Subject Ph. D.

The digital transformation of industry is now at the core of the Industry 4.0 paradigm and cyber-physical systems, where industrial platforms (smart buildings, robotic production lines, flexible manufacturing systems, etc.) are becoming increasingly instrumented, interconnected, and data-driven. These environments are characterized by significant structural and functional dynamism: reconfiguration of production lines, workload variability, robot mobility, human-machine interactions, and energy fluctuations. In such a context, the ability to continuously monitor the state of the system with a high degree of accuracy is essential to ensure performance, operational reliability, predictive maintenance, and energy optimization. Furthermore, the ability to predict the future behavior of various pieces of equipment (failures, bottlenecks, etc.) is absolutely critical.

Digital Twins (DTs) are emerging as a powerful tool capable of addressing these challenges. By establishing an evolving virtual representation of a physical system, synchronized in near real time through data streams originating from the real system, a DT enables not only the supervision of the current system state but also the anticipation of its evolution, the simulation of decision- making scenarios, and the optimization of its operation. However, the effectiveness of DTs relies on a fundamental assumption: the availability of reliable, relevant, and sufficiently rich data to ensure the fidelity of the digital model. In dynamic industrial platforms, this assumption is not always satisfied.

Indeed, the data collection phase constitutes a major scientific and technological challenge. Deployed sensors are often distributed, heterogeneous, energy-constrained, and connected through wireless networks that may introduce packet losses, variable latencies, and interference. Moreover, the multiplication of measurement points generates massive volumes of data, whose transmission, storage, and processing incur significant computational and energy costs, as well as a non-negligible environmental footprint. In this context, an approach based on the exhaustive acquisition of all available data appears neither sustainable nor necessary.

The central challenge addressed by this PhD thesis is therefore that of information frugality in supplying Digital Twins of dynamic industrial systems. Information frugality is understood here as the ability to minimize the resources involved (bandwidth, energy, computing power, storage, etc.) while maintaining a level of accuracy and reliability compatible with the requirements of the physical system. The goal is to move from a volume-driven paradigm to one focused on the relevance and quality of information.

To achieve this objective, the thesis will address the following research questions:

1. Data Selection and Compression: Which data are truly necessary to ensure a faithful representation of the system within the DT? How can the most informative data be identified with respect to a given objective (diagnosis, prediction, optimization) and subsequent DT processing? This challenge may be addressed through sensor selection or the selection of data subsets, either statically or dynamically. Potential strategies may also include intelligent compression and sampling mechanisms, as well as edge computing and distributed learning approaches that enable local preprocessing at the sensor or gateway level.

2. Robust Wireless Communication: How can communication mechanisms be adapted to the constraints of industrial environments (noise, mobility, sensor density)? The selection or co-design of transmission protocols (medium access control, retransmission strategies, adaptive scheduling) must balance reliability, controlled latency, and reduced energy consumption. A joint approach involving both the application layer (DT requirements) and the network layer, together with goal-oriented communication paradigms, could enable optimization of the communication chain.

3. Uncertainty Quantification and Propagation: Every sensor measurement is affected by uncertainties arising from noise, drift, or environmental conditions. How can these uncertainties be modeled at the source? How can they be propagated throughout the DT in order to obtain not only an estimate of the system state but also an associated confidence indicator? Uncertainty may originate both from data acquisition and from data processing performed by sensors and throughout the communication process. The explicit integration of uncertainty into the data collection and modeling process, while taking the DT task into account, constitutes a key lever for improving decision-making robustness.

Beyond these research directions, the objective of this thesis is to develop a comprehensive methodological framework integrating data selection, transmission, compression, and uncertainty modeling within a coherent multi-level approach. The aim is to propose algorithmic and architectural mechanisms that jointly optimize the informational quality of the DT and the frugality of the resources employed.

Ultimately, this thesis seeks to overcome scientific challenges related to data quality and management in industrial Digital Twins by proposing an integrated approach to information frugality. The expected contributions are both theoretical and practical: the formalization of relevance metrics adapted to dynamic industrial systems, the development of robust and efficient communication strategies, contextual compression methods that incorporate modeling requirements, and probabilistic frameworks for uncertainty quantification. These contributions will be validated on representative use cases (e.g., smart buildings or robotic cells) in order to assess improvements in bandwidth usage, energy consumption, and Digital Twin accuracy.