Fault recognition of pump failure based on feature fast construction and CNN

JIAO Han-hui1, HU Ming-hui1,2, WANG Xing3, FENG Kun2, SHI Bao-hu3
(1.Key Lab of Engine Health MonitoringControl and Networking of Ministry of Education, Beijing University of
Chemical Technology, Beijing 100029, China; 2.Beijing Key Laboratory of Highend Mechanical Equipment
Health Monitoring and SelfRecovery, Beijing University of Chemical Technology, Beijing 100029, China;
3.SINOPEC Marketing South China Company, Guangzhou 510180, China)

Abstract: Aiming at the problems of the huge amount of computation due to the conversion of onedimensional signal into twodimensional characteristics, when the convolutional neural network (CNN) is applied to the vibration signal analysis, the influence of input construction and different construction methods of convolution neural network on the performance of neural network was studied. Based on the characteristics of the pump vibration signal analysis, a new fast construction method was proposed to convert onedimensional vibration signals into twodimensional features. Based on the feature fast construction method and the principle of CNN, an intelligent recognition model of the pump failure was constructed. Using the data of bearing fault and unbalance fault in a petrochemical field, the fault model was tested and compared with other signal conversion methods and fault recognition models. The results indicate that they can quickly converge, and the fault identification accuracy rate is more than 95%. The model has significant advantages over other models in the accuracy of fault recognition and training efficiency.
Key words: convolutional neural network (CNN); feature fast construction; vibration signal analysis; fault diagnosis; pump failure