Failure prognostics by support vector regression of time series data under stationary/nonstationary environmental and operational conditions ; Prédiction de données de séries chronologiques avec des méthodes basées sur la régression à vecteurs de support dans des conditions environnementales et opérationnelles stationnaire/non-stationnaire

Open Access

Abstract

This Ph. D. work is motivated by the possibility of monitoring the conditions of components of energy systems for their extended and safe use, under proper practice of operation and adequate policies of maintenance. The aim is to develop a Support Vector Regression (SVR)-based framework for predicting time series data under stationary/nonstationary environmental and operational conditions. Single SVR and SVR-based ensemble approaches are developed to tackle the prediction problem based on both small and large datasets. Strategies are proposed for adaptively updating the single SVR and SVR-based ensemble models in the existence of pattern drifts. Comparisons with other online learning approaches for kernel-based modelling are provided with reference to time series data from a critical component in Nuclear Power Plants (NPPs) provided by Electricité de France (EDF). The results show that the proposed approaches achieve comparable prediction results, considering the Mean Squared Error (MSE) and Mean Relative Error (MRE), in much less computation time. Furthermore, by analyzing the geometrical meaning of the Feature Vector Selection (FVS) method proposed in the literature, a novel geometrically interpretable kernel method, named Reduced Rank Kernel Ridge Regression-II (RRKRR-II), is proposed to describe the linear relations between a predicted value and the predicted values of the Feature Vectors (FVs) selected by FVS. Comparisons with several kernel methods on a number of public datasets prove the good prediction accuracy and the easy-of-tuning of the hyperparameters of RRKRR-II. ; Ce travail de thèse est motivée par la possibilité de surveiller l'état des composants de systèmes d'énergie pour leur utilisation prolongée et sécuritaire, conformément à la pratique correcte de fonctionnement et des politiques adéquates de maintenance. La motivation est de développer des méthodes basées sur la régression à vecteurs de support pour la prédiction de données de séries chronologiques dans des conditions environnementales ...