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

Paper Number: 167289

Fault Diagnosis in Wind Turbine Gearboxes: Enhancing Reliability With Sensor Fusion 

Wind turbines are vital to renewable energy, yet gearbox failures remain a significant challenge, often occurring within the first half of their expected lifespan. Bearings, particularly in the high-speed and planetary stages, suffer from fatigue, micro-pitting, and white-etching cracks, while planetary gears experience load imbalances, tooth cracks, excessive wear, and pitting. These failures lead to costly repairs and extended downtime, making early fault detection and classification essential. In this study, a 40/60 kW wind turbine drivetrain test rig was developed from a decommissioned small-scale turbine for Prognostics and Health Management. The system, driven by a 60 HP motor, simulates input torque and integrates five degrees of freedom load application device to replicate real-world wind shear and rotor weight effects. The drivetrain was tested under healthy conditions, and different bearing faults were induced in the outer race of planet gears. At the same time, strain gauge sensors were placed on the ring gear of the planetary gearbox, an acoustic sensor was placed on the frame of the drivetrain, and accelerometers were positioned at the top and side of the gearbox and generator to read the generated signals, with data sampled at 5 kHz, adhering to Shannon-Nyquist sampling theory​. While strain gauges are essential for detecting torsional stress and structural deformations, relying on a single strain gauge sensor limits the ability to capture the complete stress distribution across the gearbox. Additionally, strain gauge sensors are generally less susceptible to the superposition of various machine-induced phenomena. Similarly, accelerometers mounted on the generator can effectively detect global vibration anomalies associated with bearing and gear defects. However, accelerometers lack precision in measuring localized stress due to the superposition of multiple vibration patterns. Fusing these two sensors compensates for their limitations, ensuring a comprehensive fault detection approach. Hence, this study aims to answer the research question, “Can fusion of limited multimodal sensor data enhance fault diagnostics in wind turbine gearboxes?” The objectives of this paper are (1) to Identify the most sensitive sensor locations and (2) to classify faults based on severity. A short-time Fourier transform was employed for time-frequency representation to improve signal analysis, enabling the detection of transient faults. The power spectrum, derived from frequency domain analysis, identified dominant fault-related frequencies, enhancing diagnostic accuracy. A fuzzy entropy-based adaptive weighting approach was introduced for sensor fusion, dynamically assigning weights to strain and vibration features. Unlike traditional fusion methods, this approach ensures that features contributing more to fault signatures receive higher importance, creating a robust health indicator. For classification, multiple machine learning (ML) algorithms such as support vector machines, decision trees, random forests, and k-nearest neighbors are compared, improving early failure prediction and optimizing maintenance scheduling. Combining adequate signal processing, fuzzy entropy-based weighting, and comparative ML classification, this fusion-based methodology enhances wind turbine fault detection to reduce maintenance costs and extend the turbine’s lifespan.

Presenting Author: Omodolapo Ajayi Texas Tech University

Presenting Author Biography: ...

Authors:

Omodolapo Ajayi Texas Tech University
Stephen Ekwaro-Osire Texas Tech Univ
Onur Can Kalay Texas Tech University
Fisseha Alemayehu West Texas A&M University
Olympio Belli Texas Tech University