Performance of soft sensors based on stochastic configuration networks with nonnegative garrote

In this study, stochastic configuration networks (SCNs) and nonnegative garrote (NNG) algorithm are employed to develop a soft-sensing technique that infers difficult-to-measure variables with easy-to-measure variables in industrial processes. The proposed method consists of two stages, that is, performing industrial data modeling with SCNs and applying NNG algorithm for shrinking input weights and removing some redundant input variables from the well-trained leaner model. Cross-validation and the Akaike information criterion are employed to determine the optimal shrinkage parameter for the NNG. A numerical example and real industrial data are used to validate the performance of the proposed algorithm. Several state-of-the-art feature selection schemes for neural networks are tested. Comparative results demonstrate that the proposed soft-sensor outperforms others in terms of the prediction accuracy.

Automated summary

This study develops a soft-sensing technique using SCNs and NNG algorithm to infer difficult-to-measure variables with easy-to-measure variables in industrial processes. The proposed method consists of two stages: using SCNs for industrial data modeling and applying NNG algorithm for model optimization. Experimental results demonstrate that the proposed soft-sensor performs better in terms of prediction accuracy.