HAL Arts et Métiers
Not a member yet
14127 research outputs found
Sort by
Cadre méthodologique pour le diagnostic robuste des défauts process et produit en fonction des connaissances modèles et données disponibles
No full textThis thesis proposes an integrated framework to design, adapt, and evaluate fault detection–diagnosis systems for complex industrial processes. It begins with a structured analysis of the production system to identify potential faults and inventory available resources. On this basis, it selects and adapts the most appropriate family of methods given the constraints and the actual availability of information; in the industrial case study, this leads to semi-supervised approaches, enhanced by leveraging internal variables from model predictive control and by coupling an autoencoder with a classifier to overcome certain diagnostic limitations. The thesis then adapts the detection threshold using conformal prediction to formally control the false alarm rate without strong distributional assumptions. Finally, it revisits evaluation to ensure post-deployment durability: beyond the sole detection rate and false alarm rate, the false discovery rate is introduced and controlled to reduce alarm fatigue. The overall approach is validated on both an industrial and an academic case study.Cette thèse propose une démarche intégrée pour concevoir, adapter et évaluer des systèmes de détection-diagnostic dans les procédés industriels complexes. Elle débute par une analyse structurée du système de production afin d’identifier les défauts potentiels et d’inventorier les ressources. Sur cette base, elle sélectionne et adapte la famille de méthodes la plus pertinente au regard des contraintes et de la disponibilité réelle des informations ; dans le cas d’étude industriel, cela conduit aux approches semi-supervisées, enrichies par l’exploitation des variables internes de la commande prédictive et par un couplage entre autoencodeur et classifieur pour lever certaines limites diagnostiques. La thèse propose ensuite d’adapter le seuil de détection via la prédiction conforme pour contrôler formellement le taux de fausses alarmes sans hypothèses fortes de distribution. Enfin, elle réexamine l’évaluation pour garantir la durabilité après déploiement : au-delà du seul taux de détection et du taux de fausses alarmes, le taux de fausses découvertes est introduit et contrôlé afin de réduire la fatigue d’alarme. L’ensemble est validé sur un cas d’étude industriel et un cas d’étude académique
Estimating differential pistons for the Extremely Large Telescope using focal plane imaging and a residual network
No full textInternational audienceAs the Extremely Large Telescope (ELT) approaches operational status, optimising its imaging performance is critical. A differential piston, arising from either the adaptive optics (AO) control loop, thermomechanical effects, or other sources, significantly degrades the image quality and is detrimental to the telescope’s overall performance. Aims. In a numerical simulation set-up, we propose a method for estimating the differential piston between the petals of the ELT’s M4 mirror using images from a 2 × 2 Shack-Hartmann wavefront sensor (SH-WFS), commonly used in the ELT’s tomographic AO mode. We aim to identify the limitations of this approach by evaluating its sensitivity to various observing conditions and sources of noise. Methods. Using a deep learning model based on a ResNet architecture, we trained a neural network (NN) on simulated datasets to estimate the differential piston. We assessed the robustness of the method under various conditions, including variations in Strehl ratio, polychromaticity, and detector noise. The performance was quantified using the root mean square error (RMSE) of the estimated differential piston aberration. Results. This method demonstrates the ability to extract differential piston information from 2 × 2 SH-WFS images. Temporal averaging of frames makes the differential piston signal emerge from the turbulence-induced speckle field and leads to a significant improvement in the RMSE calculation. As expected, better seeing conditions result in improved accuracy. Polychromaticity only degrades the performance by less than 5%, compared to the monochromatic case. In a realistic scenario, detector noise is not a limiting factor, as the primary limitation rather arises from the need for sufficient speckle averaging. The network was also shown to be applicable to input images other than the 2 × 2 SH-WFS data
Prediction of weld quality in laser welding of hardmetal and steel using high-speed imaging and machine learning methods
No full textInternational audienceLaser welding of steel and hardmetal presents significant challenges due to their differing material properties. Improper laser welding parameters can result in unstable joints, ultimately leading to reduced mechanical strength of the weld. Therefore, defining an optimal process window is critical to ensuring weld quality. In addition, a continuous process monitoring method like High-Speed Imaging (HSI) is essential in real industrial applications to maintain stability and detect potential defects. Understanding plume dynamics helps identify the most important features of weld quality, but it also provides deeper insight into operational parameters that discriminate different weld types. Analysis of individual image plume frames from HSI reveals distinct statistical features that are identified as unique to each welding condition. Performing systematic feature selection using plume morphology, spatter generation and weld quality, we achieved>95 % leveraging Machine Learning (ML) classifiers. Particularly, Gradient Boosting Classifier (GBC), Linear Discriminant Analysis (LDA), Multinomial Logistic Regression (MNL-LR), Support Vector Machine (SVM), and Random Forest (RF), where the RF obtained >99 % classification accuracy of weld quality. The RF was then used in performing Recursive Feature Elimination (RFE), and with the robustness analysis, we managed to reduce the number of features from forty-nine to nine features while maintaining satisfactory performance (Accuracy = 0.981, F1-score = 0.961, AUROC = 0.997). The position of the weld plume, plume eccentricity and plume width are the most essential features that lead to the improvement of node purity and classification accuracy.</div
ADAPT-MOVE : un modèle d'amélioration de l'apprentissage moteur par la réalité virtuelle
