Search

Gradual patterns that capture co-variation of complex attributes in the form “when X increases/decreases, Y increases/decreases” play an important role in many real world applications where huge volumes of complex numerical data must be handled. More recently, they have received attention from the data mining community for exploring temporal data and methods have been defined to automatically extract gradual patterns from temporal data. However, to the best of our knowledge, no method has been proposed to extract gradual patterns that always appear at the identical time intervals in the sequences of temporal data, despite the knowledge that such patterns may bring for certain applications such as e-commerce. This paper proposes to extract co-variations of periodically repeating attributes from the sequences of temporal data that we call seasonal gradual patterns. We discuss the specific features of these patterns and propose an approach for their extraction by exploiting a motif mining algorithm in a sequence, and justify its applicability to the gradual case. Illustrative results obtained from a real world data set are described and show the interest for such patterns.

Over the last few decades, action recognition applications have attracted the growing interest of researchers, especially with the advent of RGB-D cameras. These applications increasingly require fast processing. Therefore, it becomes important to include the computational latency in the evaluation criteria. In this paper, we propose a novel human action descriptor based on skeleton data provided by RGB-D cameras for fast action recognition. The descriptor is built by interpolating the kinematics of skeleton joints (position, velocity and acceleration) using a cubic spline algorithm. A skeleton normalization is done to alleviate anthropometric variability. To ensure rate invariance which is one of the most challenging issues in action recognition, a novel temporal normalization algorithm called Time Variable Replacement (TVR) is proposed. It is a change of variable of time by a variable that we call Normalized Action Time (NAT) varying in a fixed range and making the descriptors less sensitive to execution rate variability. To map time with NAT, increasing functions (called Time Variable Replacement Function (TVRF)) are used. Two different Time Variable Replacement Functions (TVRF) are proposed in this paper: the Normalized Accumulated kinetic Energy (NAE) of the skeleton and the Normalized Pose Motion Signal Energy (NPMSE) of the skeleton. The action recognition is carried out using a linear Support Vector Machine (SVM). Experimental results on five challenging benchmarks show the effectiveness of our approach in terms of recognition accuracy and computational latency.

The local approach for the estimation of a Linear Parameter-Varying (LPV) model consists in an interpolation of a finite number of Linear Time-Invariant (LTI) systems called local LTI systems. Each local LTI system is obtained by performing an identification experiment at a fixed constant value of the parameter describing the dynamic variation of the LPV model. This parameter is referred in the literature as scheduling variable and the fixed values of the scheduling variable are simply called operating points or scheduling points. In order to improve the accuracy of the local method for the identification of LPV systems, the choice of the scheduling points and the inputs used at each scheduling point for the identification of the local LTI systems is addressed in this paper. To deal with this problem, an accuracy measure is first introduced. This measure is shown to be a linear combination of the classic A-optimality accuracy measure of the local LTI systems. Using this result, an algorithm is finally proposed to solve the experiment design problem.

International audience; In order to enhance the performance of electromagnetic interference (EMI) filters, it is necessary to identify high-frequency parasitic elements of their passive components, mainly those related to the coupled inductors. motivated by this issue, in this work a realistic high-frequency model is proposed for the coupled inductors. Actually, using interval analysis in particular the forward-backward contractor, a set-membership algorithm has been developed to estimate systematically the parasitic elements linked with the magnetic components. The main advantages of this algorithm compared to the fitting methods are the values of the estimated parameters are always positive and the corrupted data are taken into account. The comparison of the simulation results and the experimental data allows us to validate the proposed method.

With the availability of the recent human skeleton extraction algorithm introduced by Shotton et al. [1], an interest for skeleton-based action recognition methods has been renewed. Despite the importance of the low-latency aspect in applications, it can be noted that the majority of recent approaches has not been evaluated in terms of computational cost. In this paper, a novel fast and accurate human action descriptor named Kinematic Spline Curves (KSC) is introduced. This descriptor is built by interpolating the kinematics of joints (position, velocity and acceleration). To overcome the anthropometric and the execution rate variability, we respectively propose the use of a skeleton normalization and a temporal normalization. For this purpose, a new temporal normalization method based on the Normalized Accumulated kinetic Energy (NAE) of the human skeleton is suggested. Finally, the classification step is performed using a linear Support Vector Machine (SVM). Experimental results on challenging benchmarks show the efficiency of our approach in terms of recognition accuracy and computational latency.

