HAL Portal UTC Université de Technologie de Compiègne
Not a member yet
11652 research outputs found
Sort by
Sensitivity analysis of an imperfect maintenance policy for Proton-exchange membrane fuel cell based on geometric a semi-Markov model
No full textInternational audienceThis paper proposes a maintenance model for a Proton-Exchanged Membrane Fuel Cells system. Two types of failures are considered: non repairable internal failures and possibly repairable external failures. The maintenance involves replacing one or more cells. Therefore the maintenance can be perfect or imperfect. Moreover after each repair, the maintenance operator improves and the down time due to maintenance is reduced. A semi-Markov process is proposed to model the system's lifetime. In this framework, the reliability measures and the remaining useful lifetime are derived. The long run average maintenance cost is studied. A sensitivity analysis of model parameters and maintenance unit costs is performed and a maintenance optimization strategy is proposed.</div
Discrimination of the Lame Limb in Horses Using a Machine Learning Method (Support Vector Machine) Based on Asymmetry Indices Measured by the EQUISYM System
No full textInternational audienceLameness detection in horses is a critical challenge in equine veterinary practice, particularly when symptoms are mild. This study aimed to develop a predictive system using a support vector machine (SVM) to identify the affected limb in horses trotting in a straight line. The system analyzed data from inertial measurement units (IMUs) placed on the horse’s head, withers, and pelvis, using variables such as vertical displacement and retraction angles. A total of 287 horses were included, with 256 showing single-limb lameness and 31 classified as sound. The model achieved an overall accuracy of 86%, with the highest success rates in identifying right and left forelimb lameness. However, there were challenges in identifying sound horses, with a 54.8% accuracy rate, and misclassification between forelimb and hindlimb lameness occurred in some cases. The study highlighted the importance of specific variables, such as vertical head and withers displacement, for accurate classification. Future research should focus on refining the model, exploring deep learning methods, and reducing the number of sensors required, with the goal of integrating these systems into equestrian equipment for early detection of locomotor issues
Multi-view Evidential K-NN Classification
No full textInternational audienceMulti-view classification, aiming to classify samples represented by multiple feature vectors, has become a hot topic in pattern recognition. Although many methods with promising performances have been proposed, their practicality is still limited by the lack of interpretability in some situations. Besides, an appropriate description for the soft labels of multi-view samples is missing, which may degrade the classification performance, especially for those samples located in highly-overlapping areas of multiple vector spaces. To address these issues, we extend the K-nearest neighbor (K-NN) classification algorithm to multi-view learning, under the theoretical framework of evidence theory. The learning process is formalized, firstly, as an optimization problem, where the weights of different views, an adaptive K value of every sample and the distance matrix are determined jointly based on training error. Then, the final classification result is derived according to the philosophy of the evidential K-NN classification algorithm. Detailed ablation studies demonstrate the benefits of the joint learning for adaptive neighborhoods and view weights in a supervised way. Comparative experiments on real-world datasets show that our algorithm performs better than other state-of-the-art methods. A real-world industrial application for condition monitoring shown in Appendix F exemplifies the need to use the evidence theory and the benefits from the unique interpretability of K-NN in detail
Auto-apprentissage robuste par des étiquettes credaux
No full textMachine learning classifiers have achieved impressive success in a wide range of domains such as natural language processing, image recognition, medical diagnosis, and financial risk assessment. Despite their remarkable accomplishments, their application to real-world problems still entails challenges. One of the most common problems is the large amount of labels necessary to train a classifier. Labelling data is a time-consuming process that is prone to inconsistency because it is done by humans. In some domains (e.g. the medical domain), this process also requires specialists, which are expensive to hire and have little time to work on these activities. These issues are the main motivation to the so-calledWeakly supervised learning, a field of study in machine learning that proposes to reduce the amount of labels necessary by gathering information of unlabeled data. Among the three most studied paradigms in this field, we focus on incomplete