Construction of knowledge graph based on computer vision and ontology model for recognition of unsafe operations
FU Huimin
ZHENG Gang
Abstract：[Objective]With the rapid development of power engineering,construction site safety has become increasingly critical.Traditional manual inspection methods are time-consuming and prone to errors.In recent years,advancements in computer vision,deep learning,and knowledge graph technologies have made it possible to automatically recognize unsafe operation behaviors.However,existing computer vision methods have limitations in detecting small objects and lack high-quality databases for unsafe operation inference.To address these issues,knowledge graphs,ontology models,graph databases,and computer vision techniques were integrated to detect unsafe operations through entity detection,scene analysis,and spatial relation reasoning.An improved self-attention mechanism was also introduced to enhance small object detection capabilities.[Methods]The proposed method mainly involved ontology model construction,knowledge extraction,and knowledge reasoning.First,an ontology model of construction safety was built based on engineering documents,historical accident reports,and safety hazard reports,with information categorized into six types:entities,attributes,time,space,events,and attribute values,which were represented by normative knowledge.Second,computer vision techniques were employed to detect entities and their attributes and extract spatial relationships between entities.A Mask region-based convolutional neural network(Mask R-CNN)was used for object detection,with an improved self-attention mechanism incorporated to improve small object detection accuracy.As a result,model performance was optimized,and computational complexity was reduced.Finally,a Neo 4j graph database was utilized to store entities and their relationships,enabling automatic recognition of unsafe operations through database queries.In this way,structured reasoning for construction safety knowledge was achieved,and the intelligent level of recognizing unsafe operations was enhanced.[Results]In the experiments,a power engineering construction site was used as the test environment,and six kinds of unsafe operations that could lead to high-altitude falling were selected for simulation experiments.The simulation results indicate that the proposed method outperforms existing approaches in both detection accuracy and training efficiency.Particularly,the improved model demonstrates superior accuracy in small object detection.Additionally,scene segmentation was conducted using a feature pyramid network(FPN)and a unified perceptual parsing(UPP)method,which significantly improved the scene understanding capability of the model.Furthermore,the knowledge reasoning approach based on the Neo 4j graph database effectively integrates entity attributes and spatial relationships,enhancing the automation of unsafe operation recognition.[Conclusion]The proposed method can accurately detect unsafe operations in complex construction environments,thereby improving the intelligence level of construction site safety management.The key innovations of this research are as follows:integrating computer vision with an ontology model to enhance automation in construction safety management;improving the self-attention mechanism by modifying convolutional kernels and introducing a global max-pooling layer,which enhances the small object detection capability of the Mask R-CNN;incorporating the Neo 4j graph database for structured storage and reasoning of construction safety knowledge.This study provides an efficient and scalable solution for the automatic recognition of unsafe operations on construction sites.
Keywords：knowledge graphontology modelcomputer visionself-attention mechanismsmall object detectionunsafe operation behaviorgraph database
Publication Date：2025-07-25
Online Publishing Date：2025-09-18(First online date of this platform, not the publication date of the document)
Pages：8( 501-508 )