Key points are not available for this paper at this time.
Abstract This study focuses on an investigative approach to combine system dynamics and machine learning algorithms to develop an early warning system for the safety management of construction projects. As the construction industry is highly accident-prone, developing a decision-support system has always been a challenge for the research community. Therefore, 53 indicators that influence each other and the construction phase were included in the planning phase of the model. The system dynamics model was validated using extreme state and sensitivity tests, which showed reasonable trends in the number of accidents. For each simulated project, all indicator data was stored in one dataset, using two different accident rates: one for serious and one for fatal accidents. Consequently, two separate datasets were generated, one for serious accidents, which was balanced, and one for fatal accidents. Machine learning was applied to both datasets to predict safety performance. The datasets were pre-processed so that the features consisted only of data from the planning phase, with the target feature being occurrence of accident. The study revealed two key findings. First, the study showed the possibility of combining system dynamics and machine learning for safety predictions in cases where real project data is not available. Secondly, the results showed that it is possible to carry out projects with a higher risk of major accidents and provide an early warning of poor safety performance. The data set with serious accidents resulted in lower accuracy but higher recall values. However, the models struggled to identify fatal accidents as the values for the fatal accident dataset were too low. Therefore, it was discussed how other safety measurements could be more appropriate. Thus, the combination of system dynamics and machine learning has the potential to serve as a decision-support tool in construction projects and to disseminate knowledge about safety performance.
Nishat et al. (Thu,) studied this question.