Input variables selection criteria for data-driven Soft Sensors design

In this paper the design of a Soft Sensor to estimate the sulphur concentration in a desulphuring unit of a refinery operating in Sicily is described. In particular the problem of the input variables selection is addressed by comparing two different methods. The first method is based on the generalization of the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm to nonlinear models implemented by using Multi-Layer Perceptron (MLP) neural networks. The second one is based on the Lipschitz's quotient analysis. A comparison between the performance and the computational complexity exhibited by the two methods is discussed. The results show that the LASSO-MLP algorithm allows to construct a model with a low number of input variables, thus reducing computational complexity and measuring costs.