A systematic review and evaluation of synthetic simulated data generation strategies for deep learning applications in construction

The integration of deep learning (DL) into construction applications holds substantial potential for enhancing construction automation and intelligence. However, successful implementation of DL necessitates the acquisition of substantial data for training. The acquisition process can be error-prone, time-consuming, and impractical. For this reason, synthetic simulated data (SSD) has emerged as a promising alternative. While various strategies have been developed to generate such data, a systematic review and evaluation are lacking to aid researchers and professionals in selecting appropriate strategies for their applications. To fill this gap, this paper conducts a comprehensive literature review related to SSD generation and applications, and develops a guideline for strategy selection. Two hundred and eight articles are identified from the academic database Web of Science by using PRISMA. After thoroughly analyzing the literature, seven SSD generation strategies are identified and evaluated across six metrics. Based on the performance of each strategy, a guideline is synthesized as a decision tree. Users only need to follow the steps and answer the questions in the decision tree, and then they will get the recommended SSD generation strategy. We demonstrate the guideline’s effectiveness by comparing its recommendations with the strategies chosen by researchers in existing DL construction applications and achieve a matching rate of 82%.