Modelagem explicativa para avaliação do estado de condição de pontes rodoviárias quanto ao julgamento de insuficiência estrutural a partir de relatórios de inspeção

Abstract

Starting from the assumption of the importance of the continuous study of elements that make up urban infrastructure, such as those of Special Arts Works (OAEs), which are bridges, bridges, viaducts and underpasses in Brazil; the research proposes an explanatory/analytical modeling of information from inspections available at SGO/DNIT, with the intention of evaluating the possible relationship between the opinions of evaluators with the characteristics of the bridges in the judgment of structural insufficiency, as well as the relationship of the scores attributed to the condition state. The procedure begins by selecting the important variables (stepwise backward) to explain the evaluators' opinions regarding structural insufficiency based on a preliminary analysis using binomial logistic regression. The procedure begins by selecting the important variables (stepwise backward) to explain the evaluators' opinions regarding structural insufficiency based on a preliminary analysis using binomial logistic regression. The same procedure is carried out for the scores assigned for the condition state using ordinal logistic regression. Next, the steps of predictive analysis are carried out with 6 (six) algorithms and the verification of their accuracy based on cross-validation in the modeling of the structural insufficiency judgment. Bayesian networks were also applied to verify the relationships between the independent variables, this in the judgment of structural insufficiency, as well as in relation to the scores attributed to the condition state. It was possible to attest that out of 12 (twelve) characteristics of the bridges, 5 (five), age, type of structure, lane width and two others related to the damage found on them, are the most relevant initially to explain the judgment of structural insufficiency. The binomial logistic model indicated a correct classification for 82.73% of the records investigated. With the cross-validation method, the binomial logistic regression model had the highest accuracy of 80.76% and in the machine learning models the maximum accuracy observed was 83.40%. In the ordinal logistic model, the variables selected (stepwise backward) for modeling were: spalling of concrete with exposed reinforcement; disaggregated concrete with exposed and oxidized reinforcement; iron oxidation; age of the bridge; lane width and type of structure. The best success rates were in classifications for grades 1 (92.2%) and 2 (78.0%). Finally, using the algorithms, 20 scenarios were evaluated in which the preponderant role of damage caused by the spalling of concrete with exposed reinforcement and disintegrated concrete with exposed and oxidized reinforcement was observed, for the judgment of structural insufficiency and for the attribution of condition notes. The good quality of adjustment of the logistic regression models and machine learning algorithms was an indication that it would be possible to formulate a model with sufficient predictive capacity to assist in classifying bridges regarding structural insufficiency and assigning scores for condition status.