Application of data analytics for predictive maintenance in aerospace: an approach to imbalanced learning.
| dc.contributor.advisor | Jennions, Ian K. | |
| dc.contributor.advisor | King, Steve | |
| dc.contributor.author | Dangut, Maren David | |
| dc.date.accessioned | 2022-07-13T13:32:19Z | |
| dc.date.available | 2022-07-13T13:32:19Z | |
| dc.date.embargo | 2022-07-15 | |
| dc.date.issued | 2021-05 | |
| dc.description.abstract | The use of aircraft operational logs to predict potential failure that may lead to disruption poses many challenges and has yet to be fully explored. These logs are captured during each flight and contain streamed data from various aircraft subsystems relating to status and warning indicators. They may, therefore, be regarded as complex multivariate time-series data. Given that aircraft are high-integrity assets, failures are extremely rare, and hence the distribution of relevant data containing prior indicators will be highly skewed to the normal (healthy) case. This will present a significant challenge in using data-driven techniques to 'learning' relationships/patterns that depict fault scenarios since the model will be biased to the heavily weighted no-fault outcomes. This thesis aims to develop a predictive model for aircraft component failure utilising data from the aircraft central maintenance system (ACMS). The initial objective is to determine the suitability of the ACMS data for predictive maintenance modelling. An exploratory analysis of the data revealed several inherent irregularities, including an extreme data imbalance problem, irregular patterns and trends, class overlapping, and small class disjunct, all of which are significant drawbacks for traditional machine learning algorithms, resulting in low-performance models. Four novel advanced imbalanced classification techniques are developed to handle the identified data irregularities. The first algorithm focuses on pattern extraction and uses bootstrapping to oversample the minority class; the second algorithm employs the balanced calibrated hybrid ensemble technique to overcome class overlapping and small class disjunct; the third algorithm uses a derived loss function and new network architecture to handle extremely imbalanced ratios in deep neural networks; and finally, a deep reinforcement learning approach for imbalanced classification problems in log- based datasets is developed. An ACMS dataset and its accompanying maintenance records were used to validate the proposed algorithms. The research's overall finding indicates that an advanced method for handling extremely imbalanced problems using the log-based ACMS datasets is viable for developing robust data-driven predictive maintenance models for aircraft component failure. When the four implementations were compared, deep reinforcement learning (DRL) strategies, specifically the proposed double deep State-action-reward-state-action with prioritised experience reply memory (DDSARSA+PER), outperformed other methods in terms of false-positive and false-negative rates for all the components considered. The validation result further suggests that the DDSARSA+PER model is capable of predicting around 90% of aircraft component replacements with a 0.005 false-negative rate in both A330 and A320 aircraft families studied in this research | en_UK |
| dc.description.coursename | PhD in Transport Systems | en_UK |
| dc.description.notes | King, Steve (Associate Supervisor) | |
| dc.identifier.uri | https://dspace.lib.cranfield.ac.uk/handle/1826/18159 | |
| dc.language.iso | en | en_UK |
| dc.rights | © Cranfield University, 2015. All rights reserved. No part of this publication may be reproduced without the written permission of the copyright holder. | |
| dc.subject | Extremely rare event | en_UK |
| dc.subject | deep reinforcement learning | en_UK |
| dc.subject | imbalance learning | en_UK |
| dc.subject | predictive maintenance | en_UK |
| dc.subject | aircraft | en_UK |
| dc.title | Application of data analytics for predictive maintenance in aerospace: an approach to imbalanced learning. | en_UK |
| dc.type | Thesis | en_UK |