supervision, i.e. more than one label is a possibly correct label for each sample in the dataset. In this thesis, we focus on semi-supervised learning, where it is supposed that, for a given instance, only one (unknown) label is the correct answer. More specifically, we consider the self-supervised learning approach, where the model predicts the labels for the unlabeled part of the dataset, called pseudo-labels. The model is retrained using those pseudo-labels after, in an interactive manner. This approach has the downside to bias the model, since the pseudo-labels are mere guesses and have no guarantee to be correct. This problem is mitigated in the literature by selecting the which pseudo-labels are the most probably correct and thus used on the next iteration for the training. In this thesis, we advocate that this correction might not be enough to avoid bad pseudo-labels and the performance is better when we apply uncertainty quanti_cation on the pseudo-labels. Additionally, a good property we want to ensure is the probability calibration, where the model probability estimation is guaranteed to be closer to the real probability of the data. We propose to apply Inductive Venn Predictor (IVP) and Inductive Venn-Abers Predictor (IVAP) to better quantify the uncertainty while calibrating the model. Our experiments indicate that we achieve a better calibration with comparable accuracy in binary sets and pairwise approach, but fails to generalize well in a multi-class approach.Les classificateurs d'apprentissage automatique ont connu un succès impressionnant dans un large éventail de domaines tels que le traitement du langage naturel, la reconnaissance d'images, le diagnostic médical et l'évaluation des risques financiers. Malgré ces résultats remarquables, leur application à des problèmes réels pose toujours des problèmes. L'un des problèmes les plus courants est la grande quantité d'étiquettes nécessaires pour former un classificateur. L'étiquetage des données est un processus qui prend du temps et qui est sujet à des incohérences parce qu'il est effectué par des humains. Dans certains domaines (par exemple le domaine médical), ce processus requiert également des spécialistes, dont le recrutement est coûteux et qui ont peu de temps à consacrer à ces activités. Ces problèmes sont la principale motivation de l'apprentissage faiblement supervisé, un domaine d'étude de l'apprentissage automatique qui propose de réduire le nombre d'étiquettes nécessaires en recueillant des informations sur les données non étiquetées. Parmi les trois paradigmes les plus étudiés dans ce domaine, nous nous concentrons sur la supervision incomplète, c'est-à-dire que plus d'une étiquette est une étiquette possiblement correcte pour chaque échantillon de l'ensemble de données. Dans cette thèse, nous nous concentrons sur l'apprentissage semi-supervisé, où il est supposé que, pour une instance donnée, une seule étiquette (inconnue) est la réponse correcte. Plus précisément, nous considérons l'approche d'apprentissage auto-supervisé, où le modèle prédit les étiquettes pour la partie non étiquetée de l'ensemble de données, appelées pseudo-étiquettes. Le modèle est ensuite réentraîné à l'aide de ces pseudo-étiquettes, de manière interactive. Cette approche présente l'inconvénient de biaiser le modèle, étant donné que les pseudo-étiquettes ne sont que des suppositions et qu'il n'y a aucune garantie qu'elles soient correctes. Ce problème est atténué dans la littérature en sélectionnant les pseudo-étiquettes qui sont le plus probablement correctes et qui sont donc utilisées lors de l'itération suivante pour la formation. Dans cette thèse, nous soutenons que cette correction peut ne pas être suffisante pour éviter les mauvais pseudo-étiquettes et que la performance est meilleure lorsque nous appliquons la quantification de l'incertitude sur les pseudo-étiquettes. En outre, une bonne propriété que nous voulons garantir est le calibrage de la probabilité, où l'estimation de la probabilité du modèle est plus proche de la probabilité réelle des données. Nous proposons d'appliquer le prédicteur inductif de Venn (IVP) et le prédicteur inductif de Venn-Abers (IVAP) pour mieux quantifier l'incertitude lors de l'étalonnage du modèle. Nos expériences indiquent que nous obtenons une meilleure calibration avec une précision comparable dans les ensembles binaires et l'approche par paire, mais que nous ne parvenons pas à bien généraliser dans une approche multi-classe
Towards evaluating set-valued predictions with partial observations
No full textInternational audienceIn different situations, one will have to evaluate set-valued predictions while having set-valued/partial observations. This is the case, for instance, when experts provide weak supervision and when predictions are provided by robust classifiers. However, we still miss some principled approach to evaluate such predictions and observations. This is the goal of this paper, where we propose guidelines about such evaluations functions. We will set our study within the framework of belief functions, that seem quite adapted to the considered problem
