Research On Fault Diagnosis Of Acoustic Emission Signal Of Axle Fatigue Crack Based On Dbn

With the rapid development of China’s high speed railway business,the operating speed and load capacity of rail vehicles are gradually increased.At the same time as the high speed of development,there are also higher demands on the safety performance of trains,the first and foremost is the axle.The axle is the key load-bearing component that makes up the moving part of the train and plays a vital role in the safe and smooth operation of the rail vehicle.The axle will inevitably crack in the face of changing external environment and complicated working conditions,and axle cracks will develop to a certain extent,causing the axle to break,resulting in disastrous consequences such as train derailment.It is therefore important to be able to detect axle failures in time and accurately predict the remaining life of the axle,both to avoid the corresponding accidents and to reduce maintenance costs.This paper studies how to extract features and classify AE data of axles,as well as lifetime prediction and proposes an analysis method based on Deep belief network(DBN).Firstly,the basic principle and network structure of DBN are introduced.On this basis,the Restricted Boltzmann machine(RBM),which is a part of DBN,is studied.The influence of the number of iterations of weight updating and the learning rate on feature extraction and network computing time is analyzed.In the process of extracting the axle acoustic emission signals from DBN,the extracted multidimensional features are treated with principal component analysis(PCA)to preserve the first three principal component components,so that the features can be visualized in 3D space for easy observation.In the process of classification and recognition of axle AE signals by DBN,the change of signal recognition accuracy caused by different hidden layers is mainly analyzed.Then,by comparing with the classification accuracy of Extreme Learning Machine(ELM),a machine learning network model,the advantages of DBN in dealing with large amounts of data.For different axle cracks,DBN is used to carry out experiments in other axle acoustic emission signal datasets.For a complete axle crack acoustic emission signal,the crack is divided into four different stages according to the trend of the root mean square value of the signal,and the different stages are put into the DBN network for classification and identification.Experiments on different data sets show that the deep belief network can recognize crack signals from various AE signals of axles and has good classification ability for different stages of crack signals.Finally,a DBN-ELM-based life prediction research method is proposed,which combines the ability of automatic feature extraction by DBN with the advantages of ELM with the good generalization,and has a good prediction of the remaining life of the axle after cracking.