The Improvement And Application Of Neural Network With Eliminating Hidden Layer Redundant Information

The number of neurons in hidden layer directly affects the performance and efficiency of the entire network in neural network modeling. The problem of the network structure is described as follows:insufficient number of hidden units may lead the network unable to solve the problem it is supposed to tackle. On the other hand, when too many hidden units are used, it may lead to spurious decision boundaries or may cause over-fitting of the model and poor interpolation. Considering the redundant information produced in hidden-layer, the CPA-MLR (correlation pruning algorithm combined with multiple linear regression) is proposed to eliminate it, so as to simplify the network structure and improve the prediction performance. This method is also applied to develop naphtha dry point soft sensor, and the results show that the method can eliminate redundant information effectively, and the network after optimization is of high precision. Specific content as follows:(1) The CPA-MLR method is used to optimize the BP network structure. Firstly, an initial three layers network with the maximum nodes of hidden-layer is selected, and correlation analysis of the hidden-layer output is carried out to confirm the redundant hidden nodes. Then the redundant nodes will be deleted one by one, and a multiple linear regression model between the output of the hidden-layer and the expected input of the output-layer, which can be obtained through the inverse function of the output-layer node, is employed to obtain their optimal weight. Further, a practical example, i.e. developing naphtha dry point soft sensor, is employed to illustrate the performance of CPA-MLR. The results show that the predicting performance of the soft sensor is improved and then decreased with deleting the redundant nodes.(2) CPA-MLR, PCR and PLSR are applied to eliminate redundant information of RBF network and improve the predicting performance through optimizing the structure and the weights and bias. And developing naphtha dry point soft sensor is employed to illustrate the performances of three modified RBFN. The result shows that the predicting performances of RBFN integrated with CPA-LSR, PCR and PLSR are essentially identical. Finally, based on the related results of the soft sensors, the internal relations between PCR and PLSR is revealed from the loading vector of PLS/PCs and the weights of the output layer of the network.