Scientific Publications of the University of Toulouse II Le Mirail
Not a member yet
92205 research outputs found
Sort by
Méthode PINN-DIC intégrée pour la mesure conjointe de champs de déplacements et de propriétés matériaux
No full textNational audienceCette contribution présente une méthode de mesure de champs de déplacements et de propriétés matériaux qui combine réseaux de neurones informés par la physique (PINNs) et corrélation d'images non supervisée. Il s'agit d'une formulation mixte: le déplacement et la contrainte sont approchés par des réseaux de neurones distincts. Le recours aux PINNs est pertinent lorsque les propriétés matériaux à inférer sont de grandes dimensions. D'une part ils permettent le calcul exact et efficace des opérateurs différentiels et donc de raisonner sur un modèle dans sa forme forte, offrant un cadre très versatile. D'autre part, les paramètres matériaux à inférer sont appris en parallèle de la résolution du problème direct. Ces atouts assurent un temps de calcul raisonnable, évitent la malédiction de la dimension et garantissent une bonne robustesse et stabilité au bruit. Dans un tel contexte, la méthode proposée se différentie ainsi des approches classiques telles que FEMU, EGM ou VFM.Par ailleurs, nous intégrons une métrique de corrélation d'images pour mesurer conjointement déplacements et propriétés matériaux. Les données d'entrée nécessaires sont donc, du point de vue cinématique, des images de la structure dans son état de référence et déformé. Du point de vue statique, seule la mesure de l'effort résultant appliqué à la structure est supposée connue. Enfin, pour assurer sa convergence, notre méthode repose sur plusieurs innovations clés dans sa formulation, l'architecture des réseaux de neurones ainsi que dans son algorithme de résolution.Au total, la formulation présentée est adaptée à des contextes expérimentaux concrets et a été validée pour l'identification de champs de modules élastiques sur un cas synthétique et un cas réel d'une plaque trouée en traction jusqu'à rupture
Exact Outlier Cancellation in Discrete-Time Observers via Stubborn Redesign
No full textInternational audienceThis paper addresses the problem of robust state estimation for nonlinear discrete-time systems subject to sporadic high-magnitude measurement outliers. Given a nominal convergent observer, we propose a "stubborn" redesign by saturating the output injection term with a dynamic threshold. In contrast to previous approaches that rely on scalar filters, we introduce a higher order augmentation based on a finite-memory buffer. Our redesign ensures a deadbeat recovery property, where the influence of an outlier on the saturation threshold vanishes exactly after a finite number of steps determined by the system's observability index. This allows for perfect rejection, at the steady state, of sporadic disturbances enjoying a suitable dwell-time property. In the absence of disturbances, we recover the global convergence properties of the nominal observer. The proposed redesign is illustrated through simulation on a single-link flexible-joint manipulator
Fast and Reliable Evaluation of the Distribution of Quadratic Forms of Gaussian Random Variables
No full textQuadratic forms in Gaussian random variables yield generalized noncentral chisquare distributions which appear in many test statistics. Classical cdf methods (e.g., characteristic-function inversion, saddlepoint approximations) can be effective but are not designed for reliable finite-precision computation and offer limited structural or complexity insight. We study finite-precision cdf evaluation and its binary complexity.Building on earlier low-dimensional results, we note that relevant modified Laplace transforms of the cdf are available in closed-form and are D-finite functions. This reduces the general d dimensional cdf evaluation problem to evaluating a truncated holonomic power series with positive coefficients, together with explicit tail bounds. Next, we develop a reduced-rounding-error evaluation by switching from the standard alternating-sign scalar recurrence to an equivalent coupled recurrence in which all update coefficients are nonnegative. This "positivity feature" supports a sharper forward-error analysis and improves robustness of the finiteprecision evaluation.Finally, we propose an accelerated coefficient-generation algorithm. Using a transposed multipoint-evaluation primitive, it computes collectively-via a transposed Vandermonde/remainder-tree construction-the truncated contributions in-duced by the rational factors of the holonomic equation. The remaining step, which assembles the solution by exponentiating a truncated series formed from these aggregated transforms, is carried out with FFT-based power-series exponentiation. Overall, the method achieves quasi-linear complexity in the maximum between the number of Gaussian random variables and the truncation order
Optimizing Spreading Factor and Resources Allocation in LoRaWAN: A Hybrid Approach Using Reinforcement Learning and K-means Clustering
No full textInternational audienceWith the rapid expansion of Internet of Things (IoT) devices, current LPWAN technologies are increasingly susceptible to performance deterioration. LoRaWAN stands out as a leading LPWAN solution due to its ability to support long-range, low-power communication in massive IoT deployments. However, ensuring data reliability in dense and dynamic environments while optimizing transmission settings poses significant challenges, particularly in terms of interference mitigation and energy efficiency. To overcome these limitations, this paper proposes an innovative approach that integrates K-means clustering and leverages reinforcement learning to effectively manage gateway placement and configure transmission parameters, such as spreading factors and power levels, based on local signal conditions and environmental feedback. The model explicitly accounts for the capture effect and the imperfect orthogonality of spreading factors, allowing for more realistic and robust adaptation strategies without relying on the conventional centralized Adaptive Data Rate (ADR) mechanism. The approach is validated through a real-world experimental deployment. Furthermore, we highlighted the novelty of our approach by a detailed analysis of LoRaWAN spreading factor (SF) sensitivity thresholds, and evaluated the impact of gateway densification and SF reallocation on network performance. Both experimental and simulation outcomes show significant improvements in data extraction rate (DER) and energy efficiency, particularly in large-scale, congested network scenarios
Token-Efficient Change Detection in LLM APIs
