A Hybrid Approach Variable Selection Algorithm Based on Mutual Information for Data-Driven Industrial Soft-Sensor Applications

The development of virtual sensors predicting the desired output requires a careful selection of input variables for model construction. In an industrial environment, datasets contain many instrumentation system measures; however, these variables are often non-relevant or excessive information. This paper proposes a variable selection algorithm based on mutual information examination, redundancy analysis, and variable reduction for soft-sensor modeling. A relevance calculation is performed in the first stage to select important variables using the mutual information criterion. Then, the detection and exclusion of redundant variables are carried out, penalizing undesired variables. Finally, the most relevant variables subset is determined through a wrapper method using Mallowssans' Cp metric to assess the fitting prediction performance. The approach was successfully applied to estimate the ethanol concentration for a distillation column process using an adaptive network-based fuzzy inference system architecture as a non-linear dynamic regression model. A comparative study was performed considering the application of correlation analysis and the method proposed in this study. Simulation results show the effectiveness of the proposed approach in the variable selection providing a reduction in search of suitable models that achieve faster results for developing soft sensors oriented to industrial applications.