March 4, 2022

About the speaker

Leonel Villafranca is a Graduate Research Assistant pursuing his master’s degree in Mechanical Engineering at the University of Texas Rio Grande Valley (UTRGV). Under the supervision of his advisor Dr. Noruzoliaee, he is working with Machine Learning Algorithms to predict the remaining service life of train bearings. Born in Matamoros Mexico, Leonel came to the United States to obtain his education in 2011 starting at the high-school level. During his high-school studies he realized his affinity towards the field of math and science, which ultimately made him choose Mechanical Engineering for his Bachelor Degree also at UTRGV. During his undergraduate studies he completed three internships, worked as a tutor, and was involved in volunteering activities and school organizations. Leonel graduated summa cum laude with his Bachelor’s degree in May 2020 and is expecting to graduate in May 2022 with his master’s degree.

Additional authors include Mohamadhossein Noruzoliaee and Constantine Tarawneh.

Predicting the remaining service life of railroad bearings: Leveraging machine learning and onboard sensor data

Bearing health monitoring is a crucial aspect of the railway industry to avoid human and economic losses due to derailments, unexpected downtime for maintenance, and inefficient performance. The University Transportation Center for Railway Safety (UTCRS) at The University of Texas Rio Grande Valley has made major advancements in developing algorithms and onboard sensor modules to tackle the shortcomings of the spatially dispersed wayside bearing health monitoring systems currently in use in the U.S. rail industry. By continuously monitoring the train bearing health in terms of the temperature and vibration levels of bearings tested in a laboratory setting, statistical regression models have been developed to establish functional relationships between the sensor-acquired bearing health data (i.e., response variable) with several explanatory factors that potentially influence the bearing deterioration (i.e., input variables). This will help predict bearing failure in advance, thus, affording the railroads and railcar owners the opportunity to schedule preventive maintenance cycles rather than costly reactive ones. Despite their merits, statistical models fall short of reliable prediction accuracy levels since they entail restrictive assumptions, such as a priori known functional relationship between the response and input variables. Such assumptions could be violated given the lack of full knowledge about the complex nonlinear deterioration process of bearings, which is difficult to characterize if not impossible. To tackle, a data-driven machine learning algorithm is presented, which is capable of unraveling the nonlinear deterioration model structure purely based on the bearing health data, even when the structure is not apparent. More specifically, a Gradient Boosting Machine (GBM) is trained using the vast amount of data collected at the UTCRS over the course of more than a decade. By strategically combining multiple single decision tree models instead of fitting the best single decision tree, GBM obtains a strong ensemble prediction. The trained GBM is further compared against other models to test for accuracy in determining the remaining service life of a bearing in terms of the remaining mileage.