Variable selection based on random vector functional-link in soft sensor modeling

In soft sensor applications, the prediction using only relevant variables significantly improves model accuracy and decreases computational costs. This paper proposed a new method for variable selection based on random vector functional-link (RVFL) neural network model. This method removes input nodes from variable set according to an exclusion criterion by backward selection. Then the remaining weights are adjusted by keeping network output unchanged instead of retraining the network. Finally, the algorithm outputs a set containing the input variables which are ordered in a selection rank. Different methods are applied to several datasets. The results validates that the proposed method selects the lowest number of variables and achieves the satisfactory performance.