Back analysis of mechanical parameters based on GPSO-BP neural network and its application

Area of Science:

• Geotechnical Engineering
• Computational Mechanics
• Artificial Intelligence in Engineering

Background:

• Traditional methods for determining rock mass parameters often yield results unsuitable for numerical simulations in underground engineering.
• Back analysis, utilizing displacement monitoring, offers a promising alternative for parameter determination.

Purpose of the Study:

• To develop and validate an advanced neural network model for accurate inversion of rock mass mechanical parameters (E, μ, φ, c).
• To enhance the reliability of numerical simulations in deep-buried tunnel design and construction.

Main Methods:

• An experimental scheme using orthogonal and uniform designs to generate training data for neural networks.
• Development of a hybrid Genetic Algorithm-Particle Swarm Optimization-Back Propagation (GA-PSO-BP) neural network model.
• Application of the GPSO-BP model for inverting rock mass parameters and subsequent forward numerical simulations.

Main Results:

• The GPSO-BP model demonstrated superior convergence speed and accuracy compared to BP, GA-BP, and PSO-BP models.
• The model exhibited excellent performance with small datasets and complex problems, showing improved data fitting and rank analysis scores.
• Forward simulations using GPSO-BP derived parameters resulted in a low average error of 4.34% across four monitoring projects.

Conclusions:

• The proposed GA-PSO-BP neural network model provides an effective and accurate method for rock mass parameter inversion in underground engineering.
• This approach significantly enhances the reliability of numerical simulations, leading to better design and construction outcomes for deep-buried tunnels.