Working status prediction for a high-formwork support system using finite element model-informed deep learning and GPT-aided method

Researchers have developed an innovative approach to predict the working status of high-formwork support systems (HFSS) by combining finite element model (FEM) simulations with deep learning and large language models (LLMs). Published in *Smart Construction*, this study addresses the challenges of structural health monitoring (SHM) by leveraging a genetic algorithm (GA)-optimized FEM to generate training data for a convolutional neural network (CNN) classifier. The framework also integrates a Retrieval-Augmented Generation (RAG) model with a knowledge graph (KG) to automatically generate reasonable SHM reports for HFSS, demonstrating superior performance over common methods.

High-formwork support systems (HFSS) are critical in construction but prone to collapses due to inadequate monitoring. Common methods rely on expensive and complex experiments, limiting their practicality. This study proposes a data-driven solution using FEM simulations and deep learning to predict HFSS working statuses—normal, local instability, and fully unstable—with high accuracy.

The research begins with developing and optimizing an FEM of an HFSS using a genetic algorithm (GA) to minimize discrepancies between simulated and experimental data. The optimized FEM generates datasets for training a CNN classifier, which achieves an impressive accuracy in predicting structural statuses. Experimental validation on a full-scale HFSS confirms the CNN's superiority over support vector machines (SVM), with the CNN outperforming SVM in classification tasks.

To streamline SHM reporting, the study introduces a Retrieval-Augmented Generation (RAG) model, combining GPT-4 with a domain-specific knowledge graph (KG). The RAG model generates detailed SHM reports, evaluated using metrics like BLEU, ROUGE, and cosine similarity, demonstrating its effectiveness in producing accurate and contextually relevant reports compared to standalone GPT-4.

Key contributions of the study include:

1. A GA-optimized FEM for generating reliable training data under varied HFSS conditions.

2. A CNN classifier with high accuracy in predicting structural statuses.

3. An RAG model that automates SHM report generation, reducing reliance on subjective expertise.

The framework's potential applications extend to complex structures, with future work focusing on integrating multimodal models and edge computing for broader deployment.

This paper, "Working Status Prediction for a High-Formwork Support System Using Finite Element Model-Informed Deep Learning and GPT-Aided Method," was published in Smart Construction.

Zhao L, Mbachu J, Liu S, Zhang R. Working Status Prediction for a High-Formwork Support System Using Finite Element Model-Informed Deep Learning and GPT-Aided Method. Smart Constr. 2025(2):0015, https://doi.org/10.55092/sc20250015.