BEARING FAULT CLASSIFICATION USING MACHINE LEARNING

Abstract

Bearings are a critical component of most rotating machinery and are often responsible for significant failures. Accurate fault classification of ball bearings in high-speed machinery is crucial for ensuring the safety and reliability of modern industries. Machine Learning methods are gaining momentum as a sensible and cost-effective solution for bearing fault classification. This dissertation presents a comprehensive analysis of ML methodologies for predicting the operational state of a bearing, focusing on the classification of fault and healthy states using various Machine Learning (ML) techniques. The study utilizes the Case Western Reserve University Bearing dataset, containing data from a motor load of 2HP operating at 1750 RPM, with 10 different bearing health conditions ranging from normal to various faulty states. These faults include ball faults, inner race faults, and outer race faults of varying sizes. The vibration response of the bearing was processed to obtain time domain and frequency domain features, which were ranked using a Random Forest Classifier algorithm and subsequently fed into ML models. Four machine learning model - Support Vector Machine (SVM), Random Forest-Support Vector Machine (RF-SVM), Long Short-Term Memory (LSTM), and 1-Dimensional Convolutional Neural Network (1D CNN) were developed and trained. The analysis showed that the 1-D CNN generally outperformed the other models, achieving the highest precision, recall, and F1 scores for most classes. The RF-SVM model, utilizing the top eight most important features identified by the Random Forest, demonstrated higher accuracy than the traditional SVM model. The overall accuracies were SVM 88.45%, RF-SVM 91.58%, LSTM 95.7%, and the 1-D CNN model achieving the highest classification accuracy of 97.94%.