Search

Conférence: Confiance.ai Days, 4 au 6 octobre 2022, CentraleSupélec, Gif-sur-Yvette (France); This booklet gathers 52 papers, either in the form of articles or posters, presented during the secondedition of the Confiance.ai Days held in Saclay on October 4-6, 2022. Altogether they give a good snapshotof the research and development work done in the Confiance.ai program, an industrial and academicinitiative of the national Grand Challenge on provable and certifiable AI, launched in support of the France2030 strategy.Among these papers, a dozen are presented as « external contributions » that were selected by an ad-hoccommittee following a call for papers. All other communications belonged to one of five so-called « villages» with physical implementation in the conference hall, distributing the work done in Confiance.ai intofive topics : « End-to-end approach » ; « from Operational Design Domain to Data » ; « Explainabilityand Understanding» ; « Robustness and Monitoring » ; « Embedded AI ».After two years of activity, and complementing the Confiance.ai white paper, this document shows thediversity and the quality of the work done in the programme. Some important and up-to-date subjects areaddressed, such as – only to name a few - out-of-distribution detection, adversarial robustness, semi- orself-supervised learning, explainability by design, verification and validation, embedded AI etc. We hopethat you will enjoy reading parts of this document as much as we enjoyed preparing and attending the 2022Confiance AI days.

CAP 2019, Conférence sur l'Apprentissage Automatique, Toulouse, France, 03-/07/2019 - 05/07/2019; Les possibilités offertes en termes de collecte et de stockage de données permettent de renouveler les approches de modélisation dans le domaine du transport. L'exploitation croisée de différentes sources de données a pour vocation la création de services à forte valeur ajoutée pour l'usager. Les travaux détaillées dans cet article portent sur le développement de modèles de prévision a base de méthodes d'apprentissage notamment profond, pour la prévision court-terme de la charge (nombre de passagers) des trains. Cette prévision de l'affluence dans les trains peut servir à enrichir l'information voyageur à destination des usagers des transports collectifs qui peuvent ainsi mieux planifier leur déplacement. Elle peut également servir aux opérateurs de transport pour une régulation "à la demande" de l'offre de transport. La principale difficulté dans la prévision est liée à la variabilité intrinsèque des séries temporelles des charges à prédire, induite par l'influence de plusieurs paramètres dont ceux liés à l'exploitation (horaire, retard, type de mission...) et au contexte (information calendaire, grand évènement, météo,...). Nous proposons un modèle LSTM encodeur-prédicteur pour résoudre cette tâche de prévision. Plusieurs expérimentations sont menées sur des données réelles du réseau Transilien de la SNCF sur une durée d'un an et demi. Les résultats de prévision sont détaillées en vue de comparer les performances d'un tel modèle à plusieurs horizons temporels avec celles d'autres modèles plus classiques utilisées en prévision.

International audience; Industry-academic research partnerships are mostly considered interesting to increase industrial innovativeness, and its benefits have been discussed in the flourishing open innovation literature. However, how to create mutually beneficial partnerships seems to be a question that has not been sufficiently studied. Through this article, we discuss the goals of these partnerships by modelling different types of collaboration. We defend that their real value has to be evaluated not only by looking at the knowledge created, but also at the increase of generativity we observe, due to interactions between academia and industry. Furthermore, we propose a model based on C-K theory that can be used to design a research collaboration that increases generativity, going beyond problem solving and knowledge transfer logics. We illustrate it through a case study, which shows that value creation in an industry-research partnership is increased by a model of co-generation, instead of considering these relations as a one-way transfer. Furthermore, we show that conflicts in a partnership can be solved through a C-K based tool.

16th ASMDA Conference, LE PIREE, GRECE, 30-/06/2015 - 04/07/2015; Degradation processes modelling is a key problem to perform any type of reliability study. Indeed, the quality of the computed reliability indicators and prognosis estimations directly depends on this modelling. Mathematical models commonly used in reliability (Markov chains, Gamma processes...) are based on some assumptions that can lead to a loss of information on the degradation dynamic. In many studies, Dynamic Bayesian Networks (DBN) have been proved relevant to represent multicomponent complex systems and to perform reliability studies. In a previous paper, we introduced a, degradation model based on DBN named graphical duration model (GDM) in order to represent a wide range of duration models. This paper will introduce a new degradation model based on GDM integrating the concept of conditional sojourn time distributions in order to improve the degradation modelling. It integrates the possibility to take into account several degradation modes together and to adapt the degradation modelling in respect of some new available observations of either the current operation state or the estimated degradation level, to take into account an eventual dynamic change. A comparative study on simulated data between the presented model and the GDM will be performed to show the interest of this new approach.

