Damage identification in bridges combining deep learning and computational mechanic

Abstract

[EN]Civil infrastructures, such as bridges, are critical assets for society and the economy. Many of them have already reached their expected life and withstand loadings that exceed the design specifications. Besides, bridges suffer from various degradation mechanisms, including aging, corrosion, earthquakes, and, nowadays, the undeniable effect of climate change. This context has motivated an increasing interest in early detecting damage to prevent costly actions and dangerous failures. Structural Health Monitoring (SHM) consists of implementing effective strategies to continuously assess the health condition of structures using monitoring data collected by sensors. This dissertation focuses on the SHM problem of damage detection and identification. It is an ill-posed inverse problem that aims at inferring the health state of a structure from measurements of its response. The measurements include large amounts of noisy data affected by environmental and operational conditions, acquired with sensors of different nature. Solving such a multidisciplinary problem encompasses the use of applied mathematics, computational mechanics, and data science. In this dissertation, we exploit the potential of Deep Neural Networks in approximating complex inverse problems and employ computational parametrizations and the Finite Element Method to enrich the training phase by including damage scenarios. We explore two different approaches to the problem. In the first approach, we develop an outlier detection strategy to detect departures from the baseline condition. We only employ long-term monitoring data measured at the bridge during normal (healthy) operation. Starting from Principal Component Analysis (PCA) as a statistical data reconstruction technique, we design a specific Deep Autoencoder network that enhances PCA by adding residual connections to include nonlinear transformations. This architecture gains partial explainability by evaluating the contribution of nonlinearties over affine transformations in the reconstruction process. We also investigate the method performance when using local or global variables and evaluate the potential of combining both data sources in the damage detection task. In the second approach, we reach a higher level of damage identification by estimating its severity and location. The goal is to provide a suitable methodology for real full-scale applications that requires reasonable computational resources. We employ a calibrated computational parametrization to solve multiple Finite Element simulations under different damage scenarios. These synthetic scenarios enrich the training dataset of a Deep Neural Network that maps the response of the bridge with its health condition in terms of damage location and severity. Finally, we incorporate the effect of environmental and operational variability in the parametrization by applying a clustering algorithm to find representative samples among the entire dataset. We assume these samples cover most of the variability present in the data and consider them as starting points to generate synthetic training data. We apply the proposed methods to three main case study bridges with available monitoring data: the Beltran bridge in Mexico, and the Infante Dom Henrique bridge in Porto, and the Z24 bridge in Switzerland. Both structures resulted critical to validate and test the ability of the proposed methods and to demonstrate their applicability in the full-scale.

[ES]Esta tesis investiga la aplicación de técnicas Deep Learning y Mecánica Computacional en el ámbito de identificación de daños estructurales en puentes. En primer lugar, abordamos técnicas basadas puramente en datos, que emplean únicamente la respuesta estructural adquirida mediante un sistema de instrumentación (sensores). Estas técncias proporcionan un diagnóstico de alerta (daño- no daño). Empleamos un tipo de red neuronal conocido como Autoencoder, al que dotamos de una arquitectura particular que pretende replicar transformaciones afines (como el Análisis de Componenetes Principales) e incorporar transormaciones no lineales de forma interpretable y comprensible. Con el objetivo de alcanzar un nivel más elevado en el diagnóstico, estudiamos una metodología híbrida que incorpora la mecánica computacional como fuente adicional de datos. Mediante el uso de una parametrización de elementos finitos, obtenemos la respuesta estructural sintética ante diferentes escenarios de daño, clasificados por su localización y su grado de severidad. Esta metodología require una calibración previa de la parametrización de acuerdo a un estado de referencia, y los escenarios generados se emplean para entrenar una red neuronal profunda capaz de estimar la localización y severidad de un daño cuando se obtienen nuevas mediciones en el sistema de instrumentación.