Architectural Framework for Federated Learning in Aviation Maintenance

Designing and Analyzing Key Features for Data Sharing Acceptance with ArchiMate Modeling

More Info
expand_more

Abstract

Executive Summary
Current Status The aviation industry is transitioning from traditional maintenance practices, typically scheduled after a specified number of flight hours, to more advanced, data-driven approaches. These predictive maintenance techniques leverage machine learning models to enhance the accuracy of assessments. These models can reduce the frequency of unscheduled maintenance and optimize inventory management. However, such models rely heavily on large datasets, which are challenging to compile in the aviation sector due to the rarity of specific operational incidents and the diverse types of data collected by different companies. When collaborating, a tradeoff arises between the level of security measures, trust in partners, and ensuring the system still functions. Federated Learning emerges as a promising solution to these challenges. Federated Learning is a novel form of machine learning that allows multiple entities to collaboratively develop a shared model while keeping their data localized, thus maintaining privacy and data sovereignty. Question The primary objective of this research is to identify and validate the critical architectural features necessary for the acceptance of Federated Learning in the aviation maintenance industry. Architectural features refer to design choices such as security protocols and governance frameworks. By focusing on these aspects, this study addresses the technical and collaborative challenges that must be overcome to develop a Federated Learning system for predictive maintenance in aviation. Resulting in the following research question:
‘What architectural features should be included in the design of the ‘Federated Learning for aircrafts’ predictive maintenance system’ to be accepted by the stakeholders in the aviation industry?’ Approach This study employs a Design Science Research methodology. The sub-questions for this research follow the steps in the Design Science Research methodology. This is not the case for the demonstration phase which was deemed not possible due to the conceptual nature of the design.
1.
What are the challenges in sharing maintenance data, and how do they impact data safety and collaboration? Through a literature review and interviews with stakeholders, the study identified concerns about data privacy, security, and competition. These challenges significantly limit data-sharing initiatives. (Problem Identification and Motivation)
2.
What specific technical requirements should the Federated Learning system have to address these challenges? Based on the interviews and thematic analysis, a list of requirements was developed. These include robust privacy mechanisms, transparent governance, ensuring model explainability, and clear accountability mechanisms. (Defining the Objectives for a Solution)
3.
What architectural features should be included in the Federated Learning design to meet these requirements? ArchiMate modeling was used to design a system incorporating these requirements. Federated Learning, combined with encryption techniques and a consortium-managed system, was the result. (Design and Development)
4.
How do stakeholders perceive the acceptability of the designed system? After the validation interviews and the expert session, three changes were implemented: Switching from
5
differentiating on the model version to differentiating on the usages. Improving explainability
by switching to Trusted Executive Environments and removing differential privacy. Traditional contracts to increase trust towards each other were also added. (Evaluation)
5.
What lessons can be learned from the development and evaluation of the Federated Learning system for future improvements? The research highlights the importance of continuously building trust among stakeholders. Furthermore, the token-based reward system based on contribution is a good incentive to be adaptable and develop long-term collaboration. (Communication) Results
To build trust, this study employs traditional methods like legal contracts and appoints a consortium as a neutral party. Transparent governance and involving a neutral, trusted entity is necessary to gain stakeholder acceptance and ensure long-term collaboration. Blockchain technology enhances the transparency of the consortium's operations, ensuring all transactions and data exchanges are recorded on an immutable database...