Fault diagnosis of rolling bearing based on singular spectrum analysis and wide convolution kernel neural network. Issue 4 (December 2022)

• Record Type: Journal Article
• Authors: Zhu, Rui; Wang, Mingxin; Xu, Siyu; li, Kai; Han, Qingpeng; Tong, Xin; He, Keyuan
• Is Part Of: Journal of low frequency noise, vibration, and active control. Volume 41:Issue 4(2022)
• Page Start: 1307
• Page End: 1321
• Publication Date: 2022-12
• Subjects: bearing crack -- fault diagnosis -- singular spectrum analysis -- noise reduction -- convolution neural network; Vibration -- Periodicals; Noise -- Periodicals; Sound -- Periodicals; Damping (Mechanics) -- Periodicals; Damping (Mechanics); Noise; Sound; Vibration; Periodicals; 620.205
• DOI: 10.1177/14613484221104639
• Languages: English
• ISSNs: 1461-3484

Abstract

Research on bearing fault diagnosis generally uses a single signal processing method to extract features to a certain extent and then empirically judges the fault type at this stage. However, this approach has a poor fault feature extraction effect when the noise interference signal is large and subjective human errors. The diagnosis effect is determined by external interference and has no practical value. This paper proposes a rolling bearing fault diagnosis method based on singular spectrum analysis and a wide convolution kernel neural network, which can effectively extract the fault features of the rolling bearing crack fault signal with a strong noise interference and realize the efficient diagnosis of this kind of fault. Gaussian white noise is added to the standard bearing fault signal to construct the vibration signal with noise interference. Then, the noisy one-dimensional time series signal is preprocessed, including the data division into the training and test sets and the different kinds of numbering processing. Singular spectrum analysis is performed on the preprocessed data. Then, the denoised training set is used as input in a deep convolution neural network with a wide convolution kernel for feature extraction and model training. The trained diagnosis model is used for the fault prediction of the next test set, and the relevant diagnosis results are output. The test results show that this method can ensure the overall accuracy of more than 90% under the background of high noise, and the diagnosis rate of the model under various working conditions is stably maintained at more than 93%, without the collapse of diagnosis stability under the sudden change of noise. The advantages of model diagnosis efficiency and structural improvement fit are prominent.