The paper presents an online strategy for sensor and/or actuator fault detection and isolation applied to a dam-gallery. A recursive subspace identification algorithm is used to estimate the dam-gallery model parameters. The main contribution consists in developing a specific identification scheme, insensitive to a certain type of faults. That is, the identified parameters are invariant to the faults. A fault estimation procedure is proposed to detect potential faults. The proposed approach appears to be suitable for open channel systems for which the characteristics are not easily measurable.

Issues in Energy Efficiency of Data Centers (DC) are important, due to the cumulative effects of the increase in the DCs number and in the energy consumption per center. Developing new design recommendations to improve a cooling system efficiency, commonly quantified by the PUE metric (Power Usage Effectiveness) is one objective of the Green IT organizations. For existing DCs, without considering the optimization of the IT workload, a possible way to improve the DC's energy efficiency is to adjust the cooling setpoints. In this paper, a methodology based on predictive models is used to optimize the PUE by improving the cooling setting. The modeling approach consists in exploiting the temperatures and energy measurements at various operating conditions to predict the PUE behavior using data-driven models commonly called black box models. The optimization procedures are based on the simulation of these models in order to estimate the best working conditions.

International audience

The paper addresses the problem of minimality and identifiability of Switched AutoRegressive eXogenous (abbreviated by SARX) models. We propose a notion of identifiability and minimality for SARX models which depends only on the parameters of the model, not on data. We formulate conditions for minimality and identifiability of SARX systems. In particular, we show that SARX systems are generically identifiable.

Person identification is an important but still challenging problem in video surveillance. This work designs a completely automatic appearance-based person identification system, which has the ability to achieve new person discovery and classification. The proposed system consists of three modules: background and silhouette separation; feature extraction and selection; and online person identification. The Self-Adaptive Kernel Machine (SAKM) algorithm is used to differentiate existing persons who can be classified from new persons who have to be learnt and added. A new video database with 22 persons is created in real-life environments. The experimental results show that the proposed system achieves satisfying recognition rates of over 90% on person classification with novelty identification.

In this work, a recursive procedure is derived for the identification of switched linear models from input–output data. Starting from some initial values of the parameter vectors that represent the different submodels, the proposed algorithm alternates between data assignment to submodels and parameter update. At each time instant, the discrete state is determined as the index of the submodel that, in terms of the prediction error (or the posterior error), appears to have most likely generated the regressor vector observed at that instant. Given the estimated discrete state, the associated parameter vector is updated based on recursive least squares or any fast adaptive linear identifier. Convergence of the whole procedure although not theoretically proved, seems to be easily achieved when enough rich data are available. It has been also observed that by appropriately choosing the data assignment criterion, the proposed on-line method can be extended to deal also with the identification of piecewise affine models. Finally, performance is tested through some computer simulations and the modeling of an open channel system.

This paper presents the real-time emulation of a hydrogen-production process for assessment of an active wind- energy conversion system. The hardware-in-the-loop emulator of the electrolyzer consists of a power-electronic stage and a control stage. In the control board, the algorithmic equations of the electrolyzer modeling and its control should be implemented, as well as the emulator's power converter control. The causal ordering graph is used to model the electrolyzer and its auxiliary equipment. This model is capable of characterizing the relations among the different physical quantities and can be used to design the control system, ensuring an efficient and reliable operation of the electrolyzer. The proposed control method can manage the power and hydrogen flows. The simulation results have highlighted the variation domains and the relations among different physical quantities. The experimental results of the real-time emulation is based on a PHOEBUS advanced alkaline electrolyzer and shows the same electrical characteristics in real time with hardware.

