Risk Evaluation Model of Highway Tunnel Portal Construction Based on BP Fuzzy Neural Network

Xianghui Deng, Tian Xu, Rui WangComput. Intell. Neurosci 2018 · DOI

TL;DR

This paper introduces a novel risk evaluation model for highway tunnel portal construction by combining Fuzzy Analytic Hierarchy Process (FAHP) with a Backpropagation (BP) neural network. This hybrid approach aims to overcome the inherent subjectivity and vagueness in traditional risk assessment, leading to more accurate and reliable predictions for critical construction phases.

What Problem Does This Paper Address?

Highway tunnel construction is a complex and inherently risky undertaking. Tunnel portals, in particular, are prone to accidents due to challenging geological conditions, shallow burial depths, and susceptibility to environmental factors like rainfall. Traditional risk assessment methods often rely on subjective expert judgment, which can be imprecise and lead to inaccurate evaluations. This subjectivity, coupled with the complex and uncertain nature of risk factors, poses a significant challenge in ensuring construction safety and mitigating economic losses. The paper addresses this by developing a more robust and objective risk evaluation model.

Methods

The core of this research lies in the development of a BP Fuzzy Neural Network (BPFNN) model. This approach integrates two powerful techniques:

  • Fuzzy Analytic Hierarchy Process (FAHP): FAHP is used to structure the risk assessment into a hierarchical framework and to quantify the subjective judgments of experts into fuzzy numbers. This allows for the handling of imprecise and ambiguous data. The paper outlines a process of establishing risk levels, defining factor sets, and constructing membership matrices.
  • Backpropagation (BP) Neural Network: The BP neural network, a type of multilayer feedforward network, is employed for its ability to learn complex nonlinear mappings between input and output. In this model, the fuzzified risk factors from FAHP serve as inputs to the BP network. The network is trained on historical data to predict the risk level.

The paper details the design of the BP neural network topology, including the input, hidden (fuzzy computing), and output layers. It also discusses the determination of learning rates and the number of hidden neurons using empirical formulas.

Datasets

The study utilizes a dataset comprising historical data from 50 highway tunnel construction projects in China. This dataset was used for:

  • Training the BPFNN model: 45 of these samples were used for training the network.
  • Testing the BPFNN model: The remaining 5 samples were used as test samples to evaluate the model's prediction accuracy.
  • Developing the risk assessment index system: The data informed the hierarchical structure of risk factors and events.

Key Insights and Ideas

  • Hybrid Approach Superiority: Combining fuzzy logic with neural networks offers a powerful way to handle both the vagueness of expert opinions and the complex, nonlinear relationships inherent in risk assessment.
  • BPFNN for Tunnel Portals: The BPFNN model demonstrates high prediction accuracy (97.94%) for highway tunnel portal construction risks, outperforming traditional methods that rely heavily on subjective judgment.
  • Identification of Critical Risk Factors: The model effectively identifies and ranks risk factors. For the Tiefodian tunnel case study, "slope instability" was identified as the greatest impact index among four risk events, and "construction method" was highlighted as a significant factor.
  • Validation through Simulation: Numerical simulations (using MIDAS/GTS) before and after implementing the proposed risk control measures provide strong evidence for the model's effectiveness and the feasibility of the suggested mitigation strategies.

Key Results

  • High Prediction Accuracy: The trained BPFNN model achieved a prediction accuracy of 97.94% on test samples.
  • Tiefodian Tunnel Risk Level: The risk evaluation for the Tiefodian tunnel portal construction was determined to be Level IV (high risk).
  • Risk Factor Ranking: The study identified a hierarchy of risk factors for tunnel portals, with surrounding rock level, support parameters, rainfall and groundwater, and cross-sectional size being particularly influential.
  • Effectiveness of Risk Control Measures: Numerical simulations showed a significant reduction in vault settlement (from 54.91 mm to 6.27 mm) and horizontal convergence (from 46.22 mm to 13.32 mm) after implementing the proposed risk control measures, demonstrating their effectiveness.

Limitations and Future Work

  • Semiquantitative Nature: The authors acknowledge that the proposed method is still semiquantitative.
  • Expert Experience Dependency: Risk identification relies on the rich experience of experts, which might lead to missing less influential risk factors.
  • Sample Size: The paper suggests that the number of samples for the proposed model needs to be further expanded to improve the accuracy of risk evaluation.
  • Future Work: The authors implicitly suggest further research into refining the risk identification process and expanding the dataset to enhance the model's robustness.

Summary updated

BibTeX
@article{deng2018risk,
  author = {Xianghui Deng and Tian Xu and Rui Wang},
  title = {{Risk Evaluation Model of Highway Tunnel Portal Construction Based on BP Fuzzy Neural Network}},
  journal = {Comput. Intell. Neurosci.},
  year = {2018},
  doi = {10.1155/2018/8547313},
}
Figure