Section: Overall Objectives
Overall Objectives
monitoring, system identification, online identification and detection algorithms, statistical hypotheses testing, reflectometry, infrared thermography, non destructive testing, sensors fusion, optimal sensors placement, vibrationbased structural analysis and damage detection and localization, aeronautics, civil engineering
In Summary
The objective of this team is the development of Structural Health Monitoring techniques by intrinsic coupling of statistics and thermoaeroelastic mixing modeling for the development of robust and autonomous structural health monitoring solutions of mechanical structures. The emphasis of the team is the handling of very large systems such as the recent wind energy converters currently being installed in Europe, building on the expertise acquired by the team on bridges as an example of civil engineering structure, and for aircrafts and helicopters in the context of aero elastic instability monitoring. The necessity of system identification and damage detection systems robust to environmental variations and being designed to handle a very large model dimension motivates us. As examples, the explosion in the installed number of sensors and the robustness to temperature variation will be the main focus of the team. This implies new statistical and numerical technologies as well as improvements on the modeling of the underlying physical models. Many techniques and methods originate from the mechanical community and thus exhibit a very deep understanding of the underlying physics and mechanical behavior of the structure. On the other side, system identification techniques developed within the control community are more related to data modeling and take into account the underlying random nature of measurement noise. Bringing these two communities together is the objective of this joint team between Inria and IFSTTAR. It will results hopefully in methods numerically robust, statistically efficient and also mixing modeling of both the uncertainties related to the data and the associated complex physical models related to the laws of physics and finite element models.
Damage detection in civil structures has been a main focus over the last decade. Still, those techniques need to be matured to be operable and installed on structures in operation, and thus be robust to environmental nuisances. Then, damage localization, quantification and prognosis should be in that order addressed by the team. To be precise and efficient, it requires correct mixing between signal processing, statistical analysis, Finite Elements Models (FEM) updating and a yet to be available precise modeling of the environmental effects such as temperature through 3D field reconstruction.
Theoretical and practical questions are more and more complex. For example, in civil engineering, from handling hundreds of sensors automatically during some long period of time to localize and quantify damage with or without numerical models. Very large heavily instrumented structures are yet to come and they will ask for a paradigm in how we treat them from a renewed point of view. As the structures become large and complex, also the thermal and aeroelastic (among others) models become complex. Bridges and aircrafts are the main focus of our research. Opening our expertise on new applications topics such as helicopters and wind energy converters is also part of our priorities.
Objectives
The main objectives of the team are first to pursue current algorithmic research activities, in order to accommodate stilltobedeveloped complex physical models. More precisely, we want successively

To develop statistical algorithms robust to noise and variation in the environment

To handle transient and highly varying systems under operational conditions

To consider the impact of uncertainties on the current available identification algorithms and develop efficient, robust and fast implementation of such quantities

To consider relevant non trivial thermal models for usage in rejection based structural health monitoring and more generally to mix numerical model, physical modeling and data

To develop theoretical and software tools for monitoring and localization of damages on civil structures or instability for aircrafts

To explore new paradigms for handling very large and complex structures heavily instrumented (distributed computing)

To study the characteristics of the monitored mechanic structures in terms of electromagnetic propagation, in order to develop monitoring methods based on electrical instrumentations.