In this paper, we are concerned by the improvement of the safety and the reliability of dynamical systems subjected to slow degradations. We propose a new prognosis strategy which aims at an efficient predictive maintenance by providing an estimation of the future state of the system. The prognosis method is based on an appropriated supervision technique which consists in drift tracking of the dynamical systems using AUDyC an auto-adaptative dynamical classifier. The proposed prognosis method is compared with a prognosis method based on the ANFIS approach (Adaptive Neuro Fuzzy Inference Sytem). These two prognosis methods are implemented and applied to a temperature controller.

This paper presents a new online clustering algorithm called SAKM (Self-Adaptive Kernel Machine) which is developed to learn continuously evolving clusters from non-stationary data. Based on SVM and kernel methods, the SAKM algorithm uses a fast adaptive learning procedure to take into account variations over time. Dedicated to online clustering in a multi-class environment, the algorithm designs an unsupervised neural architecture with self-adaptive abilities. Based on a specific kernel-induced similarity measure, the SAKM learning procedures consist of four main stages: Creation, Adaptation, Fusion and Elimination. In addition to these properties, the SAKM algorithm is attractive to be computationally efficient in online learning of real-drifting targets. After a theoretical study of the error convergence bound of the SAKM local learning, a comparison with NORMA and ALMA algorithms is made. In the end, some experiments conducted on simulation data, UCI benchmarks and real data are given to illustrate the capacities of the SAKM algorithm for online clustering in non-stationary and multi-class environment.

In this paper we propose a strategy for fault detection and isolation without any fixed model of the system to be supervised. The proposed approach is based on the identification of the parameters characterizing the system without any a priori knowledge. Our contribution consists in developing a specific identification scheme that is insensitive to a certain type of faults. The identified parameters are then invariant to the presence of actuator or sensor faults. Thereafter, a fault estimation procedure is proposed in order to detect sensor or actuator faults. The paper ends with a simulation example which highlights the effectiveness of the proposed approach.

International audience; The paper deals with the modelling of switching systems and focuses on the characterization of the local functioning modes using online clustering approach. The considered system is represented as a weighted sum of local linear models where each model could have its own structure. That implies that the parameters and the order of the switching system could change when the system switches. The presented method consists in two steps. First, an online estimation method of the Markov parameters matrix of the local linear models is established. Secondly, the labelling of theses parameters is done using a dynamical decision space worked out with learning techniques, each local model being represented by a cluster. The paper ends with an example, in view to illustrate the method performances.

This paper presents the application of an online identification neural technique to a single tube heat exchanger with a constant outer surface temperature. To show the feasibility of such an identification, the response to a sequence of random temperatures at the inlet of the inner fluid is studied. In the first part, the numerical solution is given, showing that the model cannot be a first order model. Then the principles of the neural technique are presented. The standard neural architecture, which normally calculates the output of the system directly from the input, is modified. A large number of local identical networks are used, each of them modelling an elementary module. It is shown that the neural model determined from the study of the first local network is representative of all the local networks (using the actual input data). At last it is shown that, when the networks are coupled, the output of the last network is in good agreement with the values obtained by the numerical model, but in a greatly reduced time.

This paper presents the application of an online identification neural technique to the prediction of the in-situ daily performance of solar collectors. First, it is shown that the use of the Laplace transform helps to find the order of an approximated model; the input of the studied system being the solar radiation. Then it is shown that a Neural Network Output Error (NNOE) model can be accurate using the right size of the regression vector; the learning database consisting of the data obtained during a half day. Finally, it is shown that a Multiple Inputs Single Output (MISO) NNOE model can be accurate; the inputs being the solar radiation and the thermal heat loss conductance that varies with the wind velocity. In any case the differential between the actual value of the daily energy and the value computed by a neural model (SISO—Single Input Single Output) or MISO) is less than 0.5%.

A new supervision system consisting of three modules is presented. The main novelty is the first module that corresponds to a modelling task. This module, which uses the auto-adaptive and dynamical clustering (AUDyC) neural network, allows us to continuously analyse and classify the functioning state of the monitored system using a dynamical modelling of all known modes (good/bad functioning modes represent different classes). The second module exploits these models of the functioning modes in order to detect “fast” and “slow” deviations. From membership degrees and from the information extracted by the monitoring module, the third module, dedicated to the diagnostics, informs the user about the functioning conditions of the system. In this paper, the main characteristics of the AUDyC and its abilities to model on-line non-stationary data are presented. Then, the description of the supervision system is given and some experimental results stemmed from a supervision application of a hydraulic system are discussed.

