Predicting Bearing Fault Type and Depth in Rotary Machines

Fault type and depth for bearings in rotary machines were estimated from vibration data using linear regression, support vector machines, random forest, and neural network approaches. Each method was used to classify inner race, outer race and rolling element bearing faults. The accuracy of each method for differentiating between bearing fault types and estimating fault depths using vibration data envelope spectrum analysis is compared. A novel combination of envelope spectrum preprocessing with neural networks achieved prediction accuracy exceeding 99% for fault type classification, and 95% for fault depth estimation. Classification accuracy exceeding 95% was also realized using random forest methods. The results indicate that performance could be further improved with a larger neural network and additional training data.

Results:

Report for this project can be downloaded here.

Instructor: Dmitry Korkin