Research On Bearing Fault Diagnosis Method Based On Multi-scale Feature Fusio

With the development of machinery industry technology,the combination of traditional machinery industry technology with artificial intelligence,deep learning and big data technology has developed very rapidly,and these novel technologies have added a lot of fresh power to the development of the overall industry.In machinery industry equipment,bearing is a very important rotating machinery component.It supports the shaft to work properly.At the same time,the maintenance of bearings is also an important topic in the machinery industry.If the bearing fails,it will cause the machine to stop working at light level,and cause casualties in serious cases.Therefore,it is necessary to diagnose its faults,and it is also of practical significance to carry out research on bearing fault diagnosis.Due to their own limitations,traditional fault diagnosis methods and deep learning methods have very limited ability to improve the accuracy of fault diagnosis,and the actual working environment of bearings is relatively harsh,which increases the difficulty of fault diagnosis.Therefore,it is necessary to propose the model with better fault diagnosis performance to achieve accurate fault diagnosis in different environments.Some existing fault diagnosis models do not consider the fusion of multi-scale features of input data,and their diagnosis results are not very ideal.In this case,how to improve the effect of bearing fault diagnosis based on the concept of multi-scale feature fusion is the focus of this thesis.The main research contents of this thesis are as follows:(1)Aiming at the problem that the multi-scale information in the original signal is not considered when using the bearing vibration signal for fault diagnosis,and it could lead to the poor diagnosis performance.The Discrete Wavelet Transform(DWT)and the depthwise separable convolutional neural network are studied in this thesis.A bearing fault diagnosis method based on DWT-DSCNN is presented.Firstly,the discrete wavelet transform(DWT)is used to transform the original vibration signal into multi-scale decomposition,and the decomposed multi-scale information can be used as the multi-scale features of the original vibration signal.Secondly,the decomposed multi-scale information is fused to form the input sample of multi-scale feature fusion,achieving the purpose of multi-scale feature fusion from the perspective of the original data.Finally,a lightweight neural network,deep separable convolutional neural network(DSCNN),is introduced and combined with DWT to form DWT-DSCNN model.Experimental results show that this model can achieve better fault diagnosis effect.(2)Aiming at the problem that the general fault diagnosis method only considers the features of a single scale,and does not study the multi-scale of the features,which limits the performance improvement of the model,three multi-scale feature fusion models are studied:Dense Net,Inception and Res2 Net.Firstly,this thesis gives a detailed explanation of the principles of these three multi-scale feature fusion models and the process of multi-scale feature fusion of these three models,and then sets the parameters of the models and conducts experimental analysis.The experimental results could show that the multi-scale feature fusion model is a model with excellent diagnostic performance.Different from the ordinary neural network,the multi-scale feature fusion model can realize the multi-scale feature extraction and fusion of the input data,thereby achieving an ideal diagnosis effect.(3)Aiming at the problems of multi-scale feature fusion model in complex bearing working environment,such as data sets at different speeds of bearings,there is a gap in fault diagnosis accuracy and need to be further improved,and when the bearing data sets contain noise,the diagnosis effect is poor.Based on Res2 Net network,the bearing fault diagnosis methods of RES2NET-3RA and DWT-RES2NET-3RA are proposed in this thesis.Firstly,the Residual Attention(RA)mechanism can be used to weight the features so that they can better express the features to obtain the weighted features,and then fuse them with the unweighted features to further improve the effect of multi-scale feature fusion.Secondly,three RA modules are established according to the experimental analysis,namely the Res2Net-3RA model.Finally,on the basis of the Res2Net-3RA model,this thesis fuses DWT with the Res2Net-3RA model and the DWT-Res2Net-3RA multi-scale feature fusion model is proposed,this model uses DWT to form multi-scale feature fusion samples,and combines with the Res2Net-3RA model to achieve multi-scale feature fusion based on residual attention mechanism.The experimental results show that the model can achieve good diagnosis results on all different data sets.In addition,this thesis improves the DWTRes2Net-3RA model for the case where the dataset contains noise,and conducts experimental analysis on the improved model.The experimental results show that the model has good anti-noise ability.