The problem of the recursive formulation of the MOESP class of subspace identification algorithms is considered and two novel instrumental variable approaches are introduced. The first one leads to an RLS-like implementation, the second to a gradient type iteration. The relative merits of both approaches are analysed and discussed, while simulation results are used to compare their performance with one of the existing techniques. Copyright © 2004 John Wiley & Sons, Ltd.

Apres avoir rappele l’equation principale que decrit l’evolution dynamique de la temperature de sorti d’un fluide chauffe par un rechauffeur electrique, un modele equivalent du second ordre est presente. Deux techniques neuronales sont alors utilisees pour identifier un rechauffeur. La premiere methode permet une identification hors ligne. La seconde permet une identification en ligne. Il est egalement montre que l’apparition de derives pourra etre detectee a l’aide de ces techniques.

In this paper, a fault diagnosis architecture based on a dynamical clustering algorithm is developed to detect and isolate faults in wind turbines. The challenge is to deal with different kinds of faults. Constraints on the time of detection are also added in the sense that a fault must be detected as soon as possible. Also, limited historical data corresponding only to normal operating modes are available. Our methodology is based on a data-driven model and is therefore not dependent of the physical models in the wind turbine. It consists of extracting, from sensor measurements, features that are fed into a dynamical clustering algorithm. The latter learns process behaviors characterized by clusters with the ability to update, recursively, the parameters of these clusters. These parameters are used to create detection signals and health indicators used for diagnosis.

International audience; Due to the decrease of sensor and actuator prices and their ease of installation, smart homes and smart environments are more and more exploited in automation and health applications. In these applications, activity recognition has an important place. This article presents a general architecture that is responsible for adapting automation for the different users of the smart home while recognizing their activities. For that, semi-supervised learning algorithms and Markov-based models are used to determine the preferences of the user considering a combination of: (1) observations of the data that have been acquired since the start of the experiment and (2) feedback of the users on decisions that have been taken by the automation. We present preliminarily simulated experimental results regarding the determination of preferences for a user.

This paper addresses the problem of driving the state of a linear discrete-time system to zero in minimum time. The inputs are constrained to lie in a bounded and convex set. The solution presented in the paper is based on the observation that the state sequence induced by the minimum-time control sequence is the sparsest possible state sequence over a certain finite horizon. That is, the desired state sequence must contain as many zero vectors as possible, all those zeros corresponding to the highest values of the time index. Hence, by taking advantage of some recent developments in sparse optimization theory, we propose a numerical solution. We show in simulation that the proposed method can effectively solve the minimum-time problem even for multi-inputs linear discrete-time systems.

This paper proposes a method for human action classification by utilizing correlations between SURF based descriptors. This approach provides us a novel type of descriptor that can be used for action classification. The method proposed is tested using an SVM classification technique. For evaluation purposes, the KTH action recognition dataset, which is a standard benchmark for this area is used as it is one of the most well known and challenging dataset. The method proposed was able to successfully classify different action classes.

The multi-agent simulation consists in using a set of interacting agents to reproduce the dynamics and the evolution of the phenomena that we seek to simulate. It is considered now as an alternative to classical simulations based on analytical models. But, its implementation remains difficult, particularly in terms of behaviors extraction and agents modelling. This task is usually performed by the designer who has some expertise and available observation data on the process. In this paper, we propose a novel way to make use of the observations of real world agents to model simulated agents. The modelling is based on clustering techniques. Our approach is illustrated through an example in which the behaviors of agents are extracted as trajectories and destinations from video sequences analysis. This methodology is investigated with the aim to apply it, in particular, in a retail space simulation for the evaluation of marketing strategies. This paper presents experiments of our methodology in the context of a public area modelling.

