Study On Rolling Bearing Fault Diagnosis Based On Unbalanced Data Set

Rolling bearing is a widely used part in rotating machinery.Its normal operation is critical for the smooth and safe operation of the machine.However,in actual work,the time of bearing in normal state is more than that in fault state.Aiming at the problem of data imbalance due to the lack of fault data and the difficulty to obtain in the actual work of rolling bearings,this paper proposes an expansion method of fault samples combined with signal processing and image processing technology to improve the problem of data set imbalance.In addition,according to the characteristics of biological immunity,a swarm intelligence classification algorithm based on grid immune model is proposed.The effectiveness and richness of the proposed sample expansion method and the superiority of the classification method are verified by designing multiple groups of comparative experiments.First,the time-frequency analysis of the vibration signal is carried out,the vibration signal is converted into time-frequency images by short-time Fourier method,the fault sample image is input into the depth convolution generative adversarial network network for expansion,and the expanded time-frequency image is comprehensively evaluated based on the peak signal to noise ratio and structural similarity,and the expanded samples meeting the quality requirements are obtained,Then,t-SNE visualization method is used to observe the distribution of expanded samples and original samples.Secondly,according to the phenomenon that impact will occur at the fault when the bearing works,which is characterized by sudden change of pixel value in the time-frequency map,the edge detection method is used to extract the features of the time-frequency map.This paper introduces the common edge detection operators,and uses these operators to process the time-frequency map,and compares and analyzes the detection effect.The Canny operator is improved by replacing Gaussian filter with bilateral filter,which ensures image quality while realizing time-frequency image denoising.Then the processed binary graph is taken as the feature,and the dimension is reduced by linear discriminant analysis,and the linear normalization is carried out.Thirdly,the clonal selection algorithm and negative selection algorithm commonly used in biological immune process and artificial immune are introduced.On this basis,a grid immune classification model based on cell division is proposed,and the related terms and algorithm flow are introduced in detail.Finally,the experiment is designed to compare the classification effect of the mesh immune classification model and the clonal selection model.the results show that the proposed method has more significant classification effect,and verify the superiority of the proposed classification model.Then,the grid immune model is used as the classifier,and the fault samples expanded by the proposed TF-DCGAN method are added to the original samples with unbalanced data.The classification experiments are carried out under the same working condition and a small range of variable working conditions.The experimental results show that the classification accuracy has been improved to varying degrees after adding the expanded samples,with the highest improvement of6.95%,reaching 99.88%,which proves the effectiveness of the proposed TF-DCGAN method in expanding samples.