Session: 07-19-01 Machine Learning and Artificial Intelligence in Dynamics and Vibrations I

Paper Number: 69994

Start Time: Tuesday, 06:50 PM

69994 - An Empirical Study of Machine Learning and Deep Learning Methods on Bearing Fault Diagnosis Benchmarks 

Rolling element bearings are critical components regarding the reliability and safety of rotating machinery. A reliable and continuous monitoring system with high prediction accuracy prevents machine downtime, increases productivity, and reduces maintenance cost. Vibration analysis via data-driven approaches is gaining attention due to advancement of low cost sensor technologies, and novel effective machine learning methods. In recent years, deep learning methods have received increasing attention from researchers and engineers because of their inherent capability of dealing with big data, mining complex representations and overcoming the disadvantage of traditional fault classification and feature selection algorithms based on hand crafted features. Many deep learning architectures, algorithms, platforms, and frameworks are being used to solve various fault diagnosis problems that previously are deemed unsolvable. However, the literature lacks a well-structured set of rules and comprehensive evaluation of the existing methods and resources and it is not clear how to choose the best algorithm for certain situations to achieve the optimal outcome. Hence, there is a need for benchmarking infrastructure that enables a fair comparison of machine learning methods concerning the performance, time, and cost of their training and inference processes. This work evaluates traditional machine learning and recent deep learning based fault classification methods based on two benchmark rolling element bearing datasets and provides a comprehensive evaluation of the methods. Based on rolling element bearing benchmarks from Case Western Reserve University Bearing Data Center and  Paderborn University Kat Data center, respectively, five well-known traditional supervised classification algorithms (i.e. Logistic Regression, Naïve Bayes, Support vector machines, K nearest Neighbors, and Decision trees), four ensemble methods (i.e. Adaptive Boosting, Extreme Gradient Boosting, Random Forest, and Stacking), and four deep learning methods and their variants (i.e. Multi Layer Perceptron, Autoencoders, Convolutional Neural Networks, and Recurrent Neural Networks) are analyzed and compared in terms of performance, computational complexity, and inference time. We explored various hyperparameter tunning techniques and the impact of the choices on deep learning model performance. Specifically, we tried three different techniques for setting learning rate and momentum (i.e., fixed, adaptive, and cyclical). To study the effect of input types, both time frequency domain statistical features and raw inputs were used. The comparisons were made based on classification accuracy, training time, and inference time. We compared the generalization ability and robustness of the methods across various working conditions. Based on the evaluation results, we discuss technical challenges and provide suggestions for method selection and improvement.

 

Presenting Author: Behnoush Rezaeianjouybari University of Missouri

Authors:

Behnoush Rezaeianjouybari University of Missouri
Yi Shang University of Missouri