Algorithm Suite for Rotorcraft Bearing Fault Detection

To identify and classify mechanical faults, the modern tools of data science transform raw vibration data into key features from which either thresholding or learning algorithms can make predictions for damaged components. Their success relies on a complex chain of pre-processing spectral inputs. The aviation community has placed less emphasis on finding the best algorithm. We systematically evaluate 35 classifiers representing 9 major algorithmic families against features extracted from the University of Cincinnati bearing dataset. We compare the classifiers based on statistically significant differences in accuracy and execution time. Ensembles of boosted decision trees provide the best classifier family for diagnosing faulted bearings with 98% accuracy. In this study, currently favored methods which threshold single features for condition indicators offer an example of a simple one-branch partition tree, which performs significantly worse (77%) than most other classifier families.