Machine Learning for Modeling the Bearing Capacity of Prestressed Concrete Elements Damaged by Corrosion

Abstract

This work aims to study the prediction of bearing capacity of prestressed concrete beams subjected to accelerated corrosion process using Machine Learning (ML) techniques. After data collection, the results were used to model the behavior of flexural stress, and predict their final load capacity, considering position, length, and width of the cracks generated by corrosion as well as loss of bearing capacity. The study presents an analysis of 363 days old beams damaged by corrosion, connected to a galvanostat for 62, and 121 days to make faster the process. Six beams were analyzed; five of them were used to train the model, the other works as a basis to compare the results thrown by the model with the real data. After the treat, the results showed that Bagged Trees Model fits better to real data, it was seen that removing atypical data improves the correlation of predicted and real data. The actual data were compared with two different prediction analyzes; for the first one, the atypical data were not removed; in the second one, the atypical data were eliminated with a statistical analysis. Obtaining relative error percentages of 15.18%, 14.59%, presenting two predictions: final load of 1444 kg and 1126 kg. Which means a resistant moment of 650 T-m, and 506.7 T-m respectively, taking as a prediction the second value in the safe side.