To consider society concerns (damage quantification and remaining life prognosis)
Introduction to physics driven dynamical models in the context of civil engineering elastic structures
The design and maintenance of flexible structures subject to noise and vibrations is an important topic in civil and mechanical engineering. It is an important component of comfort (cars and buildings) and contributes significantly to the safety related aspects of design and maintenance (aircrafts, aerospace vehicles and payloads, longspan bridges, highrise towers... ). Requirements from these application areas are numerous and demanding.
Detailed physical models derived from first principles are developed as part of system design. These models involve the dynamics of vibrations, sometimes complemented by other physical aspects (fluidstructure interaction, aerodynamics, thermodynamics).
Laboratory and inoperation tests are performed on mockup or real structures, in order to get socalled modal models, ie to extract the modes and damping factors (these correspond to system poles), the mode shapes (corresponding eigenvectors), and loads. These results are used for updating the design model for a better fit to data, and sometimes for certification purposes (e.g. in flight domain opening for new aircrafts, reception for large bridges).
The monitoring of structures is an important activity for the system maintenance and health assessment. This is particularly important for civil structures. Damaged structures would typically exhibit often very small changes in their stiffness due to the occurrence of cracks, loss of prestressing or post tensioning, chemical reactions, evolution of the bearing behavior and most importantly scour. A key difficulty is that such system characteristics are also sensitive to environmental conditions, such as temperature effects (for civil structures), or external loads (for aircrafts). In fact these environmental effects usually dominate the effect of damage. This is why, for very critical structures such as aircrafts, detailed active inspection of the structures is performed as part of the maintenance. Of course, whenever modal information is used to localize a damage, the localization of a damage should be expressed in terms of the physical model, not in terms of the modal model used in system identification. Consequently, the following elements are encountered and must be jointly dealt with when addressing these applications: design models from the system physics, modal models used in structural identification, and, of course, data from sensors. Corresponding characteristics are given now: Design models are Finite Element models, sometimes with tens or hundreds of thousands elements, depending on professional habits which may vary from one sector to another. These models are linear if only small vibrations are considered; still, these models can be large if mediumfrequency spectrum of the load is significant. In addition, nonlinearities enter as soon as large vibrations or other physical effects (aerodynamics, thermodynamics, ..) are considered. Moreover stressstrain paths and therefore the response (and load) history comes into play.
Sensors can range from a handful of accelerometers or strain gauges, to thousands of them, if NEMS ( Nano Electro Mechanical Structures), MEMS (Microelectromechanical systems) or optical fiber sensors are used. Moreover, the sensor output can be a twodimensional matrix if electro magnet (IR (infrared), SAR, shearography ...) or other imaging technologies are used.
Multifold thermal effects
The temperature constitutes an often dominant load because it can generate a deflection as important as that due to the selfweight of a bridge. In addition, it sometimes provokes abrupt slips of bridge spans on their bearing devices, which can generate significant transient stresses as well as a permanent deformation, thus contributing to fatigue.
But it is also wellknown that the dynamic behavior of structures under monitoring can vary under the influence of several factors, including the temperature variations, because they modify the stiffness and thus the modes of vibration. As a matter of fact, depending on the boundary conditions of the structure, possibly uniform thermal variations can cause very important variations of the spectrum of the structure, up to $10\%$, because in particular of additional prestressing, not forgetting pre strain, but also because of the temperature dependence of the characteristics of materials. As an example, the stiffness of elastomeric bearing devices vary considerably in the range of extreme temperatures in some countries. Moreover, eigenfrequencies and modal shapes do not depend monotonically with temperature. Abrupt dynamical behavior may show up due to a change of boundary conditions e.g. due to limited expansion or frost bearing devices. The temperature can actually modify the number of contact points between the piles and the main span of the bridge. Thus the environmental effects can be several orders of magnitude more important than the effect of true structural damages. It will be noted that certain direct methods aiming at detecting local curvature variations stumble on the dominating impact of the thermal gradients. In the same way, the robustness and effectiveness of modelbased structural control would suffer from any unidentified modification of the vibratory behavior of the structure of interest. Consequently, it is mandatory to cure dynamic sensor outputs from thermal effects before signal processing can help with a diagnostics on the structure itself, otherwise the possibility of reliable ambient vibration monitoring of civil structures remains questionable. Despite the paramount interest this question deserves, thermal elimination still appears to challenge the SHM community.
