Novel Functional Link Neural Network with Pearson Correlation Coefficient: Application to Soft Sensing

Nowadays, with the expansion of the dimension and the scale of chemical industry, building an accurate soft sensor model becomes more difficult. Fortunately, Functional Link Neural Network (FLNN) has proven to be a dependable model for soft sensing and has been successfully implemented. Traditional FLNN ignores the fact that the input attributes have different correlations and go through function expansion blocks as a whole, which may lead to modeling accuracy can not meet the requirements. To solve this problem, a novel functional link neural network with Pearson correlation coefficient (PCC-FLNN) is proposed in this paper. The input attributes are categorized based on their Pearson correlation coefficients, with one group having positive coefficients and the other group having negative coefficients. After function expansion, these two groups of input attributes create two separate subnetworks. The proposed method has a remarkable feature: it is able to improve the modeling accuracy without increasing the training parameters. Both the UCI standard dataset and the pure terephthalic acid (PTA) process dataset are used to evaluate the efficacy of the proposed method. The findings indicate that the PCC-FLNN outperforms the FLNN in accuracy.