Semi-Real-Time Bearing Fault Diagnosis Method Combined Image Method

Abstract

<p indent=0mm>Bearings running in the high load environment for a long time often malfunction, resulting in huge losses. It can be reduced to a large extent if the fault can be detected accurately in the early stage. According to the analysis on characteristics of the bearing fault problem, a semi-real-time and high-accuracy diagnosis method is proposed. First, deep convolutional neural networks with double paths and wider kernels (DWDCNN) are used as a real-time diagnosis method. When the result looks abnormal or the bearing is in a high-noise environment, short-time Fourier transform (STFT) is used on the vibration data of the bearing to convert it to image, and smaller inception capsule net (SICN) is used for secondary diagnosis. Then a comparison experiment between the proposed models and other existing models on Case Western Reserve University (CWRU) dataset and CWRU dataset with different intensities of noise is made on the aspects of accuracy and time performance. The result shows that DWDCNN can accomplish one prediction within <sc>0.12 ms,</sc> and the accuracy can be achieved of 80.07% under the condition of SNR=<sc>−4 dB.</sc> Although using more time, the accuracy of SICN can be achieved of 95.00% under the condition of SNR=<sc>−4 dB.</sc>