Structural damage identification of existing bridges using machine learning-aided nested sequential optimization

Structural damage identification of existing bridges using machine learning-aided nested sequential optimization

Článek Scopus

The early detection of structural damage is of paramount importance for the maintenance and prolongation of the service life of existing bridges. The identification of structural damage is typically conducted using nondestructive vibration experiments in conjunction with a mathematical technique known as model updating. This paper presents the application of two distinct model updating methodologies. The first method employs a metaheuristic optimization technique with nested multilevel sampling to efficiently explore the design parameter space. The second method frames model updating as an inverse problem and employs a machine learning-based surrogate model to approximate the inverse relationship between the structural response and the structural parameters. Based on the analysis of both approaches, a new method is proposed that combines the advantages of both methods in a novel way. The method employs iterative targeting and design space reduction, in conjunction with the use of an artificial neural network, to approximate the inverse design space as accurately as possible. This is achieved by sequential training the network on the optimal data set. This procedure significantly accelerates the convergence of the optimization process and improves the accuracy of the obtained solution. All methods are applied to the damage identification of three structures: single-span and two-span steel plate girders and a riveted truss bridge. The ability to detect damage in varying locations and over different scales is assessed. The results obtained are compared in terms of the convergence speed and the accuracy of the solutions obtained. Additionally, different settings of methods varying in the number of simulations, levels and the parameter controlling the convergence rate are also studied. In all the tested cases, the superiority of the newly proposed method in terms of accuracy and computational requirements has been confirmed. The new method, while maintaining the same number of simulations, on average increases the accuracy of the obtained results by a factor of 30. It also identifies damage with equal or better accuracy at one-tenth the number of simulations than other tested methods.

The early detection of structural damage is of paramount importance for the maintenance and prolongation of the service life of existing bridges. The identification of structural damage is typically conducted using nondestructive vibration experiments in conjunction with a mathematical technique known as model updating. This paper presents the application of two distinct model updating methodologies. The first method employs a metaheuristic optimization technique with nested multilevel sampling to efficiently explore the design parameter space. The second method frames model updating as an inverse problem and employs a machine learning-based surrogate model to approximate the inverse relationship between the structural response and the structural parameters. Based on the analysis of both approaches, a new method is proposed that combines the advantages of both methods in a novel way. The method employs iterative targeting and design space reduction, in conjunction with the use of an artificial neural network, to approximate the inverse design space as accurately as possible. This is achieved by sequential training the network on the optimal data set. This procedure significantly accelerates the convergence of the optimization process and improves the accuracy of the obtained solution. All methods are applied to the damage identification of three structures: single-span and two-span steel plate girders and a riveted truss bridge. The ability to detect damage in varying locations and over different scales is assessed. The results obtained are compared in terms of the convergence speed and the accuracy of the solutions obtained. Additionally, different settings of methods varying in the number of simulations, levels and the parameter controlling the convergence rate are also studied. In all the tested cases, the superiority of the newly proposed method in terms of accuracy and computational requirements has been confirmed. The new method, while maintaining the same number of simulations, on average increases the accuracy of the obtained results by a factor of 30. It also identifies damage with equal or better accuracy at one-tenth the number of simulations than other tested methods.

Aimed multilevel sampling;ANN-AMS;Artificial neural network;FE model updating;Modal parameters;Structural damage detection;Structural vibration

Aimed multilevel sampling;ANN-AMS;Artificial neural network;FE model updating;Modal parameters;Structural damage detection;Structural vibration

ŠPLÍCHAL, B.; LEHKÝ, D.; LAMPEROVÁ, K.

29.09.2025

NEURAL COMPUTING & APPLICATIONS

37

32

Spojené království Velké Británie a Severního Irska

27301

27324

24

https://link.springer.com/article/10.1007/s00521-025-11673-w

http://hdl.handle.net/11012/255639