No full textThis study explores a virtual reality (VR) model to enhance motor learning through integrated gamification, feedback mechanisms, and adaptive functions. Traditional motor learning models emphasize feedback and repetitive practice as critical to skill acquisition; however, they lack in adapting motivation-sustaining features such as gamification and adaptive interactions, especially within immersive VR environments. This research proposes a VR-based model that unifies psychological principles with VR’s technical capabilities to create a motivating and engaging environment. The model hypothesizes that gamification elements and feedback mechanisms improve motivation and engagement, while adaptive functions optimize task performance. Methodologies involve designing VR scenarios where task-specific motor activities are augmented with interactive and gamified elements. Experiments were conducted to measure motivation, engagement, and task performance, evaluated through self-reported measures and objective data. Results indicate that the proposed model provides an enhanced learning experience by improving skill acquisition and retention across varied motor tasks, making it valuable for applications in rehabilitation, sports training, and educational settings.Cette étude explore un modèle de réalité virtuelle (RV) pour améliorer l'apprentissage moteur grâce à l'intégration de la gamification, des mécanismes de retour d'information et des fonctions adaptatives. Les modèles traditionnels d'apprentissage moteur mettent l'accent sur le retour d'information et la pratique répétitive comme étant essentiels à l'acquisition des compétences ; cependant, ils manquent d'adaptation aux caractéristiques de motivation telles que la gamification et les interactions adaptatives, en particulier dans les environnements immersifs de la RV. Cette recherche propose un modèle basé sur la RV qui unifie les principes psychologiques avec les capacités techniques de la RV pour créer un environnement motivant et engageant. Le modèle suppose que les éléments de gamification et les mécanismes de retour d'information améliorent la motivation et l'engagement, tandis que les fonctions adaptatives optimisent l'exécution des tâches. Les méthodologies impliquent la conception de scénarios de RV dans lesquels des activités motrices spécifiques à une tâche sont complétées par des éléments interactifs et ludiques. Des expériences ont été menées pour mesurer la motivation, l'engagement et la performance des tâches, évalués à l'aide de mesures auto-déclarées et de données objectives. Les résultats indiquent que le modèle proposé offre une meilleure expérience d'apprentissage en améliorant l'acquisition et la rétention des compétences dans diverses tâches motrices, ce qui le rend utile pour des applications dans les domaines de la réadaptation, de l'entraînement sportif et de l'éducation
Layer transferred UV emitting hBN/AlGaN heterostructures
No full textInternational audiencep-hBN/n-AlGaN heterojunctions were fabricated using a dry-selective lift-off/transfer of Mg-doped hexagonal boron nitride (hBN) layer on top of n-AlGaN. Electrical contacts were used as mechanical stressors to provide structural rigidity to hBN layers as well as enabling selective lift-off. These junctions exhibit a rectifying behavior with a rectification ratio of approximately 3 × 105 at 3 V. When junctions were forward biased, ultraviolet (UV) emission around 262 nm was measured. This emission corresponds to recombinations in the n-AlGaN layer, demonstrating good hole injection in the structure. Full light emitting diode (LED) structures were fabricated by integrating UV multi quantum wells (MQWs) into these junctions. Produced UV LEDs emit around 290 nm serving as a proof of concept for future layer transferred p-hBN/MQWs/n-AlGaN structures in which the Al content is increased to go toward deep ultraviolet (DUV) emission. The selective pick and place process used to build these LEDs has multiple advantages. First, it allows independent optimization of the p-side as well as of the n-side, which includes the quantum wells. Second, UV MQWs are protected from the high temperatures needed for high hBN material quality growth, and thus their thermal stability is not affected
Conception et développement du jumeau numérique pour démonstrateur de micro-réseau de bâtiment
No full textInternational audienc
Commande en commutation douce d'un convertisseur matriciel adapté aux applications de recharge embarquée des véhicules électriques
No full textInternational audienc
Calcul analytique de l'induction d'armature d'une machine électrique à aimants en surface
No full textInternational audienc
Residual stress control in large-format additive manufacturing of polylactic acid via a digital twin and in-operando imaging
No full textInternational audiencePolymer-based Large Format Additive Manufacturing (LFAM) is an extrusion-based technology that deposits large-diameter polymer beads using a robotic arm-mounted nozzle. However, slow cooling rates and heat accumulation generate technical challenges, including significant deformation that requires nozzle path adjustments and the buildup of residual stresses from thermo-chemical shrinkage that may cause debonding. This study integrates two fast modeling approaches, ScanFast (thermal) and QuadWire (mechanical), to reduce the number of degrees of freedom compared to conventional methods while maintaining accuracy. A computationally efficient digital twin of the process is developed and validated experimentally on a thin wall printed with polylactic acid. Anisotropic material properties are characterized, and in-operando temperature and displacement fields are measured using infrared thermography and backward Digital Image Correlation. The results show correlation coefficients greater than 0.80 between experimental and numerical data. The validated digital twin is then applied to assess the influence of process parameters on three key aspects: (i) the number of layers above the glass transition temperature, (ii) residual stress development, and (iii) positional offset between the nozzle and the structure. The proposed approach provides an efficient tool to optimize process parameters and nozzle trajectories, thereby enhancing the quality and manufacturability of LFAM-produced parts
Exponential stabilization of quasi-one-sided Lipschitz systems with time delay
No full textInternational audienceThis paper aims to design Observer-Based (OB) controllers that ensure the exponential stability of a class of nonlinear time-delay systems. The nonlinear part of the system satisfies a weak Quasi-One-Sided Lipschitz (QOSL) condition characterized by the matrices , as well as a QOSL condition characterized by the matrices . First, we derive a sufficient condition formulated as a Linear Matrix Inequality (LMI) via a Lyapunov–Krasovskii (LK) functional. The main advantage of this design is that the controller and observer gains are computed in a single step. However, its main drawback is that the matrices , , and are fixed rather than treated as decision variables. To overcome this limitation, we propose an improved design in which the matrices , and are treated as decision-variable matrices with a fixed structure. By using an appropriate decoupling technique, this approach provides greater flexibility in the selection of matrices and reduces conservatism. The efficacy of the developed OB controllers is demonstrated via a suitable numerical example