Intelligent Recommendation for Safety Risk Precontrol Knowledge in Mega Hydropower Engineering Construction: A Hybrid Method Driven by Textual Semantics

Abstract

Mega hydropower engineering (MHE) is mostly located in complex areas with high mountains and deep valleys. The construction environment is harsh, the operating space is limited, and the work processes are highly coupled, resulting in a high incidence of various safety risks and difficulty in precontrol. The formulation of traditional risk precontrol measures primarily relies on manual consultation of safety management standards, which is time-consuming, labor-intensive, and prone to errors. This approach struggles to meet the modern engineering requirements for real-time performance and accuracy. Furthermore, existing research has largely concentrated on postaccident investigation and response, without sufficiently leveraging the value of construction safety knowledge, such as safety management standards and historical precontrol measures, to establish a proactive risk prevention system. To fill this gap, this study first analyzes the data characteristics of construction safety knowledge and extracts its semantic features using the contrastive sentence embedding (CSE) method and general language model (GLM). A multiscale text semantic feature fusion network based on an interactive attention mechanism is developed to realize the interactive fusion of semantic features. Second, a text semantic deep matching model, CSGLM-TSDM, was constructed based on this network to quantify the semantic similarity between each safety risk factor and its precontrol knowledge. Finally, an intelligent recommendation method based on semantic similarity is developed using knowledge retrieval and inference techniques. This method can recommend the precontrol knowledge and score corresponding to each safety risk factor in real time. The results show that (1) the F1-score of the CSGLM-TSDM semantic matching model on the self-built data set is 97.57%, and the model performance is superior, and (2) the intelligent recommendation method can quickly reason out the top n precontrol knowledge and their scores with the highest similarity to any safety risk factor. The method’s average precision for recalling the top 10 pieces of knowledge reached 98.95%. This study provides a reusable computational framework for building a proactive safety prevention and control system in the global hydropower and large-scale infrastructure construction sectors. It offers significant value as a reference for the intelligent transformation of knowledge-driven engineering safety management.