Abstract:  Aiming at the defects of traditional rolling bearing operating condition identification methods that required complicated manual feature extraction and features selection of vibration signals, a method based on improved spectrum segmentation empirical wavelet decomposition (ISSEWD) and self-organizing Wasserstein network (SOWN) was proposed. Firstly, the Fourier transform was employed by the collected vibration signals of rolling bearing to obtain the frequency spectrum, then a quartile method was proposed to detect the frequency spectrum boundaries, and then the signal spectrum was adaptively segmented to decompose the rolling bearing vibration signal into several intrinsic mode functions (IMFs). Secondly, the IMFs, which can best reflect the condition characteristics of the vibration signals, were screened and reconstructed. Finally, multiple Wasserstein auto-encoders were stacked to further construct the Wasserstein network, and a self-organizing strategy was introduced and the reconstructed signals were fed into SOWN for automatic feature learning and condition identification. The results indicate that the ISSEWD-SOWN model achieves a bearing condition recognition rate of 98.98% and the standard deviation is only 0.15, which is more advantageous than other methods in terms of accuracy in condition recognition and is suitable for rolling bearing operating conditions auto identification.

Key words:  rolling bearing; improved spectrum segmentation empirical wavelet decomposition (ISSEWD); condition identification; self-organizing Wasserstein network (SOWN); intrinsic mode functions(IMFs)