Session: 15-03-01: Machine Learning and Deep Learning in Safety, Reliability, and Maintenance I

Paper Number: 166985

Comparison of Machine Learning Algorithms for Wind Turbine Drivetrain Diagnostics Considering the Effect of Non-Torsional Loads 

As of late 2024, the United States has more than 75,000 wind turbines deployed, contributing to more than 10% (425 billion kilowatt-hours in 2023) of total U.S. utility-scale electricity generation and hence monitoring their health and keep them running is paramount. Wind Turbine drivetrains are prone to early failure, way before the operational life of the turbines, which is 20 to 25 years. Early fault detections are needed to plan for proactive maintenance before a catastrophic failure ensues. To enhance the prognostics and health management of wind turbines, a decommissioned drivetrain of a 40/60 kW wind turbine was used to setup a test-rig for experimental study of fault detection and diagnostics. The test-rig is driven by a 60 Hp motor to deliver power to the grid and simulates the torsional loading in wind turbines. Moreover, to replicate the real-life loading condition on the rotor of a wind turbine, a 5 Degree-of-Freedom Load Application Device (5 DoF-LAD) is used to apply axial and radial loadings on the low speed shaft so as to apply non-torsional loads (NTLs): namely bending (about two axis) and forces in x, y, and z direction. The NTLs will simulate the wind and directional shear loadings as well as rotor weight loading on the low-speed shaft of the drivetrain. To capture the loading phenomenon, twenty-one strain-gages, four accelerometers and one acoustic sensor are used to collect data considering three health-conditions of the drivetrain namely: baseline (healthy), damaged-1 (mild planet bearing damage) and damaged-2 (severe planet bearing damage). From the twenty-six sensors listed, data was collected at a 5 kHz sampling rate, as per Shannon-Nyquist sampling criterion, considering the highest interaction of gear tooth as the highest excitation frequency in the system. Several Machine Learning (ML) algorithms such as Decision Tree (DT), Gaussian Naïve Bayes (GNB), K-Nearest Neighbor (K-NN), Support Vector Machines (SVM), and Neural Network are used to learn the data for the three health-conditions, while the drivetrain is experiencing different torsional (power loading) and NTL conditions. The NTL loading cases tend to significantly affect the outputs of the sensors which is observed using FFT of the raw data collected from several sensors. Ultimately, the collected data is also believed to influence the learning capabilities of the ML algorithms used to learn the data. This study explores the effect of different combinations of NTLs on the learning of the ML algorithms. After training, different data sets pertaining distinct conditions were feed to the ML algorithms for prediction and results are compared. Conclusions are drown using confusion matrices and accuracy readings.

Presenting Author: Fisseha Alemayehu N/A

Presenting Author Biography: Dr. Fisseha Meresa Alemayehu is an associate professor of Mechanical Engineering at the College of Engineering of West Texas A&M University, where he has been working as a faculty since June 2016. He received his B.Sc. from Addis Ababa University, Ethiopia, his M.Sc. from TU Delft, The Netherlands and his Ph.D. from Texas Tech University, all in Mechanical Engineering. He has worked as Post-Doctoral Research Associate at Texas Tech University for one year and as an Assistant Professor at Penn State Hazleton, for two years.

Authors:

Fisseha Alemayehu N/A
Fredrick Ayivor West Texas A&M University
Neal Byron West Texas A&M University
Aaron Trevizo West Texas A&M University
Fatemehsadat Tabei West Texas A&M University
Behnam Askarian West Texas A&M University