Intelligent Diagnosis Method For Small Sample Rolling Bearings Based On Optimized Parallel Two-dimensional Convolutional Neural Network

Rolling bearings are important components commonly used in rotating machinery,and their stable operation is crucial to the safety of the equipment.Deep learning methods have the advantages of efficient feature extraction and adaptive fault diagnosis,and are of high research value in bearing fault diagnosis.However,most deep learning methods are based on onedimensional vibration signals of bearings,ignoring the importance of two-dimensional images as input.In addition,there is still much room for development of fault diagnosis research under small sample conditions and complex working conditions.Therefore,this paper proposes an intelligent fault diagnosis method for rolling bearings based on an optimised parallel twodimensional convolutional neural network,and the main research contents are as follows.(1)For the problem that it is difficult to extract important features from one-dimensional vibration signals in rolling bearing fault diagnosis,a method to convert one-dimensional signals into two-dimensional high-quality images is studied.By analysing the causes of faults and the limitations of frequency domain fault diagnosis,a strategy to achieve fault diagnosis in both the time domain and the time-frequency domain is determined.Experimental analysis is carried out with CWRU bearing data,and the Gram’s angular split-field transform and the continuous wavelet transform of the cmor basis function with the best effect are selected as the time domain and time-frequency image conversion methods respectively.(2)A bearing fault diagnosis method based on parallel two-dimensional group normalized convolutional neural network is proposed to address the problem of poor diagnosis effect of existing fault diagnosis methods under small sample conditions.Firstly,the continuous wavelet transform with Gram’s angular division field transform and cmor basis function is used to convert the one-dimensional vibration signal into time-domain and time-frequency images.Secondly,two parallel two-dimensional convolutional neural network branches are constructed to extract rich fault feature information from the time-domain and time-frequency images simultaneously.Finally,the group normalisation algorithm is invoked to group normalise the features to ensure the uniformity of feature distribution and speed up the network training.The experiments show that the method has good diagnostic stability and generalization performance under the condition of small sample data.(3)To address the problem of the poor effect of rolling bearing fault diagnosis methods under complex working conditions and high noise interference,a bearing fault diagnosis method based on parallel two-dimensional depth separable residual neural network is proposed.Firstly,the one-dimensional vibration signals are converted into two-dimensional time-domain images and time-frequency images as the input of the two-branch network.Secondly,the feature representation capability and inter-channel information propagation of the network are improved by a modified convolution module and depth-separable residual blocks,and a parallel2 D depth-separable residual neural network is built to extract the fault features of both images simultaneously.Finally,the effects of the global average pooling layer,the location of the feature fusion layer and the learning rate decay strategy on the fault diagnosis performance are analysed to determine the optimal structure of the network.The experiments show that the method has good noise immunity and generalization performance.(4)From the practical application aspect of fault diagnosis research,a rolling bearing fault diagnosis system based on deep learning technology is developed.The system is developed using MATLAB software tools,and the GUI graphical user interface is used to establish three system modules: "bearing-related parameters and data pre-processing","deep learning-based fault diagnosis module" and "fault result alarm and The system can cope with different noise environments and noise levels.The system can cope with the fault diagnosis of bearings under different noise environments and different load conditions,and realizes the intelligent diagnosis of bearings,which has certain practical application value.