Vous êtes ici : Accueil » Kiosque » Annonce

Identification

Identifiant: 
Mot de passe : 

Mot de passe oublié ?
Détails d'identification oubliés ?

Annonce

25 mars 2017

Sparse regression and dimension reduction for sensor measurements and data normalization


Catégorie : Stagiaire


English version on the associated webpage

IFP Energies nouvelles propose un stage en 2017 en régression parcimonieuse et réduction de dimension appliquées à la normalisation de mesures instrumentales et de données présentant un facteur d'échelle différent. L'objectif est de l'estimer en présence de données manquante et aberrantes, et de bruits, pour des signaux courts, présentant des variations d'amplitude importantes. Cette estimation doit se faire de la manière la plus automatisée possible, en se basant sur les propriétés et des a priori sur les données (parcimonie, positivité).

 

The instrumental context is that of multiple 1D data or measurements ym related to the the same phenomenon x, corrupted by random effects nm and a different scaling parameter am, due to uncontrolled sensor calibrations or measurement variability. The model is thus:

ym(k) = am x(k) + nm(k) .

The aim of the internship is to robustly estimate scaling parameters am (with confidence bounds) in the presence of missing data or outliers for potentially small, real-life signals x with large amplitude variations. The estimation should be as automatized as possible, based on data properties and priors (e.g. sparsity, positivity), so as to be used by non-expert users. Signals under study are for instance: vibration, analytical chemistry or biological data. Of particular interest for this internship is the study and performance assessment of robust loss or penalty functions (around the l2,1-norm) such as the R1-PCA or low-rank decomposition.

Subject in short

The instrumental context is that of multiple 1D data or measurements ym related to the the same phenomenon x, corrupted by random effects nm and a different scaling parameter am, due to uncontrolled sensor calibrations or measurement variability. The model is thus:

ym(k) = am x(k) + nm(k) .

The aim of the internship is to robustly estimate scaling parameters am (with confidence bounds) in the presence of missing data or outliers for potentially small, real-life signals x with large amplitude variations. The estimation should be as automatized as possible, based on data properties and priors (e.g. sparsity, positivity), so as to be used by non-expert users. Signals under study are for instance: vibration, analytical chemistry or biological data. Of particular interest for this internship is the study and performance assessment of robust loss or penalty functions (around the l2,1-norm) such as the R1-PCA or low-rank decomposition.

Location

IFP Energies nouvelles, Rueil-Malamaison (Paris suburbs), France

Information

Webpage: Sparse regression and dimension reduction for sensor measurements and data normalization

 

Dans cette rubrique

(c) GdR 720 ISIS - CNRS - 2011-2015.