Research On The Generalization Of Deep Belief Networkand Applications

Image is an important way of information bearing.Image recognition has been widely used in face recognition,product testing and license plate recognition and other occasions to improve the level of production automation and increase the convenience and reliability of social life.Artificial neural network use the artificial animal hierarchical information processing mechanism to extract the essence of the object characteristics,with its automatic learning from big data to extract features,efficient processing methods has become the the most important method of the image processing technology.Deep Belief Network is a probabilistic generation model that simulates neural networks and uses a large amount of data to train the weights between neurons,so that the whole neural network can generate training data with the maximum probability.It relies on a large amount of data to extract enough sample features to achieve the purpose of correct identification;if the sample size is small,the feature extracted by the model is not detailed enough,will generate apoorly model,and prone to misrecognition during testing.We study the poor generalization of the model caused by the difference in the number of training samples in this paper.According to the generalization of DBN network,two improved algorithms are proposed.First,from the perspective of communication principle,an equivalent model of DBN is proposed,the characteristics of DBN is discussed in depth;the interval number in fuzzy math is introduced to expand the weights in the network,change the weights from a definite value to a range to resist the uncertainty caused by some objective changes,and compensate the network shortcomings caused by the small number of samples,so as to improve the DBN recognition performance.The other is to use the affine transformation to define the common changes of the image.According to the affine transformation formula,the formula of the backward change is deduced.use the formula to adjust the convergence weight in the DBN network,and expand the category match space of sample image,then increase the classification accuracy of the network.The improved algorithm is applied to the identification of faulty insulators in power system.By constructing DBN insulator defect recognition network,the deep belief network learning model is introduced into the fault recognition of insulators in power system to realize the automatic intelligent recognition of insulators,and improve the the correct rate of the system performance.studied the different condition of insulator recognition and various algorithms proposed in the paper are tested.Provides the new method for intelligent detection of power system.