Resumen en inglés

In statistical process control, Process Capability Indices (PCI) are a tool to quantify process diversity and analyze this diversity in relation to product requirements or specifications. The validity of the results of the PCI depends on the measures being taken from processes under control, and that the assumptions of normality and independence are fulfilled. In practice, it is difficult to satisfy these assumptions, and that is why several alternative methods have been proposed, such as the Clements percentile method, the Box-Cox potential transformation method, the Burr percentile method, and recently Machine Learning methods. Abbasi (2009) proposes to estimate the process capacity index using an artificial neural network (ANN) model, but does not consider the selection process of the input variables of the neural network, nor the optimal construction. of the network typology and, in addition, distributional assumptions are assumed. In this paper, an PCI estimation methodology is established using ANN classificationtype models, which allows categorizing the state of a process, considering different selection methods for the most relevant input variables, as well as the optimal configuration of the network. Using experimental data from non-normal distributions, the effect of different input variable selection methods on the performance of the ANN model is evaluated and compared. It is concluded that the Pearson selection method almost always provides simpler network structures than those provided by the Backward and Gedeon methods.