Predictive Aircraft Maintenance: Integrating Remaining-Useful-Life Prognostics into Maintenance Optimization

J. Lee

Research output: ThesisDissertation (TU Delft)

256 Downloads (Pure)

Abstract

Current aircraft maintenance ensures safe and reliable flight operations based on inspections repeated at fixed time intervals. The time interval between inspections is often much shorter than the average life of aircraft components, in an effort to timely detect potential failures. While this approach successfully prevents most potential failures, it is not the most efficient since airlines frequently need to ground aircraft for visual inspections. Furthermore, most inspections do not find any fault; thus, nothing is actually repaired after these frequent inspections.

Predictive aircraft maintenance (PdAM) is a newly emerging approach to maintenance which is expected to be more efficient, while providing the same or higher levels of reliability. PdAM uses the data produced by the plethora of on-board sensors installed on modern aircraft to monitor the health condition of aircraft components, without the need to ground these aircraft for visual inspections. These health condition data are analyzed to predict the Remaining-Useful-Life (RUL) of aircraft components. The core idea of PdAM is to plan maintenance tasks based on the estimated RUL. PdAM is currently not fully implemented in practice, however. Regulatory bodies have only recently started to discuss the integration of aircraft health monitoring (AHM) systems into aircraft maintenance process.

This dissertation aims to identify and address the challenges in implementing PdAM. The first challenge is the lack of mathematical models to assess the performance of PdAM. Before implementing PdAM in actual aircraft, the expected performance needs to be quantified to understand the impact on reliability and cost-efficiency. Although a few studies have proposed aircraft maintenance models, these studies only consider cost as a single performance metric. However, it is clear that aircraft maintenance should also be evaluated in terms of reliability and other key performance indicators (KPIs) representing the various (and often conflicting) interests of all stakeholders involved. In this dissertation, we construct a mathematical model of PdAM to evaluate the balance in maximizing various KPIs altogether. Our model captures the stochastic degradation and failure of aircraft components, and the interactions between stakeholders during the maintenance decision making process.

The second challenge for PdAM is the lack of optimization frameworks to plan PdAM considering RUL prognostics. In the last decades, most researchers have focused on predicting the RUL of aircraft components, but only a few studies address the question of how to actually integrate RUL prognostics into maintenance planning. Aircraft maintenance planning is a very complex process that should consider different aircraft components, a fleet of aircraft, their flight schedules, the limited hangar availability, tight safety margins, and strict regulations. Considering all these together in a single optimization framework is overly demanding. Hence, in this dissertation, the optimization of the PdAM planning is performed at three levels: component level, fleet level, and strategy level.

At the component level, we propose probabilistic RUL prognostics and a deep reinforcement learning (DRL) approach for predictive maintenance planning. The probabilistic RUL prognostics estimate the probability distribution of RUL, instead of a point-estimation. This approach quantifies the uncertainty associated with RUL prognostics. Based on the estimated RUL distribution, the DRL approach determines the optimal moment to replace an aircraft component. In the case study for the maintenance of aircraft turbofan engine, the proposed DRL approach reduces the total maintenance cost by 29.3\% and prevents 94.3\% of unscheduled maintenance, compared to the case when the point-estimation of RUL is used.

At the fleet level, PdAM is planned by simultaneously considering a fleet of aircraft having multiple components. The main interest of fleet-level PdAM is to integrate RUL prognostics and operational requirements, such as the flight schedules and the limited hangar availability. We formulate these in an integer linear programming problem that minimizes the cost of fleet-level PdAM. This approach reduces the usage of hangars by grouping the schedule of maintenance tasks when the RUL of the components are similar. Considering the maintenance of aircraft landing gear brakes for a fleet of aircraft, the total maintenance cost is reduced by 20\% compared to the traditional maintenance strategies.

At the maintenance strategy level, we optimize the design parameters of PdAM, such as safety margins and thresholds of RUL, considering multiple objectives: cost-efficiency and reliability. Since this multi-objective optimization problem is computationally intensive, we propose an efficient search algorithm using Gaussian process (GP) learning models to identify Pareto optimal design parameters of PdAM. Compared to other state-of-the-art multi-objective optimization algorithms, the proposed GP learning-based algorithm identifies more Pareto optimal solutions within the same computational time. The identified Pareto front shows that PdAM using RUL prognostics dominates traditional maintenance strategies by achieving the beneficial balance between efficiency and reliability indices. With only a 1\% reduction in the efficiency index, the Pareto optimal PdAM strategy achieves a 95\% improvement in the reliability index.

The three optimization frameworks at the three different levels of PdAM are proposed and illustrated for case studies on the maintenance of aircraft engines and landing gear brakes. These case studies show three main benefits of PdAM: 1) the maintenance cost is minimized by scheduling maintenance tasks only when necessary; 2) failures and unscheduled maintenance are prevented by considering RUL prognostics; and 3) Pareto optimal performance is achieved considering the balance between reliability and efficiency.

Finally, this dissertation identifies the emerging challenges associated with the introduction of PdAM. Such challenges are often attributable to the introduction of new technologies, such as aircraft health monitoring systems, RUL prognostics algorithms, and decision support systems to plan PdAM. Based on structured brainstorming sessions with domain experts and end-users, three major challenges of future PdAM are identified: 1) the (often unknown) reliability of new technologies, 2) the timeliness and accuracy of communication between the stakeholders of the new PdAM, and 3) the end-users' trust in the new technologies.

Throughout this dissertation, we have focused on decision support systems of PdAM, in the form of optimization frameworks. These frameworks provide substantial support for the implementation of PdAM in practice. Even so, it remains future work to build users' trust in PdAM, to integrate it into strict aviation legislation, and to adopt PdAM at the business level. The strongest support for trust, legislation, and business regarding PdAM should be based on mathematical models and optimization frameworks. Therefore, this dissertation is a starting point for an informed discussion on the future of predictive aircraft maintenance.
Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • Delft University of Technology
Supervisors/Advisors
  • Mulder, M., Supervisor
  • Mitici, M.A., Advisor
Thesis sponsors
Award date7 Dec 2022
Print ISBNs978-94-9329-948-1
DOIs
Publication statusPublished - 2022

Funding

This research work is part of ReMAP project, which received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 769288.

Keywords

  • Aircraft Maintenance
  • Predictive Maintenance
  • Remaining-Useful-Life Prognostics
  • Scheduling
  • Optimization
  • Modeling and simulation

Fingerprint

Dive into the research topics of 'Predictive Aircraft Maintenance: Integrating Remaining-Useful-Life Prognostics into Maintenance Optimization'. Together they form a unique fingerprint.

Cite this