Enhanced semi‐supervised ensemble machine learning approach for earthwork construction simulation activity sequence automatically updating driven by weather data

Abstract

Construction activity sequence changes are among the most crucial considerations in establishing construction simulation models. However, conventional simulation models are designed with a fixed sequence of simulation activities, which is unable to update automatically. Existing machine learning methods require abundant time for manual labelling and are unsuitable for construction data with high‐dimensional and heterogeneous characteristics. The motivation for this work is to develop a labour‐saving activity sequence classification model for earthwork construction simulation which realizes flexibly modifying simulation activities according to different weather conditions. Three heterogeneous semi‐supervised classifiers with complementary characteristics were ensembled to reduce the workload of manual labelling and enhance the generalization ability of activity sequence classification based on weather data. Furthermore, Dempster–Shafer‐based evidence reasoning improved by a security filtering mechanism was adopted to enhance the accuracy of semi‐supervised classification. The proposed enhanced ensemble semi‐supervised activity sequence classification model was embedded in an earthwork construction simulation model, which was evaluated in a case study of rockfill dam construction. The proposed classifier outperformed four common semi‐supervised methods and two common supervised methods in terms of accuracy and generalization ability. Additionally, the proposed simulation method outperformed conventional simulation methods in terms of the construction schedule, construction intensity and consistency of the simulated activity sequence with the true values by 65.55%, 28.47% and 88.15%, respectively.