Research On Application Of Graph Neural Network In Bearing Fault Diagnosis

Rolling bearing fault diagnosis is closely related to equipment safety,and has always been a research topic that has been paid much attention.In practical engineering applications,because the bearing is subjected to various pressures,the damage degree of each part is also different,and the vibration signals of the bearing faults often do not have good statistical characteristics as the stationary signals.In addition,the bearing is affected by factors such as working conditions,equipment types,and working environment.The formation mechanism of bearing compound fault is complicated,and it is difficult to collect compound fault data.Incomplete data type or data which missing some fault characteristics are often collected.It is too difficult to train the fault diagnosis model by incomplete fault data.Compound fault data containing unknown features and the lack of data of certain fault categories bring difficulties to the researches of fault diagnosis models.At the same time,faults are able to cause machine faults and even accidents that are difficult to recover,so faults are not allowed,and data on service faults cannot be obtained at all,which hinders the rapid development of bearing service fault diagnosis research.When faced with problems such as compound fault data containing unknown features and missing data of certain fault categories,a graph convolutional network bearing fault diagnosis model is designed.First of all,convert the original vibration signal by wavelet transform into the time-frequency figures with both time and frequency information,and then input them into the graph convolutional layers for learning.Furthermore,fine-tune and re-adjust the trained model with a small amount of untrained compound fault data which contains unknown features,and perform the re-adjust on the model parameters to construct a model for bearing fault diagnosis.Experiments show that the model has a higher accuracy in diagnosing bearing faults than other existing methods.In practical engineering application,it is difficult to obtain service fault data,aiming at the difficult problem of bearing service fault diagnosis,the bearing fault diagnosis model based on transfer learning is supposed to effectually learn fault knowledge from artificially simulated fault with sufficient data,and transfer it to real bearing service faults,and improve the accuracy of bearing service fault diagnosis.Specifically,the original vibration signals of the source and target domains are converted into time-frequency figures with both time and frequency information through wavelet transform,and input them into the graph convolutional layers for learning.The fault feature representations of the source and target domains are effectually extracted.And then calculate the Wasserstein distance between the source domain data and the target domain data distribution.By minimizing the difference in data distribution,a fault diagnosis model which is able to diagnose bearing service faults is constructed.At the same time,the model also has ability to transfer from one working condition to another,and perform the intelligent diagnosis between different component types and different working conditions.Aiming at the difficult problem of compound fault and service fault diagnosis,a complete bearing fault diagnosis system is designed,mainly for fast diagnosis of compound faults and bearing service faults with scarce data,providing users with a simple,convenient,fast and good use experience.The system adopts a hierarchical design to realize the main functions of user and administrator login,user management,data preprocessing,model training,and model prediction,which is able to quickly predict the type of faults and save training resources.