Research On Bearing Fault Diagnosis Based On Improved Welch Power Spectrum

Rolling bearing is one of the most important components in rotating machine.The functions of rolling bearing are irreplaceable in industry.Usually,the major accidents and equipment downtime due to the failures of rolling bearing.Accurate and real-time diagnosis of bearing is always the key to the smooth operation of the equipment.With the advent of intelligentized era of big data,real-time diagnosis and accurate monitoring equipment symbolized the intelligent bearing fault diagnosis has entered the big data era.Therefore,the diagnostic performance of some traditional diagnostic methods which used feature extraction with classification algorithm cannot be guaranteed and it cannot meet the needs of the era of big data.In this paper,we study how to improve the diagnostic accuracy and success rate of bearing fault diagnosis.It is found three main problems in traditional bearing fault diagnosis methods:(1)Information lost in signal processing: usually,feature values are according to the specific problem and keep the important information for diagnosis.This process will eliminate some useless and not sensitive information to improve classification accuracy.But in a sense,this process is also a kind of reduction of the contained information in signal.The lack of other information in the signal results in a decrease in diagnostic performance when other problems are encountered or diverse data is mixed.(2)Small quantity of training samples: Because,some projects in early stage were in poor working conditions,it is difficult to obtain some relevant data.This means that the quantity of training data is insufficient,and the diagnostic model does not have enough training,so the models cannot achieve the ideal diagnostic effect.(3)Noise interference: In actual production,the majority of equipment is performed by Multi-component collaborative operation.Therefore,the vibration signal produced by the bearing in operation is often disturbed through the vibration of other parts.These noise can distort the original signal produced by the bearing,it may mask,hide or even tamper with the parts of original signal that need to be identified.For above problems,in this paper,we research in how to avoid or reduce the information lost in signal processing,reduce the number of training samples needed,reduce environmental noise interference and so on.In this study,an improved method of Welch power spectrum was firstly proposed to transform the chaotic time-domain data into the orderly power spectrum data,which can be directly used in the intelligent classifier,and reduces the interference of environmental noise to the diagnosis process.Then,two effective bearing fault diagnosis models(W-RBFNN,W-CNN)were constructed power spectrum of Welch and suitable intelligent classifiers(Radial Basis Function,RBFNN & Convolutional Neural Networks,CNN).In this paper,both models are researched in the database of Case Western Reserve University(CWRU),and tested by mixed load,different Numbers of training samples and anti-noise diagnosis.The test results show that the two bearing fault diagnosis models(W-RBFNN?W-CNN)have outstanding diagnostic performance in the above problem,which compared with some excellent bearing fault diagnosis models in recent years.In the diagnosis of mixed load data,the diagnostic accuracy of the two models all reached 100%.Under the condition that there was only one training sample in each bearing state,the average diagnostic accuracy of W-RBFNN reached 95.89% and that of W-CNN reached99.76%.Meanwhile,the recognition rates of anti-noise diagnosis was better than some diagnostic models proposed in recent years.