Research And Application Of RBF Neural Network Modeling Method For Complex Chemical Industry

As the result of the development of the computer technology,a variety of artificial intelligence methods(expert system,genetic algorithm,neural network etc.)are widely applied for the simulation of production process as well as prediction models,and have achieved good results.Radial Basis Function(RBF)neural network,which is a simple single-hidden-layer feed-forward neural network with excellent learning performance,which has the advantages of fast implementation and best global approximation performance.Because the complexity of industrial data has characteristics of high dimensions,strong coupling and redundant information,it is difficult to enhance the modeling ability of the RBF and we cannot accurately predict their production status based on the RBF,it is necessary to extract the characteristics of the original data as the network input data to improve the analyze and predict ability of RBF neural network on complex data.However,when the original input data is not easy to do artificial extraction,or the effect of artificial extraction is difficult to guarantee,it is necessary to change the structure of the prediction or classification model,such as increasing the number of hidden layers and constructing the deep network structure.Therefore,this paper proposes two different ways to improve the learning performance of RBF neural network for different task requirements and different types of raw input data.The main research contents are as follows:1.Because the complexity of industrial data has characteristics of high dimensions,strong coupling and redundant information,it is difficult to enhance the modeling ability of the RBF neural network,integrating fuzzy C-Means algorithm(FCM),we propose an improved RBF based on analytic hierarchy process(AHP)(FAHP).The FCM algorithm is used to cluster the input attributes of the high dimensional data,and then the redundant information are filtered using the AHP based on the entropy weight to achieve the feature extraction of original input data.The feasibility and validity of the FAHP algorithm to improve the modeling accuracy of the RBF neural network are verified by the standard data set(University of California Irvine,UCI).2.The RBF neural network model improved by FAHP algorithm has a good effect on the modeling of some complex data,but it is necessary to extract the features in other ways for modeling the original input data which is not easy or not suitable for processing with FAHP algorithm.In general,feature extraction of data can be achieved by changing the neural network model structure,such as adding hidden layers.Compared with the single hidden layer neural network,the deep neural network with multiple layers of hidden layers has stronger ability to characterize the data,but with the deepening of the network structure,the neural network training by traditional learning mechanism of neural network is more difficult.Learning from the Deep learning theory to solve this problem,the paper try to use the RBF-based feature extraction module to construct a multi-layer learning structure to achieve the feature learning and advanced representation of input information.The construction of deep RBF neural network(DRBF)model is completed,and the training steps of the network model are described in detail.The feasibility of the DRBF network model is verified by experiments on UCI datasets.3.The FAHP-RBF and DRBF prediction models are applied to the production forecast and the production equipment level analysis in different industries in the complex chemical industry,such as ethylene industry and purified terephthalic acid(PTA).From the prediction accuracy and the time required for modeling,the paper analyzes the characteristics of the two modeling methods and the applicable scenarios.Then,design and construction the Web prototype system based on B/S architecture-the prototype system of complex chemical energy efficiency analysis.