Detecção de dano em estruturas baseada em técnicas de aprendizado híbrido

Abstract

Engineering structures are often subject to changes in their dynamic properties due to various factors such as deterioration of construction due to time of use, the lack of maintenance, adverse environmental and operational effects, unexpected loads, structural reinforcement procedures, among others. In most cases, these modifications may be associated with structural damage. Continuous monitoring of the dynamic behavior of a structure makes it possible to investigate and analyze changes to ensure their integrity, ensuring greater comfort and safety for users and enabling more effective repair and maintenance actions, directly impacting the reduction of operating costs. For such, monitoring systems must have computational tools able to evaluate the acquired information in real time and continuously. In this dissertation, a new approach based on the direct analysis of acceleration measurements to detect structural damage and modifications using computational intelligence techniques is presented. The approach is to use together supervised (RNA) and unsupervised (k-means) classification methods to construct a hybrid classifier. The goal is to detect not only known damage states but also unidentified structural states. The idea is thus to allow the real-time monitoring of structural integrity, providing answers automatically and continuously, based only on vibrational testing environments with the structure in operation. To attest to the robustness of this approach, data from numerical simulations and experimental tests performed in the laboratory and in situ are used. The results show promising performance of the proposed methodology