Slump Estimation of Concrete Flow Discharged through a Mixer Truck Chute : A Machine Learning Approach Using 3D Flow Tracking and Time-Series Analysis = 믹서트럭 슈트를 통해 경사면으로 배출되는 콘크리트 유동의 슬럼프 추정:3차원 유동 추적 및 시계열 분석 기반 머신러닝 접근

다국어 초록 (Multilingual Abstract)

The slump test has long been established as the most widely used field method for evaluating the workability of concrete. However, since it relies on manual procedures, the results can vary depending on the operator’s skill level and environmental conditions. Moreover, as it is based on sampling, the test fails to fully capture the overall flowability and quality of the entire concrete being placed. To address these limitations, various automated techniques that combine vision-based sensing with machine learning have been proposed. However, most of these approaches focus on measuring concrete during the mixing or transportation stages. They primarily utilize sensors or video data from inside the mixer drum. Consequently, they fall short of providing the real-time information necessary for assessing and controlling concrete quality during the final placement stage. Notably, on-site, real-time sensing technologies that enable both producers and contractors to jointly assess and verify the flowability of fresh concrete under on-site conditions remain underdeveloped. In response, this study proposes a novel non-contact, non-intrusive vision-based method for estimating slump by observing the flow of concrete discharged from a mixer truck chute. Instead of relying on physical molds or barriers, the method analyzes the flow along an inclined plane, enabling real-time assessment under varying discharge rates and chute angles. To achieve this, a depth camera and optical flow algorithm are used to reconstruct the time-series 3D surface trajectory, from which position-based, kinematic, and physics-informed features are extracted and normalized using the Froude number to account for flow condition differences. These features are then used to train and evaluate several deep learning models, including CNNs, LSTMs, and Transformers, for both classification and regression tasks. Models that incorporated positional, kinematic, physical, and non-dimensional features showed the highest performance. Under laboratory conditions, the CNN-based classification model achieved a Macro-F1 score of 91.10%, and the LSTM-based regression model recorded an RMSE of 11.31 mm. Under field conditions, the CNN classifier achieved 99.88% Macro-F1 accuracy, while the LSTM regressor reduced RMSE to 2.55 mm, demonstrating both the practical applicability and robustness of the proposed method. Lastly, permutation importance analysis was conducted to quantitatively assess the contribution of each feature group. The results showed that non-dimensionalized and energy-based features played a critical role in enhancing prediction accuracy. This analysis lays the groundwork for developing lightweight models in future applications. Overall, this study presents an interpretable, non-contact framework for high-precision, real-time evaluation of concrete workability, demonstrating its potential for integration into intelligent construction and automated quality control systems.

국문 초록 (Abstract)

슬럼프 시험은 콘크리트의 워커빌리티(workability)를 평가하기 위한 가장 널리 사용되는 현장 시험 방법으로 오랫동안 활용되어 왔다. 그러나 수작업 절차에 의존하므로, 시험 결과는 작업자의 숙련도나 환경 조건에 따라 달라질 수 있으며, 샘플링 방식이라는 한계로 인해 전체 콘크리트의 유동성과 품질을 충분히 반영하지 못한다. 이러한 한계를 보완하고자, 영상 기반 센싱과 머신러닝을 결합한 다양한 자동화 기술이 제안되어 왔다. 하지만 대부분은 혼합 또는 운반 단계에 국한되며, 믹서 드럼 내부 센서나 영상 데이터를 활용하는 방식이 주를 이루어 최종 타설 단계에서 품질을 실시간으로 평가하고 제어하는 데에는 한계가 있다. 특히, 제조자와 시공자가 현장에서 신선한 콘크리트의 유동성을 함께 관찰하고 판단할 수 있는 실시간 계측 기술은 아직까지 충분히 개발되지 않았다. 이에 본 연구는 믹서트럭 슈트를 따라 흘러내리는 콘크리트 흐름을 영상으로 관측하여 슬럼프를 추정하는 비접촉식·비구조 간섭형 계측 기법을 제안한다. 물리적 몰드나 장치를 사용하지 않고, 경사면을 따라 흐르는 유동만을 분석함으로써 다양한 배출 속도 및 슈트 기울기 조건에서도 실시간 평가가 가능하다. 이를 위해 뎁스(depth) 카메라와 광학 흐름 알고리즘을 활용해 콘크리트 표면의 시계열 3차원 궤적을 복원하고, 위치, 운동학, 물리 기반의 특성을 추출한 뒤, 프로드수(Froude number)를 이용한 무차원화 과정을 통해 조건 차이를 보정하였다. 이러한 특성은 CNN, LSTM, Transformer 모델에 입력되어 분류 및 회귀 과제로 각각 학습·검증되었으며, 네 가지 특성을 모두 포함한 모델이 가장 높은 예측 성능을 보였다. 실험실 조건에서 CNN 기반 분류 모델은 Macro-F1 91.10%를, LSTM 기반 회귀 모델은 RMSE 11.31 mm를 기록하였고, 현장 조건에서는 각각 Macro-F1 99.88%, RMSE 2.55 mm로 더 높은 정확도를 나타냈다. 또한 permutation importance 분석을 통해 각 특성 그룹의 기여도를 정량적으로 평가한 결과, 무차원화 특성과 에너지 관련 특성이 예측 정확도 향상에 핵심적인 역할을 하는 것으로 나타났다. 이 결과는 계산 효율성을 고려한 경량화 모델 개발의 기반을 제공한다. 본 연구는 콘크리트 워커빌리티를 고정밀·실시간으로 평가할 수 있는 해석 가능한 비접촉 프레임워크를 제시하며, 지능형 건설 및 자동화 품질 관리 시스템으로의 확장 가능성을 보여준다.

