Failure diagnosis and physical interpretation of journal bearing for slurry liquid using long-term real vibration data

Abstract

Pumps are important machines used in rivers, social infrastructure, and industrial facilities. During long-term operation, journal bearings that support the pump shaft are subject to wear and peeling caused by liquids, including slurry. Wear and peeling can change the characteristics of journal bearings and cause abnormal shaft vibration. If wear and peeling progress, it can severely damage the pump. Thus, periodic maintenance and replacement are required. However, the frequency of periodic maintenance should be reduced as much as possible from a cost standpoint. Therefore, it is desirable to monitor the condition of the machine and perform maintenance only when necessary. In this study, the long-term vibration of a submerged journal bearing with slurry-containing water was monitored and recorded to identify the features that are important for condition monitoring and diagnosis and to interpret their contributions. First, an experimental test rig for a rotating shaft system was developed and long-term vibration data and changes in wear were recorded. A machine learning model (support vector machine (SVM)) was trained to predict the wear and damage conditions of the bearings, and its effectiveness was verified. In addition, two important features were selected as major contributors to the wear and peeling phenomena of journal bearings. These important features were interpreted using partial dependence (PD), Individual Conditional Expectation (ICE), and SHapley Additive exPlanations (SHAP), and the degree of contribution and characteristics of these features were clarified. Later, a reduced SVM model was trained using only these important features, and its effectiveness was clarified using another bearing’s data of wear and peeling processes.