基於深度學習之混凝土坍度模型研究

Abstract

混凝土坍度為衡量新拌混凝土工作性之關鍵指標，直接影響施工順利性與成品品質。傳統坍度試驗雖為現場常見的檢測方式，但其具備一次性、延遲性及人工誤差等限制，難以滿足現代工程對即時性與精準性的需求。為克服上述問題，本研究提出一套結合影像處理與深度學習之混凝土坍度自動預測系統，透過攝影機擷取拌合過程中之混凝土影像，並以卷積神經網路（CNN）為核心，建構基於 ResNet 架構之回歸預測模型。 本研究共蒐集42部不同坍度條件之拌合影片，進行影像裁切、資料標註與增強後，建立完整訓練、驗證與測試資料集。模型訓練中透過超參數組合、Early Stopping 與混合精度訓練等技術提升效能，並使用多種資料增強策略優化泛化能力。實驗結果顯示，最佳模型在測試集上預測平均絕對誤差最小可達0.7757cm，驗證其應用於即時坍度監控之潛力。 為提升模型可解釋性，本研究進一步導入 Grad-CAM視覺化技術，分析模型對於影像中關注區域與實際混凝土特徵之關聯性。結果指出，模型多聚焦於水泥漿與粗細骨材交界區域，能夠捕捉與坍度相關之視覺線索，惟於含有攪拌機設備背景或過度動態畫面時預測準確性略有下降。 本研究成功展示深度學習於混凝土坍度預測之可行性與實務應用潛力，未來可望應用於智慧工地即時品管與自動化混凝土生產控制系統中，提升施工效率與品質穩定性。

Slump is a key indicator for evaluating the workability of fresh concrete, directly impacting construction efficiency and the quality of the final product. While the traditional slump test is commonly used on construction sites, it suffers from limitations such as one-time measurement, delays, and human error, making it inadequate for the modern engineering demands of real-time and accurate monitoring. To address these challenges, this study proposes an automatic concrete slump prediction system that integrates image processing and deep learning. Using a camera to capture images during the mixing process, a regression model based on a convolutional neural network (CNN) architecture—specifically ResNet—is developed. A total of 42 mixing videos under varying slump conditions were collected. After image cropping, data annotation, and augmentation, complete training, validation, and testing datasets were established. During model training, techniques such as hyperparameter tuning, early stopping, and mixed-precision training were employed to enhance performance, along with various data augmentation strategies to improve generalization. Experimental results showed that the best model achieved a minimum mean absolute error (MAE) of 0.7757 cm on the test set, demonstrating its potential for real-time slump monitoring. To improve model interpretability, Grad-CAM visualization was introduced to analyze the relationship between the model’s attention regions and actual concrete features. Results indicate that the model primarily focuses on the interface between cement paste and aggregates, successfully capturing visual cues related to slump. However, prediction accuracy decreases slightly in images containing background mixing equipment or overly dynamic scenes. This study successfully demonstrates the feasibility and practical potential of applying deep learning to concrete slump prediction, with promising applications in smart construction sites for real-time quality control and automated concrete production, ultimately enhancing construction efficiency and consistency.