Abstract: Using deterministic values of input variables is desirable for process design. However, some of these input variables may present uncertainty, which may drive the designed process to unwanted responses and, consequently, generating large economic damages. This manuscript proposes a methodology for avoiding the scenario earlier described. The methodology considers three steps: (1) deterministic process design, (2) elimination of non-influential input variables using global sensitivity analysis, and (3) classification of the influential input variables using least squares support vector machines (LS-SVM) classifier, whose parameters are tuned through particle swarm optimization (PSO). The proposed methodology was applied in the design of mineral concentration circuits. The results show that the elimination of non-influential input variables from training data helps to improve the accuracy and to prevent the overfitting of LS-SVM classifier. The methodology allows classifying input variables and knowing what combinations will drive the designed process to unwanted conditions. Thus, the proposed methodology could be useful for fault detection and diagnosis in large size processes operating under uncertainty.

Keywords: Least squares support vector machines, Particles swarm optimization, Hybrid kernel, Global sensitivity analysis, Process design, Flotation, Fault detection