Prediction of cement-stabilized recycled concrete aggregate properties by CNN-LSTM incorporating attention mechanism

Abstract

Cement-stabilized recycled concrete aggregate (CSRCA), as an eco-friendly and cost-effective building material, has its mechanical properties and fatigue resistance influenced by multiple factors. Machine learning-based prediction methods offer a new solution for the rapid assessment of CSRCA performance. To efficiently and accurately predict the mechanical properties and fatigue resistance of CSRCA, this study proposes a machine learning-based CSRCA performance prediction model. By collecting data on various influencing factors such as mix ratios, aggregate characteristics, and compaction, and combining convolutional neural networks (CNN) with long short-term memory networks (LSTM), the study introduces an attention mechanism (ATT) to optimize feature weighting, thus constructing a CNN-LSTM-ATT prediction model. Additionally, Pearson correlation analysis and random forest algorithms are used to investigate the key influencing factors and feature importance of CSRCA's mechanical and fatigue performance. The Shapley Additive Explanation (SHAP) is employed to enhance the model's interpretability. The results indicated that each input parameter exhibited a significant nonlinear relationship with both compressive strength and fatigue life. Specifically, increasing cement dosage, degree of compaction, and 19mm sieve passing rate contributed positively to the mechanical properties and fatigue life of CSRCA. Conversely, the recycled aggregate crushing value had a significant negative impact on compressive strength but minimal effect on fatigue life. Notably, recycled aggregate dosage demonstrated opposing effects on compressive strength and fatigue life. The prediction accuracy of the CNN-LSTM-ATT model, incorporating the attention mechanism, surpassed that of the conventional CNN-LSTM model, enhancing the accuracy, robustness, and generalization ability of the prediction framework. The prediction metrics for compressive strength were R² = 0.993, MAE = 0.101, and RMSE = 0.144, while those for fatigue life were R² = 0.994, MAE = 55.05, and RMSE = 73.43, the prediction ability is significantly higher than other prediction models of the same type. Consequently, the CNN-LSTM-ATT model provides a scientific foundation for the performance prediction and proportioning design of CSRCA, holding significant implications for engineering practice.