The aviation sector is rapidly advancing towards the next generation of aircraft, expected by 2035, propelled by Artificial Intelligence (AI), cloud computing, and cybersecurity technologies. The Digital Twin is at the forefront of these advancements as it aims to integrate all these technologies. However, significant challenges arise in implementing Digital Twins, particularly for in-service aircraft with limited computational resources. This research aims to develop a power-eficient Digital Twin framework tailored for predictive maintenance. After an extensive review of the latest Digital Twin research, considerable efort was focused on creating a high-fidelity model of a critical aircraft component. Based on this model, a methodology for fault data simulation has been developed. The simulated data, generated through fault injection in a multi-physics model, will be the foundation for designing an efective machine-learning algorithm for predictive maintenance. Finally, the algorithm will be deployed on a device with limited computational resources without compromising the system's reliability.
Digital Twins are increasingly gaining traction across various industries, notably in the aerospace
sector [
Currently, the industry relies mainly on time-based maintenance protocols, scheduling maintenance
at fixed intervals regardless of the actual condition of the aircraft components. While systematic, this
method often results in either excessive maintenance, which is costly, or insuficient maintenance,
which can compromise safety. In 2022, airlines worldwide spent about $76.8 billion on maintenance,
repairs, and overhauls, accounting for 10.9% of their total operational costs [
Based on these developments, this research aims to address the challenges associated with implementing Digital Twin technology for predictive maintenance in aerospace. The following sections will introduce the open problems and detail how this research seeks to resolve them.
CEUR
ceur-ws.org
Despite the significant potential of this technology discussed in the previous section, Digital Twin is
still largely unexplored, and many challenges must be addressed:
• Lack of formal definition of Digital Twin : At the time of writing, many diferent definitions
of Digital Twins existed; however, there was no standardized methodology to efectively create a
Digital Twin for predictive maintenance in the aircraft domain. The lack of consensus complicates
the development and deployment of Digital Twin technologies, leading to variations in
implementation, functionality, and expectations. A standardized definition and framework are essential for
ensuring interoperability and consistency across diferent systems and applications [
The aforementioned challenges are at the base of the two primary approaches for building Digital
Twins: the model-driven approach and the data-driven approach [
The model-driven approach, also known as the physics-based approach, relies on creating detailed
mathematical models based on the fundamental physical principles governing the system. These
models are developed using domain knowledge and are typically derived from first-principle equations,
empirical relationships, and high-fidelity simulations. One of the significant strengths of the
modeldriven approach is its high interpretability; the mathematical equations and relationships used in the
model are based on well-understood physical laws, making the results easy to interpret and validate.
Furthermore, these models possess strong predictive power as they can forecast the system’s behavior
under various conditions, including those not directly observed in the data. The consistency and stability
of model-driven approaches are also noteworthy, as they are generally stable and consistent across
diferent scenarios due to their basis in physical principles. However, the model-driven approach is not
without its drawbacks. Developing accurate physics-based models can be time-consuming and costly,
requiring significant expertise and computational resources. These models often lack flexibility and
struggle to adapt to new data or unforeseen conditions that were not included in the original equations.
Additionally, scalability issues can arise as the complexity of the system increases, rendering the models
extremely complex and computationally intensive [
The data-driven approach leverages large datasets and advanced statistical methods, such as machine learning, to build models that learn patterns and relationships directly from the data without relying heavily on prior domain knowledge. The proposed approach is highly adaptable, with models that can quickly adjust to new data and mutable conditions. The eficiency of data-driven methods allows for relatively quick development and deployment, provided there is suficient data and computational power. Furthermore, these methods can handle large-scale data and complex systems, making them suitable for applications involving big data. Nevertheless, data-driven models have their limitations. Many of these models, particularly those based on deep learning, can act as ”black boxes,” providing little insight into how decisions are made. The accuracy and reliability of data-driven models heavily depend on the quality and representativeness of the training data, meaning that imbalanced or noisy data can significantly impair model performance. Additionally, without careful management, data-driven models can overfit the training data, leading to poor generalization of new or unseen data.
Given the complementary strengths and weaknesses of model-driven and data-driven approaches, a hybrid approach seeks to integrate both methodologies to create more robust and versatile Digital Twins. By incorporating physical principles and data-driven capabilities prediction, hybrid models can ofer better interpretability than purely data-driven models. The integration allows the model to leverage the adaptability of data-driven methods while maintaining the consistency and stability of physics-based models.
