Research On Fault Diagnosis Of Rolling Bearing Based On Deep Transfer Learning

As one of the key components of rotating machinery,the rolling bearing will interfere with the normal working of rotating machinery once the faults occur under the complex condition of high speed and heavy load for a long time,and further cause economic losses and even casualties.Therefore,the rolling bearing fault diagnosis is particularly important and urgent.At present,rolling bearing fault diagnosis mainly adopts the traditional fault diagnosis method or deep learning fault diagnosis method.However,the traditional fault diagnosis method needs to select fault features artificially by constructing an extra algorithm based on understanding certain signal processing technology.And the deep learning-based fault diagnosis method requires a large number of labeled sample data,but the bearing fault samples are few and scarce in reality.Hence,the generalization ability of deep learning is insufficient when the bearing condition changes,and the application of deep learning in bearing fault diagnosis is limited.In order to solve the problem of few-shot,two bearing fault diagnosis methods based on time-frequency image of rolling bearing vibration signals are proposed in this paper,including few-shot transfer learning and metric learning methods.The model performance is tested on the bearing fault diagnosis benchmark dataset of Case Western Reserve University and Paderborn University.The specific research contents are as follows:(1)Considering the nonlinearity and environmental noise of rolling bearing vibration signals,extracting fault features from bearing vibration signals will influence the result of bearing fault recognition.Therefore,time-frequency images are used in this paper to quantify vibration signals to improve classification accuracy.From the vibration signal of rolling bearing to the time-frequency image,it can be divided into three stages.Firstly,the genetic algorithm is used to determine the penalty factor and modal number adaptively in the variational modal decomposition algorithm.Secondly,combined with the correlation coefficient,the useless modes are removed and the modes that are more relevant to the original signal are retained.Then,the instantaneous frequency characteristics of each mode of bearing vibration signal are obtained by using the pseudo WignerVille distribution,and the pseudo Wigner-Ville distribution of each mode is plotted in the same time-frequency coordinate.Finally,the time-frequency image of the rolling bearing vibration signal is obtained.(2)A classifier with good performance is often lacking in the diagnosis process because of insufficient bearing fault samples.Therefore,aiming at the problem of few-shot,this paper constructs a new relation network with deep coding ability and attention mechanism based on relational network,meta-learning,and deep transfer learning.The experimental results show that the method has better multi-task learning ability in meta-learning and better classification performance in bearing fault diagnosis.(3)Aiming at the problems of domain adaptive state recognition and artificial fault to real fault transfer of rolling bearings under variable working conditions,this paper proposes a fault diagnosis model of bearings in variable working conditions based on domain adversarial.The model takes the time-frequency image of bearing vibration signal as input,which consists of a feature encoder,a metric,and a domain discriminator.The experimental results show that the proposed model can extract the features applicable to the cross-domain of bearings under different working conditions in few-shot scenarios,and has a good performance in the case of the bearing artificial fault to the real fault transfer.