Identification par les techniques des sous-espaces : application au diagnostic

Abstract : This thesis focus on MIMa system identification and sensor fault detection and isolation with subspace methods. At the beginning, linear system identification is tackled. First, a presentation of classical subspace methods based on the estimation of the observability matrix and/or state sequence matrix is presented; then a new approach based on a FIR modelling is proposed. This approach is based on the estimation of the Markov parameters which allows the determination of a minimal and balanced realisation using algorithm ERA. Three methods for identification of linear dynamic systems with coloured noises are proposed. ln the context of switching systems identification, a multiple model modelling with binary weighting functions are used. Two methods are proposed to cope with this unsupervised learning problem: the first carries out a classification withtlchange detection methods, but the second performs this classification by finding hyperplanes which represent the local models in the inputs and outputs space. Both methods estimate a minimal and balanced realisation of each local model by least squares method with the knowledge of the weighting functions obtained by classification. A multiple model approach is used to deal with nonlinear system identification. The weights of the local models are supposed to be known and a method giving a minimal and balanced realisation of each local model is proposed. Finally, sensor fault detection and isolation in linear systems is performed without the knowledge of a model. A method only based on the knowledge of the inputs and outputs is proposed. Since no estimated mode! is used, the proposed method does not suffer from parametric uncertainties. By weighting the state with the high power of the state matrix which is supposed to be stable, the state influence is removed and the input influence is suppressed by orthogonal projection, then, we obtained a structured residual. This residual allows the detection and the isolation of the sensor faults.
Document type :
Theses
File URL :
http://docnum.univ-lorraine.fr/prive/INPL_T_2004_PEKPE_K_M.pdf
Complete list of metadatas

https://hal.univ-lorraine.fr/tel-01750053
Contributor : Thèses Ul <>
Submitted on : Thursday, March 29, 2018 - 12:36:45 PM
Last modification on : Tuesday, February 12, 2019 - 4:50:07 PM

Identifiers

  • HAL Id : tel-01750053, version 1

Collections

Citation

Komi Midzodzi Pekpe. Identification par les techniques des sous-espaces : application au diagnostic. Autre. Institut National Polytechnique de Lorraine, 2004. Français. ⟨NNT : 2004INPL111N⟩. ⟨tel-01750053⟩

Share

Metrics

Record views

15