Periodic Reporting for period 2 - DREAM-ON (Structural damage: robust, real-time, and data-driven adaptive modeling for online control)
Periodo di rendicontazione: 2022-12-01 al 2024-05-31
The DREAM-ON project deals with the design of smart self-aware mechanical structures able to perform online monitoring of their integrity, and autonomously take anticipated actions during service before downtime or failure occurs. This refers to modern Structural Health Monitoring (SHM) which is a critical issue in several domains such as transport energy, for enhanced durability and performance, with optimized maintenance and capability to pursue operation in degraded mode. It is also a societal challenge in terms of reliability and safety.
To reach the goal and implement such an advanced integrated SHM technology, the innovative concept which is developed in the project is a synergistic and real-time dialog between the mechanical structure of interest and an on-board customized digital twin. It involves the use of advanced in situ sensing techniques, high-fidelity modeling and simulation tools, as well as powerful numerical methods, in order to perform early damage detection, accurate diagnosis and prognosis, and eventually feedback control and safe decision-making on operating range for complex engineering systems. A specific aspect of the project is on the design of consistent and efficient numerical strategies, to achieve real-time and portability constraints (edge computing), but also robustness and accuracy with respect to various uncertainty sources (modeling, measurements, environment).
The overall objective of the project is thus to develop, implement and test a light and flexible computational twining platform that combines all these previous features and dynamically interacts with an evolving mechanical structure, for integrated simulation-based monitoring capabilities and extended operational efficiency. Activities are focused on data assimilation, prediction of damage evolutions, and command synthesis, using physics-guided but also data-enhanced numerical approaches. They are conducted in line with up-to-date computational and experimental capabilities provided by micro-sensors, micro-processors, micro-controllers, and artificial intelligence.
To reach the goal and implement such an advanced integrated SHM technology, the innovative concept which is developed in the project is a synergistic and real-time dialog between the mechanical structure of interest and an on-board customized digital twin. It involves the use of advanced in situ sensing techniques, high-fidelity modeling and simulation tools, as well as powerful numerical methods, in order to perform early damage detection, accurate diagnosis and prognosis, and eventually feedback control and safe decision-making on operating range for complex engineering systems. A specific aspect of the project is on the design of consistent and efficient numerical strategies, to achieve real-time and portability constraints (edge computing), but also robustness and accuracy with respect to various uncertainty sources (modeling, measurements, environment).
The overall objective of the project is thus to develop, implement and test a light and flexible computational twining platform that combines all these previous features and dynamically interacts with an evolving mechanical structure, for integrated simulation-based monitoring capabilities and extended operational efficiency. Activities are focused on data assimilation, prediction of damage evolutions, and command synthesis, using physics-guided but also data-enhanced numerical approaches. They are conducted in line with up-to-date computational and experimental capabilities provided by micro-sensors, micro-processors, micro-controllers, and artificial intelligence.
The research work performed so far was mainly conducted through the supervision of 3 ongoing PhD theses (which respectively started in Sept. 2021, Oct. 2021, and Nov. 2022).
A first component of the work was dedicated to the design and implementation of an effective physics-guided data assimilation strategy, coupling deterministic and stochastic techniques. This strategy permits the continuous-in-time updating of the simulation, from data acquired on-the-fly, so that it remains an accurate virtual representation (digital twin) of the real system for further decision-making. The performance was compared to alternative approaches of the literature, in terms of numerical efficiency and robustness to measurement noise. The strategy was successfully applied to simple but representative structures which were manufactured in the lab. For these test cases, real data were given by distributed optic fiber sensors (DOFS), using the experimental equipment purchased in the project.
Coupled data assimilation, we also investigated multiscale adaptive modeling, that is the selection of a relevant physical model with respect to experimental information, in order to optimize the compromise between accuracy and computational cost.
A second component of the work addressed data-based model enrichment using relevant deep learning techniques. This refers to the notion of hybrid twins with correction of physics-based ignorance from data. As a main model bias in the project context is related the material response, we designed some efficient tools that involve neural networks to learn thermodynamics potentials that represent the constitutive law. The strategy, validated on various constitutive models and with various data sources, enables to dynamically enrich the initial physical model and keep a high-fidelity simulation model, which is a major requirement for accurate prediction and safe decision-making.
Eventually, a last component of the work (which started about one year ago) was about feedback control and on-the-fly command synthesis for online monitoring. Here again, AI tools were beneficially used in addition to classical control techniques (Model Predictive Control) to reach real-time computations.
The proposed control strategy, leaning on diagnosis and prognosis tools described above, was tested on representative cases with increasing complexity, and we recently started its transfer for application with a Stewart platform (multi-axial actuator) which was also purchased for the purpose of the project.
A first component of the work was dedicated to the design and implementation of an effective physics-guided data assimilation strategy, coupling deterministic and stochastic techniques. This strategy permits the continuous-in-time updating of the simulation, from data acquired on-the-fly, so that it remains an accurate virtual representation (digital twin) of the real system for further decision-making. The performance was compared to alternative approaches of the literature, in terms of numerical efficiency and robustness to measurement noise. The strategy was successfully applied to simple but representative structures which were manufactured in the lab. For these test cases, real data were given by distributed optic fiber sensors (DOFS), using the experimental equipment purchased in the project.
