Research On Bearing Fault Diagnosis Based On Convolutional Neural Network And Transfer Learning

In modern industrial production,rotating machinery is one of the most important equipment.As the core component of many rotating mechanisms,fault diagnosis of rolling bearings can ensure smooth operation of the equipment,effectively improve production efficiency and reduce economic losses.Therefore,rolling bearing fault diagnosis is a more and more important and popular research topic in the area of mechanical fault diagnosis.Currently,most of the bearing fault diagnosis studies take vibration signal as the data to be analyzed.In order to achieve accurate fault diagnosis,feature extraction is a very important preprocessing step.Traditional bearing fault feature extraction is often based on time-domain and frequencydomain signal analysis to extract the features of typical bearing faults.However,this process relies heavily on manual and prior knowledge.In recent years,as the development of intelligent fault diagnosis technology,the automatic feature extraction technology based on deep learning has been widely used in bearing fault diagnosis.If the labeled data is sufficient,a promising diagnosis performance can be obtained.But in practical industrial application,for each machine in the new working condition,it is difficult to collect enough typical fault samples.For this end,the transfer learning aims to use a large number of labeled source domain data and few unmarked target domain data to implement the adaptation of model across different domains.In order to solve the problem of feature extraction for bearing faults when the labeled data in the new environment is insufficient,this paper mainly carried out the following studies by integrating the knowledge of convolutional neural network and transfer learning:1.Bearing fault diagnosis relies heavily on the extraction of fault features.In order to realize more direct fault diagnosis based on raw data,a fault diagnosis method based on dense convolution network is proposed.The method firstly normalizes the original one-dimensional data,and then input it into the dense block network to adaptively learn the effective characteristics of the bearing vibration signal and to realize classification with the inclusion of the softmax classification layer.The bearing fault database provided by Western Reserve University is used in the experiments to verify the effectiveness of the proposed network.The results are compared with those of other commonly used models for bearing fault diagnosis.It is shown that the proposed network by using bearing vibration data can realize apparently higher accurate fault type classification than the other networks.Especially for the vibration data mixed with some noises,this network can still provide satisfactory diagnosis accuracy.2.The vibration data of bearings are often distributed differently in different working conditions,leading to low fault diagnosis accuracy.To this end,a novel deep adaptation network was proposed for cross-domain bearing fault diagnosis.First,Fourier transform is used to transform the original vibration signals in the time domain into the corresponding signals in the frequency domain.After that,the classification features are extracted by deep feature extractor.Second,Maximize Mean Discrepancy(MMD)is used for aligning the marginal distribution of the deep features.Finally,Wasserstein metric network is used to match the category structure of labeled data of the source domain with that of unlabeled data of the target domain,that is,to align the category condition distribution of different domains.By this way,the distribution of data feature can be best aligned in different domains,resulting in better diagnosis accuracy of the trained model in the unlabeled target domain.In the experiment,the model migrations under two kinds of cross-domain conditions are designed and tested by using the bearing fault dataset published by Case Western Reserve University.It was verified that the network could provide high diagnosis accuracy in different migration scenarios and was superior to other deep adaptation networks.