MODIFIED RADIAL BASIS FUNCTION NEURAL NETWORK INTEGRATED WITH MULTIPLE REGRESSION ANALYSIS AND ITS APPLICATION IN THE CHEMICAL INDUSTRY PROCESSES

The construct of a radial basis function neural network (RBFNN) plays an important role in predicting performance. However, determining the optimal construct is difficult. A modified RBFNN integrated with correlation pruning algorithm-least squares regression (CPA-LSR) was proposed to optimize the number of hidden neurons as well as the weights and bias. First, an initial RBFNN was built by superposing each center to a training set point. This RBFNN was then trained. Next, CPA-LSR was applied to eliminate the redundant information of the initial network and to improve the predicting performance by optimizing the structure as well as the weights and bias. Finally, the developed naphtha dry point soft sensor and the industrial oxidation of p-xylene to terephthalic acid were employed to illustrate the performances of the modified RBFNNs. The result reveals an improvement in the predicting performances of the RBFNNs integrated with CPA-LSR. Conclusively, RBFNNs integrated with CPA-LSR are recommended because the redundant neurons are effectively eliminated, and the optimal structure of the RBFNN is obtained by CPA-LSR. Moreover, such RBFNNs are intuitive and eliminate the need for parameterization.