No full textRemote change detection in LLMs is a difficult problem. Existing methods are either too expensive for deployment at scale, or require initial white-box access to model weights or grey-box access to log probabilities. We aim to achieve both low cost and strict black-box operation, observing only output tokens.Our approach hinges on specific inputs we call Border Inputs, for which there exists more than one output top token. From a statistical perspective, optimal change detection depends on the model's Jacobian and the Fisher information of the output distribution. Analyzing these quantities in low-temperature regimes shows that border inputs enable powerful change detection tests.Building on this insight, we propose the Black-Box Border Input Tracking (B3IT) scheme. Extensive in-vivo and in-vitro experiments show that border inputs are easily found for non-reasoning tested endpoints, and achieve performance on par with the best available grey-box approaches. B3IT reduces costs by 30× compared to existing methods, while operating in a strict black-box setting
Landscape k-complexity of isotropic centered Gaussian fields
No full textIn large dimension, we study the asymptotic behavior of the mean number of critical points with index k below a level u for an isotropic centered Gaussian random field defined on a family of subsets of depending on d. We prove the existence of three regimes depending on the speed of growth of the volume the parameter set. In the first regime the mean number of critical points decreases exponentially with the dimension. For the second regime, there exists a critical level such that the mean number of critical points with index k below a level u with increases exponentially with the dimension d independently of the index k and decreases exponentially with d when . In the third regime, there exists a layered structure depending on the level u considered and on the index k of the critical points. This behavior is similar to the one encountered on the sphere by Auffinger et al. [5]. In the particular case of the Bargmann-Fock field, only two regimes coexist
Uncertainty handling in human-AmI interaction for opportunistic software composition
No full textInternational audienceIn Ambient Intelligence (AmI), seamless interaction between humans and AI systems is crucial for the effective utilization of smart environments. This paper investigates human–AI interaction in AmI, focusing on the management of uncertainty—situations where the AI is unsure how to choose between multiple possible outputs—within the context of Opportunistic Software Composition. The Opportunistic Composition Engine (OCE) dynamically constructs applications or assemblies from available software components, learning from human feedback to tailor applications to user preferences and context. We design, implement, and evaluate three approaches to uncertainty handling: explicitly asking the user to resolve uncertainty, offering a choice between multiple assemblies, and using a visual gradient to indicate the AI’s confidence. Each approach is implemented in a distinct version of OCE and assessed through a user study ( N = 121) measuring usability and learning performance in a 2D online simulated ambient environment. Our results show that providing clear visual or informational cues about uncertainty improves user satisfaction, while forcing users to resolve uncertainty themselves reduces usability and can hinder learning. Importantly, the other interaction strategies do not impair the system’s learning, suggesting that lightweight, system-guided handling of uncertainty can effectively support both user experience and AI adaptation. Beyond these conclusions, this work demonstrates the utility and general applicability of simulation-based technologies for prototyping and evaluating AmI applications
The lithic assemblage from the 700,000 year-old butchery site of Kalinga (Luzon Island, Philippines): New insights on technological variability in the Early Palaeolithic in Island Southeast Asia
No full textInternational audienceDespite a long history of archaeological and paleoanthropological research, our understanding of the Early Palaeolithic in Island Southeast Asia remains limited, as only a few sites with clear stratigraphic contexts - Mata Menge (Flores Island), Ngebung 2 (Java Island), Calio and Talepu (Sulawesi Island) - have been documented. This scarcity of well-stratified assemblages has led to a poor characterisation of technological variability and cultural sequences, which mainly rely on surface collections and early typological classifications/facies (e.g., Pacitanian, Cabengian, Cabalwanian, Liwanian, Arubian, etc). These classifications, mostly based on a handful of cobble artefacts, have long obscured the technological diversity of early hominin settlements in the region. The Kalinga site (Luzon Island, Philippines), dated to ca. 709 ka, provides a rare opportunity to examine an Early Palaeolithic assemblage in a secure stratigraphic context. Previous studies offered only a brief description of the lithic assemblage. In this article, we present the first comprehensive technological and structural analysis of the Kalinga lithic assemblage from two archaeological layers (units F and Y), applying analytical readings developed for West Eurasian Early Palaeolithic contexts. We also reassessed and questioned two previously defined Philippine “lithic facies” - Cabalwanian and Liwanian - based on surface finds in the Cagayan Valley since the 1930s. By comparing the technological characteristics of Kalinga artefacts with these assemblages, our aim is to clarify their variability, evaluate the validity of typological distinctions, and recontextualized Philippine lithic industries within broader Asian Palaeolithic frameworks, with particular focus on small flake technologies. This study contributes new data on early hominin technological behaviors in Southeast Asia and provides a basis for revising cultural classifications built on limited evidence
From Ranging to Sensing: Toward Robust Human Detection with UWB Measurements
No full textInternational audienceUltra-Wideband (UWB) radio has recently gainedattention as a low-power and centimeter-accurate technology forindoor localization and device-free sensing. This paper investi-gates whether standard time-of-flight (ToF) ranging measure-ments can be repurposed for reliable human-presence detectionwithout any hardware modification. A unified signal-processingpipeline was implemented on the LocURa4IoT testbed usingDecawave DWM1001 modules, combining Savitzky–Golay filter-ing for analytical precision with a lightweight moving-averagescheme for real-time multi-anchor operation. The frameworkcomputes temporal derivatives and adaptive range deviationsto detect human-induced perturbations in UWB links. Exper-iments conducted in mixed line-of-sight (LOS) and non-LOSindoor conditions demonstrate that human motion producesreproducible fluctuations exceeding the intrinsic ranging noise(3–7 cm) by more than a factor of two. The results confirmthat standard UWB ranging infrastructures can be effectivelyleveraged for robust device-free human detection and motivatefuture extensions toward CIR-based fine-grained sensing