Study On The Theory And Technology Of Neural Network Structure With The Applications In Process Simulation And Control

The backpropagation neural network(BPN) is the most widely used network paradigm for solving chemical engineering problems. Here some novel theories and techniques to solve the important problems about network structure optimization, process modelling and control are proposed. A developed intelligent system integrating BPN structure optimization(BPNSO) and special BPN dynamic adaptive control(SBPNDAC) is presented.Firstly, this dissertation gives a review on the developed history of artificial neural networks(ANNs) and its current application situations in process modelling, simulation and control. After analysed the basic properties and principles of neural networks, the ideas of improving the learning algorithm and establishing a process simulation and control system using special BPN are proposed. The size and properties of samples, initial weights and its adjustment way are disceussed, which influence the stability and convergence of the network. The dual convergent criterion based on the training data and testing data is used and dynamic adaptive adjustment strategy based on the training error for learning rate η and momentum factor a is proposed.The relationship between the topology and convegence of networks is approached and the reason causing overtraining and overfitting is analysed. Then the redundant weights and nodes in the network are defined by using improved activity function and a new algorithm called the BPNSO for determining the network structure for any application is proposed, which use dual optimization objective function combining the mean squre error of the network output with network complexity penalty function. A novel adjustment strategy for penalty factor is also proposed, which based on the learning error and weight magnitude. The paper provides a detailed procedure for BPNSO.In the dissertation, some parameters in the BPNSO are defined and analysed. The capability of BPNSO is demonstrated by application both in steady-state and dynamic modeling cases. The results show that the redundant weights and nodes not only in the hidden layer but also in the input layer can be identified and deleted. It...