Trascrivere l’intero corpo in lingua dei segni e gesti co-verbali con Typannot
No full textInternational audienceSign Languages (SL) and co speech gestures (CSG) convey meaning by moving one or more body articulators (hands, arms, face, etc.) in a coordinated manner — though not necessarily simultaneously. Meaning thus arises not only from the movement itself but also from how it unfolds within a temporal dynamic. These elements must be considered when developing a transcription system for body articulation, yet this is rarely the case.The development of Typannot, a transcription system for SL and CSG, required designing an articulatory model of the body. Inspired by Boutet's kinesiological approach, this model accounts for variations in the degrees of freedom of all body segments, offering a coherent and comprehensive description of whole-body movements. These movements are then embedded in a temporal dynamic, allowing for an accurate representation of their coordination.A transcription system must be precise, readable, queryable, and compatible with linguistic analysis software. This requires handling the intrinsic linearity of word-processing tools. The proposed solution — developed through collaboration between designers, linguists, and computer scientists — represents temporal relationships between movements of different articulators within a linear extent.However, manually transcribing whole body movement requires significant effort, limiting researchers’ ability to process data. To address this, a computer assisted transcription tool was developed, directly associating movement data — automatically extracted using a pose estimator (e.g., AlphaPose) — with Typannot characters. This transcription methodology was implemented within the ANR LexiKHuM project, where head and torso movements of both deaf and hearing speakers were analyzed to determine their role in expressing epistemicity.We aim to show how whole-body movements can be transcribed linearly without losing the connection to the multilinearity and temporality of bodily dynamics. We will also demonstrate how motion capture speeds up and improves the transcription process, illustrating our points with a concrete example related to the study of bodily markers of epistemicity.Les Langues des Signes (SL) et les gestes co-verbaux (CSG) sont produits par la mise en mouvement d’un ou plusieurs articulateurs corporels (mains, bras, visage etc.). Ces articulateurs agissent de manière coordonnée, mais pas nécessairement simultanée, pour produire du sens. Le sens émane donc non seulement du mouvement lui-même mais aussi de la manière dont celui-ci s’inscrit dans une dynamique temporelle. Ces éléments doivent être pris en compte lors du développement d’un système de transcription articulatoire du corps, mais c’est rarement le cas.Pour le développement de Typannot, un système de transcription de la LS et de la CSG, a donc été développé un modèle articulatoire du corps : inspiré de l’approche kinésiologique de Boutet, il permet de prendre en compte les variations de degrés de liberté de l’ensemble des segments corporels, offrant une description cohérente et exhaustive de tous les mouvements du corps. Ces mouvements sont ensuite inscrits dans une dynamique temporelle permettant de rendre compte de leur coordination.Un système de transcription doit être à la fois précis, lisible, requêtable et compatible avec les logiciels d’analyse linguistique : cela implique de composer avec la linéarité intrinsèque des outils de traitement de texte. La solution proposé, fruit de la collaboration entre designers, linguistes et informaticiens, permet de représenter dans un extension linéaire les relations temporelles entre les mouvements des différents articulateurs qui produisent un signe ou un geste.Toutefois, la transcription à la main d’un objet aussi complexe que le mouvement de l’ensemble du corps nécessite un temps d’élaboration conséquent, qui peut limiter la capacité des chercheurs de traiter leurs données. On a donc développé un outil de transcription assisté par ordinateur, qui permet d’associer de manière directe des données sur le mouvement, extraites automatiquement grâce à un estimateur de pose (e.g., AlphaPose), aux caractères ce système de transcription. Cette méthodologie de transcription a été mise en œuvre dans le cadre du projet ANR-LexiKHuM, où ont été analysés les mouvements de la tête et du buste de locuteurs sourds et entendants afin de déterminer la manière dont ces deux articulateurs contribuent à l’expression de l’épistémicité.Nous souhaitons montrer comment les mouvements de l’ensemble du corps peuvent être transcrits de manière linéaire sans perdre le lien avec la multilinéarité et la temporalité de la dynamique corporelle. Nous démontrerons également comment la capture de mouvement accélère et améliore le processus de transcription, en illustrant nos propos par un exemple concret lié à l'étude des marqueurs corporels d'épistémicité
