Inherent subjectivity, inefficiencies, and the substantial cost related to human-based visual inspection of high-pressure turbine (HPT) blades has driven research in alternative automated techniques. The
combination of computer vision (CV) and deep learning (DL) provides a compelling alternative. However, as DL models increase in capability, their large parameter spaces require substantial data to
achieve robust training. Therefore, this study explores the use of two Auxiliary Classifier Generative Adversarial Networks (AC-GANs) to augment proprietary datasets for two failure modes: (a) a 3-
channel red-green-blue (RGB) model for obstructed holes, and (b) a 4-channel RGB plus depth (RGBD) model for foreign object damage, with the depth reconstructed using monocular depth estimation. A Differential Evolution Optimizer (DEO) was used for hyperparameter optimization of a ResNet-18 classification target model. In the obstructed hole dataset, GAN-based augmentation showed significantly improved accuracy, recall, F1, and AUC-ROC (p < 0.05), competing with traditional methods using less augmentation. In the foreign object damage dataset, the inclusion of depth information significantly enhanced accuracy, precision, F1, and AUC-ROC performance (p < 0.05). However, the RGBD augmentation mainly resulted in a trade-off between precision and recall, without statistically significant differences (p > 0.05).

Proton-exchange Membrane (PEM) fuel cells are essential systems for hydrogen-electric powertrains in aviation, aiming to meet climate-neutral goals. However, their integration faces challenges, particularly regarding power density, reliability, and durability. This research addresses PEM fuel cell durability through Prognostics and Health Management to predict Remaining Useful Life (RUL) under static and quasi-dynamic conditions. We propose a prognostic method utilising Echo State Networks (ESNs) to manage the chaotic time-series data of PEM fuel cells, extending the Prognostic Horizon (PH) to 125 hours. Our approach involves decomposing stack voltage and current time-series data into trend, seasonal, and residual components via Seasonal and Trend decomposition using LOESS, and predicting these iteratively with ESNs optimised through Bayesian optimisation using Optuna. A comparative study found that ESNs perform best at predicting trends in single-input, single-output forecasts of current time-series, while Long Short-Term Memory networks are better at capturing seasonality and residuals. Additionally, while empirical and semi-empirical models assessed PEM fuel cell membrane health, their effectiveness in predicting RUL in combination with predicted stack voltage was limited by average degradation across cells. This study presents a robust and universal prognostic approach for PEM fuel cells, facilitating their reliable integration into aviation applications.

Recent advancements in deep learning for aircraft engine fault detection have been predominantly focused on research using simulated datasets. Despite significant progress, the gap between simulated and real-world data underscores a pressing need for models that are more applicable and adaptable to the aerospace industry. This discrepancy stems from factors such as water washes, maintenance activities, noise, and nuanced variations in operating conditions. Further complicating this issue is the lack of failure data leading to class imbalance and limiting the performance of fault classification models. In response to these challenges, this study uses Generative Adversarial Networks (GANs) to augment real-world failure data from General Electric Next Generation (GEnx) aircraft engines. New synthetic data are generated using a Wasserstein GAN with Gradient Penalty (WGAN-GP) and convolutional layers. Evaluation of GAN-generated data remains an active area of research. Accordingly, we also introduce a novel validation method based on a GEnx Gas Path Analysis model. This evaluation step revealed that the GAN could effectively generate gas path response variables that were physically meaningful and consistent with the operating conditions. Furthermore, integrating the GAN-generated data into the original dataset improved the baseline fault detection model’s F1-score by an average of 2.8%. This research also highlights the GAN’s ability to learn and reproduce degradation patterns applicable across different engine units, emphasizing its potential to overcome the challenges between engine unit-to-unit variations. Additionally, this work can potentially be extended to other engine families that require synthetic data to improve maintenance strategies.