Toward a multidisciplinary approach
Unlike previously mentioned blind approaches, successful endeavours to eliminate the temperature from subspacebased damage detection algorithms prove the relevance of relying on predictive thermomechanical models yielding the prestress state and associated strains due to temperature variations. As part of the CONSTRUCTIF project supported by the Action Concertée Incitative Sécurité Informatique of the French Ministry for Education and Research, very encouraging results in this direction were obtained and published. They were substantiated by laboratory experiments of academic type on a simple beam subjected to a known uniform temperature. Considering the international pressure toward reliable methods for thermal elimination, these preliminary results pave the ground to a new SHM paradigm. Moreover, for onedimensional problems, it was shown that real time temperature identification based on optimal control theory is possible provided the norm of the reconstructed heat flux is properly chosen. Finally, thermomechanical models of vibrating thin structures subject to thermal prestress, prestrain, geometric imperfection and damping have been extensively revisited. This project led by Inria involved IFSTTAR where the experiments were carried out. The project was over in July 2006. Note that thermomechanics of bridge piles combined with an ad hoc estimation of thermal gradients becomes of interest to practicing engineers. Thus, I4S's approach should suit advanced professional practice. Finite element analysis is also used to predict stresses and displacements of large bridges in HongKong bay .
Temperature rejection is the primary focus and challenge for I4S's SHM projects in civil engineering, like SIMS project in Canada, ISMS in Danemark or SIPRIS in France.
A recent collaboration between Inria and IFSTTAR has demonstrated the efficiency of reflectometrybased methods for health monitoring of some civil engineering structures, notably external posttensioned cables. Based on a mathematical model of electromagnetic propagation in mechanical structures, the measurement of reflected and transmitted electromagnetic waves by the monitored structures allows to detect structural failures. The interaction of such methods with those based on mechanical and thermal measurements will reinforce the multidisciplinary approach developed in our team.
Models for monitoring under environmental changes  scientific background
We will be interested in studying linear stochastic systems, more precisely, assume at hand a sequence of observations ${Y}_{n}$ measured during time,
$\left\{\begin{array}{ccc}{X}_{n+1}& =& A{X}_{n}+{V}_{n}\\ {Y}_{n}& =& H{X}_{n}+{W}_{n}\end{array}\right.$  (1) 
where ${V}_{n}$ and ${W}_{n}$ are zero mean random variables, $A$ is the transition matrix of the system, $H$ is the observation matrix between state and observation, and ${X}_{n}$ the process describing the monitored system. ${X}_{n}$ can be related to a physical process (for example, for a mechanical structure, the collection of displacements and velocities at different points). Different problems arise
1/ identify and characterize the structure of interest. It may be possible by matching a parametric model to the observed time series ${Y}_{n}$ in order to minimize some given criterion, whose minimum will be the best approximation describing the system,
2/ decide if the measured data describe a system in a so called "reference" state (the term "reference" is used in the context of fault detection, where the reference is considered to be safe) and monitor its deviations with respect of its nominal reference state.
Both problems should be addressed differently if
1/ we consider that the allocated time to measurement is large enough, resulting in a sequence of ${Y}_{n}$ whose length tends to infinity, a requirement for obtaining statistical convergence results. It corresponds to the identification and monitoring of a dynamical system with slow variations. For example, this description is well suited to the longterm monitoring of civil structures, where records can be measured during relatively (to sampling rate) large periods of time (typically many minutes or hours).
2/ we are interested in systems, whose dynamic is fast with respect to the sampling rate, most often asking for reaction in terms of seconds. It is, for example, the case for mission critical applications such as inflight control or realtime security and safety assessment. Both aeronautics and transport or utilities infrastructures are concerned. In this case, fast algorithms with samplebysample reaction are necessary.
The monitoring of mechanical structures can not be addressed without taking into account the close environment of the considered system and their interactions. Typically, monitored structures of interest do not reside in laboratory but are considered in operational conditions, undergoing temperature, wind and humidity variations, as well as traffic, water flows and other natural or manmade loads. Those variations do imply a variation of the eigenproperties of the monitored structure, variations to be separated from the damage/instability induced variations.
