머신러닝을 활용한 콘크리트 압축강도 예측

다국어 초록 (Multilingual Abstract)

Objectives : The central objective of this research is to build, implement, and evaluate a Machine Learning tool that can predict concrete compressive strength with a high degree of accuracy. Specific research goals include:
· Selection and development of appropriate Machine Learning algorithms for predictive modeling.
· Creation of a comprehensive and representative dataset for training and validation purposes.
· Optimization of the model's hyperparameters to achieve optimal predictive performance

Methods : In this study, a Machine learning model was developed using a Python program to correlate W/B ratio, Water content per unit volume or concrete, sandcoarse aggregate ratio, fly ash, superplasticizer, blast furnace slag, 7 input properties and output values of concrete compression strength, which affect concrete compressive strength. Scikit-learn, Keras, and TensorFlow open source were used to simplify the development process. And, data was preprocessed and characterized with ANN, Random Forest Regressor, LGBM, and Catboost, With grid search, the model was optimized and verified, Finally, the model verification and evaluation process were performed using the K-fold method. By training and validating the K-fold cross-validation model using 10 datasets several times, an accurate model was made possible, and a model
with a small RMSE(root mean square error) was created with 20% of the evaluation data not included in the training.

Conclusions : Among the Machine Learning models proposed in this study, Catboost was evaluated as a more accurate model than ANN, Random Forest Regressor, and LGBM, although large errors occurred in some predictions. Although the performance of the models based on ANN and Random Forest was low, they reached a fitting coefficient of 0.913 and 0.917 respectively, indicating a MAE(mean absolute error) value similar to that of other models. As a result, the four ensemble models for concrete compression strength prediction are sufficiently guaranteed in performance and are considered to be applicable. As there is a lot of data, it is expected that machine learning models can predict more accurately, and adding more variables such as atmospheric temperature and humidity not covered in this study can also develop models with better predictive performance.

목차 (Table of Contents)

• Ⅰ. 서 론 1
• 1. 개요 및 배경 1
• 2. 연구목적 1
• Ⅱ. 문헌 고찰 2
• 1. 도입 2
• 2. 인공지능, 머신러닝 및 딥러닝 개요 3
• 1) 인공지능(AI) 3
• 2) 머신러닝(ML) 5
• 3) 딥러닝(DL) 6
• 3. 머신러닝, 딥러닝 모델을 이용한 콘크리트 특성 예측 7
• Ⅲ. 머신러닝 모델 개발 9
• 1. 개요 9
• 2. 머신러닝, 딥러닝 기법 10
• 1) 인공신경망(ANN) 11
• 2) 랜덤 포레스트 회귀분석기(RFR) 12
• 3) Light Gradient Boosting Machine(LGBM) 13
• 4) Catboost Regressor 14
• 3. 데이터 세트(Dataset) 15
• 4. 모델 구성 17
• 1) 데이터 전처리 및 특징 18
• 2) 그리드 검색 알고리즘을 통한 모델 최적화 20
• 3) K-fold 교차 검증을 이용한 모델 검증 22
• 4) 예측 성능 평가 지표 23
• Ⅳ. 결과 분석 24
• 1. 교차 검증을 통한 모델성능 평가 24
• 2. 모형의 예측능력 평가 25
• Ⅴ. 결 론 28
• 참고문헌 30
• 부록 : 데이터셋 36