Resarch On BP Neural Network Optimization Algorithm And Its Application In Fault Diagnosis

Classification refers to the process of obtaining the corresponding functions on the same categories based on the original data and modeling so that the unknown class can be classified reasonably.The developments of society and science are getting faster,researches on classification are also more and more deep and wide.In practice,the data is getting bigger,the problems are getting more complicated.With these facts,these ordinary methods always have low forecast efficiency and accuracy.The BP neural network algorithm is used to train the samples over and over again,and amend the error at the same time.It has excellent multidimensional function mapping ability,which effectively improves the forecast accuracy of the model,and has wide applications and recognition.However,the conventional BP neural networks have no rigorous theoretics on designing the network structures and obtaining the number of hidden layer nodes,and the randomicity of the weight and threshold directly affect the generalization and prediction accuracy.In view of the problems above,we have proposed a new method for obtaining the number of hidden layer nodes based on the previous researches firstly,and then proved its effectiveness and practicality.Secondly,for the problem that the randomicity of initial weight and threshold slow down its convergence speed,we have proposed a modified parameter optimization method,theory and its procedures based on genetic algorithm.Comparative experiments reflected that the improved method effectively accelerate the convergence.Finally,based on the adjusted BP neural network forecast model,we solved the problems of railway turnout failure diagnosis by classifying and forecasting and reached better prediction results.The comparative experiments showed that the improved BP neural network model is reasonable and effective and significantly improves the forecast accuracy.