Imputação de dados ausentes em séries temporais de carbonatação dos concretos

Imputing missing data in time series of concrete carbonation

Abstract

The growing use of contemporary models for predicting carbonation fronts, such as artificial neural networks, requires greater rigor in terms of database completeness. Treating carbonation depth databases as time series is a favorable alternative for quality assurance. Therefore, the aim of this article is to identify the best technique for imputing missing data in time series of carbonation depths of concrete with different compositions. The database used was information collected from concretes subjected to natural carbonation over 20 years of exposure belonging to the GEDur/UFG. Ten imputation techniques were tried, such as foward fill, moving average, interpolation and Kalman filter. All the techniques and analyses were implemented using the Python programming language in an integrated development environment. Based on the performance metrics and visual analysis, it was found that monotonic cubic spline interpolation captured the pattern of the carbonation depth curve as a function of time with greater precision and accuracy, achieving a performance index of 0.998 and RMSE between 0.106 mm and 0.863 mm depending on the concrete sample.