关键词: 滚动轴承, 小样本, 短时傅里叶变换, 同步压缩小波变换, 二维卷积神经网络, 支持向量机, t-SNE可视化

Abstract: In order to solve the problem of low fault diagnosis accuracy of small sample rolling bearings,the STFT-SWT-2DCNN-SVM model was proposed. Combining the fixed window characteristics of STFT and the multi-scale analysis ability of SWT,the model converts the signal into a time-frequency graph,extracts the fault features and optimizes the model through 2DCNN,and then fuses the training model of STFT and SWT for verification,and finally outputs the results with SVM classifier and visualizes them with t-SNE. The PT400 experimental platform was used to carry out the relevant experimental verification. The results show that the bearing fault classification and recognition accuracy rate of this method can reach 98.666 7 % under the condition of 300∶600 training set∶test set,and under the premise that the total number of training samples is 100 and the average number of single-class samples is 5.6,the bearing fault classification and recognition accuracy rate can still reach 97.45 %,which is 15.32 %,39.23 % and 44.91 % higher than that of STFT-SWT-2DCNN,SWT-2DCNN-SVM and STFT-2DCNN-SVM,respectively. It is proved that the research model has good robustness and industrial practicability.

Key words: rolling bearing, small samples, short-term Fourier transform, synchronous compression wavelet transform, 2DCNN, SVM, t-SNE visualization