GSS-RiskAsser: A Multi-Modal Deep-Learning Framework for Urban Gas Supply System Risk Assessment on Business Users

Xuefei Li, Liangtu Song, Liu Liu, Linli Zhou

TL;DR

This paper introduces GSS-RiskAsser, a novel multi-modal deep learning framework for assessing urban gas supply system risks. By combining visual pipeline data with expert-scored risk factors, GSS-RiskAsser achieves superior risk prediction accuracy compared to single-modal approaches, offering a more robust solution for urban infrastructure safety.

What Problem Does This Paper Address?

Urban gas supply systems are critical infrastructure, but their complex networks and the inherent risks of natural gas pose significant threats to public safety. Traditional risk assessment methods often rely on manually constructed mathematical models using single-modal data (like pipeline parameters), which can be subjective, incomplete, and struggle to capture the full picture. This paper addresses the need for a more comprehensive and data-driven approach to urban gas supply system risk assessment, particularly at the user level, by leveraging the power of deep learning and multi-modal data. The authors define a new task, Urban Gas Supply System Risk Assessment (GSS-RA), to tackle this challenge.

Methods

The core of this work is the GSS-RiskAsser framework, designed to process multi-modal inputs. It comprises three main components:

  • Visual Embedding Transformer (VET): This module processes visual information from images of gas supply systems. It combines a Convolutional Neural Network (CNN) for initial feature extraction with a Transformer architecture. The Transformer uses self-attention mechanisms to refine visual features, effectively capturing spatial relationships and details within the images.
  • Score Matrix Transformer (SMT): This module handles non-visual, expert-derived risk information. The input is a score matrix representing various accident and consequence-related factors. Similar to VET, SMT employs a Transformer architecture with self-attention to process these structured risk attributes. Crucially, the VET and SMT encoders are trained independently, acknowledging the different nature of their input data.
  • Multi-Modal Fusion (MMF): This module is responsible for effectively combining the features extracted by VET and SMT. It utilizes a cross-attention mechanism, allowing the two modalities to exchange information and learn from each other. This fusion strategy is designed to create a richer, more comprehensive representation for final risk prediction.

Datasets

The paper introduces a new, domain-specific dataset:

  • GSS-20K: This dataset is specifically created for the GSS-RA task. It contains multi-modal data, including:
  • Visual Information: RGB images of gas supply systems from individual users.
  • Non-Visual Information: A score matrix derived from expert judgments on accident-related and consequence-related factors.
  • The dataset also includes hierarchical risk values (possibility of rupture, consequence severity, and final risk) annotated by experts.
  • To address data imbalance, a Semi-Composited Multi-Modal Generation (SCMMG) strategy is employed, using data augmentation for images and neighbor sampling/Gaussian noise for score matrices to generate synthetic positive samples.

Key Insights and Ideas

  • Multi-modal data is crucial for gas supply system risk: The research strongly suggests that relying on either visual or expert-scored data alone is insufficient. Combining both provides a more accurate and robust assessment.
  • Transformer architectures excel at both modalities: The paper demonstrates the effectiveness of Transformer-based models (VET and SMT) for processing both image data and structured score matrices, highlighting their versatility.
  • Cross-attention is key for effective fusion: The proposed cross-attention mechanism in the MMF module is shown to be superior to simpler fusion methods like concatenation or sum fusion, enabling deeper interaction between modalities.
  • User-level assessment is feasible and valuable: By focusing on individual users, the framework can provide granular risk insights, which is more practical for targeted interventions than broader urban-level assessments.

Key Results

  • Superior Performance: GSS-RiskAsser achieved the lowest Mean Absolute Error (MAE) and Mean Square Error (MSE) across all three risk metrics (possibility, consequence, and final risk) compared to various single-modal and multi-modal baselines.
  • For example, GSS-RiskAsser achieved MAE values of 0.054, 0.040, and 0.036 for possibility, consequence, and final risk, respectively, outperforming even its ablated versions (VET only and SMT only).
  • Multi-modal advantage: Multi-modal methods consistently outperformed single-modal methods, with GSS-RiskAsser showing a significant improvement.
  • Practical Applications: The framework was successfully deployed for:
  • Gas Supply System Risk Heatmap Monitoring: Visualizing risk distribution across urban areas.
  • Gas Supply System Risk Relationship Graph: Illustrating how neighboring users' risks can influence each other, providing insights into localized risk propagation.

Limitations and Future Work

  • Dataset Size: While GSS-20K is a significant contribution, the paper notes that larger datasets would further enhance deep learning model performance. The SCMMG strategy was employed to mitigate this.
  • Generalizability: The current framework is trained on data from a specific city. Future work could explore its generalizability to different urban environments and gas supply system configurations.
  • Real-time Prediction: While the framework is efficient, further optimization for real-time monitoring and prediction in dynamic urban environments could be explored.
  • Additional Modalities: The authors suggest incorporating other data sources (e.g., sensor data, weather information) to further enrich the risk assessment.

Summary updated

Abstract

Gas supply system risk assessment is a serious and important problem in cities. Existing methods tend to manually build mathematical models to predict risk value from single-modal information, i.e., pipeline parameters. In this paper, we attempt to consider this problem from a deep-learning perspective and define a novel task, Urban Gas Supply System Risk Assessment (GSS-RA). To drive deep-learning techniques into this task, we collect and build a domain-specific dataset GSS-20K containing multi-modal data. Accompanying the dataset, we design a new deep-learning framework named GSS-RiskAsser to learn risk prediction. In our method, we design a parallel-transformers Vision Embedding Transformer (VET) and Score Matrix Transformer (SMT) to process multi-modal information, and then propose a Multi-Modal Fusion (MMF) module to fuse the features with a cross-attention mechanism. Experiments show that GSS-RiskAsser could work well on GSS-RA task and facilitate practical applications. Our data and code will be made publicly available.
BibTeX
@inproceedings{li2021gss,
  author = {Xuefei Li and Liangtu Song and Liu Liu and Linli Zhou},
  title = {{GSS-RiskAsser: A Multi-Modal Deep-Learning Framework for Urban Gas Supply System Risk Assessment on Business Users}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=CLpX1V7-qN3},
}
Figure