Research On Axle Fault Diagnosis Method Based On Transfer Learning

As the main load and running part of rail vehicles,axle is prone to failure under long-term high load operation.Therefore,it is of great significance to the fault detection and maintenance of axle.Nowadays,the axle fault diagnosis methods have a series of problems.This paper took acoustic emission(AE)signals as the research object and proposed a series of axle fault diagnosis method based on transfer learning,which achieved effective extraction and accurate recognition of axle fault features.The main work of this paper is as follows:(1)Aiming at the problem of feature extraction difficulty and small amount of data of original one-dimensional AE signal,this paper used Case Western Reserve University bearing vibration fault signals as source domain data,and used one-dimensional convolutional neural network to extract local features without losing temporal features,which constructed a deep neural network model(PTL-WDCNN)as a feature extractor.Then the parameters of the model are transferred to the target domain AE fault diagnosis task with less data.(2)Aiming at the problem of missing some important features in AE signal processing.Firstly,the continuous wavelet transform is introduced to transform the original onedimensional acoustic emission fault signal into a two-dimensional time-frequency image,which was used to obtain the multidimensional feature information of the original signal.Secondly,the pre-trained Efficient Net B0 network model was introduced and the conditional parametrized convolutional structure was added to extract 2D time-frequency graph features of acoustic emission.Finally,support vector machine is used to replace Soft Max as a classifier to train the features extracted from the model.(3)Aiming at the problem of the inconsistent feature distribution of AE signal data under different working conditions,the maximum mean difference of joint discriminant probability(DJP-MMD)for different working conditions is calculated to minimize the difference of probability distribution between source domain and target domain,and it is added into the objective function of model training so that the distributions of the two domains in the same feature space are identical.Thus,the model obtained by training the ae signal data in one working condition is used to diagnose the ae signals in other working conditions.Through experimental demonstration,the proposed method can effectively analyze the characteristics of the AE signals of the axle fault,and improve the accuracy of the axle fault diagnosis under different working conditions with a small sample size.It not only greatly reduced the training time and calculation cost,but also enlarged the application range of existing fault diagnosis methods,and provided a new idea for the intelligent fault diagnosis of axle.