Fault Diagnosis Of Rolling Bearings Based On Adversarial Transfer Learning

Rolling bearings as the core parts of rotating machinery equipment,once failure,may cause major economic losses and casualties,so it is of great significance to carry out fault diagnosis research on rolling bearings to ensure their stable operation.Existing rolling bearing fault diagnosis methods usually rely on a large number of labeled data,but in practical industrial applications,it is difficult to collect a large number of labeled data,so there are certain defects.Migration of as a kind of new machine learning method,learning through knowledge migration from the source domain to target domain can effectively solve the problem of insufficient label sample,but the existing migration study methods usually trained by adopting the method of adaptive or fine-tuning,influenced by load operation environment,the training method can not be effectively aligned distribution characteristics of different data sets.In order to solve the above problems,this paper proposes two improved transfer learning methods with rolling bearings as the research object.The specific research contents are as follows:(1)By combining adversarial learning with transfer learning,a new rolling bearing fault diagnosis method is proposed.The method introduces the idea of generative adversarial network on the basis of transfer learning,reduces the distribution difference between source domain data and target domain data through the adversarial training between feature extractor and discriminator,and solves the problem of insufficient label samples.In the process of constructing the model,the feature extractor of transfer learning is replaced by a one-dimensional convolution structure in order to realize the effective processing of vibration signals in time domain.(2)Based on the above research,an improved adversarial transfer learning rolling bearing fault diagnosis method was designed by introducing maximum mean deviation.In this method,one-dimensional Inception structure is adopted to expand the depth of feature extractor on the basis of the original one-dimensional convolution layer,so as to better extract vibration signal features.On the other hand,considering that the original adversarial transfer learning method could not effectively align the feature distribution of source and target domain data,the maximum mean deviation was used to match the difference of edge distribution to improve the model diagnosis result.Finally,for the reasonableness of the experiment,the noise with different SNR was added to the bearing vibration signal to simulate the actual working environment.The proposed method was applied to two rolling bearing fault diagnosis cases,and compared with the traditional transfer learning method and machine learning method,the classification accuracy,recall rate and F1 score were taken as indicators,and the effect of the model was verified by combining t-SNE data distribution and thermal diagram.The application results show that the method designed in this paper has good anti-interference ability and robustness.