The Study Of ANN BP Algorithm And Its Application In Industrial Detection

BP Neural Network is one of the representative Neural Network Models. It has been applied in a lot of fields. Because of it's inherent defaults, many scholars and experts devote themselves to improving it's characters, and have made considerable headway for BP Algorithm. However, most study of BP Neural Network was based on the improvement of BP Algorithm in one aspect, then compared the improved BP Algorithm with the basic BP Algorithm, few summary documents are reported. And Neural Network technology has not been used in the quality detection of glass products. This paper will make a systematic summary and some elementary researches on HP Algorithm, and apply it to the computer vision detection system on-line of the glass products in order to enable the detection system to possess the capabilities of auto-learning and auto-adapting.The paper is made up of two sections. The first section makes a simple summary to NeuralNetwork from it's basic characters, developing course, compositions, types and applications etc..What has done in this paper is presented in the second section, which includes the study of HP NeuralNetwork's structure and BP Algorithm and the application of BP Algorithm in the computer visiondetection system on-line of the glass products.As we know, it is difficult to determine the hide nodes of BP Neural Network. A simple direct method of estimating the appropriate hide nodes is proposed in this paper based on experiments and previous experience. This method has simplified the determination of hide nodes and achieved good result.In the learning course of the basic HP Algorithm and many improved HP Algorithms, the controlling means are always as: give a fixed permitted convergence error or learning times, or both of this two are given at the same time. These means have their faults as follow: learning times is too long, even the learning can't end, or the learning accuracy is not high enough sometimes. In order to improve the speed and the accuracy of HP Algorithm, an improved HP Algorithm, namely HP Algorithm with two times adaptive adjust of learning parameters is proposed in this paper. Besides the adaptive adjust of learning rate and momentum factor, this algorithm can give appropriate permitted convergence error with adaptive adjust in the course of learning. So it overcomes the faults of the controlling means used in traditional BP Algorithm. The experiment results show that the learning accuracy and speed of BP Algorithm with two times adaptive adjust of learning parameters is high than that of traditional HP Algorithm. So it possesses better value of application.Crack detection system of glass bottles is a subsystem of computer vision detection system on-line of glass products. The whole system was developed by the cooperation of the Electronic Technique Research Institute of GuangXi Normal University and GuiLin Glass Factory. In the course of test, the writer found that the subsystem possesses worse adaptability of adapting the change of the environment and worse tolerance to mistakes of the new information. The main cause is that the algorithm of the subsystem does not possess the capability of auto-learning and auto-adapting. In this aspect, the Neural Network provide us with a completely new way by learning the experiential samples and storing the results of learning into the network respectively using the form of weights-2-and thresholds. It is more important that the Neural Network possesses very strong capabilities of auto-learning and auto-adapting. In this paper, the author applied BP Algorithm in this subsystem, and a model of crack detection of glass bottle is described. This model enables the detection system to enhance the capabilities of auto-learning and auto-adapting greatly by overcoming the weakness of the worse adaptability of adapting the change of the environment and the worse tolerance to mistakes of the new information. The author has realized BP Algorithm with Visual C++. The experiment results show that the detecting result of B...