This paper deals with a data mining approach applied on Bike Sharing System Origin-Destination data, but part of the proposed methodology can be used to analyze other modes of transport that similarly generate Dynamic Origin-Destination (OD) matrices. The transportation network investigated in this paper is the Vélib’ Bike Sharing System (BSS) system deployed in Paris since 2007. An approach based on Latent Dirichlet Allocation (LDA), that extracts the main features of the spatio-temporal behavior of the BSS is introduced in this paper. Such approach aims to summarize the behavior of the system by extracting few OD-templates, interpreted as typical and temporally localized demand profiles. The spatial analysis of the obtained templates can be used to give insights into the system behavior and the underlying urban phenomena linked to city dynamics.

The optimization of maintenance strategies has become a key issue in the railway industry but also in most industrial fields. To address this challenge, many studies dealt with the estimation of optimal maintenance parameters. But what commonly happens when the degradation process suddenly changes? The operator has to face an unexpected, increasing number of severe defects (and then a strong drop of its availability). These changes are generally due to either: a new component, introduced in the system for obsolescence reasons; or changing operating conditions. Based on the dynamic Bayesian networks (DBN), formalism that has been proved relevant to perform reliability analysis can easily represent complex system behaviors. This paper introduces a dynamic maintenance strategy, able to detect these drifts and to evaluate their impacts on the rolling stock doors system’s behavior.

A non linear regression approach which consists of a specific regression model incorporating a latent process, allowing various polynomial regression models to be activated preferentially and smoothly, is introduced in this paper. The model parameters are estimated by maximum likelihood performed via a dedicated expecation-maximization (EM) algorithm. An experimental study using simulated and real data sets reveals good performances of the proposed approach.

In most industrial fields, and particularly in the railway industry, the optimization of maintenance policies has become a key issue. Dynamic Bayesian networks (DBN) have been proved as relevant to perform reliability analysis as they can easily represent complex systems behaviors. Based on this formalism, graphical duration models (GDM) were developed by (Donat, et al., 2009) to set all kind of sojourn time distributions for each state of the system. Unlike to some Markovian approaches that impose exponential behavior, this approach could better model the exact degradation dynamic of real industrial systems. But, what commonly happens when the degradation process suddenly changes? The operator has to face with an unexpected increasing number of severe defects (and then a strong drop of its availability). These changes are generally due to either new component, introduced in the system for obsolescence reasons, or to changing operating conditions. The aim of the study introduced in this paper, focusing on Dynamic Maintenance Strategies, is to detect these drifts and to evaluate their impacts on the system’s behavior.

Considering availability purposes for train transportation, passenger accesses (doors and steps) are often designated as critical systems. To improve global availability of its rolling stock, Bombardier Transportation (BT) aims at reinforcing its maintenance procedure by introducing predictive diagnosis. The SURFER project has been initiated to develop online and in-cars tools to early detect and prevent faults. In this paper, an overview of achieved progress with respect to online predictive diagnosis will be introduced. For this purpose, many signals are recorded using a test bench by BT: electrical motor intensity current, door displacement, binary indicators as door closed and locked. The paper focuses on designing a semi-Supervised discriminative probabilistic model that take into account contextual variables (train inclination or constraints due to passengers affluence) to perform a robust predictive diagnosis. The main steps of the proposed method are the followings: the segmentation of the provided signals into opening and closing phases, the extraction of relevant features from opening/closing phases, the setting of the discriminative diagnosis model based on statistical semi-supervised learning. The proposed approach is tested on signals collected from regional trains fleeting around Paris. It allows the earlier detection of anomalies, for instance, those due to maladjustments. The practical implementation of this approach will be detailed together with its preliminary results.

This article introduces a state-space model for the dynamic modeling of curve sequences within the framework of railway switches online monitoring. In this context, each curve has the peculiarity of being subject to multiple changes in regime. The proposed model consists of a specific latent variable regression model whose coefficients are supposed to evolve dynamically in the course of time. Its parameters are recursively estimated across a sequence of curves through an online Expectation-Maximization (EM) algorithm. The experimental study conducted on two real power consumption curve sequences from the French high speed network has shown encouraging results.

International audience; Assessing the operating state of the railway infrastructure and rolling stock using condition measurements acquired through embedded sensors has become a powerful decision-making support for preventive maintenance strategies. This article introduces a dynamic approach for the online monitoring of railway switch operations. The method is based on modeling the power consumption curves acquired during successive switch operations using conjointly five polynomial regression models whose coefficients are dynamically estimated across a sequence of curves. The experimental study conducted on two real power consumption curve sequences from the French high speed network has shown encouraging results in terms of characterization of the temporal evolution of railway switch operations.

