In factory production of Modular Integrated Construction (MiC), traditional scheduling methods based on manual experience often suffer from decision-making delays and limitations in achieving the global optimum, easily leading to supply chain bottlenecks due to high variability in order changes and the dynamic scheduling of heterogeneous resources. Our "IoT-BIM Intelligent Scheduling Platform Driven by Large Language Models and Knowledge Graphs," developed in collaboration with Hong Kong Polytechnic University and the University of Hong Kong, is designed to solve this complex combinatorial optimisation problem.
The platform is deeply integrated with the Domain-Adaptive MiC Large Language Model (MiC-LLM), the Dynamic Graph Knowledge Base (GKB), and the IoT Platform. Three core technologies have constructed an innovative paradigm of "semantic-driven decision-making." Through its unique TextSchedule Functionally, the system utilises Low-Rank Adaptation (LoRA) fine-tuning technology and the Retrieval-Augmented Generation (RAG) mechanism to accurately parse managers' natural language instructions, transform unstructured requirements into structured constraints, and invoke heuristic optimisation algorithms to generate the optimal production Gantt chart.
To address production disruptions, the platform has built an LLM-based Multi-Agent System (LLM-MAS) that employs a hierarchical reinforcement learning framework. It can complete the closed-loop dynamic rescheduling within 1 minute, from problem construction and constraint extension through model verification, to achieve precise resource matching.
This technology not only automates scheduling decisions but also uses NVIDIA Isaac Sim to build a high-fidelity digital twin environment, enabling real-time visualisation of the entire production, transportation, and assembly processes. Currently, the platform is undergoing pilot deployment at the HaiChuang Intelligent Manufacturing Factory of China Overseas Innovation & Technology Limited. These pilots aim to bridge the semantic gap between standardised Multi-trade Integrated Mechanical, Electrical, and Plumbing (MiMEP) products, BIM geometric models, and production work orders. In the future, this innovation will drive the transformation of the construction supply chain towards data-driven "intelligent production scheduling," accelerating the in-depth implementation of Construction 2.0.
Figure 1：Scheduling: Intelligent Production Planning
Figure 2: TextSchedule, extracts key scheduling parameters from natural-language descriptions