The goal of this paper is to contribute to the realization of a system able to recognize people in video surveillance images. The context of this study is to classify a new frame including a person into a set of already known people, using an incremental classifier. To reach this goal, we first present the feature extraction and selection that have been made on appearance based on features (from color and texture), and then we introduce the incremental classifier used to differentiate people from a set of 20 persons. This incremental classifier is then updated at each new frame with the new knowledge that has been presented. With this technique, we achieved 92% of correct classification on the used database. These results are then compared to the 99% of correct classification in the case of a nonincremental technique and these results are explained. Some future works will try to rise the performances of incremental learning the one of non-incremental ones.

We consider the problem of identifying a switched nonlinear system from a finite collection of input-output data. The constituent subsystems of such a switched system are all nonlinear systems. We model each individual subsystem as a sparse expansion over a dictionary of elementary nonlinear smooth functions shaped by the whole available dataset. Estimating the switched model from data is a doubly challenging problem. First one needs, without any knowledge of the parameters, to decide which subsystem is active at which time instant. Second, the representation of each nonlinear subsystem over the considered basis shall be performed in a high dimensional space. We tackle both tasks simultaneously by sparse optimization. More specifically, we view the switched nonlinear system identification problem as the problem of minimizing the l 0 norm of an error vector. We subsequently relax it into an l 1 convex minimization problem for which powerful numerical tools exist.

The analysis of digital video content is of fundamental importance for efficient browsing, indexing and retrieval of video database in order to facilitate user's access to relevant data. An essential first step is the parsing of the video content into visually-coherent segments, called shots. In this paper we propose an efficient approach for shot change detection and shot modeling based on a new Switched AutoRegressive (SAR) model identification technique. We make the assumption that pixel intensities of all the frames obey a SAR model where each linear sub-model of the SAR model corresponds to a shot and each discrete state corresponds to a different event in the video. Finally, experimental results on three different video sequences show the performance and the feasibility of the proposed approach.

In the context of evolutionary data classification, dynamical modeling techniques are useful to continuously learn clusters models. Dedicated to on-line clustering, the AUDyC (Auto-adaptive and Dynamical Clustering) algorithm is an unsupervised neural network with auto-adaptive abilities in nonstationary environment. These particular abilities are based on specific learning rules that are developed into three stages: “Classification”, “Evaluation” and “Fusion”. In this paper, we propose a new densities merge mechanism to improve the “Fusion” stage in order to avoid some local optima drawbacks of Gaussian fitting. The novelty of our approach is to use an ambiguity rule of fuzzy modelling with new merge acceptance criteria. Our approach can be generalized to any type of fuzzy classification method using Gaussian models. Some experiments are presented to show the efficiency of our approach to circumvent to AUDyC NN local optima problems.

This paper presents a kernel-based clustering algorithm called SAKM (Self-Adaptive Kernel Machine) that is developed to learn continuously evolving clusters from non-stationary data. Dedicated to online clustering in multi-class environment, this algorithm is based on an unsupervised learning process with self-adaptive abilities. This process is achieved through three main stages: clusters creation (with an initialization procedure), online clusters adaptation and clusters fusion. Thanks to a new specific kernel-induced similarity measure, the SAKM algorithm is attractive to be very computationally efficient in online applications. At the end, some experiments illustrate the capacities of our algorithm in non-stationary environment.

In the context of pattern recognition area, a small number of clustering techniques are dedicated to the on-line classification of non-stationary data. This paper presents a new algorithm designed with specific properties for the dynamical modeling of classes. This algorithm, called AUDyC (Auto-adaptive and Dynamical Clustering), is based on an unsupervised neural network with full auto-adaptive abilities. The classes modeling is obtained using Gaussian prototypes. Thanks to specific learning strategies, prototypes and classes are created, adapted or eliminated in order to incorporate new knowledge from on-line data. To do that, new learning rules have been developed into three stages: "Classification", "Fusion" and "Evaluation". The results show the real abilities of the AUDyC network to track classes and then to model their evolutions thanks to the adaptation of the prototypes parameters.

This paper proposes an improvement of the Cluster Detection and Labeling Neural Network. The original classifier criterion has been modified by introducing Elliptical Basis Functions (EBF) as transfer function of the hidden neurons. In the original CDL network, a similarity criterion is used to determine the membership to prototypes and then to classes. By introducing EBF, we have introduced degrees of membership leading to elliptic shape of classes. In this paper, the functioning of the original CDL network is summarized. Then, the improvements of the architecture in terms of network architecture, neuron activation function and learning stages are described. We present the improvement with EBF and the modification of the auto-adaptation neural network abilities. As validations of our architecture, we illustrate its benefits in comparison with the original CDL network.

