A novel statistical pattern recognition model is proposed in order to solve specific industrial diagnosis problems. The representation of discriminative parameters as a function of the operating point parameters enables accurate operating mode distance rejection. Adapted distance rejection options are presented in order to deal with unknown classes. These methods have been applied to a real world diagnosis problem: the monitoring of car catalytic converters.

On real processes many functional states are usually observed. however some of them may represent difference significant levels (or rates) of the same functional mode. They do not represent by themselves a functional mode, they are only sub-modes of the same functional mode (e.g. one of the functional mode of my car is tank empty; some functional states observed may be 10% empty tank, 50% empty tank, 96% empty tank). Thus the interest is not only to diagnose the functional mode (e.g. tank empty, sleeping driver) but also to highlight the gravity level of this functional mode (e.g. the reaction in the middle of desert, confronted to a 4% empty tank is not the same as confronted to a 98% empty tank). The aim of this paper is to present a diagnosis of the current functional mode of a process and its gravity level. Usually in pattern recognition area, a membership function is a monotonic decreasing function of a Euclidean distance between two objects. Those objects represent two states of the process and a distance here is a dissimilarity measure between those states. Such a distance is defined in all the fuzzy subset associated with this membership function. So that function is monotonic decreasing in all subset directions. In this paper directional membership functions are proposed. In this case the distance is defined only by reference to a path describing the evolution from one functional state to another one. Then the obtained membership function is oriented according to this path and do not decrease identically within all directions in the subset. Such membership functions are then suitable in order to diagnose the state associated with data evoluting between known functional modes. An application to the French telephone network illustrates this method.

This paper introduces a new method for transforming the images of a document into its layout structure by the use of an analysis system and several models. The analysis system is independent of the document, and can recognize various documents with the help of some generic models. The definition of a model is based on a document's generic layout structure. The models contain rules necessary for identifying layout structure components.

In this paper, a new approach to the tracking and matching processes is proposed. These two usually distinct operations are merged into a single one. The two binocular image sequences are seen as being a unique entity. This is done by using a clustering-based approach. We discuss about the determination of the best fitted invariant 3D region descriptors and the corresponding feature space. An optimal configuration of the binocular system is deduced from these considerations. This tracking and matching strategy allows the simultaneous processing of 2 X N binocular frames and the handling of occlusions phenomena and nonrigid motion segmentation and object deformation.

It is the authors' belief that edge detection should crucially depend on the forms of smoothing filters. After an investigation of different filter forms for step edge model, the authors give two propositions for ramp edge detection. By applying them to a simulated image with ideal ramp edges and real images that seem to have more ramp edges, the two propositions are confirmed.>

The deduction of structure from motion using projective geometry is considered. It is shown that, linear camera motion, 2D feature matching and 3D feature recovery can be performed by a Hough transform scheme. Some experimental results on synthesized and real image sequences are illustrated.>

A loose-pattern process approach to clustering sets consists of three main computations: loose-pattern reject option, tight-pattern classifcation, and loose-pattern assigning classes. The loose-pattern rejection is implemented using a rule based on q nearest neighbors of each point. Two clustering methods, GLC and OUPIC, are introduced as tight-pattern clustering techniques. The decisions of loose-pattern assigning classes are related to a heuristic membership function. The function and experiments with one set is discussed.

The concept of a dependable complexity index for a system is presented. The proposed complexity index reflects the cost of testing for a failed component in a system made up of several components and equipped with some detectors. The complexity measure is applied to determine an optimal number of detectors in a system.

A multivariate nonparametric pattern recogiton system is described for the surveillance of a high-flux isotope reactor. Two nonparametric methods are worked out: one using the Bayes rule with the Rosenblatt-Parzen estimator for the probability law, and one using the k-nearest neighbor rule. Performances are evaluated by comparing the probability of misclassification between the two chosen clases: the first corresponds to a nonaction of the reactor operator on its power and the second to an action of the pilot. Processing is perfonned on the power sigal of the reactor which is an observation corrupted by noise. The system has been tested on several experiences and implemented to work in real time on the reactor. The aim Is to conceive a computer-aided decsion system for the reactor's pilot.