Hybrid Algorithm Of BP Neural Network And PSO And Its Application

BP neural network, a kind of nonlinear complex network system which simulates cerebrum information processing, has strong nonlinear mapping ability, fault tolerance, distributed information storage, parallel processing and adaptive learning. It has been widely applied in pattern recognition, intelligent control, business management, market analysis and other fields. PSO algorithm is a optimization algorithm based on the theory of swarm intelligence, and searches in the guidance of swarm intelligence optimization through the cooperation and competition among the particles, which has high convergence speed, high robustness, and global search ability, and does not require special information itself. When the standard BP neural network starts learning and training, the initial weights and threshold of the network are given randomly, which matter to learning is whether it can reach the global minimum and converge. PSO algorithm is used to optimize the distribution of initial weights and threshold of BP neural network, the introduction of PSO aims at avoiding the local minimum.The main contribution of this paper is generalized as follows:(1) The hybrid algorithm of BP Neural Network and PSO. PSO algorithm is used to optimize the distribution of initial weights and threshold of BP neural network, the introduction of PSO aims at avoiding the local minimum.(2) The improved hybrid algorithm is applied to the energy consumption prediction for steelmaking process in iron&steel industry. Increasing a momentum term in the adjustment formula of weights and threshold and adaptive learning rate, the method is adopted to improve the network training speed of BP algorithm. Using linear time-varying inertia weight, the method is adopted to adjust the search capability of PSO algorithm. The test results show that:the average training time of the hybrid algorithm is 1.42 seconds, which is 2.46 seconds shorter than the standard BP algorithm. The average prediction error of hybrid algorithm is 3.67%, which decreases by 2.32% than the standard BP algorithm.(3) The improved hybrid algorithm is applied to surface defect classification of cold-rolled strip in iron&steel industry. With increasing a momentum term in the adjustment formula of weights and threshold and steepness factor in the activation function, the method is adopted to improve the network training speed of BP algorithm. Through increasing shrinkage factor, the method is adopted to improve the convergence of PSO algorithm. The experimental results show that:the training time of hybrid algorithm is 89 seconds, which is 28 seconds shorter than the standard BP algorithm. The classification rate of hybrid algorithm reaches 93.33%, which increases by 14.66% than the standard BP algorithm.(4) Based on results of theoretical studies, a cold-rolled strip surface defect detection and classification system is developed using VC++ language.