A Novel Variable Selection Algorithm for Soft Sensor Using Gated Recurrent Unit and Garrote Variable Attention Mechanism

This paper proposed a variable selection algorithm for soft sensor applications using the garrote variable attention mechanism and gated recurrent unit (GRU). Firstly, a well-trained GRU deep neural network is obtained by hyperparameter optimization. Secondly, a set of variable attention coefficients was designed based on the maximal information coefficient and embedded in the nonnegative garrote (NNG) algorithm constraints to overcome the disadvantage of biased coefficient estimation in this algorithm. Finally, the GRU network input weights are sparsely optimized using the optimized NNG algorithm to eliminate redundant variables and simplify the model structure. A numerical simulation case verifies the effectiveness of the proposed algorithm.