Study On Intelligent End-to-End Fault Diagnosis Of Rolling Bearing Under Variable Speed Condition

Rolling bearings are important components in rotating machinery.High operating speeds and variable operating frequencies lead to rolling bearings being extremely prone to failure.Compared to rolling bearing fault diagnosis under constant rotational speed,rolling bearing fault identification under variable rotational speed is more practical and difficult.The bearing fault diagnosis method based on signal processing relies on professional background knowledge to extract complex features and achieve fault classification.However,it is difficult and inefficient to extract specific fault features under complex rotational speed variations.The method based on deep neural network can adaptively extract features from time-frequency map without complex signal processing.These extracted features not only include well-known fault features,but also may have some potential features that can be used to identify faults.Therefore,based on deep neural network theory,this thesis studies the intelligent end-to-end fault diagnosis of rolling bearings at different rotational speeds.The specific content is as follows:(1)Under different rotational speed operating conditions,the fault characteristic frequency of rolling bearings is also different.To solve this problem,a variable speed rolling bearing fault diagnosis method based on Efficient Net is proposed.The bearing vibration signal is collected as the original signal,and the time-frequency maps with fault characteristics are obtained through short-time Fourier transform.Utilizing the advantages of Efficient Net network in image feature extraction and detail learning,the features of bearings at different rotational speeds are sensed and learned,thereby classifying different faulty bearings.Case Western Reserve University bearing dataset demonstrates the effectiveness of this method.In addition,our laboratory has conducted experiments on bearing faults at different rotational speeds,and the experimental results show that the method is effective in diagnosing rolling bearing faults at different rotational speeds.(2)In practical engineering,bearings are often operated at time-varying rotational speeds,the fault characteristic frequency which varies with time increases the difficulty of fault diagnosis.To solve this problem,a fault diagnosis method for time-varying rotational speed rolling bearings based on Efficient Net V2 and migration learning is proposed.This method focuses on real-time changing fault characteristics through SE module in Efficient Net V2,achieving end-to-end bearing fault diagnosis.In addition,the Efficientnet V2-s pre-training weight is introduced for training,and the method based on parameter migration is used to improve training efficiency.Two time-varying speed bearing datasets demonstrate that this method can effectively extract the characteristics of bearing faults even when the fault characteristics are weak and there are composite faults.This method can classify faulty bearings with 98% accuracy under time-varying speed conditions.(3)Combining the characteristics of fault diagnosis for rolling bearings with different rotational speeds and time-varying rotational speeds,to achieve higher model learning efficiency,an end-to-end fault diagnosis method for rolling bearings based on Light Fused-MBconv(LFMB)model was proposed.This method introduces the progressive learning strategy,designs and builds a LFMB model,which can perform intelligent fault diagnosis for rolling bearings at constant and time-varying speeds.Experiments on a constant rotational speed rolling bearing dataset and two time-varying rotational speed rolling bearing datasets show that the diagnostic accuracy of the method can reach 99% or more,and the model has strong generalization.Comparative experiments have shown that this method has advantages in end-to-end rolling bearing fault diagnosis methods,and has achieved a competitive rapid and accurate diagnosis effect.(4)We summarized the work and made a prospect for the related researches.This thesis has 60 figures,24 tables,112 references.