Multi-stage Prediction of Bearing Failure

• Main Author: Lim, Adrian Jern Ee
• Format: Final Year Project / Dissertation / Thesis
• Published: 2019
• Subjects: TJ Mechanical engineering and machinery
• Online Access: http://eprints.utar.edu.my/3463/1/ME%2D2019%2D1402577%2D1.pdf; http://eprints.utar.edu.my/3463/

Bearing is a commonly used item in many fields to reduce frictional force in rotating machineries such as manufacturing field, aerospace field and so on. Therefore, bearing plays an important role in rotating machineries where its lifespan is crucial. A bearing, which is changed way too early before it fails results in wastage of materials whereas a bearing that fails during operation causes losses due to unplanned breakdown. However, a large percentage of industries are still relying on human experience prediction to predict the lifespan of bearings, which is inefficient and inconsistent. Therefore, this work provides a series of techniques which predicts the lifespan of bearing. This series of technique includes signal processing, diagnosis as well as prognosis. It was noted that the sensors used in this project includes acoustic emission (AE) sensor, thermocouple as well as accelerometer. The main emphasize of this work would be on the steps of feature selection as well as prognosis. For feature selection, two main elements were included which is neighbourhood component analysis (NCA) as well as recursive feature elimination (RFE). RFE is mainly to exclude the unimportant features and provide carefully analysed weightage for each unique feature. NCA then make use of the weightage computed through RFE to produce a health indicator. On the other hand, prognosis uses support vector regression (SVR) to further predict the remaining useful life (RUL) of bearing. Firstly, SVR uses the health indicator to predict the RUL of each individual training test. The result generated will be then compiled into combined training test. Finally, when the combined training test is matured, the training data will be used to predict online test. This work also emphasizes on the technique used for grid search as well as cross-validation to tune the parameters. By carrying out the series of technique mentioned above, the online test conducted achieve accuracy as high as 81.95 %.