Zhuojin Pipe Fitting Co., LtdSteel Pipe Pile Bearing Capacity Prediction Using Improved BKA Optimized RF-MLP Model
Literature Overview
This study develops a predictive model for the bearing capacity of driven steel pipe piles, which are widely used in bridge foundations, offshore platforms, and high-rise building foundations. The research combines data analysis techniques—specifically a Random Forest-Modified Multilayer Perceptron (RF-MLP) hybrid model—with an improved Black-Knight Algorithm (BKA) for hyperparameter optimization. The goal is to achieve more accurate and reliable bearing capacity predictions compared to traditional empirical methods and conventional data analysis approaches.
Methodological Framework
The predictive model architecture involves:
- Feature selection and engineering: Input parameters include soil layer properties (cohesion, friction angle, unit weight, SPT N-values), pile geometry (diameter, wall thickness, length), driving conditions (number of blows, driving resistance), and environmental factors (water table depth, temperature).
- Random Forest feature importance ranking: The RF component identifies the most influential input parameters and eliminates redundant features, reducing model complexity and improving generalization.
- MLP regression model: The MLP serves as the primary predictor, with its architecture (number of hidden layers, neurons per layer, activation functions) optimized by the improved BKA.
- Improved BKA optimization: The enhanced Black-Knight Algorithm searches the hyperparameter space more efficiently than standard optimization methods, avoiding local optima and converging to globally optimal model configurations.
Performance Comparison
| Model | RMSE (kN) | R² | MAPE (%) | Training Time (s) |
|---|---|---|---|---|
| Empirical formula (API RP 2A) | 285 | 0.72 | 18.5 | — |
| Standard MLP | 142 | 0.88 | 9.8 | 3.2 |
| RF-MLP (standard optimization) | 108 | 0.93 | 7.2 | 4.1 |
| RF-MLP (improved BKA) | 78 | 0.96 | 5.1 | 5.3 |
| Proposed model (this study) | 78 | 0.96 | 5.1 | 5.3 |
The proposed model demonstrates superior predictive accuracy compared to both empirical methods and conventional data analysis models. The improved BKA optimization contributes significantly to model performance by identifying optimal network architectures that balance complexity and generalization ability.
Engineering Practice Integration
For practical engineering application, the model requires:
- Data collection protocols: Standardized procedures for collecting soil investigation data and pile driving records to ensure consistent input quality.
- Model deployment: Integration into design software or standalone applications for rapid bearing capacity estimation during preliminary design phases.
- Uncertainty quantification: The model should provide confidence intervals around predictions to support risk-informed decision-making.
- Validation requirements: Site-specific calibration using available load test data to verify model predictions before final design adoption.
The study also discusses the model's applicability across different soil types and pile driving conditions, noting that performance may vary for very soft soils (clays with undrained strength below 20 kPa) or very stiff soils (rock or hard sand) where the training data may be limited.
Study Reflections
This research represents a promising application of advanced optimization algorithms to geotechnical engineering problems. The integration of feature selection (RF) with regression modeling (MLP) and intelligent optimization (improved BKA) creates a robust predictive framework that addresses the inherent variability in soil-pile interaction. For steel pipe pile designers, this model offers a valuable supplementary tool that can accelerate preliminary design and provide sanity checks on empirical estimates. However, engineers should remain mindful that data analysis models are only as good as their training data and should not replace rigorous geotechnical investigation and load testing for critical projects. The approach demonstrates how computational intelligence can enhance traditional engineering practice without replacing the need for engineering judgment.