Zhuojin Pipe Fitting Co., LtdAxial Compressive Bearing Capacity of Hollow Sandwich Concrete-Filled Steel Tube Columns Using PSO-Optimized BP Neural Network
Literature Overview
This paper by Zhao Junhai and colleagues, published in Progress in Steel Building Structures (2024, Vol. 26, No. 9, pp. 45–52), investigates the axial compressive bearing capacity of circular concrete-filled double-skin steel tube (CFDST) columns using a particle swarm optimization (PSO)-optimized backpropagation (BP) neural network model. The research was funded by the National Natural Science Foundation of China (51878056) and was conducted at Chang'an University. The study addresses the complex interaction between the outer steel tube, inner steel tube, and core concrete in CFDST columns, which makes traditional analytical methods insufficient for accurate prediction.
Core Technical Content
CFDST columns feature a unique double-skin steel tube configuration with an inner steel tube, an outer steel tube, and a concrete core filling the annular space between them. This configuration provides superior mechanical performance compared to conventional single-skin CFST columns, including higher load-bearing capacity, improved ductility, and enhanced resistance to local buckling. However, the complex interaction between the three components—outer tube, inner tube, and concrete—makes the stress state highly nonlinear and difficult to predict using traditional analytical methods.
PSO-BP Neural Network Model Architecture
| Component | Description |
|---|---|
| Input parameters (8 factors) | Outer tube diameter, outer tube wall thickness, inner tube diameter, inner tube wall thickness, concrete compressive strength, outer tube yield strength, inner tube yield strength, column slenderness ratio |
| Output parameter | Axial compressive bearing capacity |
| Training dataset | 167 experimental data points |
| Optimization algorithm | Particle Swarm Optimization (PSO) |
| Neural network type | Backpropagation (BP) neural network |
The study first analyzes the limitations of the traditional BP neural network model, which suffers from slow convergence, tendency to get trapped in local minima, and sensitivity to initial weight settings. The PSO algorithm is then applied to optimize the initial weights and thresholds of the BP neural network, resulting in a PSO-BP model with significantly improved prediction accuracy.
Comparative Results
| Model/Method | Prediction Accuracy | Key Limitation |
|---|---|---|
| Traditional BP neural network | Moderate | Slow convergence; local minima |
| PSO-BP neural network | High | Slightly higher computational cost |
| GB 50935-2014 (Chinese standard) | Lower than PSO-BP | Simplified assumptions; conservative |
| Eurocode 4 | Lower than PSO-BP | Different interaction model |
| CIDECT | Lower than PSO-BP | Empirical coefficients; limited database |
The PSO-BP neural network model demonstrates superior prediction accuracy compared to three existing design standards (GB 50935-2014, Eurocode 4, and CIDECT), and also outperforms the traditional BP neural network model. This indicates that the PSO optimization effectively addresses the limitations of the traditional BP model, resulting in a more reliable prediction tool for CFDST column design.
Interpretation of Technical Points
Structural Behavior of CFDST Columns
The double-skin steel tube configuration in CFDST columns creates a unique confinement mechanism. The outer tube provides external confinement to the concrete core, while the inner tube acts as a core that resists local buckling and provides additional confinement to the concrete in the annular space. The interaction between the outer tube and inner tube through the concrete core creates a complex load-sharing mechanism that depends on the relative stiffness and strength of the two tubes.
Under axial compression, the concrete core expands laterally, exerting outward pressure on the outer tube and inward pressure on the inner tube. The outer tube resists this expansion through its hoop strength, while the inner tube resists the inward pressure through its compressive strength. The effectiveness of this confinement depends on the slenderness ratio of both tubes, as well as the concrete strength and the tube wall thicknesses.
PSO Optimization of BP Neural Network
The PSO algorithm is a population-based optimization technique inspired by the social behavior of bird flocking or fish schooling. In the context of BP neural network optimization, PSO is used to find the optimal initial weights and thresholds that minimize the prediction error. The PSO algorithm operates by maintaining a population of candidate solutions (particles), each of which moves through the search space based on its own velocity and the positions of the best solutions found so far.
The advantages of PSO over traditional optimization methods include:
- No gradient information required
- Fewer hyperparameters to tune
- Faster convergence compared to genetic algorithms
- Robustness to local minima
By applying PSO to optimize the initial parameters of the BP neural network, the resulting PSO-BP model achieves faster convergence, higher prediction accuracy, and greater robustness compared to the traditional BP model.
Engineering Practice Relevance
For engineers designing CFDST columns, the following considerations are important:
| Design Parameter | Typical Range | Influence on Capacity |
|---|---|---|
| Outer tube diameter | 200–600 mm | Larger diameter increases capacity |
| Outer tube wall thickness | 4–20 mm | Thicker wall increases confinement |
| Inner tube diameter | 100–300 mm | Larger inner tube reduces concrete volume |
| Inner tube wall thickness | 3–15 mm | Thicker wall increases core resistance |
| Concrete compressive strength | 30–80 MPa | Higher strength increases capacity |
| Outer tube yield strength | 235–460 MPa | Higher strength increases tube contribution |
| Inner tube yield strength | 235–460 MPa | Higher strength increases core contribution |
| Slenderness ratio | 5–30 | Higher slenderness reduces capacity |
The PSO-BP neural network model provides a powerful tool for predicting the axial compressive bearing capacity of CFDST columns. However, engineers should be aware of the limitations of any data-driven model, including:
- The model is only as good as the training data
- Extrapolation beyond the training data range may lead to inaccurate predictions
- The model does not capture the underlying physical mechanisms
- The model requires validation against independent test data
Study Insights and Implications
This research demonstrates the potential of hybrid optimization-neural network approaches for predicting the structural behavior of complex composite members. The PSO-BP model represents a significant improvement over traditional analytical methods and existing design standards for CFDST columns, offering higher prediction accuracy and greater flexibility.
The use of 167 experimental data points for training the neural network is a reasonable dataset size for this type of structural analysis. However, engineers should be aware that the model's accuracy depends on the representativeness and quality of the training data. Future research should aim to expand the dataset to include a wider range of geometric parameters, material properties, and loading conditions, which would improve the model's generalization capability.
From a practical standpoint, the PSO-BP model can be integrated into design software to provide engineers with a quick and accurate estimation of CFDST column capacity. This would complement traditional analytical methods and design standards, providing an additional check on the design and potentially leading to more efficient and cost-effective solutions.
The research also highlights the importance of considering the interaction between the outer tube, inner tube, and concrete core in CFDST column design. Traditional analytical methods often simplify this interaction through empirical coefficients or simplified assumptions, which can lead to conservative or non-conservative predictions. The PSO-BP model, by capturing the complex nonlinear interactions through data-driven learning, provides a more accurate representation of the actual structural behavior.
In conclusion, this paper presents a novel and effective approach for predicting the axial compressive bearing capacity of CFDST columns using a PSO-optimized BP neural network. The model demonstrates superior accuracy compared to traditional methods and existing design standards, offering a valuable tool for engineers designing CFDST columns. The integration of optimization algorithms with neural networks represents a promising direction for structural engineering research, and this study makes a meaningful contribution to that field.