TL;DR
This paper introduces a novel deep learning approach, SE-ConvNeXt, to accurately detect the substitution rate of steel slag sand in construction materials. By integrating a Squeeze-and-Excitation (SE) attention mechanism into the ConvNeXt architecture and employing transfer learning, the method achieves a high accuracy of 92.64%, addressing the inefficiency of traditional detection methods and improving the quality control of steel slag utilization in concrete.
What Problem Does This Paper Address?
The construction industry faces increasing demand for natural sand, leading to environmental depletion and rising costs. Steel slag, an industrial waste product, offers a promising substitute. However, the proportion of steel slag sand significantly impacts the volumetric stability and workability of concrete. Accurately controlling this substitution rate is crucial to prevent engineering risks and maximize steel slag utilization. Traditional detection methods are often inefficient and lack representative sampling, necessitating a faster, more accurate, and automated solution.
Methods
The core of this work is the SE-ConvNeXt model. It builds upon the ConvNeXt architecture, a modern convolutional neural network known for its strong performance. The key innovation is the integration of a Squeeze-and-Excitation (SE) attention mechanism. This mechanism allows the model to dynamically recalibrate channel-wise feature responses, effectively focusing on the most informative color features of the steel slag sand mix.
To further enhance performance, the researchers employed transfer learning. This involved pre-training the SE-ConvNeXt model on a subset of the mixed sand data and then fine-tuning it on the complete dataset. This approach leverages learned features from the initial training, accelerating convergence and improving accuracy, especially when dealing with variations in sand types.
Datasets
The study utilized a custom-created dataset of mixed sand samples. Three types of mixed sand (A, B, and C) were prepared by combining steel slag sand with standard sand and two types of natural sand, at substitution rates of 0%, 20%, 40%, 60%, 80%, and 100%. For each substitution rate and sand type, 100 images were captured using a professional camera. These images were then cropped, resulting in a dataset of 7200 images in total. The dataset was split into training (60%), validation (20%), and testing (20%) sets.
Key Insights and Ideas
- Color is Key: The study highlights that the color characteristics of steel slag sand are a strong indicator of its substitution rate. As the proportion of steel slag increases, the overall mix tends to become darker.
- Attention Matters: The SE attention mechanism proved crucial in helping the ConvNeXt model better capture these subtle color variations, leading to improved feature extraction and classification accuracy.
- Transfer Learning Pays Off: Pre-training the model significantly boosted its performance, demonstrating the effectiveness of transfer learning for this specific task, even when dealing with diverse sand types.
- Low Substitution is Tricky: The model struggled more with accurately identifying lower substitution rates (e.g., 20% and 40%) due to the less pronounced color differences compared to higher rates.
Key Results
- The proposed SE-ConvNeXt model achieved an average accuracy of 87.99% on the test set without transfer learning.
- After applying transfer learning, the accuracy of the SE-ConvNeXt model increased to 92.64%, a significant improvement of 4.65%.
- The SE-ConvNeXt model outperformed the original ConvNeXt network by 1.74% in accuracy, with only a marginal increase in parameters (5.84%).
- When compared to other convolutional neural networks (EfficientNet V2, ShuffleNet V2, VGG-16, MobileNet V3, ResNet-50), SE-ConvNeXt demonstrated superior performance, achieving the highest accuracy (97.0% on validation, 87.99% on test) with a relatively efficient parameter count.
- The model showed good stability, with a standard deviation of 0.902% across five cross-validation runs.
Limitations and Future Work
- Low Substitution Rate Challenges: The paper acknowledges that accurately identifying lower substitution rates (20% and 40%) remains a challenge, particularly for certain sand types (like sample C) where color differences are less distinct. This is attributed to the subtle visual cues at these lower percentages.
- Transfer Learning for Diverse Sands: While transfer learning improved accuracy, the improvement was less pronounced for sample C compared to samples A and B. This suggests that further research might be needed to optimize transfer learning strategies for highly varied sand compositions.
- Future Work: The authors suggest that their method can be used for the rapid and accurate detection of steel slag sand substitution rates, implying its potential for real-world implementation in construction quality control. Further research could focus on improving accuracy at lower substitution rates and exploring more advanced transfer learning techniques for even greater efficiency and robustness.
Summary updated