An Automated Fault Diagnosis Technique For Varying Speed Bearings Using Convolutional Neural Network And Synchrosqueezing Transform

Rolling element bearing is an indispensable fundamental component in a rotating machine.Its stable operation and health condition have a great impact on the reliable operation of a machine and even the safety of the operator(s).It is thus critical to detect an incipient bearing defect as early as possible and to take remedy measures on time to prevent a component defect to become a machine functional failure.In practical applications,bearings often operate under variable speed conditions and can get damaged due to lack of lubrication or poor working environments.The bearing condition monitoring signal acquired under variable speed conditions is much more complex than that of constant speeds.The fault feature information of bearings is hard to extract using the traditional signal processing techniques based on the stationary assumption.Therefore,this thesis utilizes the advantage of Convolutional Neural Network(CNN)in image recognition and feature extraction,and applies the technique together with Synchrosqueezing Transform(SST)for an automated bearing fault diagnosis under variable speed conditions.It is the first time that CNN and SST are combined for an accurate bearing fault diagnosis under varying speed conditions.Particular efforts are devoted to improving the training speed and efficiency of CNN network using the highly energy concentrated time-frequency fault feature extracted from SST.To achieve the goal lays out in the study,the research presented in this thesis is undertaken in the following aspects:(1)Acquiring various bearing fault signal data from an experimental bearing test rig under variable operation speed conditions;(2)Using Peak-Hold-Down-Sample(PHDS)algorithm and other algorithms to reduce the size of the original acquired data for a faster and efficient fault diagnosis process in the subsequent analysis;(3)Applying Synchrosqueezing Transform technique together with Short Time Fourier Transform(STFT)to resolve the problem of insufficient time-frequency resolution of using STFT alone,and to convert one-dimensional time-domain data into two-dimensional time-frequency representation graphs.The time-frequency graphs are then used as input samples to train the CNN network;and(4)Adaptive Moment Estimation(Adam)algorithm and Batch-Normalizing(BN)algorithm are used to improve the training speed and recognition accuracy of CNN,and to achieve a final recognition rate of 100%.Results obtained from this study show that the proposed fault diagnosis technique can successfully realize the automated fault diagnosis of bearings under variable speed conditions.The technique has several advantages than the existing techniques such as one-dimensional CNN and STFT on the following aspects:(1)it successfully overcomes the difficulty in fault feature extraction under varying speed conditions,(2)it has a faster convergence speed,and(3)it can produce a much more accurate recognition rate.