Paramétrisation et classification de signaux en contrôle non destructif. Application à la reconnaissance des défauts de rails par courants de Foucault

Abstract : The work presented in this report deals with a device for the rail head defect detection and recognition. In the first section, a non contact eddy current inspection system dedicated to rail head non destructive testing in the exploitation situation is presented. The main conception options (differential measurement, bi-frequency, shielding...) are described and validated by in-situ experimental tests. A list of defect classes has been established and a representative data base has been also constituted to elaborate the processing system. The first part of the defect recognition process concerns the representation mode of sensor output signals. The main properties required for the parametrization are a great descriptive potential as well as a strong insensitivity to problem invariants (as play-back operation, scale factor, lift-off). An original parametrization procedure referred to as "Modified Fourier Descriptors" has been elaboreted and compared to parametrization of an autoregressive type. A parameter selection must then be carried out in order to maintain only the parameters relevant to class separability. For the parameter classification, the orthogonalization method and the sequential backward and forward procedures are compared. In order to select a subset of parameters, many stop criterions are also presented. The application of those methods to our data base is illustrated. The last section of this report is devoted to a supervised neuronal classification by means of multilayer perceptron and radial basis function. For these two types of networks, both global and partitioned approachs are exposed. In the first case, the muliclass problem is solved simultaneously while in the partial case, the problem of classification is subdivised into subproblems. The classification performance are given for the two approaches and it will be shown that the partitionning results are better and bear more relevance to our application in correspondance with a small learning dataset size.
Complete list of metadatas

Cited literature [38 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00006600
Contributor : Patrice Aknin <>
Submitted on : Wednesday, July 28, 2004 - 4:40:45 PM
Last modification on : Wednesday, June 26, 2019 - 9:56:02 AM
Long-term archiving on: Friday, April 2, 2010 - 8:45:35 PM

Identifiers

  • HAL Id : tel-00006600, version 1

Collections

Citation

Latifa Oukhellou. Paramétrisation et classification de signaux en contrôle non destructif. Application à la reconnaissance des défauts de rails par courants de Foucault. Traitement du signal et de l'image [eess.SP]. Université Paris Sud - Paris XI, 1997. Français. ⟨tel-00006600⟩

Share

Metrics

Record views

935

Files downloads

2912