Characteristics Analysis And Improvements Of Generalized Congruence Neural Networks

In recent years, theoretical research of neural network (NN) has made a great progress and many outputs have been achieved in several application fields. As a classic NN model, Back Propagation Neural Network (BPNN) has also made a rapid progress and been used widely. Meanwhile, it also shows some shortcoming such as low train efficiency, slow convergence rate and difficulty on digital implementation. With pepoles' higher demand in real-time and large-scale of NN, it is importand and necessary to solve these problems.Former research studies on Generalized Congruence Neural Network (GCNN) shows that the convergence rate of GCNN is faster than that of BPNN. It also shows that GCNN with single hidden layer has consistent approach ability.The paper studies on characteristics analysis and improvements of GCNN. The chief contents include:Ⅰ. GCNN is improved by two ways, one is bayes' regularization and the other is stop ahead of schedule. It proved that the improved GCNN has some merits such as better generalization ability, higher learning rate, smaller relative error and more convenient to operate. Practical examples show that the improved GCNN is feasible.Ⅱ. The paper analyzes the performance of the improved GCNN such as the approach ability, the generalization ability, the effect of neurons' threshold on learning convergence, the effect of the initial value on training rate and precision, the effects of the numbers of hidden layer and neurons in hidden layer on network learning and work. Moreover, an example is present to prove the conclusion.Ⅲ. An example is used to compare the performance between BPNN and GCNN is provided and the result shows that network structure and learning algorithm of the improved GCNN are effective and GCNN has a higher rate and a better classify precision than the normal BPNN.At last, GCNN is successfully applied to tool state monitoring. Through analyzing the date from different work condition, the paper draws a same conclusion that GCNN can be used effectively to diagnose the wear of tool in aptitude. Meanwhile, as the classification tool of wear, the network can not only accomplish pattern design and classification effectively but also solve fault diagnosis process intelligently.