Advancing System-level Prognostics with Multimodal Data Integration and Uncertainty Quantification

**Advancing System-Level Prognostics with Multimodal Data Integration and Uncertainty Quantification**:

- Réf **ABG-128835**
- Sujet de Thèse
- 24/02/2025
- Financement public/privé
- Khanh Nguyen
- Lieu de travail- Tarbes - Occitanie - France
- Intitulé du sujet- Advancing System-Level Prognostics with Multimodal Data Integration and Uncertainty Quantification
- Champs scientifiques- Sciences de l’ingénieur
- Informatique
- Mots clés- Prognostics; graph neural network; multimodal learning; complex system

**Description du sujet**:
**Advancing System-Level Prognostics with Multimodal Data Integration and Uncertainty Quantification**

**1. Scientific context**

One of the primary challenges in SLP lies in effectively modeling the dependencies and interactions among system components, which significantly influence degradation and failure modes. Conventional methods, such as model-based methods, have addressed these dependencies by leveraging physical prior knowledge of system dynamics. For example, the Inoperability Input-Output Model (IIM) integrates mission profile effects and online parameter estimation, offering dynamic RUL predictions for complex systems like the Tennessee Eastman Process [3]. Similarly, state-space models with diffusion coefficient matrices effectively capture the coupling effects of degradation in multi-component systems, employing techniques like Kalman filtering and Monte Carlo simulations [4]. However, as systems grow increasingly and sensor data become more high-dimensional, these models often face scalability challenges.

To complement model-based methods, data-driven approaches have gained prominence for their ability to integrate diverse data sources. Bayesian Networks (BNs), for example, provide a robust framework for modeling probabilistic relationships while incorporating historical data, real-time sensor inputs, and expert knowledge. This capability is crucial for evaluating the health of complex systems where interactions between components play a critical role [5]. However, BNs also face limitations, including reliance on high-quality historical data, computational complexity, and scalability issues in handling intricate systems with numerous variables.

To address the limitations of model-based and data-driven methods, hybrid approaches have emerged as a promising solution by leveraging the strengths of both. For instance, Eker _et al._ [6] emphasize that hybrid methodologies enhance robustness in prognostics by accommodating variability and uncertainties in real-world operational conditions. Similarly, Li _et al._ [7] highlight the integration of deep learning algorithms, such as convolutional neural networks, with physics-based models to improve RUL estimation. However, combining data-driven and model-based methods requires meticulous calibration and validation to ensure the hybrid framework accurately captures system dynamics and degradations. Another critical gap is the lack of comprehensive uncertainty quantification at the system level. While Nguyen _et al._ [8] proposed a probabilistic deep learning methodology combining probabilistic models with deep recurrent neural networks to predict RUL distributions of components and derive system-level reliability, their study assumes independent component degradation and does not account for interactions. This simplification limits the model’s applicability to systems with complex interdependencies, underscoring the need for further research to address these limitations in hybrid prognostic frameworks.

**2. Thesis objectives**

This thesis aims to address the critical challenges in SLP by developing **advanced hybrid approaches** that allow enhancing the robustness, scalability, and reliability of prognostics algorithms, ensuring their effectiveness and adaptability in increasingly intricate engineering systems and dynamic industrial environments. Building on the foundation of prior research, the proposed methods will introduce transformative strategies for_ robust data integration, efficient modeling of component interactions, and rigorous uncertainty management_. Ultimately, the goal is to establish more accurate and scalable prognostic solutions capable of adapting to increasingly complex engineering systems and dynamic industrial environments.

In summary, the outcomes of this thesis will **primarily contribute to advancing the theoretical foundation** of hybrid prognostic modeling by bridging gaps between data-driven and physics-based approaches. This work aims to establish new paradigms in the scientific understanding of system-level prognostics and to contribute significantly to the broader research community.

**References**

[1] Khanh T.P. Nguyen, Kamal Medjaher, Do T. Tran (2023). A review of artificial intelligence methods for engineering prognostics and health management with implementation guidelines. Artificial Intelligence Review, Volume: 56, Issue: 4, Pages: 3659- 3709.

[2] Ferhat Tamssaouet, Khanh T.P. N