Acute toxicity of atrazine degradation solutions via advanced oxidation processes: An experimental and in-silico approach based on mass spectrometry data
No full textInternational audienceAlthough atrazine has been banned, residues of the compound and its metabolites are still detected in certain areas of French groundwater. In response to the European Green Deal, this study focuses on the degradation of atrazine in tap water and the toxicological evolution of different treatments. The degradation efficiency of atrazine in deionized water by high-voltage electrical discharge (HVED) and high-frequency ultrasound (HFUS) was higher than that in tap water with the same initial atrazine concentration, likely due to competition from impurities in tap water, such as ions and other organic compounds. Additionally, acute toxicity testing using Daphnia magna was conducted to assess the toxicological impact of the degradation processes. Acute toxicity (LC50) values were used to evaluate the toxicity of the different treatment solutions initially containing atrazine. High-resolution mass spectrometry coupled with liquid chromatography (LC-HRMS) was employed to identify and compare the transformation products generated by each degradation technique. Furthermore, an in-silico toxicity evaluation using the QSAR model was proposed and compared to the experimental data. The in-vivo assessment revealed that atrazine solutions treated with Fenton oxidation (LC50 < 0.1 mg/L) were significantly more toxic than those treated with HFUS (LC50 > 5 mg/L) or HVED (1.5 < LC50 < 2.5 mg/L). However, in the in-silico evaluation, Fenton oxidation appeared less toxic, as residual toxicity from ferrous ions and hydrogen peroxide was not considered. Consequently, the degradation of atrazine by HVED and HFUS proved to be more efficient and environmentally favorable.</div
Proximity-centred accessibility – A conceptual debate involving planning practitioners worldwide
No full textInternational audienceIn recent years, the concept of proximity has garnered increasing attention in both transportation research and practice, albeit under various terms and interpretations. Among these, the concept of the 15-minute city has catalysed attention in planning practice, with recent evolution to the x-minute city and city of proximities. In research, proximity-centred accessibility has been offered as an umbrella term to express the ability to reach activities and destinations at short distances. Regardless of the terminology used, the essence of proximity lies in the ease with which one can access desired activities and destinations within reasonable travel times, independent of speed-enhancing transport modes most notably through walking.This research investigates the nuanced meanings ascribed to proximity-centred accessibility by planning practitioners globally, spanning diverse regional and local contexts. For this, we used an online survey, disseminated among over 9000 practitioners from 22 countries across 5 continents, which generated over 1300 responses. The survey explored the preferred terms for proximity-centred accessibility and their definitions, specifically emphasizing time and distance thresholds and the identification of relevant activities. By juxtaposing our findings with an earlier survey of accessibility researchers, this study also contributes to the groundwork for a conceptual framework for proximity-centred accessibility
Catalyst deactivation of biomass-derived biochars during tar cracking: Role of porosity
No full textSource Agritrop Cirad (https://agritrop.cirad.fr/613530/)International audienceTars present in syngas pose major challenges to the development of thermochemical processes. Catalytic cracking of tars using biochar is a promising option, both technically and economically. To elucidate the role of biochar porosity, heterogeneous tar cracking tests were carried out in an inert atmosphere. The variation in biochar porosity during long-term tar cracking tests was studied to gain a deeper understanding of biochar deactivation. Physically activated biochar derived from beech wood biomass was used as a catalyst, with toluene serving as a model molecule for tars. The results showed that the activated biochar exhibited good initial catalytic activity for toluene cracking, compared to homogeneous thermal cracking. However, due to polymerization reactions during toluene cracking, coke was deposited on the inner pore surfaces of the biochar, leading to its deactivation. After a while, the toluene conversion rate decreased to levels comparable to those achieved during thermal cracking. This was accompanied by a decrease of the specific surface area of the biochar. It has been shown that heterogeneous toluene conversion primarily occurs in the micropores, which serve as the active sites, while mesopores are essential for the diffusion of toluene into the micropores. Biochar deactivation was exacerbated by the obstruction of mesopores. Furthermore, increasing the mesopore content in biochar could slow down the process of biochar deactivation