Search

International audience; Developing algorithms able to predict daily feed intake is essential for implementing precision-feeding strategies in real time. Given the lack of a mechanistic model to predict feed intake in lactating sows, a new approach that combined real-time prediction with off-line learning of sow feeding behaviours was developed. A database of 39,090 lactations from 6 farms that contained the first 20 post-farrowing feed intake values was used to (i) identify groups of sows with similar feeding behaviour and (ii) test three functions to predict feed intake. The homogeneity of clusters obtained by off-line learning was assessed according to the Silhouette and Calinski-Harabasz scores. The prediction functions were evaluated by calculating mean error (ME) and root mean square error (RMSE) per day and per sow. The clusters with the best homogeneity were obtained by dividing the database into two groups. The trajectory of feed intake of the first group increased continuously during lactation, while that of the second plateaued from day 10 onwards. The ME per sow obtained for these two trajectories using the best function was-0.08 kg/d, and the corresponding RMSE was 1.06 kg/d. Although individual variability was high, the use of trajectories improved the prediction of feed intake. In practice, learning of trajectories may be recalculated regularly, while the real-time prediction function, which requires fewer computing resources, could be embedded into the smart feeder.; Le développement d'algorithmes de prédiction de la consommation alimentaire journalière est essentiel à la mise en oeuvre des stratégies d'alimentation de précision en temps réel. Compte tenu de l'absence de modèle mécaniste de prédiction de la consommation chez la truie en lactation, une nouvelle approche est proposée combinant une prédiction en temps réel avec un apprentissage « hors ligne » des comportements alimentaires de la truie. Une base de données de 39 090 lactations, provenant de six exploitations et contenant les consommations des 20 jours après mise-bas, a été utilisée pour (i) identifier des groupes (clusters) de truies présentant un comportement alimentaire similaire et (ii) tester trois fonctions de prédiction de la consommation. L'homogénéité des clusters obtenus par apprentissage « hors ligne » a été évaluée selon les indices Silhouette et Calinski-Harabasz. Les méthodes de prédiction ont été évaluées avec l'erreur moyenne et l'erreur quadratique moyenne (RMSE) déterminées par jour et par truie. Les clusters les plus homogènes sont obtenus lorsque la base est divisée en deux groupes. La trajectoire du premier groupe est caractérisée par une augmentation continue de la consommation au cours de la lactation, et la seconde, par un plateau atteint à partir du 10 ème jour. L'erreur moyenne par truie, obtenue en utilisant deux trajectoires et la meilleure fonction de prédiction, est de-0,08 kg/j, avec une RMSE de 1,06 kg/j. Bien que la variabilité individuelle soit élevée, l'utilisation des trajectoires améliore la prédiction de la consommation. En pratique, l'apprentissage des trajectoires peut être renouvelé régulièrement, tandis que la procédure de prédiction, peu gourmande en puissance de calcul, peut être intégrée dans le système d'alimentation de précision.

We address the problem of diagnosing complex discrete-event systems such as telecommunication networks. Given a flow of observations from the system, the goal is to explain those observations by identifying and localizing possible faults. Several model-based diagnosis approaches deal with this problem but they need the computation of a global model which is not feasible for complex systems like telecommunication networks. Our contribution is the proposal of a decentralized approach which permits to carry out an on-line diagnosis without computing the global model. This paper describes the implementation of a tool based on this approach. Given a decentralized model of the system and a flow of observations, the program analyzes the flow and computes the diagnosis in a decentralized way. The impact of the merging strategy on the global efficiency is demonstrated and illustrated by experimental results on a real application.

Detection and isolation of failures in large and complex systems such as telecommunication networks are crucial and challenging tasks. The problem considered here is that of diagnosing the largest French packet switching network. The challenge is to be as efficient as the existing expert system while providing greater generality and flexibility with respect to technological and reconfiguration changes in the network. The network is made up of interconnected components each of which can send, receive and transmit messages via its ports. The problem we are faced with is to follow the evolution of the network on the basis of the stream of time-stamped alarms which arrive at the supervision center. We have decided to use model-based techniques which are recognized to be more adapted to evolutive systems than expertise-based approaches are. This paper starts with a description of how we model the global behavior of this discrete-event system by using communicating finite state machines. It goes on to explain how this model is used for analyzing the stream of alarms and diagnosing the network. Our work is based on the diagnoser approach proposed by Sampath et al. (1995). Starting from a model of the network adapted to simulate faults, this approach transforms it into a finite state automaton, called a diagnoser, in order to analyze the stream of alarms. The approach described in Sampath et al. (1995; 1996) proved to be grounded on certain basic hypotheses which were too restrictive for our application. This paper extends Sampath’s proposal to communicating finite state machines. The difficulties we had to cope with are outlined and the way we overcome them is presented. A major difficulty is the huge size of the global model of the system. To solve this problem we take advantage of the hierarchical structure of the network and rely on a generic model of the system for building a generic diagnoser.

