Interpretable decision support system for tower crane layout planning: A deep learning-oriented approach

Concerning the deployment of heavy on-site machinery to transport construction components, tower crane layout planning (TCLP) has an essential impact on construction safety and efficiency. The decision-making process for TCLP mainly relies on the construction managers’ experience, leading to inconsistent design quality. Insufficient attention has been given to making the TCLP evaluation results interpretable and providing real-time feedback to facilitate the decision support processes that may compensate engineers’ potential inexperience and inability to address site complexities. Currently, deep learning methods are extensively employed in novel tasks to extract patterns from datasets. Therefore, this study introduces a deep learning-based interpretable decision support system for TCLP (IDSS-TCLP) to real-time assess selected TCLP and provide users with specialized guidance via an interpretable mechanism. This system originates from the TCLP decision process, sequentially connecting four decision engines for the Checker, Indicator, Corrector, and Improver. The Checker is responsible for evaluating essential parameters for crane type selection. The Indicator is designed to assess the lifting safety and efficiency performance. The Corrector aims to identify common design issues, and the Improver is tasked with proposing a more proper TCLP given the current input. The Checker employed mathematical equations to filter out unqualified parameters, while the Indicator and Corrector leveraged various deep neural networks to fulfill their respective functions. The generative adversarial networks (GAN) framework was employed within the Improver to generate an appropriate TCLP. The Indicator selected ResNet-50 and Inception-v3 to predict the lifting safety and efficiency scores based on accuracy rate. The Corrector encompasses both ResNet-101 and Inception-v3 to identify common design problems. Optimal TCLP outcomes were achieved by the Improver sequentially applying neural networks with λ values of 100 and 10, guided by improvement rate and success rate results. Furthermore, a graphical user interface (GUI) for this IDSS-TCLP was developed to present the evaluation process. An interpretable mechanism was introduced to integrate decision engines with the GUI, facilitating human–computer interaction through interpretable decision suggestions. A real construction project was used as validation, revealing the applicability and reasonableness of IDSS-TCLP. This proposed toolkit integrating deep learning neural networks and an interpretable mechanism will catalyze further investigation in developing accessible and scalable deep learning-based tools supporting on-site construction management.