For example, in civil engineering, an essential problem for inoperation health monitoring of civil structures is the variation of the environment itself. Unlike laboratory experiments, civil structure modal properties change during time as temperature and humidity vary. Traffic and comparable transient events also influence the structures. Thus, structural modal properties are modified by slow low variations, as well as fast transient non stationarities. From a damage detection point of view, the former has to be detected, whereas the latter has to be neglected and should not perturb the detection. Of course, from a structural health monitoring point of view the knowledge of the true load is itself of paramount importance.
In this context, the considered perturbations will be of two kinds, either
1/ the influence of the temperature on civil structures, such as bridges or wind energy converters : as we will notice, those induced variations can be modeled by a additive component on the system stiffness matrix depending on the current temperature, as
We will then have to monitor the variations in ${K}_{struct}$ independently of the variations in ${K}_{T}$, based on some measurements generated from a system, whose stiffness matrix is $K$.
2/ the influence of the aeroelastic forces on aeronautical structures such as aircrafts or rockets and on flexible civil structures such as longspan bridges : we will see as well that this influence implies a modification of the classical mechanical equation (2)
where $(M,C,K)$ are the mass, damping and stiffness matrices of the system and $Z$ the associated vector of displacements measured on the monitored structure. In a first approximation, those quantities are related by (2). Assuming $U$ is the velocity of the system, adding $U$ dependent aeroelasticity terms, as in (3), introduces a coupling between $U$ and $(M,C,K)$.
Most of the research at Inria for a decade has been devoted to the study of subspace methods and how they handle the problems described above.
Model (2) is characterized by the following property (we formulate it for the single sensor case, to simplify notations): Let ${y}_{N}\cdots {y}_{+N}$ be the data set, where $N$ is large, and let $M,P$ sufficiently smaller than $N$ for the following objects to make sense: 1/ define the row vectors ${Y}_{k}=({y}_{k}\cdots {y}_{kM}),\leftk\right\le P$; 2/ stack the ${Y}_{k}$ on top of each other for $k=0,1,\cdots ,P$ to get the data matrix ${\mathcal{Y}}_{+}$ and stack the column vectors ${Y}_{k}^{T}$ for $k=0,1,\cdots ,P$ to get the data matrix ${\mathcal{Y}}_{}$; 3/ the product $\mathscr{H}={\mathcal{Y}}_{+}{\mathcal{Y}}_{}$ is a Hankel matrix. Then, matrix $\mathscr{H}$ on the one hand, and the observability matrix $\mathcal{O}(H,F)$ of system (2) on the other hand, possess almost identical left kernel spaces, asymptotically for $M,N$ large. This property is the basis of subspace identification methods. Extracting $\mathcal{O}(H,F)$ using some Singular Value Decomposition from $\mathscr{H}$ then $(H,F)$ from $\mathcal{O}(H,F)$ using a Least Square approach has been the foundation of the academic work on subspace methods for many years. The team focused on the numerical efficiency and consistency of those methods and their applicability on solving the problems above.
There are numerous ways to implement those methods. This approach has seen a wide acceptance in the industry and benefits from a large background in the automatic control literature. Up to now, there was a discrepancy between the a priori efficiency of the method and some not so efficient implementations of this algorithm. In practice, for the last ten years, stabilization diagrams have been used to handle the instability and the weakness with respect to noise, as well as the poor capability of those methods to determine model orders from data. Those methods implied some engineering expertise and heavy post processing to discriminate between models and noise. This complexity has led the mechanical community to adopt preferably frequency domain methods such as Polyreference LSCF. Our focus has been on improving the numerical stability of the subspace algorithms by studying how to compute the least square solution step in this algorithm. This yields to a very efficient noise free algorithm, which has provided a renewed acceptance in the mechanical engineering community for the subspace algorithms. Now we focus on improving speed and robustness of those algorithms.
Subspace methods can also be used to test whether a given data set conforms a model: just check whether this property holds, for a given pair {data, model}. Since equality holds only asymptotically, equality must be tested against some threshold $\epsilon $; tuning $\epsilon $ relies on socalled asymptotic local approach for testing between close hypotheses on long data sets — this method was introduced by Le Cam in the 70s. By using the Jacobian between pair $(H,F)$ and the modes and mode shapes, or the Finite Element Model parameters, one can localize and assess the damage.
In oder to discriminate between damage and temperature variations, we need to monitor the variations in ${K}_{struct}$ while being blind to the variations in ${K}_{T}$. In statistical terms, we must detect and diagnose changes in ${K}_{\mathrm{struct}}$ while rejecting nuisance parameter ${K}_{T}$. Several techniques were explored in the thesis of Houssein Nasser, from purely empirical approaches to (physical) model based approaches. Empirical approaches do work, but model based approaches are the most promising and constitue a focus of our future researches. This approach requires a physical model of how temperature affects stiffness in various materials. This is why a large part of our future research is devoted to the modeling of such environmental effect.
This approach has been used also for flutter monitoring in Rafik Zouari's PhD thesis for handling the aeroelastic effect.