Research on Construction and Application of Data-driven Incipient Fault Detection Model for Rotating Machinery

Abstract

The traditional online monitoring system could not realize incipient fault warning, and the fixed threshold grading alarm method which is used to evaluate the machine degradation status exist many false alarms and missed alarms. Excessive false alarm rate and missed alarm rate are difficult to guide the enterprises to carry out predictive maintenance of rotating machinery, and are difficult to guarantee its running safety, reliability and utilization. In order to meet the needs of engineering applications, a data-driven incipient fault detection and warning model has been built based on the technologies such as wavelet packet decomposition (WPD), dynamic kernel principal component analysis (DKPCA), T2 statistical analysis, Beta distribution control limit and so on. The incipient fault detection model has been validated by the rolling bearing vibration data from Center for Intelligent Maintenance Systems (IMS) of University of Cincinnati and by the "run to failure" online monitoring vibration data from P3409A centrifugal pump bearing of PetroChina certain hydrocracking unit. Compared with the traditional fixed threshold grading alarm method, the verified results show that the model can detect the incipient fault of rolling bearings and can realize accurate incipient fault warning, and can reduce the false alarm rate and missed alarm rate effectively. The incipient fault detection and warning model is driven by real-time vibration signals, and it works only need the historical data collected under normal operating status of key components of rotating machinery.