In most industrial fields, and particularly in the railway industry, the optimization of maintenance policies has become a key issue. To address this problem, predictive maintenance seems to be one of the most effective known approaches. It consists of driving the maintenance process anticipating the evolution of the system state. A prediction process, called "prognosis" could also be introduced and it lays on the online estimation of the remaining useful life (RUL). Dynamic Bayesian networks (DBN) have been proved as relevant to perform reliability analysis as they can easily represent complex systems behaviors. Based on this formalism, graphical duration models (GDM) were developed by (Donat 2009) to set all kind of sojourn time distributions for each state of the system. Unlike to some Markovian approaches that impose exponential behavior, this approach could better model the exact degradation dynamic of real industrial systems. In our study, where time and states space are discretized, we consider rail degradation whom process is assumed to be monotone increasing. The real state of rails is unknown but periodically observed by an ultrasonic vehicle, characterized by its good detections rates, false alarms rates and non detection rates. This paper introduces an online RUL estimation algorithm based on the use of graphical duration models. The objective is both to evaluate the RUL and to adjust it, in real time, when a new observation is available. Then, the algorithm is applied to the rail degradation example to evaluate the quality of the RUL estimations. This paper will also provide some comparative results in respect of the considered degradation model (Markovian approach vs GDM). Finally, these RUL computations, and the associated estimations of the future system behavior, will be subsequently used to optimize the rail maintenance strategy and its diagnosis schedule.

National audience; Cet article aborde le problème de la classification des données temporelles en utilisant un mélange dynamique de lois gaussiennes dont les moyennes sont considérées comme des variables latentes qui évoluent suivant des marches aléatoires.

A non linear regression approach which consists of a specific regression model incorporating a latent process, allowing various polynomial regression models to be activated preferentially and smoothly, is introduced in this paper. The model parameters are estimated by maximum likelihood performed via a dedicated expecation-maximization (EM) algorithm. An experimental study using simulated and real data sets reveals good performances of the proposed approach.

International audience; The problem of durability of fuel cell technology is central for its spreading and commercialization. There is therefore a growing need to build accurate diagnosis tools which can give the operating state of the fuel cell during their use. When supervised machine learning approaches are used to build such diagnosis tools, they generally require a large amount of labeled data. Collection and annotation of data can be either difficult to perform or time consuming. In this paper, authors are interested in the monitoring of fuel cells in an unsupervised framework, meaning that no labels are required to learn the diagnosis model. The aim is to build a monitoring tool able to easily visualize the State Of Health of full cells fromelectrochemical impedance spectroscopy measures, showing thus its evolution from fault free case ("normal" behaviour) to defective classes such as drying or flooding. The proposed approach is based on Self Organizing Maps (SOM) which have shown their performance to solve fault detection and prediction in many industrial systems. By automatically visualizing the data into a two-dimensional space, the interpretation of the results have become easy and instinctive. The approach also allows the clustering of the data into different groups of classes, thus enabling the classification of new observations. Experimental results carried out on real data sets have shown the efficiency of the proposed approach with respect to standard supervised and unsupervised classification approaches.

National audience

WCRR 2011 - 9th World Congress on Railway Research, LILLE, FRANCE, 22-/05/2011 - 26/05/2011; The remote monitoring of the railway infrastructure and particularly the switch mechanism is of great interest for railway operators. The problem consists in detecting earlier the presence of defects in order to alert the concerned maintenance service before a breakdown occurs. For this purpose, this paper introduces a new probabilistic-based approach to dynamically modeling the evolution of condition measurements acquired during switch operations. It consists of two steps. The feature extraction from the electrical power consumption signals which aims at summarizing each signal by a low dimensional feature vector. Then, a specific autoregressive model is proposed to model the dynamical behavior of the switch mechanism.

National audience

International audience; Les équipements employés dans les milieux industriels, tels que les chaînes de production, l'ingénierie ou le transport, se présentent généralement sous la forme de systèmes complexes multi-composants et multi-états. L'état de ces systèmes est souvent affecté par les conditions d'usage. La compréhension de l'évolution au cours du temps des états de fonctionnement de ce type de système, nécessite l'élaboration de modèles dynamiques adaptés, généralement stochastiques. En complément de ces modèles de dégradation, la politique de maintenance déployée agit directement sur la dynamique d'apparition des états dégradés, selon que l'on choisit d'agir préventivement ou curativement. L'objectif de ce travail est, d'une part, de proposer, en se basant sur une approche de modélisation par réseaux bayésiens dynamiques, une fonction de coût permettant d'évaluer des politiques de maintenance. D'autre part, de mettre en oeuvre un algorithme d'optimisation de type génétique en vue de retenir la politique de maintenance préventive optimale. Dans cet article, on propose une fonction de coût associée à un cas d'étude (un système de distribution de fluides à 3 vannes) et un algorithme d'optimisation de type algorithme génétique. Le critère d'optimisation proposé utilise une modélisation stochastique de la dégradation du système basée sur une structure particulière des réseaux bayésiens dynamiques nommée Modèle Graphique de Durée (MGD). Ce critère de coût fait intervenir bien sûr les paramètres essentiels de la maintenance, tels que la probabilité d'effectuer une action de maintenance, son coût et sa durée, la disponibilité du système et les utilités associées aux états du système.