Analysis of an unreplicated 22 factorial experiment performed in a continuous process

Abstract

This paper presents a tentative analysis method for unreplicated factorial designs where regular statistical experimental analysis cannot be used. The methodology is demonstrated through the analysis of an unreplicated two-level, two-factor factorial experiment performed in a continuous production process where the process was not in statistical control and where changes in the experimental design made conventional experimental analysis impossible. The first step of the analyses included screening of the sampled data. Principal component analysis and factor analysis were then used to create an overview of how the various responses and experimental factors were related. Carbon monoxide efficiency was selected as the most important parameter to be analysed further. Elastic net regression was used as a screening tool to remove non-significant factors, interaction, and covariates. Finally, the carbon monoxide efficiency variation was modelled using an intervention analysis. Two experimental factors were found to actively influence the response. The experiment that from other perspectives can be considered to be unanalysable, did thus reveal causal effects. The results imply that for processes where the process dynamics may be monitored, observations of the process dynamics may reduce the needs for repeated experimental runs, thus reducing the experimental costs.