Identificação de danos em elementos estruturais lineares utilizando metodologias de modelo híbrido baseado na física e orientado a dados, machine learning e digital twin

Abstract

This work presents a novel structural health monitoring framework for damage identification in linear steel structural elements (compressed bars and beams), in the context of dynamical systems. The framework integrates a hybrid physics-based and data-driven model (that can result in a generalizable, accurate, interpretable and computationally efficient model) and supervised machine learning methods, to construct a digital twin. The governing equations of motion of the healthy structure, discovered by hybrid modeling, are used to simulate the response of the system with damage at different locations and intensities. From these simulations a dataset is constructed to train the machine learning classifiers, considering the beam scenarios with damage and undamaged. The constructed digital twin relates the inputs of the physical twin to specific damage scenarios to quickly warn if there is damage, where it is located and what is its severity, supporting engineering decisions. The DT framework was evaluated in different application configurations, demonstrating the potential contribution to the establishment of structural integrity monitoring systems. Support Vector Machine was the best performing classifier, with a precision of 93.37% for the cantilever bar and 80.33% for the fixed-end bar. The damage to the cantilever and fixed-end bars, considering axial vibration, was identified and robust for certain noise levels. The identification of damage in beams, considering transverse vibration, proved promising, with a precision of 84.49% for the Support Vector Machine classifier, using the displacement feature, and 99.98% for the Quadratic Discriminant Analysis classifier, using the displacement and acceleration features.