Research On Fault Diagnosis Method For Rolling Bearings Based On Time-Frequency Image Recognition

Rolling bearing is one of the most commonly used key parts in all kinds of mechanical equipment,and it is also a vulnerable part.The running state of rolling bearing directly affects the performance of the whole equipment,as well as its safety and reliability.Therefore,it is of great scientific significance and engineering application prospect to carry out the fault diagnosis research of rolling bearing to ensure the safe and reliable operation of mechanical equipment and avoid the occurrence of major and catastrophic accidents.Due to the changes of load,speed and other factors during operation,the vibration signal of rolling bearing shows non-stationary characteristics,which will become more obvious with the occurrence of fault.Compared with traditional time-domain analysis and frequency-domain analysis,time-frequency analysis can better reveal the amplitude/energy distribution and local time-varying characteristics of non-stationary signals in time-frequency domain.Therefore,the time-frequency images of vibration signals constructed by time-frequency analysis contain more abundant information about rolling bearing states.Thus,through in-depth analysis and recognition for time-frequency images,rolling bearing fault diagnosis can be better realized.The common fault diagnosis methods based on time-frequency image recognition can be roughly divided into three categories: artificial diagnosis method,traditional machine learning-based method and deep learning-based method.Among them,the artificial diagnosis method is the method that the professionals complete the fault diagnosis by analyzing the time-frequency characteristics revealed by time-frequency images,while the other two kinds of diagnosis methods are realized by using computers to automatically classify and recognize time-frequency images.Obviously,the artificial diagnosis method is time-consuming and laborious,and its diagnosis results are affected by human subjective factors.Also,because whether time-frequency images are accurate or not has a great influence on the analysis and judgment of professionals,this kind of method seriously depends on the performance of time-frequency analysis.Whereas the other two methods can liberate the professionals from the complicated image analysis and recognition,and the diagnosis results are more objective.However,the artificial diagnosis method has the advantage of low data requirement,while the other two methods need a certain amount of data to train the models,especially the deep learning-based method which needs a lot of data for training.In addition,the recognition rate and stability of these three methods need to be further improved.All in all,the three kinds of methods have their own advantages and disadvantages respectively,and there are still some problems to be further studied.The research works mainly include the following three aspects:(1)Study on time-frequency analysis methods of rolling bearing vibration signal.The artificial diagnosis method relies heavily on the performance of time-frequency analysis,while the commonly used time-frequency analysis methods have the problems of non-sparse analysis results,low resolution and cross-term interference.Therefore,a first-order primal dual algorithm based sparse time-frequency analysis method(STFA-PD)is proposed.Based on the sparse representation theory,the sparse time-frequency analysis model is constructed,and the first-order primal dual algorithm is used to solve the model.Because the proposed STFA-PD is a linear time-frequency analysis method,and the sparse constraint is introduced,the time-frequency image constructed by this method has high time-frequency resolution and time-frequency concentration,and there is no cross-term interference.The simulation results show that the STFA-PD can overcome the shortcomings of traditional time-frequency analysis method,such as low resolution,cross-term interference,and has better denoising performance.In addition,experiments are carried out on the popular Case Western Reserve University(CWRU)bearing data.The results show that,compared with the other method,this method can obtain time-frequency images of bearing vibration signal with better sparseness,time-frequency resolution and concentration,which can accurately reflect the time-frequency characteristics of the signal and make the fault impact characteristics more obvious,and provide more accurate information for professionals to complete the fault diagnosis.(2)Study on rolling bearing fault diagnosis method based on texture feature transfer of time-frequency images.Aiming to address the problem of the diagnosis performance degeneration of the traditional machine learning-based method caused by the feature distribution differences of data in different working conditions,a rolling bearing fault diagnosis method based on texture feature transfer of time-frequency images is proposed.Firstly,the vibration signals are converted to time-frequency images by using time-frequency analysis,and then the gray level co-occurrence matrix(GLCM)texture features of time-frequency image are extracted as the feature vectors reflecting the bearing states.After that,the joint distribution adaptation(JDA)algorithm is used to map the GLCM texture features of data in different working conditions to a low-dimensional potential space.Finally,the common transfer features with smaller distribution differences are employed as the input of the nearest neighbor classifier to realize the accurate recognition of the rolling bearing state under different working conditions.Because the JDA is employed to conduct feature transfering,this method can use data in a working condition for training to accurately classify and diagnose the data in other working conditions.The performance of the method is verified by the CWRU bearing data.The results show that the proposed method can extract the common transfer features of the data in different working conditions,and thus significantly improve the diagnosis performance.(3)Study on deep transfer learning-based rolling bearing fault diagnosis method.In order to solve the problem of performance degradation caused by working condition differences and the problem that a small amount of data in a single working condition can not effectively train a deep network,a deep transfer learning-based fault diagnosis method of rolling bearing is proposed.In this method,time-frequency analysis is used to convert vibration signals to time-frequency images.Combined with the theories of deep transfer learning and residual learning,a transfer deep residual convolutional neural network(TDRCNN)is proposed to automatically learn the features of time-frequency images and complete the classification diagnosis.Through introducing the deep transfer learning method,the TDRCNN can make full use of the knowledge learned from a large number of source domain data in a working condition and a small number of target domain data in another working condition to accurately classify and diagnose the target domain data.And,by using the residual structure,the TDRCNN also overcomes the problems of training difficulty and performance degradation existing in traditional convolutional neural networks(CNNs).Experiments are conducted on the popular CWRU bearing dataset to validate the effectiveness and superiority of the proposed method.The results show that the TDRCNN is superior to the traditional CNNs and the transfer networks without residual structure in training,fault classification,visual feature clustering and separation.Also,the experimental results verify the effectiveness of the proposed method with different time-frequency analysis methods and variable working condition differences.