurocomputing solver is elucidated in a concurrent screening of three foam characteristics which are encountered in popular solar-collector assembly operations. Seven controlling factors were profiled and it was found that the temperature of the polyol additive is the sole statistically predominant effect. Finally, through our industrial paradigm, we illustrate the superiority of the fusing principle for downsizing stochastically multiple characteristics and thus gaining faster and more accurate perceptron predictions. We show that the proposed method outperforms the outcomes obtained by the desirability analysis. We identify the points of superiority to the crisper resolution in locating effect dominance accompanied with recovered stochastic significance.

Figure optionsDownload as PowerPoint slide