This paper focuses on the feasibility of online identification of thermal systems. The transfer function is not looked for, but a black box model is obtained. In the first part, the principles of online identification are reminded. This leads to the definition of the regression vector and of the regressors. Then these principles are applied to neural based techniques which are adapted from standard ARX (AutoRegressive structure with eXtra inputs) and OE (Output-Error) models. For the Neural Network ARX (NNARX) model, only one example is given, which leads to a not fully satisfactory identification. This identification is based on the response of the system to random heat rates during random times. The validation is based on the response to another set of random heat rates and on the response of the system to a step function. For Neural Network OE (NNOE) model, the influence of the number of regressors is presented along with the influence of the number of neurons on the hidden layer. It is shown that many architectures lead to a good identification, but that some particular models may lead to a very poor result. To make the comparison possible between the proposed models, a distance criterion is computed. This leads to the choice of the best adapted architecture.

International audience; This paper is concerned with the identification of piecewise linear MIMO state space systems in a recursive way. The proposed method summons up benefits of recursive parameters estimation, on-line switching times detection and on-line order estimation. A structured identification scheme which applied on-line, allows to track both the extended observability matrix range space and its dimension. This method is used on-line to blindly identify switching systems and to label the different submodels. Since subspace identification methods rely on batch data block matrices, a minimum dwell time in each discrete state is necessary to achieve good performances. Simulation results comfort this point and illustrate the abilities and the benefits of the proposed approach.

International audience; The problem of the recursive formulation of the MOESP class of subspace identification algorithms is considered and two novel instrumental variable approaches are introduced. The first one leads to an RLS-like implementation, the second to a gradient type iteration. The relative merits of both approaches are analysed and discussed, while simulation results are used to compare their performance with the one of existing techniques

National audience

Editeurs invités du numéro spécial Identification des systèmes - Nouveaux développements et applications, sous la direction de : Marion Gilson, Laurent Bako, Fransisco Carrillo, Stéphane Lecoeuche, Guillaume Mercère; National audience; Incontournables de nombreuses disciplines scientifiques et technologiques (physique, chimie, biologie, économie...). La modélisation permet en effet de formaliser le comportement du processus étudié à l'aide d'une représentation, baptisée " modèle ", à partir de laquelle il est possible de comprendre, commander ou améliorer le fonctionnement du procédé analysé. Il est important de noter que ce champ thématique à caractère pluridisciplinaire (automatique, traitement du signal, statistique, analyse numérique, génie des procédés...) trouve ses applications dans des domaines très variés allant des processus de fabrication aux systèmes de transport, en passant par les processus environnementaux. L'objectif de ce numéro est de rendre compte des travaux récents dans le domaine de la modélisation et de l'identification des systèmes. Il est constitué de dix articles sélectionnés par le comité scientifique parmi les 43 communications présentées lors des troisièmes Journées Identification et Modélisation Expérimentale (JIME'2011) organisées sous l'égide du groupe de travail Identification de Systèmes du GdR MACS à l'École des Mines de Douai en avril 2011. Ces journées avaient pour objectifs de rassembler les acteurs francophones du domaine de l'identification des systèmes et de proposer une image de la recherche en identification et en modélisation expérimentale, grâce à des présentations orales, des sessions posters et des démonstrations logicielles. Les dix articles sélectionnés ont été retravaillés par les auteurs puis ont suivi le processus de relecture de JESA afin de constituer ce numéro. Les articles ainsi réunis permettent de mettre en lumière les derniers développements théoriques dans le domaine de l'identification des systèmes, ainsi que leurs nombreuses applications et interactions avec d'autres communautés scientifiques. Les avancées récentes concernent notamment, le choix optimal du signal d'excitation, l'identification de systèmes non linéaires, l'identification de systèmes bouclés, l'identification de modèles à temps continu pour des domaines variés tels que la robotique ou les bassins versants.