Is this project an undergraduate, graduate, or faculty project?
Graduate
individual
What campus are you from?
Daytona Beach
Authors' Class Standing
graduate
Lead Presenter's Name
Muhammad Ali Najjar
Faculty Mentor Name
Dumindu Samaraweera
Abstract
Predictive maintenance is a critical component of aviation safety, yet the development of accurate data-driven models is often hindered by the severe class imbalance inherent in real-world flight data; healthy flights are abundant, while flights preceding a failure are rare. This imbalance biases machine learning models, leading to poor detection of critical maintenance needs. This project addresses this challenge by leveraging the NGAFID aviation maintenance dataset to develop and validate an integrated predictive framework. We propose training a Time-series Generative Adversarial Network (TimeGAN) to synthesize high-fidelity, realistic "pre-maintenance" flight data. This synthetic data is used to create a balanced training set for a deep learning classifier (e.g., LSTM), with the goal of significantly improving its predictive performance as measured by the F1-score. Furthermore, to enhance model reliability and trustworthiness for this safety-critical application, we aim to incorporate Uncertainty Quantification (e.g., Monte Carlo Dropout). This allows the model to not only predict maintenance needs but also to provide a confidence score for its predictions. The expected outcome is a robust and reliable predictive maintenance framework that demonstrates improved fault detection and provides essential uncertainty metrics for real-world decision-making.
Did this research project receive funding support from the Office of Undergraduate Research.
No
Enhancing Aviation Safety: A Framework for Predictive Maintenance using Generative Data Augmentation and Uncertainty Quantification
Predictive maintenance is a critical component of aviation safety, yet the development of accurate data-driven models is often hindered by the severe class imbalance inherent in real-world flight data; healthy flights are abundant, while flights preceding a failure are rare. This imbalance biases machine learning models, leading to poor detection of critical maintenance needs. This project addresses this challenge by leveraging the NGAFID aviation maintenance dataset to develop and validate an integrated predictive framework. We propose training a Time-series Generative Adversarial Network (TimeGAN) to synthesize high-fidelity, realistic "pre-maintenance" flight data. This synthetic data is used to create a balanced training set for a deep learning classifier (e.g., LSTM), with the goal of significantly improving its predictive performance as measured by the F1-score. Furthermore, to enhance model reliability and trustworthiness for this safety-critical application, we aim to incorporate Uncertainty Quantification (e.g., Monte Carlo Dropout). This allows the model to not only predict maintenance needs but also to provide a confidence score for its predictions. The expected outcome is a robust and reliable predictive maintenance framework that demonstrates improved fault detection and provides essential uncertainty metrics for real-world decision-making.