Estimating missing values in compressive strength of cementitious materials: A machine learning and statistical approach with irregular data

Abstract

This study focuses on predicting missing compressive strength values in cementitious materials during the curing process, utilizing time-domain reflectometry (TDR) measurements. TDR is conducted at 30 different curing times, but compressive strength data is available only at 13 intervals due to sample limitations. The study employs statistical models (ARIMA, Kalman filter, MICE) and machine learning models (LSTM, BiLSTM) to predict missing values based on the available data. Data is categorized into a single variable (compressive strength only) and multiple variables (including TDR measurements). The Kalman filter exhibits the lowest error ratio for single-variable predictions, while the MICE model proves most effective under multiple-variable conditions. This demonstrates that integrating the MICE model with TDR measurements can effectively estimate missing compressive strength values, with the Kalman filter serving as a viable alternative for single-variable scenarios.