Machine learning for detecting structural changes from dynamic monitoring using the probabilistic learning on manifolds

Christian Soize (1) , André Orcesi (2)

1 MSME - Laboratoire Modélisation et Simulation Multi-Echelle
2 MAST-EMGCU - Expérimentation et modélisation pour le génie civil et urbain

This paper presents a machine learning approach for detecting structural stiffness changes of civil engineering structures considered as dynamical systems, using only an experimental database constituted of a small number of records related to the experimental first eigenfre-quency of the structure and a set of measured temperatures. Since the number of records in the experimental database is assumed to be small, the "small data" case must be is considered and consequently, the most classical methods of machine learning, which require "big data", cannot be used. The method of the probabilistic learning on manifolds recently introduced for analyzing small data is thus used. The validation of this method is performed on a box-girder bridge for which its dynamic monitoring has generated an experimental database. The proposed approach can be used for other similar problems.

Mots clés: bridge; detection; dynamic monitoring; structural change; probabilistic learning; machine learning; small data; identification; detection of disorders

Domaines: Génie civil; Statistiques [stat]; Machine Learning [stat.ML]; Probabilités [math.PR]

Christian Soize, André Orcesi. Machine learning for detecting structural changes from dynamic monitoring using the probabilistic learning on manifolds. Structure and Infrastructure Engineering, 2021, 17 (10), pp.1418-1430. ⟨10.1080/15732479.2020.1811991⟩. ⟨hal-02931147⟩