Research On Bearing Fault Diagnosis Algorithm Based On Spatial Decoupling Strategy And Deep Learning

Rotating machinery plays an important role in the process of industrial production and manufacturing.Rolling bearing,as an important part of rotating machinery transmission system,its operation condition directly affects industrial production efficiency and economic benefits.Therefore,the rolling bearing condition monitoring and fault diagnosis is of great significance.In recent years,the research on bearing fault based on data-driven and deep learning has developed rapidly.However,the distribution of fault information features in the data collected by sensors is not concentrated,and the nonlinear coupling of fault proximity features in space is strong,which affects the diagnostic accuracy and efficiency of deep learning algorithm.Therefore,in this paper,three bearing fault diagnosis algorithms based on spatial decoupling of data preprocessing and deep learning are designed for the effective extraction feature information and efficient diagnosis.The contents are as follows.(1)Most of the existing data-driven deep learning methods for bearing fault diagnosis focus on the complexity of the network,ignoring the spatial resolution of the time-series fault data and the discrimination of the fault category,which increases the pressure of the networks and affects the timeliness of the diagnosis.Aiming at the problem of fault proximity features decoupling and efficient diagnosis,a bearing fault diagnosis algorithm based on data cascade spatial projection(CSP)decoupling method and convolutional neural network(CNN)is proposed.The algorithm introduces the color names(CN)to map the fault data to subdivide the fault characteristics,and uses the PCA algorithm to project the data mapped by CN again to the low-dimensional subspace.So as to realize the function of reducing the dimension on the basis of extracting the main fault information,and realize the decoupling of the spatial distribution of fault data.The data preprocessed by cascade mapping is converted into two-dimensional images.And the CNN network is input to train the bearing fault diagnosis model to realize fast and accurate fault diagnosis.(2)Fault data is collected by multiple sensors from different channels,which contains multiple single-channel faults.The ability to extract complete and representative fault features from massive and high-dimensional fault data affects the diagnostic performance.At the same time,the ability of deep network to extract the spatial characteristics of fault data affects the diagnostic accuracy.Aiming at the problem of difficult feature extraction and low diagnostic accuracy of single channel data,a singlechannel bearing fault diagnosis algorithm based on kernel space decoupling and 3D deformable convolution network(3D-DCN)is proposed.BLBP and MKPCA are successively used for spatial projection preprocessing of single-channel fault data to improve the intra-class discrimination and inter-class separation of fault data and realize the spatial decoupling of fault data.The preprocessed data are converted into twodimensional graphics and input into 3D-DCN network for fault diagnosis model training.Experimental results show that the proposed algorithm can achieve 100 % diagnostic accuracy on the datasets of CWRU and XJTU-SY.(3)Under the actual operating conditions of rolling bearings,the collected fault data inevitably contain noise factors,which affects the accuracy of fault diagnosis.Aiming at the problem of the effect of mixed noise on the effectiveness of bearing fault diagnosis,a bearing fault diagnosis algorithm based on modified kernel principal component analysis(MKPCA)and residual network with deformable convolution(DC-Res Net)is proposed.MKPCA performs kernel space projection on fault signals with different signalto-noise ratios to extract the main information,eliminate some noise effects and decouple fault categories.DC-Res Net network can further filter noise effects and fully extract fault features by training preprocessed data.The experimental results show that the proposed algorithm can achieve high precision fault diagnosis under different SNRs.