Using Random Forest to Interpret Out-of-Control Signals

Statistical quality control procedures have become essential practices to ensure competitiveness in any manufacturing process. Since the quality of manufactured goods usually depends on several correlated characteristics, statistical multivariate techniques are needed to detect and analyze out-of-control situations. The difficulties in the interpretation of those out-of-control observations in multivariate control charts have motivated the development of different techniques in order to determine the variable or variables that have motivated the changes in the process and, in case of more than one variable as responsible of the change, to evaluate their contribution. Specifically, these techniques are mainly based in two alternatives, one that considers the T-2 decomposition and other related to the application of classification techniques. The application of this latest techniques includes increasingly sophisticated methods, being the most usual alternative based on the application of Artificial Neural Networks. In this paper, we propose Random Forest as a powerful classification technique in statistical process control, considering a wide range of different situations in the function of the type of change and the magnitude of the correlation coefficient between variables. Moreover, the performance of Random Forest is analyzed in comparison with the results obtained from the application of Artificial Neural Networks to try to find out in which cases the superiority of Random Forest can be supported.