Coupled data assimilation, we also investigated multiscale adaptive modeling, that is the selection of a relevant physical model with respect to experimental information, in order to optimize the compromise between accuracy and computational cost.
A second component of the work addressed data-based model enrichment using relevant deep learning techniques. This refers to the notion of hybrid twins with correction of physics-based ignorance from data. As a main model bias in the project context is related the material response, we designed some efficient tools that involve neural networks to learn thermodynamics potentials that represent the constitutive law. The strategy, validated on various constitutive models and with various data sources, enables to dynamically enrich the initial physical model and keep a high-fidelity simulation model, which is a major requirement for accurate prediction and safe decision-making.
Eventually, a last component of the work (which started about one year ago) was about feedback control and on-the-fly command synthesis for online monitoring. Here again, AI tools were beneficially used in addition to classical control techniques (Model Predictive Control) to reach real-time computations.
The proposed control strategy, leaning on diagnosis and prognosis tools described above, was tested on representative cases with increasing complexity, and we recently started its transfer for application with a Stewart platform (multi-axial actuator) which was also purchased for the purpose of the project.
SHM and smart structures have been an active research topic for several decades. However, achievements are still largely academic due to difficulties of transfer to real systems. Through the research work performed so far in the project, using unconventional approaches, we already achieved major advances compared to the state-of-the-part to envision this transfer.
First, contrary to many works that use artificial intelligence alone, we developed a consistent hybrid twin strategy that promotes physics-guided but also data-enriched methods for data assimilation and control, thus making benefit of all available knowledge stored in models and data. For instance, a priori physics knowledge is inserted in the neural network architectures and loss functions which are used, which leads to much higher performance. Moreover, all methods are bias-aware, which means that they naturally integrate uncertainties contained in models and data. Eventually, we introduced on-the-fly adaptive modeling with continuous evolution of the model complexity depending on the richness of experimental information, thus leading to a smart dynamical management of computing resources. All this enables to alleviate drawbacks of purely data-based approaches in terms of interpretability, prediction capability, and robustness. It also allows for high efficiency and accuracy without considering computationally demanding models or large amounts of data, which are limited as difficult and expensive to obtain in practical engineering applications.
Related to experimental aspects, we did not work only with synthetic data, but also with real data coming from optic fiber sensing, successfully inferring noisy heterogeneous sequential measurements streams. Such an advanced high-resolution sensing technology is increasingly used in industry, it was thus important to consider it to show transfer capabilities.
Eventually, we addressed both data assimilation and command synthesis in parallel, which is almost never done in the literature as it refers to separate scientific communities. In the project, we develop multidisciplinary interactions and we mix a variety of tools and methods which come from these separate communities (computer science, mechanical engineering, data science).
It is expected that the overall methodology be validated on a representative and original lab experiment, with real-time damage tracking and monitoring to preserve the integrity of a composite structure under controlled multi-axial loading. It will use the purchased experimental equipment (DOFS interrogator & Stewart platform) for online adaptation of the operational plan to prevent the structure from operating outside a safety regime. This proof-of-concept will show the interest of the project outcomes for the design of the next-generation of SHM systems.
First, contrary to many works that use artificial intelligence alone, we developed a consistent hybrid twin strategy that promotes physics-guided but also data-enriched methods for data assimilation and control, thus making benefit of all available knowledge stored in models and data. For instance, a priori physics knowledge is inserted in the neural network architectures and loss functions which are used, which leads to much higher performance. Moreover, all methods are bias-aware, which means that they naturally integrate uncertainties contained in models and data. Eventually, we introduced on-the-fly adaptive modeling with continuous evolution of the model complexity depending on the richness of experimental information, thus leading to a smart dynamical management of computing resources. All this enables to alleviate drawbacks of purely data-based approaches in terms of interpretability, prediction capability, and robustness. It also allows for high efficiency and accuracy without considering computationally demanding models or large amounts of data, which are limited as difficult and expensive to obtain in practical engineering applications.
Related to experimental aspects, we did not work only with synthetic data, but also with real data coming from optic fiber sensing, successfully inferring noisy heterogeneous sequential measurements streams. Such an advanced high-resolution sensing technology is increasingly used in industry, it was thus important to consider it to show transfer capabilities.
Eventually, we addressed both data assimilation and command synthesis in parallel, which is almost never done in the literature as it refers to separate scientific communities. In the project, we develop multidisciplinary interactions and we mix a variety of tools and methods which come from these separate communities (computer science, mechanical engineering, data science).
It is expected that the overall methodology be validated on a representative and original lab experiment, with real-time damage tracking and monitoring to preserve the integrity of a composite structure under controlled multi-axial loading. It will use the purchased experimental equipment (DOFS interrogator & Stewart platform) for online adaptation of the operational plan to prevent the structure from operating outside a safety regime. This proof-of-concept will show the interest of the project outcomes for the design of the next-generation of SHM systems.