AI Spots Dangerous Aggregate Clumps in Fresh Dam Concrete Before Vibration

Engineers building high arch dams may soon have a new digital assistant watching the concrete as it is poured. Researchers report in Case Studies in Construction Materials that an improved Siamese neural network, trained with a metric-learning strategy for small datasets, can recognize coarse aggregate clustering on freshly paved concrete surfaces with an F1-score of 97.72 percent on their test set. The work addresses a stubborn construction problem: before a concrete lift is vibrated, workers must judge by eye whether patches of the surface are crowded with coarse aggregate, and that judgment is inherently subjective.

Coarse aggregate clustering is a form of uneven aggregate distribution in fresh concrete. Concrete is a composite of aggregates and mortar, and the uniformity of its components strongly influences the performance, quality, and durability of the finished structure. In high arch dam construction, concrete passes through multiple transfer stages during transport, which inevitably reduces uniformity. After the concrete reaches the placement area and is spread by pavers, areas of clustered coarse aggregate can appear. If vibration proceeds directly over such regions, the outcome can be poor: areas without clustering may be over-vibrated while clustered regions receive insufficient vibration.

The current specification relies on manual, on-site determination of clustering followed by manual dispersal of the aggregates. The authors note that subjectivity makes this difficult to solve reliably before vibration. Core samples taken after the concrete has hardened can assess segregation, but such post-construction testing is not timely and offers limited guidance for the crews working the surface. What is needed, the researchers argue, is a timely judgment of whether clustering exists after paving and before vibration, so problems can be corrected before they compromise the vibration process and overall construction quality.

Existing approaches to concrete uniformity have significant limitations for this setting. Traditional experimental methods such as sampling and content extraction analysis are complex, labor-intensive, and not highly accurate. Newer techniques including CT, gamma-ray imaging, nuclear magnetic resonance, and ground penetrating radar mostly require strict laboratory environments and specialized equipment. Numerical simulations with discrete element or finite element methods have studied concrete in the mixer or immediately after production, but the multiple transportation and transfer stages between production and placement mean those results are hard to apply at the actual work face. Digital image methods, meanwhile, have largely analyzed cross-sections of hardened concrete, which is a post-analysis with poor real-time performance.

The practical constraints of a dam site shaped the study’s design. The time interval between paving and vibration is very short, and the construction environment offers limited space for data acquisition, so collecting sufficient on-site image data is challenging. This pushed the team toward few-shot learning, and specifically metric learning, which judges category by measuring similarity between samples rather than by training a large supervised classifier on fixed categories. The model’s purpose is to pull embeddings of the same category closer together and push different categories farther apart, so a query image can be assigned to the category of the nearest support sample.

The dataset was built from an actual high arch dam construction site in southwest China. Two smartphones, an OPPO Reno6 Pro+ and a Huawei P70 Pro, were used to record handheld videos across multiple construction warehouse surfaces and at different times of day, covering three concrete states: immediately after discharge from the cable crane bucket, after paving was completed, and after vibration was largely finished. Frames extracted every ten frames at 1280 by 720 resolution were cropped to a uniform 500 by 500 pixels, yielding 398 images: 131 of coarse aggregate clustering, 131 of normal aggregate distribution, and 136 of post-vibration concrete. The data were split into training, validation, and test sets of 266, 66, and 66 images, with random cropping, horizontal flipping, and color jittering applied for augmentation.

Because frame extraction and cropping can produce near-duplicate images, the team screened all training and test data with perceptual hashing, structural similarity, and ResNet-50 cosine similarity. Under the joint criteria of a pHash distance of 5 or less, SSIM of 0.95 or higher, and ResNet-50 cosine similarity of 0.99 or higher, the number of near-duplicate image pairs was zero. When only the looser ResNet-50 cosine similarity threshold of 0.95 was applied, the flagged proportions were 0.29 percent for clustering images, 0.50 percent for normal concrete, and 1.79 percent for post-vibration concrete, an increase the authors attribute to the surface becoming more uniform as vibration proceeds.