Starting from these open challenges and approaches described in the previous sections, the paper wants
to tackle these challenges by developing a hybrid Digital Twin framework (see Figure 2) delineated as
follows:
1. Design a high-fidelity model and robust methodology for faulty data simulation : The
ifrst step involves creating a detailed model of a critical aircraft component to simulate various
fault conditions. The objective is twofold: firstly, to align with the current trends in the aerospace
industry for designing and validating aircraft systems, and secondly, to enhance existing fault
simulation methodologies. The high-fidelity model will replicate real-world conditions as closely
as possible, ensuring that the simulated data accurately reflects potential faults and their impact
on the system.
2. Develop a tool for data augmentation: The step focuses on using data augmentation techniques
to improve the quality of the dataset. The goal is to make these enhanced techniques accessible
to engineers who possess strong domain expertise but have limited knowledge of artificial
intelligence. By doing so, we aim to bridge the gap between domain experts and AI, enabling
more efective utilization of AI-driven insights in the design and maintenance processes. Based
on the distribution of the model-based data, the tool will generate diverse and representative data
that can improve the training and validation of predictive maintenance models.
3. Develop and evaluate a predictive maintenance supervisor: The step focuses on designing
a predictive maintenance supervisor using state-of-the-art machine learning algorithms. The
supervisor will monitor the health of the aircraft systems in real-time, predicting potential failures
before they occur [
By using this hybrid framework, we can leverage the decades-long experience of the aeronautical
industry with model-driven design and high-fidelity software tools such as CATIA, ANSYS, Autodesk,
and Simulink. The integration will facilitate the generation of analytically created and deterministic
fault data. Consequently, this deterministic faulty and healthy data will serve as a foundation for
generating additional data through Generative AI, thereby enhancing the overall quality and reliability
of the synthetic datasets. The proposed approach wants to address the current limitations posed by the
scarcity of data, allowing traditional machine-learning techniques to perform better than deep-learning
techniques [
Our preliminary work has focused on validating system failure modes through fault simulation as a
foundational step toward developing an efective Digital Twin framework for predictive maintenance.
Fault simulation has recently gained attention for its potential to enhance predictive maintenance
strategies. However, the field is still in its early stages, and comprehensive fault libraries and standardized
methodologies are not yet fully developed. A generic methodology (see Figure 3), applicable to diferent
aircraft parts, has been proposed [
The initial results from these models demonstrated promising accuracy and reliability, suggesting their potential application in real-world scenarios. The fault simulation framework successfully replicated various failure modes, ofering valuable insights into the system’s resilience and identifying potential areas for improvement. It is important to note that the proposed methodology tightly depends on the quality of the model, and developing such high-fidelity models can be a costly and lengthy process. However, the expertise required to develop these models is well-established in the aerospace industry.
Based on the work presented in the previous section, our future work will mainly focus on the Datadriven part of the framework: 1. Enhancing Data Quality and Generating Explainable and Reliable Data: The first step will involve generating explainable and reliable data. We will develop a pipeline that engineers with limited AI knowledge but strong domain expertise can easily use, bridging the gap between modeldriven and data-driven approaches. Difusion models, a type of Generative AI, will be employed to create synthetic datasets that augment the existing data. These models are particularly efective in generating high-fidelity data that can simulate a wide range of fault conditions, thereby addressing the scarcity of failure data, which hampers traditional machine learning models. 2. Creating an Explainable Predictive Maintenance Algorithm: Subsequent eforts will focus on creating an explainable predictive maintenance algorithm to monitor the health state. Explainable AI techniques will be integrated to show the transparency and interpretability of the machine learning models used. The model’s explainability is crucial for gaining the trust of pilots, maintenance engineers, and regulatory bodies, as it allows for a clear understanding of how predictions are made and facilitates better decision-making processes. Moreover, if the algorithm is not performing well, the explainability will make it easier to identify which part of the model is incorrect, allowing for targeted adjustments and improvements to the predictive maintenance supervisor. 3. On-Board Deployment with Limited Computational Capabilities: Finally, we will deploy the proposed solution on hardware with limited computational capabilities. We will explore edge and split computing techniques to minimize space and energy requirements during in-flight operations.
For scenarios where the solution needs to operate locally without in-flight connectivity, data gathered
during flight can be securely transferred using a secure transfer protocol at the end of the mission [