O uso de aprendizado de máquina para o monitoramento de estruturas da construção civil.

Abstract

Structural Health Monitoring (SHM) is defined by the use of techniques for the gathering and assessment of data from engineering structures to identify and characterize their damage, aiming to ensure their safety and integrity. Although the costs associated with SHM are considered high, in the long-term, early detection of damage can represent significant savings in resources and accident prevention. Thus, this paper attempts to develop a civil engineering structures monitoring method so that machine learning algorithms are used, and, thereby, the relative displacement and aid in repair decisions are characterized. To this purpose, the main machine learning algorithms were compared, by creating predictive models based on the use of data from the structural health monitoring of the church of Vila Nova de Foz Côa, in Portugal. Were analyzed the performance metrics obtained with the forecasts made in the training database and, mainly, in the base of the test as criteria for evaluating the models. In this sense, those created from recurrent neural network algorithms such as Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) indicated eminent results. Therefore, the results obtained suggest which algorithms best adapted to the data and, consequently, presented superior predictive capacities given the database used in the experiments, which is configured as a relevant study in the elaboration of an aid tool in the structural health monitoring of civil engineering structures.