Automating predictive maintenance: A framework for aircraft anomaly detection and forecasting
Year: 2025
Type: dissertation
Abstract: This research aims to improve the accessibility of developing predictive maintenance (PdM) tools, so that those created by domain specialists with less technical expertise in computer science can match the capabilities of those created by data scientists. This will in turn optimise one component in PdM development workflow, resulting in a greater number of tools for maintaining aircraft readiness and reducing aircraft downtime. This research asks how PdM tools can overcome industry challenges through greater automation, and whether automated approaches can match traditional methods in performance while enabling non-experts to develop effective solutions. The research community has extensively applied a wide variety of machine learning models and methods to PdM problems. The implementation of these models contrasts to the level of computer-science expertise possessed by the domain specialists who require and request these tools the most. The lack of available public data has also led to a strong bias towards turbofan engines to validate these models, opening a research gap as other important complex aircraft components such as landing gear are neglected outside of the industry. Through a combined case study and experimental approach, a novel framework for building PdM tools is proposed. Six synthetic datasets were generated using DoppelGANger models for public releases, and comparative experiments evaluated traditional methods against AutoML tools. This research demonstrate that the H2O AutoML tool can perform comparably to traditional methods for PdM, as demonstrated through an experimental comparison with the methods identified within Airbus and state-of-the-art models. The proposed framework was instantiated to adhere to Airbus components and conditions, and was validated through three Airbus supplied case studies using a combination of proprietary and synthetic datasets. This thesis concludes that with the framework integrated into the Airbus workflow, AutoML tools applied to data augmented by synthetic data can achieve similar results to traditional methods used by data scientists, and requires minimal machine learning experience to attain a suitable model for the given data. The increased training times and computational demands are negligible concerns, as this framework is designed to processes in the background, enabling domain specialists without coding experience to focus on their expertise while still producing PdM models efficiently. Several additional areas for further automation were identified, including environmental noise reduction and the use of inter-connected databases for automated data collection, which could further enhance the development of PdM tools in the industry.
Author: Stanton, Izaak
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204332016