Machine learning models have improved Prognostics and Health Management (PHM) in aviation, notably in estimating the Remaining Useful Life (RUL) of aircraft engines. However, their 'black-box' nature limits transparency, critical in safety-sensitive aviation maintenance. Explainable AI (XAI), particularly Counterfactual (CF) explanations, offers a way to explain model decisions by suggesting alternative scenarios for different outcomes. Additionally, Bayesian models enhance predictions by quantifying uncertainty, yet the combination of CF explanations and Bayesian methods is largely unexplored. This study investigates counterfactual methods within a Bayesian framework to improve the explainability of RUL estimation and improve model performance. For this, a Bayesian Long Short-Term Memory (LSTM) model was applied to the C-MAPSS data-set. This research uniquely applies CF explanations in two ways, with the goal of offering insights into how varying operational conditions could affect the RUL, and to improve the model's performance by generating additional augmented data with reduced uncertainty for the model to train on. Preliminary results show that CF explanations are able to provide insights and suggestions for RUL improvement. Also, the addition of the augmented data using the CF uncertainty reduction method has shown to improve the models predictive performance, confirming the viability of this approach as a data augmentation method.

For the remaining useful life (RUL) prediction of bearings across varying operating conditions, transfer learning models have demonstrated high accuracy. To make use of the maximum amount of bearing degradation information, multi-source domain adaptation models have been developed to enable predictions across operating conditions and bearing types. These models typically make use of vibration data in the time, frequency, and time-frequency domains. However, in practice, some systems only collect data in the frequency domain, such that prediction models developed for these systems lack access to degradation features from the time or time-frequency domains. To address this, this study proposes a multi-source domain adaptation (MSDA) model for bearing RUL prediction across operating conditions and bearing types using only frequency domain data as input data. The model employs a common feature extractor to capture domain invariant features, and domain specific regressors for RUL prediction. Domain adaptation is done by minimising the maximum mean discrepancy between features from the source and target domains, while the domain specific RUL predictions are aligned by minimising the distance between the RUL predictions. Experimental results across three scenarios demonstrate that the MSDA model achieves accurate RUL predictions. Even with only frequency domain data, the RMSE and Score are comparable to those of other advanced algorithms that use both frequency and time domain data.

This research aims to investigate and experiment on a state-of-the-art problem to treat the non-monotonic behavior of fault progression trends in predictive maintenance. Well-established algorithms and literature are researched for fault progression prognostics, however, not considerable attention has been given to
monotonic constraints at a preprocessing stage. A non-monotonic trend carries complex information which has outliers and nonessential signal values. The goal of the project is to motivate the usage of monotonic constraints to treat non-monotonic signals of a degrading component. The problem is presented as follows: Determining if the monotonic constrained method at a preprocessing step shall assist prognostics to estimate the remaining useful life of a component accurately. To explore this research, a monotonic constraint - Average Conditional Displacement (ACD) is used at the preprocessing step of a model, in comparison with regular preprocessing methods. The model is experimented on the NASAs simulated C-MAPSS datasets of turbofan engine and modelled with two prognostics algorithms. The model performance is measured
with performance metrics. The results showcase that by treating non-monotonic trends with monotonic constraints does improve the prognostics. However, they are not significantly advanced compared to other preprocessing steps.

In this research, a sustainability and cost assessment of battery health management strategies applied to Lithium batteries of an electric Unmanned Aerial Vehicle (eUAV) is performed. A mission-based strategy is proposed with the aim to elongate battery lifetime. With this strategy, the battery is charged to the estimated State of Charge (SOC) level required to complete the next flight. The mission-based strategy is compared to two other strategies: the SOC 100% strategy that always fully charges the battery before flight, and, the SOC 80% strategy that charges that battery to 80% before flying. The three strategies are tested for a variety of flight distances. The battery model is simulated using Python Battery Mathematical Modelling (PyBaMM). A Monte Carlo (MC) simulation is run to review the response to uncertainties in initial battery compositions and operating conditions. Ultimately, the strategies are evaluated on environmental impact, financial costs and flying efficiency. The results show that the mission-based strategy outperforms the SOC 100%, yielding lower emissions and costs and higher flying efficiency performance. However, depending on the range flown, the SOC 80% shows environmental, cost and flying efficiency benefits that challenge the relevance of implementing a mission-based battery health management strategy.

