A machine learning-based approach for detection of whirl instability and overheating faults in journal bearings using multi-sensor fusion method

This study proposes a novel multi-sensor fusion-based monitoring technique for the detection of overheating and oil whirl instability faults. By using this method, multi-faults can be detected as soon as possible with the minimum number of sensors with high accuracy. This technique of monitoring uses two eddy current-type proximity probes in X-Y configuration and an embedded load cell in the housing under the Babbitt layer for measuring the fluctuations of oil pressure load. A test rig consisting of a rotor with a hydrodynamic journal bearing was built. For collected data, the parameters in the time and frequency domain are extracted for four conditions including health, oil whirl fault, overheating fault, and both faults simultaneously. In this study, the t-distributed stochastic neighbor embedding (t-SNE) method is used for feature reduction, and then, obtained data are classified by a multi-class Naive Bayes classification model. Finally, the accuracy and sensitivity of the classifier are investigated and concluded that the proximity probe sensor is useful for the detection of overheating and oil whirl faults with high accuracy, but the load cell sensor just can accurately detect oil whirl fault. Accordingly, the proximity probe sensor can be used for overheating fault detection without a thermocouple sensor. Although load cell cannot detect overheating, it can be used for oil whirl fault detection.