FU Yan, LIU Zhihao, YE Ou. A method for constructing a knowledge graph of unsafe behaviors in coal mines[J]. Journal of Mine Automation,2024,50(1):88-95. DOI: 10.13272/j.issn.1671-251x.2023060014
Citation: FU Yan, LIU Zhihao, YE Ou. A method for constructing a knowledge graph of unsafe behaviors in coal mines[J]. Journal of Mine Automation,2024,50(1):88-95. DOI: 10.13272/j.issn.1671-251x.2023060014

A method for constructing a knowledge graph of unsafe behaviors in coal mines

  • Although knowledge graphs have been widely applied in various fields, there is relatively little research on coal mine safety, especially in the area of unsafe behavior underground. A bottom-up knowledge graph of unsafe behaviors in coal mines has been constructed. Firstly, a combination of traditional machine learning and deep learning algorithms is used for named entity recognition. RoBERTa is used for word vectorization. The bidirectional long short term memory network (BiLSTM) is used to annotate the vectors, improving the network model's capability to capture contextual features. To solve the problem of insufficient data volume in the dataset of unsafe behaviors in coal mines, a multi-layer perceptron (MLP) is used. The conditional random field (CRF) model is adopted to solve the problem of unrecognized word relationships and capture full-text information and prediction results. Secondly, based on the structural characteristics of the statements, a dependency syntax tree structure based on the knowledge "entity - relationship - entity" triplet is designed to extract and represent knowledge resources in the field of unsafe behavior underground. Finally, a knowledge graph of unsafe behaviors underground is constructed. The experimental results show that the RoBERTa-BiLSTM-MLP-CRF model has good recognition performance for four types of entity categories: results, violating behavior, erroneous behavior, and careless behavior, with accuracy rates of 86.7%, 80.3%, 80.7%, and 77.4%, respectively. ② Under the same dataset, the accuracy, recall, and F1 value of the RoBERTa-BiLSTM-MLP-CRF model training are improved by 1.6%, 1.5%, and 1.6%, respectively, compared to the RoBERTa-BiLSTM-CRF model.
  • loading

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return