| Rolling bearings are an important part of large machinery.Under the condition of high speed and high strength,the failure of rolling bearings is unavoidable.Once the failure occurs,it will not only bring great influence to the normal operation of equipment,but also bring huge economic losses.Therefore,it is necessary to conduct in-depth fault diagnosis and analysis of rolling bearings.The application of traditional fault diagnosis technology in fault diagnosis is restricted because it needs a lot of expert knowledge and complex feature extraction.And deep learning technology can be extracted from 2 d image data of key information,therefore,based on the rolling bearing vibration signals transfer image data sets the research object,the key for rolling bearing vibration signal feature extraction,deep learning model is trained,bearing fault sample number less,problems such as unbalanced data sets,Deep convolutional neural network and improved generative adversarial network are used to study these problems.The specific work is as follows:(1)Aiming at the problem that deep convolutional neural network is not able to extract subtle features of one-dimensional vibration data,the pre-processing method of transforming vibration data into image data is adopted,and the transformed wavelet time-frequency map is input into 16-layer Visual Geometry Group Networks(VGGNet)for training.The bearing fault data collected by the bearing fault diagnosis laboratory and the open data set of bearing fault diagnosis from Case Western Reserve University were used for experimental verification.The results show that the average diagnostic accuracy of VGGNet network is 98.96% and99.52% respectively,which reflects the superiority of VGGNet network.(2)In order to solve the problems of incomplete training and long training time of Deep model,the rolling bearing fault diagnosis method based on Deep Residual Network(Res Net)based on Transfer Learning was adopted.Firstly,Image Net large image database was used to pre-train most network parameters of Res Net at layer 34,and then the rolling bearing time-frequency image set was input into the network for secondary training.Finally,the rolling bearing data set was used for experimental verification,and the average fault recognition rate of TL-RESNET34 model was 98.89%.The feasibility of the proposed method is verified.(3)Aiming at the problem that the fault data in the actual operation of mechanical equipment is far less than the normal data,the fault data-oriented fault diagnosis method based on Generative Adversarial Networks(GAN)is adopted.First,Wasserstein distance was used to replace JS divergence of adversity-generating network to establish the basic model,then gradient penalty(GP)was used to optimize the model,the distribution of original data was learned,and the fault sample set that could be used for training was further expanded.Finally,The validity of data driven and generative adversarial network in rolling bearing fault diagnosis is illustrated by setting unbalanced data set and data enhancement experiments. |
