Automating Pipe Spool Fabrication Shop Scheduling for Modularized Industrial Construction Projects Using Reinforcement Learning

Abstract

Industrial projects are primarily constructed using a modularized and prefabricated approach. Modules are produced in an offsite fabrication shop and then transported to the construction site for installation. Thus, timely and sequence-specific delivery of preassembled construction elements is essential to prevent delays and ensure a smooth construction progress. As such, fabrication shop schedules are crucial for the success of the entire construction project. Unlike a manufacturing fabrication shop, a construction fabrication shop fabricates unique engineer-to-order products, resulting in challenging shop schedules that involve several conditions and constraints, including material availability, processing time, resource availability, and due dates. Further, the manual iterative nature of the scheduling process makes it laborious and time-consuming, especially when it happens on a frequent basis. This paper presents a deep reinforcement learning (DRL) model for automating the scheduling process. The scheduling process is formulated as a Markov decision process (MDP); then, DRL is used to solve the MDP efficiently for a fabrication shop with large state space. The model is tested on a data set from a pipe spool fabrication shop located in Alberta, Canada; the results show that the DRL outperforms the most popular dispatching rules. This study serves as a first attempt, to our best knowledge, to automate the scheduling process using DRL, thus creating a solid foundation for future advancement in automating and optimizing construction scheduling. Industrial projects primarily use prefabricated modules built in fabrication shops, which are then transported to sites for installation. Among these modules are pipe spools, which consist of pipes and other components assembled into a piping system for industrial projects. Scheduling the fabrication process of these highly customized pipe spools is challenging due to factors like material availability, readiness of spool to be manufactured, fluctuating processing times, resources availability, and due dates. As such, the scheduling process is conducted manually in an attempt to fulfill the previously mentioned factors or in some cases a dispatching rule is followed; however, the results are not satisfactory. Accordingly, this research develops an automated scheduling approach that relies on training a model on a historical data set; after learning from it, we feed the model with the data it has never seen to test its capabilities. We validate our approach using a case study of a fabrication shop that is located in Alberta, Canada. Industry professionals note that this automated method shows great promise in improving the scheduling process by effectively handling the complexities involved in the process.