Overall economic assessments (OEAs) can provide a sound basis for decision-making in the areas of investments in new technologies and the application of existent technologies or operating practices. However, due to their long time horizons and complex nature, OEAs often contain many uncertain inputs, making a deterministic simulation insufficient to reflect the true value of the output. In order to incorporate these uncertainties, a systematic and efficient approach for uncertainty analysis is required. This paper sets out such a process, which consists of an iterative Uncertainty Quantification (UQ) based on importance measures for each uncertainty obtained from a Global Sensitivity Analysis (GSA). Methods for UQ and GSA are generally actively researched and well established in theory, but are infrequently applied on actual problems due to the computational and organisational complexity associated with integrative uncertainty assessments. To address this issue, the process is demonstrated on an interdisciplinary problem, namely the economic valuation of Engine Wash (EW) procedures using the cost-benefit tool LYFE. It is concluded that with this iterative uncertainty quantification procedure, the total uncertainty in the output distribution, measured using the 2.5th and 97.5th percentiles and expressed in terms of the Delta Net Present Value, is reduced from $45K - $983K to $78K - $584K. To achieve this reduction, additional modelling was carried out for only the two most important of the six uncertainties, determined using the GSA results, which illustrates the efficient allocation of modelling resources.

The demand for online shopping has grown tremendously in the last couple of years. Picnic, a major player in the online grocery industry, is struggling to achieve long-term growth within its current operations. Scheduling and planning are key drivers for maintaining operational efficiency. The Fulfilment Centre (FC) and Distribution Centre (DC) costs heavily depend on efficient operations. This research focuses on improving the scheduling process in an e-grocery FC and DC. The main objective of the model is to maximise the quality of the schedule, which is achieved through two lexicographical objectives. First, the make span is minimised to improve efficiency and to calculate the number of employees required to fulfil the workload. The make span of a schedule is defined as the time between the first scheduled activity i and the last activity j in a schedule s. Next, the number of switches between activities is reduced. The second objective is to increase overall productivity since switching moments cause slack in the operations. Two solution methods are proposed to solve the Flexible Resource Constrained Project Scheduling Problem (FRCPSP). The first solution method is a Mixed Integer Linear Programming (MILP) formulation that is solved with a Branch & Cut (B&C) algorithm. Next, a meta-heuristic is proposed named Variable Neighbourhood Search (VNS). The initial solution is computed by solving the MILP for one objective, minimising the make span. Next, the VNS uses nested neighbourhoods to modify the answer resulting in fewer switches per schedule. For small instances, the exact formulation outperforms the meta-heuristic in most cases. Conversely, the meta- heuristic features a higher efficacy and efficiency when tackling more significant instances, being the only solution method capable of yielding feasible solutions for real-world scheduling problems. Despite the effectiveness of the proposed meta-heuristic, some operational adjustments are still required before implementing the proposed decision-making tool.

The design space of power plants of naval surface vessels is ever expanding. As a result, the design space has become too large for a human design team to fully explore. Therefore, a concept exploration tool is introduced to generate and evaluate all possible concepts within the design space. An optimal solution is chosen based on ship characteristics and client preferences. This tool helps designers find the optimal solution and visualises the design space to enable a less constrained design methodology.

Impact on composite structures shows a different damage behaviour compared to metal structures. Due to the current short operational life time of composite aircraft the risks of impact damages on composite structures are unknown. This paper proposes a new method for quantitative risk analysis of low velocity impact damages on composite aircraft structures by combining a conventional risk analysis with the probability of detection of the damages. Real damage data of metal structures is used to estimate impactors and to predict damages on composite structures by means of an aircraft impact damage model. To create a large set of impact events and to conduct a more accurate analysis impactor data estimated from the real metal damage data is augmented. Three fuselage sections with a significant difference in amount of damages have been selected to compare the different risk results. The outcome of a conventional risk analysis and the outcome of the probability of detection of damages show that the section with the highest amount of damages results into the highest risk. However, the proposed method by combining the probability of detection of the damages with a conventional risk analysis shows different and more revealing results. The fuselage section with nearly 50% less damages compared to the section with the highest damages appears to be the highest risk section but the difference with the other fuselage sections is small. The proposed risk analysis method intends to be a useful tool for aircraft maintenance organisations for a different approach of assessing the risks of impact damages on composite structures.