Subway fire risk stops relying on hand-picked nodes: a graph network selects 22 from a 666-node knowledge graph, lifting high-casualty probability from 6.58% to 53.2%
Synopsis
Subway fire risk assessment has long depended on experts manually choosing Bayesian network nodes, which is subjective and hard to adapt. This pipeline extracts 215 fire keywords and six risk labels from accident records and fire codes with natural language processing, builds a 666-node knowledge graph, classifies nodes with a three-layer graph convolutional network, and combines confidence weighting with K-means clustering to select 22 representative nodes for the Bayesian network. Under coupled risks such as malicious passenger behavior, smoke exhaust failure and delayed alarm, the high-casualty probability rises from a 6.58% baseline to 53.2%.
Interpretation
The nodes of a subway fire Bayesian network are now selected automatically from data rather than specified entirely by experts. Earlier Bayesian networks were built from manually selected nodes, carried strong subjective bias, and could not be dynamically reconstructed for a specific subway fire setting. The automatic selection procedure picks 22 representative nodes from the 666-node knowledge graph, covers all six risk labels, and forms the main framework of the Bayesian network.
Node selection rests on how the graph neural network classifies each node's risk category and how confident that classification is. A three-layer graph convolutional network aggregates neighbor information layer by layer, maps the 215-dimensional features into low-dimensional vectors, and outputs a probability distribution over six labels through Softmax. Training uses 16 subway fire incident records from 1958 to 2025 and two fire safety codes, with nodes split 70%, 15% and 15% into training, validation and test sets; validation accuracy stabilizes around 0.9 after 500 epochs.
Expert scoring supplies the Bayesian network's probability parameters so the network can report casualty probabilities under different risk combinations. Automatically selected nodes fix only the network structure; the occurrence probability of each risk state still needs prior and conditional probabilities. Seventeen subway fire safety experts scored 27 core questions through the Delphi method, with the coefficient of variation measuring opinion dispersion and Cronbach's alpha measuring overall consistency; means are adopted as final parameters when alpha exceeds 0.8.
Perspective
The model targets quantitative risk simulation and accident retrospective analysis for subway station fires; its inputs are historical incident records, fire code text and expert scoring, and its outputs are state probabilities for risk nodes and changes in casualty probability. The knowledge graph remains limited in scale, with data drawn mainly from web crawling, literature review and subway regulatory documents, and it lacks a comprehensive dataset from specific subway projects and real-time operational data.
Real-time dynamic monitoring of equipment status, personnel behavior and environmental parameters is absent, which limits gains in intelligent maintenance efficiency. The constrained graph scale and the missing comprehensive dataset from specific subway projects are conditions the authors state explicitly.