Development and experimental test of data-driven approaches for physics-informed wind-turbine digital twins

In a context of energy crisis, health monitoring of wind turbines is a major economic issue. This monitoring allows costs to be reduced by anticipating maintenance operations and optimising operating time. Classical approaches propose to exploit the measurements of a few sensors placed on the wind turbine. For example, OMA (Operational Modal Analysis) approaches allow to follow a change in the natural frequencies or modal deformations of the structure. This can be used to diagnose the presence or absence of simple defects (overall loss of stiffness, mass imbalance) which, if left untreated, can lead to accelerated fatigue of the turbine. However, these approaches are blind to the physics. Indeed, the data alone can only diagnose what the sensors "see". For example, for a wind turbine instrumented only with accelerometers, a purely data-based approach does not allow to trace back to more complex quantities such as the stresses experienced by the structure – stresses that directly influence material fatigue and thus the total lifetime. In a mixed approach combining statistical processing of measurements with a physical model, the availability of the physical model makes it possible to analyse, complete and extrapolate the measurement data. The challenge is then to have a correctly recalibrated model, able to link for example acceleration measurements to stress values. 

The aim of this thesis is to develop methodologies for monitoring the health of a wind turbine by coupling a physical model of the turbine to data measured by various sensors. The objective is to be able to map the system at any time, analyse its performance in real time, detect anomalies, etc., to anticipate maintenance operations and the remaining lifetime of the asset. This mixed approach is interesting for the engineer because it "forces" the data to follow basic physical principles. One of the major difficulties is the need to manage the uncertainty linked to the necessarily limited modelling of reality. One of the key points of this approach is therefore calibration: it is necessary to guarantee that the initial model is consistent with the measurement data, and that the instrumentation (number, type, and positioning of the sensors) of the real system effectively allows reliable monitoring of the performance metrics. An experimental validation in the laboratory on a reduced model of a wind turbine will allow to demonstrate the efficiency and the relevance of the algorithms developed during the thesis.

This thesis follows on from an IFPEN/Inria PhD thesis ending in 2023 on data analysis and the consideration of uncertainties related to noise and measurement procedures. This new thesis is dedicated to the integration of modelling aspects for an efficient and reliable health monitoring of wind turbines. The objective is to develop coupled approaches to create a wind turbine digital twin allowing the link between a simple statistical model from sensor data and a complex physical model from simulation. The thesis will be structured around the following four axes:
• Integration of the sources of uncertainty linked to the modelling: the objective here is to list the sources of uncertainty present in the numerical model, in particular the modelling errors (structural links, material properties, etc.), to study the sensitivity of the system and classify these sources of uncertainty.
• Methods of recalibration between model and data: the methods developed will make it possible to recalibrate the experimental model to the measured data (model updating) but will also be able to locate and quantify precisely the defects present in the wind turbine to monitor its fatigue state.
• Optimal placement of sensors: this axis aims to estimate the minimum needs in terms of sensors necessary to reliably associate the real system and its digital twin, and will also make it possible to know a priori the performance of a given choice of sensors for the detection of specific faults in the physical model (for a given set of sensors, it is possible to estimate which faults are identifiable or, conversely, for a certain type of fault, it is possible to estimate what instrumentation is necessary to monitor it).
• Notion of virtual sensor: this axis aims at fully exploiting the physical model to extrapolate quantities measured at points of the structure into other quantities estimated at other points of the structure (for example one can estimate a stress in the foundations of the wind turbine from vibration data recorded at the top of the mast)

The PhD student will operate a wind turbine model parameterised in DeepLines WindTM or OpenFAST software. The methods developed and/or used during the thesis will be applied first on synthetic data for validation purposes, then on the concrete case of an experimental set-up to test their robustness when "going real". It will thus be possible to test the model recalibration methods in the first instance, and then to identify faults deliberately introduced into the set-up in the second instance, while varying the instrumentation setup.

This thesis is part of a collaborative project between Inria (Institut National de Recherche en Sciences et Technologies du Numérique) and IFPEN (IFP Energies nouvelles) on the structural health monitoring of wind turbines, within the broader context of development of innovative products for industrial needs. The PhD student will have the opportunity to interact with researchers and engineers from both institutes. The supervision will be ensured by M.R. El Amri (IFPEN), E. Denimal (Inria), L. Mevel (Inria, thesis director), and J-L Pfister (IFPEN). The thesis will be carried out between IFPEN in Rueil-Malmaison and Inria in Rennes, being mainly based in Rueil-Malmaison. The PhD student will thus depend on the solid mechanics department at IFPEN and on the I4S team at Inria.

Keywords: wind energy, structural health monitoring, system identification, vibration analysis

Academic supervisor    Laurent MEVEL (senior researcher), Laboratoire I4S (Inria Rennes & Université Gustave Eiffel)
Doctoral School    N°601 Mathématiques, télécommunications, informatique, signal, systèmes, électronique
IFPEN supervisor     Jean-Lou PFISTER (PhD) , Solid Mechanics Department jean-lou.pfister@ifpen.fr
PhD location    Centre IFPEN de Rueil-Malmaison
Duration and start date    3 years, starting in fourth quarter 2023
Employer    IFP Energies nouvelles, Rueil-Malmaison, France
Academic requirements    Excellent Master/BAC+5/French Grandes écoles qualifications; with a spe-cialisation in mechanical engineering, applied mathematics, computer and information sciences, or other similar specialisations.
Language requirements    Fluency in French or English, willingness to learn French 
Other requirements     Familiarity with and interest in coding in Python or similar languages. Interest in combining mechanical and data sciences.