Intelligent Bayesian analysis Based on Knowledge Graph and Graph Neural Network for Subway Fire Risk

Technical Summary: Intelligent Bayesian Analysis Based on Knowledge Graph and Graph Neural Network for Subway Fire Risk

Problem Statement
Subway systems, characterized by narrow spaces and dense crowds, face significant fire safety challenges. Existing risk assessment methodologies suffer from distinct limitations:

1. Knowledge Graphs (KG): Traditional KGs rely solely on node-edge topology, ignoring entity attribute correlations, which results in poor reasoning capabilities.
2. Bayesian Networks (BN): Classical BNs are typically constructed via manual node selection by experts, introducing strong subjective bias and limiting adaptability to complex, dynamic systems.
3. General Limitations: Current approaches often lack real-time, data-driven dynamic risk assessment, struggle with implicit relationships between entities, and fail to dynamically construct BNs tailored to specific subway fire incidents.

Methodology
The study proposes an integrated KG-GNN-BN framework that transitions from structured data extraction to probabilistic reasoning through four key stages:

1. Knowledge Graph Construction:

   • Data Sources: Historical fire records (1958–2025) and regulatory documents were analyzed.
   • NLP Processing: Natural Language Processing (using the Jieba library) extracted 215 fire characteristic keywords and defined 6 risk accident labels (Event/Year, Management System, Human Behavior, Equipment/Structure, Environment/Airflow, Aftermath/Improvement).
   • Graph Structure: A subway fire KG was built containing 666 nodes and 931 edges, embedding feature and label attributes into the topology to form a complete dataset.

2. Graph Neural Network (GNN) Training:

   • Architecture: A three-layer Graph Convolutional Network (GCN) with residual connections and batch normalization was employed.
   • Process: The model performed node embedding extraction and feature learning. It utilized a message-passing mechanism to iteratively update hidden states based on neighbor aggregation.
   • Optimization: The model was trained using cross-entropy loss and the Adam optimizer with a cosine annealing learning rate scheduler.

3. Automatic Node Selection for BN:

   • Confidence & Clustering: The trained GNN generated node embeddings. A confidence score was calculated via Softmax normalization. K-means clustering (K=6, matching the tag categories) was applied to ensure diversity and avoid feature coverage bias.
   • Scoring: A comprehensive scoring formula integrated classification confidence and intra-cluster representativeness (distance to cluster centroid).
   • Selection: Based on these scores, 22 representative nodes were automatically selected from the 666 nodes to form the core causal chain of the BN. This process utilizes automated scoring but involves manual adjustment of weight coefficients and thresholds to balance representativeness and prevent overfitting.

4. Bayesian Network Parameterization:

   • Structure: The 22 selected nodes formed the BN structure.
   • Probabilities: Prior and conditional probabilities were determined using the Delphi method, aggregating feedback from 17 subway fire safety experts. Consistency was verified using Cronbach's alpha (>0.8) and the coefficient of variation.

Key Results

• Model Performance: The GNN achieved high training accuracy (approaching 1.0) and validation accuracy (stabilizing around 0.9), demonstrating strong generalization. The confusion matrix indicated high discriminative capability across the 6 tag categories.
• Sensitivity Analysis: Using Netica, the study identified critical risk nodes.
  • Casualties: The 'Initial Alarm' node showed the highest sensitivity coefficient (2.05), followed by 'Evacuation Efficiency' (1.9).
  • Causal Chains: 'Fire Spread' was most influenced by 'Material Flame Retardancy'; 'Smoke Diffusion' by 'Open Fire Occurrence'; and 'Initial Alarm' by 'Duty Arrangement'.
• Scenario Simulation:
  • Under a baseline state, the probability of severe casualties was 6.58%.
  • In a composite risk scenario (simulating the 2003 Daegu arson case with malicious behavior, equipment failure, and delayed alarms), the probability of severe casualties surged to 53.2%.
  • 'Smoke Diffusion' in a high-failure state caused the most significant escalation in casualty probability (Low casualty probability dropped from 77.8% to 47.7%).
• Reverse Tracing: The analysis identified 'Initial Alarm', 'Evacuation Efficiency', and 'Fire Spread' as the core risk transmission chain, with variation values exceeding 13%. 'Passenger Behavior' was identified as the primary driver for 'Open Fire Occurrence'.

Significance and Claims
The paper claims that the proposed KG-GNN-BN model offers a data-driven pipeline that enhances the objectivity of BN construction. Its significance lies in:

• Objectivity: It reduces the subjectivity of traditional manual node selection by using an automated, GNN-trained scoring approach, though final selection involves manual threshold tuning, and probability parameters still rely on expert scoring via the Delphi method.
• Integration: It seamlessly transforms structured regulatory semantics and historical accident data into a probabilistic reasoning network.
• Applicability: The model provides a quantitative tool for both dynamic prediction and retrospective analysis of subway fire risks. It supports not only general risk analysis but also the construction of customized KGs for specific operational contexts.
• Strategic Insight: The study provides specific, data-backed prevention strategies, emphasizing the need to prioritize rapid alarm response, stable evacuation efficiency, and controlled smoke diffusion to mitigate high-consequence risks.

Limitations Acknowledged
The authors note that the current KG scale is constrained, relying primarily on web crawlers, literature, and regulatory documents rather than real-time operational data from specific subway projects. Consequently, the model currently lacks real-time dynamic monitoring capabilities for equipment status, personnel behavior, and environmental parameters.