Research On Fault Diagnosis Method Of Rolling Bearing Based On Improved Residual Neural Network

Rolling bearing plays an important role in the working process of rotating machinery and is one of the most important parts of rotating machinery.Because the working environment of rolling bearings in rotating machinery is generally closed and complex,there are problems such as aging and damage during the application process,and the failure can cause safety accidents and huge economic losses.Therefore,it is very important to diagnose the faults of rolling bearings.Because traditional methods are accompanied by some subjective factors in the process of fault diagnosis,which limit the complete extraction of fault feature information by such methods,and it is difficult to achieve better generalization performance when it deals with complex faults.However,deep learning methods can combine fault feature extraction and classification,and automatically extract representative features from the original signal data,thereby the influence of manual experience on feature extraction is eliminated.Therefore,this thesis studies the fault diagnosis method of rolling bearing based on residual neural network.The main research contents of this thesis are as follows:(1)Aiming at the problem of poor fault diagnosis caused by complex and changeable working conditions of rolling bearings,large environmental noise interference and insufficient effective data samples,a fault diagnosis method for rolling bearings based on residual neural network is proposed.First,we use three consecutive convolutional layers to construct an improved data pooling layer.The purpose is to effectively extract the fault feature information from the vibration signals and reduce the amount of calculation of the parameters in the residual neural network.Then,we design a dilated residual block that combines dilated convolution and residual block to learn fault feature information.The advantage is that the method has a strong feature learning ability by expanding the receptive field.Finally,we introduce the Dropout method to discard a certain proportion of neurons,which can effectively avoid the negative impact of overfitting.The bearing data of Case Western Reserve University set is used to verify that the proposed method in this chapter has higher diagnostic accuracy,stronger noise resistance and generalization ability in rolling bearing fault diagnosis.(2)Aiming at the problem that the traditional residual neural network is difficult to effectively extract the fault characteristic information of rolling bearing in the variable working conditions,an improved residual neural network fault diagnosis method is proposed.The method takes the time domain signals of rolling bearing as input.Aiming at the strong time-varying characteristics of time-domain signal,an improved data pooling layer based on perception module is constructed.The pooling layer is constructed by three 3×3 small convolution layers in series and stacked,and residual connection is added.The feature information can be extracted effectively.The jump connection line is constructed in the residual block,and a residual block with jump connection line is designed,which enhances the learning efficiency of the residual block on the feature information.Taking advantage of the ability of the dilated convolution to expand the receptive field,the ordinary convolution in the residual block with skip connection lines is replaced with the dilated convolution,and a hollow residual block with skip connection lines is designed.The two designed residual blocks are connected end-to-end to construct a neural network.The effectiveness of the method is verified by the data sets under 4 different working conditions published by Case Western Reserve University.(3)Because one-dimensional signals of rolling bearing are transformed into twodimensional signal,some characteristic information will be lost.To solve this problem,a fault diagnosis method of rolling bearing based on one-dimensional residual neural network is proposed.The proposed method takes one-dimensional time-domain signals of rolling bearings as input.First,a zigzag dilated convolution is constructed in the convolutional neural network to effectively improve the receptive field of the convolutional layer,and then a multi-level residual connection structure with different weight coefficients is designed to make the lower layer features which can be transferred to the upper layer,which improves the feature learning ability of the method.Finally,an attention module SE block is added after each sub-residual block to enhance useful feature channels and weaken useless feature channels,thereby it can enhance the method’s ability to identify fault characteristics.Experiments with data set of Case Western Reserve University show that this method can effectively diagnose rolling bearing faults under variable operating conditions and has a higher fault diagnosis accuracy than other deep learning methods.