Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Mechanical Engineering

Committee Chair/Advisor

Dr. Phanindra Tallapragada

Committee Member

Dr. Amin Bibo

Committee Member

Dr. Javad Velni

Committee Member

Dr. Meghashyam Panyam

Committee Member

Dr. Suyi Li


A significantly increased production of wind energy offers a path to achieve the goals of green energy policies in the United States and other countries. However, failures in wind turbines and specifically their gearboxes are higher due to their operation in unpredictable wind conditions that result in downtime and losses. Early detection of faults in wind turbines will greatly increase their reliability and commercial feasibility. Recently, data-driven fault diagnosis techniques based on deep learning have gained significant attention due to their powerful feature learning capabilities. Nonetheless, diagnosing faults in wind turbines operating under varying conditions poses a major challenge. Signal components unrelated to faults and high levels of noise obscure the signature generated by early-stage damage. To address this issue, we propose an innovative fault diagnosis framework that utilizes deep learning and leverages cyclostationary analysis of sensor data. By generating cyclic spectral coherence maps from the sensor data, we can emphasize fault-related signatures. These 2D color map representations are then used to train convolutional neural networks capable of detecting even minor faults and early-stage damages. The proposed method is evaluated using test data obtained from multibody dynamic simulations conducted under various operating conditions. The benchmark test cases, inspired by an NREL study, are successfully detected using our approach.

To further enhance the accuracy of the model, subsequent studies employ Convolutional Neural Networks with Local Interpretable Model-Agnostic Explanations (LIME). This approach aids in interpreting classifier predictions and developing an interpretable classifier by focusing on a subset range of cyclic spectral coherence maps that carry the unique fault signatures. This improvement contributes to better accuracy, especially in scenarios involving multiple faults in the gearbox that need to be identified. Moreover, to address the challenge of applying this framework in practical settings, where standard deep learning techniques tend to provide inaccurate predictions for unseen faults or unusual operating conditions, we investigate fault diagnostics using a Bayesian convolutional neural network. This approach incorporates uncertainty bounds into prediction results, reducing overconfident misclassifications. The results demonstrate the effectiveness of the Bayesian approach in fault diagnosis, offering valuable implications for condition monitoring in other rotating machinery applications.

Author ORCID Identifier




To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.