Investigation on EWS detection and classification through Deep Neural Networks

Details

Supervisors

Massart, Estelle ; Absil, Pierre-Antoine

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électromécanicien, à finalité spécialisée: mécatronique

Abstract

enForecasting sudden changes in complex systems is an important but complex task. Until now multiple methods have been developped with varying accuracy. Recently, deep learning has been used for bifurcation detection and classification and has proven its capabilities to outpass the generic methods. One benchmark in this field is the work of Bury et al. which had promising results for the detection and classification of codimension-1 bifurcations, namely Hopf, fold and branch (transcritical and pitchfork) bifurcations for continuous-time systems thanks to a CNN- LSTM architecture. But transcritical and pitchfork bifurcations represent fundamentally different dynamical mechanisms and should not be grouped together. This study aims to improve the work of Bury et al. to classify separetly pitchfork from transcritical bifurcations.