Skip to Main content Skip to Navigation

Reconnaissance des formes dans un environnement dynamique appliquée au diagnostic et au suivi des systèmes évolutifs

Abstract : Several systems are evolving, i.e. their behavior is dynamic and it leads to changes in their functioning characteristics. The monitoring of the evolving systems functioning modes is a major problem for the diagnosis methods. Indeed, in these conditions it is necessary to use or to develop methods taking into account the new characteristic information of the current system behavior and permitting to adapt the known functioning modes. We have chosen to use pattern recognition methods for their ability to work on applications for which they only have observations. Several dynamic pattern recognition methods have been proposed in order to take into account of the changes realized in patterns and classes characteristics according to the time. The method Dynamic Fuzzy Pattern Matching (DFPM) has been developed to consider the gradual change of classes characteristics after the classification of each new pattern. The method integrates several mechanisms such as indicators of the patterns usefulness, a residual permitting to follow classes evolutions, and splitting and merging procedures in order to adapt dynamic classes. Several versions of the Dynamic Fuzzy K-Nearest Neighbours (DFKNN) have also been proposed: Supervised-DFKNN (S-DFKNN) and Semi-Supervised DFKNN (SS-DFKNN). These DFKNN methods use procedures to detect and confirm classes evolutions. Then, they realize the adaptation of the evolved classes by using the most characteristics patterns of the current trend of a system. The proposed methods permit to online detect the evolution of a system behavior, to validate this evolution and to proceed to the adaptation of a class when its characteristics have changed. Two others approaches of pattern recognition (structural and mixed) have also been proposed to deal with dynamic patterns. The structural approach is based on a segmentation method which does not necessitate the definition of an approximation error threshold, and on an adaptive number of primitives defined according to each characteristic phase of a dynamic pattern. Once, the dynamic patterns segmentation is realized, the method estimates their trend using the selected primitives. Then, the classification of all patterns can be realized using a measure of similarity. The mixed pattern recognition proposed method is based on the use of statistical and structural characteristics to realize the classification of dynamic patterns. A mixed signature is obtained for each dynamic pattern treated. This signature permits to obtain characteristic and interpretable information. A mixed similarity measure, based on structural and statistical characteristics, is then calculated from the mixed signature, to measure the membership of a pattern to a class. This measure also permits to quantify the evolution a pattern can realize after a change in the characteristics of a system. The set of proposed methods has been used on several simulated and real applications. These applications concerned the industrial domain (detection of a welding's quality, detection of a leak in a steam generator) and the clinical domain (characterization of the inter-segmental coordination of the gait patterns for hemiparetic patients).
Complete list of metadatas

Cited literature [78 references]  Display  Hide  Download
Contributor : Laurent Hartert <>
Submitted on : Wednesday, December 22, 2010 - 3:47:38 PM
Last modification on : Friday, May 17, 2019 - 11:38:34 AM
Long-term archiving on: : Thursday, June 30, 2011 - 1:46:20 PM


  • HAL Id : tel-00549782, version 1



Laurent Hartert. Reconnaissance des formes dans un environnement dynamique appliquée au diagnostic et au suivi des systèmes évolutifs. Automatique / Robotique. Université de Reims - Champagne Ardenne, 2010. Français. ⟨tel-00549782⟩