Zhuojin Pipe Fitting Co., LtdAxial Compression Bearing Capacity of Hollow Sandwich CFDST Columns Using PSO-BP Neural Network
Literature Overview
This paper, authored by Zhao Junhai, Hua Linwei, and Wang Yu from Chang'an University, was published in "Progress in Steel Building Structures" in 2024 (Vol. 26, Issue 9, pp. 45-52). The study investigates the axial compression bearing capacity of circular concrete-filled double-skin steel tube (CFDST) columns using a particle swarm optimization (PSO) enhanced backpropagation (BP) neural network algorithm. The research was supported by the National Natural Science Foundation of China (51878056). CFDST columns, also known as hollow sandwich steel tube concrete columns, consist of an outer steel tube, an inner steel tube, and a concrete core confined between them, offering unique structural advantages including enhanced confinement, improved fire resistance, and increased energy absorption capacity.
Core Technical Content
The authors collected a database of 167 experimental data sets from published literature to train and validate the neural network model. Eight input parameters were selected as the input layer of the network, including the outer steel tube diameter, outer steel tube wall thickness, inner steel tube diameter, inner steel tube wall thickness, outer steel tube yield strength, inner steel tube yield strength, concrete compressive strength, and column length-to-diameter ratio. The axial compression bearing capacity was used as the single output parameter.
Input Parameters for the Neural Network Model
| Parameter | Symbol | Typical Range | Units |
|---|---|---|---|
| Outer tube diameter | D_o | 100 - 400 | mm |
| Outer tube wall thickness | t_o | 3 - 20 | mm |
| Inner tube diameter | D_i | 50 - 200 | mm |
| Inner tube wall thickness | t_i | 2 - 10 | mm |
| Outer tube yield strength | f_y,o | 235 - 460 | MPa |
| Inner tube yield strength | f_y,i | 235 - 460 | MPa |
| Concrete compressive strength | f_c | 20 - 80 | MPa |
| Length-to-diameter ratio | L/D_o | 2 - 15 | - |
The authors first analyzed the deficiencies of the traditional BP neural network model, which include slow convergence, tendency to fall into local minima, and sensitivity to initial weight values. The PSO algorithm was then introduced to optimize the initial weights and biases of the BP network, resulting in a PSO-BP hybrid model that exhibits faster convergence and higher prediction accuracy.
Model Performance Comparison
| Model | RMSE (kN) | MAE (kN) | R² | Average Error (%) |
|---|---|---|---|---|
| Traditional BP Neural Network | 85.3 | 52.1 | 0.92 | 8.5 |
| PSO-BP Neural Network | 42.7 | 28.5 | 0.98 | 4.2 |
| GB 51248-2017 Code Formula | 120.5 | 78.3 | 0.88 | 12.3 |
| CECS 280-2020 Code Formula | 98.2 | 65.4 | 0.91 | 10.1 |
| Eurocode 4 Formula | 110.8 | 72.6 | 0.89 | 11.5 |
The PSO-BP neural network model demonstrated significantly superior prediction accuracy compared to the traditional BP network and three code-based formulas. The R² value of 0.98 indicates an excellent correlation between predicted and experimental values, and the average error of 4.2% is well within the acceptable range for engineering design purposes.
Methodological Analysis
The PSO algorithm works by simulating the social behavior of bird flocking, where each particle (representing a potential solution) adjusts its position in the search space based on its own best-known position and the best-known position of the entire swarm. In the context of neural network optimization, the particles represent the initial weights and biases of the BP network, and the fitness function is the mean squared error between predicted and actual bearing capacities.
The key advantages of the PSO-BP approach over the traditional BP network include:
- Global optimization capability: The PSO algorithm explores the solution space more thoroughly, reducing the risk of converging to a local minimum.
- Faster convergence: The social learning mechanism of PSO allows the population to converge to the optimal solution more rapidly than the gradient-descent-based training of BP networks.
- Reduced sensitivity to initial conditions: Since the initial weights are optimized by PSO rather than randomly initialized, the model is less prone to training failures.
Training and Validation Process
| Stage | Data Set Size | Purpose |
|---|---|---|
| Training | 117 samples (70%) | Model parameter optimization |
| Validation | 25 samples (15%) | Hyperparameter tuning and early stopping |
| Testing | 25 samples (15%) | Unbiased performance evaluation |
The 70-15-15 split of the database into training, validation, and testing sets is a standard practice in data analysis and provides a reliable assessment of model performance. The validation set was used to monitor the training process and implement early stopping to prevent overfitting, while the testing set was reserved for final performance evaluation.
Engineering Practice Implications
The PSO-BP neural network model provides a powerful tool for predicting the axial compression bearing capacity of CFDST columns, which is particularly valuable given the complex interaction between the outer tube, inner tube, and concrete core. Traditional analytical formulas, such as those provided in GB 51248-2017 and Eurocode 4, often rely on simplified assumptions that may not accurately capture the multi-axial stress state of the confined concrete or the interaction between the two steel tubes.
For engineering design, the model can be used as a preliminary estimation tool during the conceptual design phase, allowing engineers to rapidly evaluate different geometric and material configurations. The model can also be used for quality control purposes, where the predicted bearing capacity can be compared with the experimentally determined capacity of test specimens to verify that the manufacturing and construction quality meets design requirements.
However, it is important to note that the model's performance is limited by the range of the training data. Extrapolation beyond the parameter ranges represented in the database should be avoided, as the neural network may not generalize accurately to unseen configurations. Engineers should exercise caution when applying the model to novel geometries or material combinations that fall outside the training domain.
Study Insights and Reflections
This study demonstrates the potential of data analysis techniques for solving complex structural engineering problems that are difficult to address with traditional analytical methods. The PSO-BP neural network approach is particularly well-suited for CFDST columns, where the interaction between multiple structural components creates a highly nonlinear behavior that is challenging to model with closed-form expressions.
The comparison with three code-based formulas provides valuable context for understanding the relative accuracy of the neural network model. The fact that the code formulas exhibit higher errors (10-12%) is not surprising, as they are developed based on simplified assumptions and are intended to be conservative for safety. The neural network model, trained on experimental data, captures the actual behavior more accurately but must be used with appropriate caution regarding its generalization capability.
I note that the study uses a database of 167 samples, which is a moderate size for neural network training. While the model achieves good performance, increasing the database size through additional experimental data or finite element simulation results could further improve the model's robustness and generalization capability. Additionally, the model could be extended to predict other performance metrics such as peak strain, failure mode, and energy absorption capacity, which are important for seismic design applications.
The methodology employed in this study — combining experimental data with data analysis optimization — represents a promising direction for structural engineering research. As computational resources and data availability continue to grow, similar approaches can be applied to a wide range of structural problems, including steel-concrete composite beams, welded joints, and bolted connections. The key challenge remains ensuring that the models are validated against independent experimental data and used within their domain of applicability, which requires both technical rigor and professional judgment.