Intelligent prediction of compressive strength of concrete based on CNN-BiLSTM-MA

Various factors affect the compressive strength of concrete, and performing experiments on this is both time-consuming and costly. Constructing predictive models for the strength of concrete at different ages can significantly enhance the efficiency of testing and serve as a foundation for engineering quality analysis. Currently, concrete strength prediction using traditional machine learning models exhibits suboptimal accuracy, while the performance enhancement achieved by single deep learning models is limited. This study utilizes 8 variables associated with concrete strength as inputs for the predictive model, focusing on the compressive strength of concrete at different ages. In this study, we propose a hybrid model, referred to as CNN-BiLSTM-MA, which combines convolutional neural network (CNN), bidirectional long short-term memory neural network (BiLSTM), and multi-head attention mechanism (MA). The CNN module is employed to extract data feature relationships, while the BiLSTM module captures deep features within the concrete compressive strength data. Additionally, the MA module is utilized to focus on important features influencing concrete compressive strength, further improving the model’s predictive accuracy. To determine optimal parameters for the model, five mainstream machine learning models, including Random Forest, Adaboost, GBDT, XGBoost, and LightGBM, were tuned using a random search algorithm and 5-fold cross-validation. Subsequently, these models were contrasted with our proposed model using the experimental dataset. The results indicate that the CNN-BiLSTM-MA model achieved the best predictive performance and generalization with RMSE=1.21, MSE=1.57, MAE=0.85, and R 2 =0.97.