Research And Application On Interval Neural Network

Information is the source of understanding of the world and transforming the world. However, due to the diversity, complexity, time-variation of the world, the information people realized is often uncertain and imprecise. The phenomenon of imprecision of the measurement data caused by the limit of sensor precision or external disturbance also appear in the complex industrial process. So the first issue is how to describe the uncertain information and establish mathematical model of uncertain information. Granule is a concept to describe uncertain, imprecise, fuzzy information. There are various kinds of information granulation methods. One of them is set theory and interval analysis theory which has excellent universality and outstanding description ability of inaccuracy data. Neural network has a wide range of applications in terms of complex system modeling. Combining interval theory and neural network consisting of interval neural network can well solve the problem of inaccurate information modeling. So the research on the structure and learning algorithm of interval neural network have a very important theoretical and practical significance.Based on the interval neural network theory, structure and learning algorithm of interval feedforward neural network and interval recurrent neural network are further studied in this thesis through extensive review of relevant documents. Aim at the disability of Jordan neural network to model k order system, an improved Jordan neural network is proposed. Then extend this neural network to interval form and applied it to glutamic acid fermentation process modeling. The main research work is as follows:First of all, the interval feedforward neural network can be divided into three categories, and the relevant model is established and deduced gradient descent algorithm is used in three layer interval neural network. Then using data experiment show that the different performance between each type of interval neural networks and conventional neural network. Considering the gradient descent algorithm easy to fall into local optimum problem, by using the branch and bound algorithm based on interval optimization which is a global optimization algorithm to solve the problem of interval weights and threshold value of neural network learning, and good results are obtained.Secondly, use Jordan neural network as example to study the structure and learning algorithm of interval recurrent neural network. Considering the ordinary Jordan neural network can’t to order k system modeling, this thesis improve the structure of neural network by joining the feedback of feedback and context level. Modified Jordan neural network prove to have the characteristic similar to the PID controller of proportion, integral and differential features. Using the experimental data proved that the improved Jordan neural network has better dynamic characteristics and k order system modeling capabilities.Finally, respectively use BP neural network, Jordan neural network and improved Jordan neural network to model glutamic acid fermentation process. Then draw a conclusion that interval recurrent neural network has higher accuracy and faster speed of learning compared with the feedforward neural network.