A Novel Hybrid Neural Network And Its Applications

As a class of intelligent systems, artificial neural networks (ANNs) are used to simulate human neural networks to handle information. Not being the prefect models of the human neural networks, ANNs can be widely used to obtain new knowledge from data. Therefore, ANNs can solve some difficult problems, such as speech / image recognition and comprehension, knowledge processing, combinational optimization, intelligent control. Now, ANNs have become one of the most fascinating research topics in the field of artificial intelligence.First, this thesis reviewed the development and applications of ANNs, and it presented some typical ANNs which are widely used. Especially, it analyzed multi-layer feed-forward neural networks (FNNs) and their traditional learning algorithm, Back Propagation (BP) algorithm, because FNNs are the most common used ANNs.To overcome the limitations of general FNNs and BP algorithm, this thesis introduced a hybrid feed-forward neural network, which is composed of a linear model and a general multi-layer FNN, and proposed a new learning algorithm for the hybrid FNN. The learning algorithm is only based on linear least squares, and no iterative learning processes are needed. According to the demand of model precision, the algorithm determines the best weights of network and the minimal number of hidden nodes automatically. Compared with the well-known BP algorithm, simulation results show that the new learning algorithm for the hybrid FNN is more efficient in model precision, rate of convergence and generalization ability.To verify the effectiveness of the proposed hybrid FNN, this thesis addressed the estimation problem for the frozen point of light cyclic oil in a fluidized catalytic cracking unit (FCCU) in a refinery. Based on the sample data collected from the industrial unit, we built a soft sensor model by using an above hybrid FNN.Furthermore, the online self-learning problem of soft sensor model was also discussed. Finally, the built model was used in an industrial FCCU. The predictive precision of the model is satisfying, and it can meet the actual demand of the FCCU. Application results show that the hybrid FNN can be widely used as a basic model of soft-sensors.