Research On The Structure Design Of Feedforward Neural Network And Its Application To Modeling Complex Chemical Processes

With the fast development of the industrial processes, more and more processes tend to have the features of high nonlinearity, multiple input variables, high complexity, high synthesis, and so on. So it is more and more difficult to build accurate models using the first-principle based methods, i.e. the mechanism analysis methods. Recent years, advanced sensor technologies enable the process data more easily to be collected and stored. The historical process data contain much knowledge of the processes. Hence, the data-driven strategies based methods play an increasingly important role in modeling complex industrial processes.Among the data-driven approaches, the artificial neural networks (ANNs) have been successfully applied to many fields, such as modeling, control, optimization, and so on, because ANNs have salient abilities in the learning, the parallel computing and the nonlinear mapping. As a kind of ANNs, the feedforward artificial neural networks (FANNs) with simplicity, easiness, and other good features have attracted more and more attention from researchers. However, the traditional FANNs cannot achieve required performance in modeling modern complex industrial processes. Therefore, studying and establishing the FANN models with perfect performance is of great significance to enrich the neural network models and further promote the applications of the neural network technology to modeling complex industrial processes.In this thesis, we studied two feedforward neural networks of the hierarchical neural network (HNN) and the extreme learning machine (ELM) from the two perspectives of the hierarchical structure design and the double parallel structure design. And then the two kinds of feedforward neural networks are applied to modeling complex chemical processes. HNN is good at dealing with high-dimensional data; however, the subnet structure design of HNN is difficult. ELM with a fast learning speed and good generalization ability is one of the hot researches in the field of machine learning. However, facing the noise and the collinearity problems in the process data, some problems still exist in ELM:1, the low performance in processing noise; 2, the performace limitation with the traditional three-layer structure; 3, the bad effect of the collinear data on the network performance. In this thesis, these problems are solved one by one, aiming at establishing some reliable models for the complex chemical process with a specific problem. The main obtained research results are summarized as follows:(1) In order to solve the problem of the subnets design in HNN, a subnet design method based on the input attributes space division is proposed. So an input attributes space division based HNN model is built as a reliable model for the complex chemical processes with large numbers of input parameters. This proposed design method avoids using the complicated expert knowledge. Firstly, an advanced extension classification algorithm is adopted to cluster the high-dimensional space of the input attributes. Then the number of the subnets is determined according to the number of the clusters. And the inputs of each subnet are determined according the attributes in the corresponding cluster. The proposed design method simultaneously solve the two problems of determining the number of the subnets and the inputs of each subnet, which provides a simple and effective subnet design method for HNN.(2) In order to enhance the performance of ELM in processing the noise, an ELM model with a hierarchical structure is proposed. In the proposed hierarchical ELM model, the original inputs are not directly put into the ELM. The original input data are firstly put into the auto-associative neural network for removing the noise and reducing the high dimension. Then the outputs of the hidden layer nodes of the auto-associative neural network are put into the ELM. So the bad effect of the noise on ELM is avoided. Simulation results using the process data with noise illustrate the effectiveness and feasibility of the proposed model.(3) In order to address the limitation problem of the three-layer structure in ELM, the parallel structure design is adopted to enhance the performance. Although the traditional double parallel architecture can well solve the structure limitation problem, the traditional double parallel architecture brings the other two problems:1, increase the complexity; 2, bring more collinear information. By studying the parallel structure theory and the Pearson correlation coefficient, an ELM with an input-output Pearson correlation coefficient oriented parallel structure is proposed to solve the first problem. In this proposed model, according to the correlation coefficient between the input attributes and the output attributes, the input attributes are separated into two categories:the positive input attributes and the negative input attributes. As a result, a parallel structure with a positive and negative independent input attributes is established. Simulation results show that, compared with the extreme learning machine with the traditional double parallel structure and the traditional extreme learning machine, the proposed model achieves a simpler structure and faster response.(4) ELM cannot well solve the collinear data information in the traditional double parallel structure. In order to solve this problem, a robust double parallel extreme learning machine based on the partial least square learning is proposed. In this proposed method, the output weights are calculated using the partial least square method but not the original generalized inverse learning method. On the one hand, the collinear information in the original data and the outputs of the hidden layer is removed using the partial least square method. On the other hand, the difficulty in determining the optimal number of hidden layer nodes is avoided by the partial least square method through selecting the latent variables. Simulation results show that the performance of this proposed model is stable and robust, which provides a reliable model for the complex chemical process.