Performance Evaluation of Daubechies Wavelet-Based Feature Extraction for Multi-State Remaining Useful Life Prediction in Roller Bearings Using Machine Learning Algorithms

Abstract

Determining the Remaining Useful Life (RUL) in roller bearings is of utmost importance in rotary machinery. Knowing the present state and acting before a failure occurs is the most important aspect in industrial setups. This research presents an effective methodology to determine the RUL state of roller bearings by successfully using different combinations of Daubechies order and decomposition levels of Wavelet Transforms and applying machine learning methods. A dataset comprising temperature and vibration signals collected from a roller bearing test rig was developed for this study. These signals were then filtered using Butterworth bandpass filter for vibration signal filtering and moving average filter for temperature signal filtering followed by splitting the signal into overlapping windows. Then the signals are subjected to Wavelet Packet Transform followed by statistical feature extraction. In the classification phase, machine learning models such as the Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Decision Tree (DT) and Naive Bayes (NB) were used to classify the RUL state in roller bearing. While analyzing different wavelet types (db1 to db10) through seven decomposition levels, this research determined that a db4 wavelet at the third level was identified as optimal for detecting RUL state in roller bearing. The results show that Support Vector Machine (SVM) classifier achieved maximum classification accuracy of 97.68 ± 0.64%, which is higher than the other classification models used in this study. These results show that the careful calibration of wavelet parameters, combined with an efficient machine learning model can provide a reliable solution for real-time machine health monitoring and predictive maintenance of rotating equipment.