목차 (Table of Contents)

• Contents
• Chapter 1. Introduction 1
• 1.1 Research Background 1
• 1.2 Problem Description 4
• 1.3 Research Objectives and Scope 7
• 1.4 Dissertation Outline 10
• Chapter 2. Literature Review and Theoretical Background 14
• 2.1 Chapter Introduction 14
• 2.2 Concrete Production, Transportation, and Placement Process 16
• 2.3 Slump Test: Principle, Procedure, and Limitations 18
• 2.4 Current Status and Limitations of Slump-Estimation Research 20
• 2.4.1 Concrete Slump Monitoring During the Manufacturing Stage 20
• 2.4.2 Concrete Slump Monitoring During Transport 22
• 2.4.3 Concrete Slump Monitoring After Arrival on Site 25
• 2.5 Fluid Flow Measurement Techniques and Applicability to Concrete 27
• 2.6 Applicability of Optical Flow to Concrete Flow Analysis 30
• 2.7 Dimensionless Scaling and Physics-Based Interpretation 33
• 2.6 Summary 37
• Chapter 3. Laboratory Experiment Methodology 39
• 3.1 Chapter Introduction 39
• 3.2 Laboratory Experimental Setup 41
• 3.2.1 Camera Accuracy Comparison: Monocular RGB vs. Depth Sensors 41
• 3.2.2 Structural Analysis of a Concrete Mixer Truckand Design Considerations 44
• 3.2.3 Design and Fabrication of Experimental Apparatus 47
• 3.3 Generating 3D Tracking Data by Optical Flow and Depth data 54
• 3.4 Development of Dynamic Features 60
• 3.5 Summary 70
• Chapter4. Field Experiment Methodology 73
• 4.1 Chapter Introduction 73
• 4.2 Field Setup & Instrumentation 75
• 4.3 Experimental Matrix & Procedure 79
• 4.4 Data Acquisition & Pre-processing 83
• 4.5 Summary 86
• Chapter 5. Machine Learning Model Construction & Development 88
• 5.1 Chapter Introduction 88
• 5.2 Common Techniques for Models 90
• 5.2.1 Normalization Methods 90
• 5.2.2 Data Preparation for Variable-Length Sequences 92
• 5.2.3 Temporal Encoding: Avoiding Sequence Length Bias 97
• 5.2.4 Pooling Mechanism 100
• 5.3 Model Architecture 103
• 5.3.1 CNN-based Model (Single Point and Multi Points) 103
• 5.3.2 LSTM-based Model (Single Point and Multi Points) 106
• 5.3.3 Transformer-based Model (Single Point and Multi Points) 109
• 5.4 Training Configuration & Hyper‑parameters 114
• 5.5 Permutation Importance Methodology 119
• 5.6 Summary 121
• Chapter 6. Laboratory Experiment Results 123
• 6.1 Chapter Introduction 123
• 6.2 Laboratory Experiment Model Comparison 125
• 6.2.1 Laboratory Experiment Classification Model Performance 125
• 6.2.2 Laboratory Experiment Regression Model Performance 130
• 6.3 Laboratory Experiment Feature Scenario Performance Comparison 134
• 6.3.1 Overview: Feature Scenario Structure and Evaluation (Laboratory) 134
• 6.3.2 Single Feature Group Analysis (Laboratory) 141
• 6.3.3 Combined Feature Group Analysis (Laboratory) 144
• 6.3.4 Performance of Combined Original and Non-Dimensionalized Features (Laboratory) 148
• 6.5 Laboratory Experiment Permutation Importance Analysis 152
• 6.6 Summary 159
• Chapter 7. Field Experiment Results 163
• 7.1  Chapter Introduction 163
• 7.2 Field Experiment Model Comparison 165
• 7.2.1 Field Experiment Classification Model Performance 165
• 7.2.2 Field Experiment Regression Model Performance 168
• 7.3 Field Experiment Feature Scenario Performance Comparison 171
• 7.3.1 Overview: Feature Scenario Structure and Evaluation (Field) 171
• 7.3.2 Single Feature Group Analysis (Field) 177
• 7.3.3 Combined Feature Group Analysis (Field) 181
• 7.3.4 Performance of Combined Original and Non-Dimensionalized Features (Field) 184
• 7.5 Field Experiment Permutation Importance Analysis 186
• 7.6  Summary 191
• Chapter 8. Conclusion & Future Work 193
• 8.1  Research Results 193
• 8.2  Research Contributions 195
• 8.3  Limitations & Future Directions 197
• References 200
• Korean Abstract 211