L’utilisation de techniques de raisonnement sur les modeles permet une representation des connaissances plus facile a acquerir qu’une expertise directe. Neanmoins, la complexite combinatoire associee a ces techniques ne permet pas d’envisager un systeme temps reel s’appuyant directement sur celles-ci. Un systeme de supervision des reseaux de telecommunication est propose s’appuyant, d’une part sur la generation hors-ligne .de sequences temporelles d’alarmes, aussi appelees chroniques ou scenarios, et d’autre part sur la reconnaissance en ligne et en temps reel de ces sequences. Seule la premiere partie de ce systeme, la generation hors-ligne est exposee. Cet apprentissage s’appuie sur une modelisation structurelle ainsi que fonctionnelle du reseau. Les delais de propagation des alarmes tout comme les probabilites de masquage et de perte de certains messages sont pris en compte.

This paper extends the decentralized approach proposed by Pencole (2000) for diagnosing discrete-event systems. The incrementality issue, which is a crucial point in the context of online diagnosis, is discussed and two solutions are proposed.

Detection and isolation of failures in large and complex systems, as telecommunication networks, are crucial and challenging tasks. We propose in this article a model-based diagnosis system for telecommunication network breakdown based on two steps: Off-line step: the telecommunication management network is, first, modeled by temporal and communicating finite state machines. Then, a set of breakdown situations is simulated in the model. The simulation task generates for each simulated situation a set of all possible sequences of alarms received by the supervisor. And a learning database is built associating breakdown situations to different sequences of alarms. Finally, a learning module computes discrimination between all simulated breakdown situations; On-line step: results of off-line step are used to recognize on-fly breakdown situations from the stream of alarms arriving at the supervisor.

The authors address the use of model-based techniques for diagnosing telecommunication networks. The aim is to design a system for the supervision of packet switching in the French telecommunication network, in such a way as to be both generic and efficient. i.e., to support technological evolution and to meet real time constraints. The basic idea of their work is to use the model for simulation and to rely on a discrimination phase to build characteristic scenarios (i.e. patterns of events satisfying temporal constraints) for all interesting failure modes of the network. The architecture proposed in the GASPAR system is made up of four modules: a modeling module; simulation module; discrimination module; and a recognition module.

A lot of useful hidden information is included among everyday data, which often can't be extracted manually. Therefore, machine learning is used to get knowledge by generalizing examples. When all data is already stored, a common approach used is batch learning, but when learning comes to deal with huge amounts of data, arriving at different instances of time in sets of examples, incremental learning is the best solution. To improve the performances of supervised incremental learning algorithms specific examples are stored and used for future learning steps. In this paper is described an incremental learning framework that generates classification predictions, a partial memory approach (specific border examples stored), for the "Manage YourSelf" project, that deals with the diagnosis and monitoring of embedded platforms. This approach extends the AQ21 algorithm to solve a specific learning task. It was chosen because it is one of the latest developed and it has many useful features. The incremental solution is proposed because we are dealing with a data flow and partial-memory approaches were searched because here is no need to re-examine all the instances at every learning step. Furthermore, those notably decrease memory requirements and learning times.

Service-Oriented Computing is based on dynamic composition of web services to meet the demand of a user. A major challenge in conditioning actual use of web services is to monitor their performance and enable them to react to unexpected malfunctioning. This can be done using the mechanisms of exception handling. But they do react in a predefined manner and local issues have to be planned at the services design time. However, in dynamic environments like the Internet, web services may be subject to unexpected malfunctioning which may not be handled with repair mechanisms defined at design time. In addition, local management ignores errors during the interactions between services, which limit their effectiveness. Such failures may also propagate through the services before being detected, and the key is to find the problem at the source of the malfunction and repair the service. In this context, this work is dedicated to study a distributed but coordinated and dynamic management of repair mechanisms. The difficulty is that repairs are carried out locally, but a global approach must be ensured to take into account interactions between different services. Our objective is to propose a diagnostic-repair architecture and mechanisms for this feature in detail.