The detection model itself is an improved Siamese neural network using ResNet-10 as the backbone feature extractor, a lightweight variant derived from ResNet-18 by reducing the number of basic residual blocks in each of the four residual stages from two to one. Output channels of 64, 128, 256, and 512 are retained, and the change lowers the total number of residual blocks from eight to four, reducing depth and computational cost while keeping residual learning and multilevel feature extraction suited to a small dataset. The fully connected classification layer is removed, global average pooling produces a 512-dimensional feature vector, and an embedding projection maps it into a compact 128-dimensional space. Euclidean distances between paired embeddings are converted to a similarity score in the interval from 0 to 1 using an exponential transform, where larger values mean greater similarity. Training used an episode-based N-way K-shot paradigm with 2 categories per episode, 1 support sample and 3 query samples per category, 100 training and 50 validation episodes per epoch, the Adam optimizer with an initial learning rate of 2 times 10 to the negative fifth decayed by cosine annealing, weight decay of 1 times 10 to the negative third, dropout of 0.3, 60 epochs, and a contrastive loss with a margin of 1.9 and a negative-pair weight of 2.0, run on a consumer-grade NVIDIA GeForce RTX 4060.

Against four comparison methods on the test set, the proposed Siamese ResNet-10 achieved accuracy of 0.9851, precision of 0.9690, and an F1-score of 0.9772. LBP-SVM reached 0.6443 accuracy and 0.5952 F1, a pure ResNet-10 CNN baseline 0.7841 and 0.7477, a Relation Network 0.7949 and 0.7601, and a ResNet-18 transfer-learning baseline 0.9310 and 0.9021. Confusion matrix results showed the proposed method correctly recognized 0.98 of negative samples and 0.99 of positive samples, with false-positive and false-negative rates of only 0.02 and 0.01, compared with misclassification of 43 percent of negative samples by LBP-SVM and roughly 30 percent by the CNN and Relation Network baselines.

In application testing, images of freshly paved concrete taken after paving and before vibration were compared against clustering and normal reference images by relative proximity of similarity scores. One clustering image scored 0.654, 0.686, and 0.719 against three other clustering images but only 0.238 and 0.228 against two normal paving images, and another clustering-normal pair scored 0.205, consistent with actual conditions. When query images were brightened or darkened by one exposure value in linear RGB space, same-class similarities under low illumination ranged from 0.661 to 0.711 and different-class similarities from 0.211 to 0.245, close to the original results, while under high illumination same-class similarities dropped to between 0.553 and 0.570, a mean decrease from 0.686 to 0.561, possibly because increased exposure weakens texture information such as aggregate edges and local light-dark variation. Even so, within the illumination range tested, all similarity results remained consistent with the actual conditions. The authors present the work as a demonstration of feasibility for real engineering sites rather than a finished standard; notably, current specifications still lack an explicit, quantifiable threshold for identifying clustering, and the similarity-based framework is proposed as a step toward filling that gap.

Subject of Research: A metric-learning-based Siamese neural network method for identifying coarse aggregate clustering in freshly paved concrete at high arch dam construction sites under small-sample conditions.

Article Title: A metric-learning-based method for identifying coarse aggregate clustering in high arch dam construction under small-sample conditions

Article References: Chen, P., Qu, X., Guan, T., Pang, B., Wang, D., Guo, Z., & Xu, Y. (2026). A metric-learning-based method for identifying coarse aggregate clustering in high arch dam construction under small-sample conditions. Case Studies in Construction Materials, 25, Article e06643. https://doi.org/10.1016/j.cscm.2026.e06643

Image Credits: AI Generated

DOI: 10.1016/j.cscm.2026.e06643

Keywords: coarse aggregate clustering, high arch dam, concrete uniformity, metric learning, Siamese neural network, few-shot learning, ResNet-10, fresh concrete, construction quality control, image similarity, N-way K-shot, computer vision