Life of Research
Appearance

Artificial Intelligence and Machine Learning for Predictive Aircraft Maintenance

  • Dominic Namachila
  1. University of Cape Town

Thesis / Dissertation, Technical University of Munich (TUM)

LOR-00203

Download PDF (1.2 MB) Cite

Cite

Artificial Intelligence and Machine Learning for Predictive Aircraft Maintenance

Read

Abstract

Unscheduled maintenance events remain one of the largest controllable sources of cost, delay and risk in commercial aviation. Conventional maintenance regimes are either calendar- and usage-based, which tends to replace components that still have useful life, or reactive, which tolerates failures that disrupt operations. Condition-based and predictive maintenance promise a middle path, but their adoption in certified, safety-critical environments is constrained by the need for accurate, calibrated and explainable prognostics that can be defended to regulators and maintenance engineers alike.

This thesis investigates artificial intelligence (AI) and machine learning (ML) methods for predicting the remaining useful life (RUL) of aircraft systems, with a primary focus on turbofan engines and a secondary focus on fusing unstructured maintenance records with sensor data. Four research questions are addressed:

How do modern deep sequence models compare with classical feature-engineered approaches under a common, leakage-free evaluation protocol?
How can predictive uncertainty be quantified and calibrated so that it can be used directly in risk-based maintenance decisions?
How can prognostic models be made interpretable to domain experts?
How well do models transfer across operating conditions, fault modes and fleets when labelled failure data are scarce?

The thesis makes four contributions:

A reproducible benchmarking framework and evaluation protocol for RUL prognostics that enforces engine-level data splitting and reports both symmetric and asymmetric (safety-weighted) metrics.
A hybrid convolutional-attention architecture with heteroscedastic output heads and conformal calibration that produces prediction intervals with finite-sample coverage guarantees.
An attribution analysis combining SHAP and attention rollout to relate model behaviour to physical degradation mechanisms.
A domain-adaptation and federated-learning study addressing data scarcity and data-sharing restrictions between operators.

Finally, a cost-aware decision layer is proposed that converts probabilistic RUL forecasts into maintenance scheduling actions. The experimental design is built around the NASA C-MAPSS and N-CMAPSS turbofan degradation datasets, and the methodology is intended to generalise to other line-replaceable units.

Keywords: predictive maintenance; prognostics and health management; remaining useful life; deep learning; uncertainty quantification; explainable AI; transfer learning; federated learning; aviation safety.

Files (1)

Cite this record

Namachila, D. (2024). Artificial Intelligence and Machine Learning for Predictive Aircraft Maintenance [Thesis, Technical University of Munich (TUM)]. Life of Research. https://lifeofresearch.com/theses/203

APA 7th edition
Export to a reference manager
BibTeX
@phdthesis{namachila2024artificial,
  author       = {Namachila, Dominic},
  title        = {{Artificial Intelligence and Machine Learning for Predictive Aircraft Maintenance}},
  year         = {2024},
  month        = jun,
  date         = {2024-06-07},
  school       = {Technical University of Munich (TUM)},
  type         = {Thesis},
  url          = {https://lifeofresearch.com/theses/203},
  abstract     = {Unscheduled maintenance events remain one of the largest controllable sources of cost, delay and risk in commercial aviation. Conventional maintenance regimes are either calendar- and usage-based, which tends to replace components that still have useful life, or reactive, which tolerates failures that disrupt operations. Condition-based and predictive maintenance promise a middle path, but their adoption in certified, safety-critical environments is constrained by the need for accurate, calibrated and explainable prognostics that can be defended to regulators and maintenance engineers alike.

This thesis investigates artificial intelligence (AI) and machine learning (ML) methods for predicting the remaining useful life (RUL) of aircraft systems, with a primary focus on turbofan engines and a secondary focus on fusing unstructured maintenance records with sensor data. Four research questions are addressed:

How do modern deep sequence models compare with classical feature-engineered approaches under a common, leakage-free evaluation protocol? How can predictive uncertainty be quantified and calibrated so that it can be used directly in risk-based maintenance decisions? How can prognostic models be made interpretable to domain experts? How well do models transfer across operating conditions, fault modes and fleets when labelled failure data are scarce?

The thesis makes four contributions:

A reproducible benchmarking framework and evaluation protocol for RUL prognostics that enforces engine-level data splitting and reports both symmetric and asymmetric (safety-weighted) metrics. A hybrid convolutional-attention architecture with heteroscedastic output heads and conformal calibration that produces prediction intervals with finite-sample coverage guarantees. An attribution analysis combining SHAP and attention rollout to relate model behaviour to physical degradation mechanisms. A domain-adaptation and federated-learning study addressing data scarcity and data-sharing restrictions between operators.

Finally, a cost-aware decision layer is proposed that converts probabilistic RUL forecasts into maintenance scheduling actions. The experimental design is built around the NASA C-MAPSS and N-CMAPSS turbofan degradation datasets, and the methodology is intended to generalise to other line-replaceable units.

Keywords: predictive maintenance; prognostics and health management; remaining useful life; deep learning; uncertainty quantification; explainable AI; transfer learning; federated learning; aviation safety.},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}

Using Zotero or Mendeley? Their browser connector saves this record from this page, with its PDF when there is one. On search and browse pages it lists every result so you can save many at once.

Guides What a complete citation needs What CC BY 4.0 allows