CrushNet: A deep learning and image analysis-based method for predicting the crushing value of recycled aggregates

The acceleration of urbanization has significantly increased the annual production of recycled aggregates (RA). However, the multi-source nature of RA raw materials poses challenges of high cost and time consumption for quality assessment. Among these, the crushing value, a key strength indicator, currently relies on destructive tests for determination, which fails to meet the urgent demand in the construction industry for large-scale, batch-processing, and non-destructive testing of RA crushing values. To address this issue, this study proposes an intelligent detection method based on a convolutional neural network (CNN) model, pioneering the use of image recognition technology for the intelligent prediction of RA crushing value. By collecting 2778 image samples covering natural aggregates (NA), recycled concrete aggregates (RCA), recycled brick aggregates (RBA), and recycled mixed aggregates (RMA), and simultaneously conducting crushing value tests, a specialized deep learning dataset was constructed using image data augmentation techniques. Furthermore, based on the multi-scale feature extraction mechanism of the Inception module, an RA crushing value detection model (CrushNet) was constructed, and AlexNet, VGGNet16, ResNet18, EfficientNet and ConvNeXtV2-Atto were used as benchmark models for comparison and verification. In addition, to interpret the model's prediction mechanism, we designed a multi-task learning comparative experiment. This experiment involved the simultaneous regression of aggregate gradation and crushing value, followed by a visualization of the attention distribution in both single-task and multi-task models via Grad-CAM. Finally, a visual analysis was conducted to elucidate the decision-making process based on the evolution of cross-layer attention in the CNN. The results show that CNN can predict aggregate crushing values with high accuracy by identifying aggregate type, particle size, angularity, void content and other features. CrushNet significantly outperformed all benchmark models on the test set (test set MAE = 0.0073, R 2 = 0.952), providing a versatile solution for the intelligent detection of aggregate crushing values. This study provides a viable approach for enabling rapid quality control of recycled aggregate (RA), reducing batch testing costs, and promoting high-throughput processing of RA in construction engineering projects. • First application of deep learning and image recognition to predict the crushing value of recycled aggregates. • CrushNet achieves superior accuracy ( MAE = 0.0073, R 2 = 0.952), outperforming six benchmark models. • Interpretability analysis via Grad-CAM reveals focus on stress-concentration and structural defects. • Provides a rapid, image-based solution for quality assessment of recycled aggregates in engineering.