Hybrid Variable Selection and Support Vector Regression for Gas Sensor Optimization

The improvement of combustion processes in industry, especially in the automotive branch, is of great importance to maintain the environmental permitted limits. Carbon monoxide concentration in the exhaust gases can give an insight into the efficiency of the combustion taking place and for this reason, it is important to have sensors that can measure it accurately. First results of a long term study with one of the leading sensor manufactures showed high performance using genetic programming. However, this expensive approach is difficult to apply in real-world settings. Therefore a hybrid optimization that combines support vector regression (SVR) with variable pre-selection is proposed. Three different methods for variable selection are compared for this application, a genetic algorithm, and two methods from Bayesian statistics: statistical equivalent signatures and projection predictive variable selection. Furthermore, a multi-objective approach using the same hybrid definition is implemented for the cases in which several sensors need to be considered simultaneously. Our results show that the hybrid model is an improvement compared to the previous study, while delivering good performance when dealing with a multivariate formulation. Genetic algorithms in combination with SVR lead to enhanced variation on the groups of selected variables.