Classification Analysis of Bearing Contrived Dataset under Different Levels of Contamination

Abstract

Bearings are a common failure component found in roto-dynamic equipment. As a bearing fails, tell-tale signs in collected data indicate progressing damage, depending on the operating conditions and bearing failure mode. This paper classifies bearing damage under different damage levels and operating conditions for contamination failure and focuses on differentiating the collected signals between different contamination levels against the baseline data. A contaminate was measured and mixed into the bearing grease before applying it to the rolling elements. An increasing amount of contamination was mixed into the bearing grease to simulate progressing damage and failure mode. Five classifiers are used to diagnose the condition: Random Forest, Multilayer Perceptron, K-Nearest Neighbor, Decision Tree, and Naive Bayes. The algorithms are compared using four different metrics: weighted average, Precision, Recall, and F-Measure. The algorithms are trained to diagnose failures over multiple operating conditions to circumvent possible operation changes in the real world. The algorithms were trained on the training dataset, and the model was deployed on unseen test data to evaluate the performance of the classifiers. Random forest classifier provided the best classification results with an overall accuracy of 96 % for the test data.