• 한국컴퓨터정보학회논문지
• /
• 제28권4호
• /
• pp.41-51
• /
• 2023

본 논문에서는 보행자의 걸음걸이로부터 분노 감정 검출을 위한 다중 시간 윈도 특징 추출 기술을 제안한다. 기존의 걸음걸이 기반 감정인식 기술에서는 보행자의 보폭, 한 보폭에 걸리는 시간, 보행 속력, 목과 흉부의 전방 기울기 각도(Forward Tilt Angle)를 계산하고, 전체 구간에 대해서 최솟값, 평균값, 최댓값을 계산해서 이를 특징으로 활용하였다. 하지만 이때 각 특징은 보행 전체 구간에 걸쳐 항상 균일하게 변화가 발생하는 것이 아니라, 때로는 지역적으로 변화가 발생한다. 이에 본 연구에서는 장기부터 중기 그리고 단기까지 즉, 전역적인 특징과 지역적인 특징을 모두 추출할 수 있는 다중 시간 윈도 특징 추출(Multi-Time Window Feature Extraction) 기술을 제안한다. 또한, 제안하는 특징 추출 기술을 통해 각 구간에서 추출된 특징들을 효과적으로 학습할 수 있는 앙상블 모델을 제안한다. 제안하는 앙상블 모델(Ensemble Model)은 복수의 분류기로 구성되며, 각 분류기는 서로 다른 다중 시간 윈도에서 추출된 특징으로 학습된다. 제안하는 특징 추출 기술과 앙상블 모델의 효과를 검증하기 위해 일반인에게 공개된 3차원 걸음걸이 데이터 세트를 사용하여 시험 평가를 수행했다. 그 결과, 4가지 성능 평가지표에 대해서 제안하는 앙상블 모델이 기존의 특징 추출 기술로 학습된 머신러닝(Machine Learning) 모델들과 비교하여 최고의 성능을 달성하는 것을 입증하였다.

• 대한임베디드공학회논문지
• /
• 제14권4호
• /
• pp.207-218
• /
• 2019

The purpose of this study is to develop a model that can systematically study the whole learning process of machine learning. Since the existing model describes the learning process with minimum coding, it can learn the progress of machine learning sequentially through the new model, and can visualize each process using the tensor flow. The new model used all of the existing model algorithms and confirmed the importance of the variables that affect the target variable, survival. The used to classification training data into training and verification, and to evaluate the performance of the model with test data. As a result of the final analysis, the ensemble techniques is the all tutorial model showed high performance, and the maximum performance of the model was improved by maximum 5.2% when compared with the existing model using. In future research, it is necessary to construct an environment in which machine learning can be learned regardless of the data preprocessing method and OS that can learn a model that is better than the existing performance.

• Geomechanics and Engineering
• /
• 제29권3호
• /
• pp.249-258
• /
• 2022

A disc cutter is an excavation tool on a tunnel boring machine (TBM) cutterhead; it crushes and cuts rock mass while the machine excavates using the cutterhead's rotational movement. Disc cutter wear occurs naturally. Thus, along with the management of downtime and excavation efficiency, abrasioned disc cutters need to be replaced at the proper time; otherwise, the construction period could be delayed and the cost could increase. The most common prediction models for TBM performance and for the disc cutter lifetime have been proposed by the Colorado School of Mines and Norwegian University of Science and Technology. However, design parameters of existing models do not well correspond to the field values when a TBM encounters complex and difficult ground conditions in the field. Thus, this study proposes a series of machine learning models to predict the disc cutter lifetime of a shield TBM using the excavation (machine) data during operation which is response to the rock mass. This study utilizes five different machine learning techniques: four types of classification models (i.e., K-Nearest Neighbors (KNN), Support Vector Machine, Decision Tree, and Staking Ensemble Model) and one artificial neural network (ANN) model. The KNN model was found to be the best model among the four classification models, affording the highest recall of 81%. The ANN model also predicted the wear rate of disc cutters reasonably well.

• Geomechanics and Engineering
• /
• 제37권6호
• /
• pp.539-554
• /
• 2024

This study explores development of prediction model for seismic site classification through the integration of machine learning techniques with horizontal-to-vertical spectral ratio (HVSR) methodologies. To improve model accuracy, the research employs outlier detection methods and, synthetic minority over-sampling technique (SMOTE) for data balance, and evaluates using seven machine learning models using seismic data from KiK-net. Notably, light gradient boosting method (LGBM), gradient boosting, and decision tree models exhibit improved performance when coupled with SMOTE, while Multiple linear regression (MLR) and Support vector machine (SVM) models show reduced efficacy. Outlier detection techniques significantly enhance accuracy, particularly for LGBM, gradient boosting, and voting boosting. The ensemble of LGBM with the isolation forest and SMOTE achieves the highest accuracy of 0.91, with LGBM and local outlier factor yielding the highest F1-score of 0.79. Consistently outperforming other models, LGBM proves most efficient for seismic site classification when supported by appropriate preprocessing procedures. These findings show the significance of outlier detection and data balancing for precise seismic soil classification prediction, offering insights and highlighting the potential of machine learning in optimizing site classification accuracy.

• 한국공간구조학회논문집
• /
• 제23권3호
• /
• pp.87-94
• /
• 2023

Chloride is one of the most common threats to reinforced concrete (RC) durability. Alkaline environment of concrete makes a passive layer on the surface of reinforcement bars that prevents the bar from corrosion. However, when the chloride concentration amount at the reinforcement bar reaches a certain level, deterioration of the passive protection layer occurs, causing corrosion and ultimately reducing the structure's safety and durability. Therefore, understanding the chloride diffusion and its prediction are important to evaluate the safety and durability of RC structure. In this study, the chloride diffusion coefficient is predicted by machine learning techniques. Various machine learning techniques such as multiple linear regression, decision tree, random forest, support vector machine, artificial neural networks, extreme gradient boosting annd k-nearest neighbor were used and accuracy of there models were compared. In order to evaluate the accuracy, root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE) and coefficient of determination (R2) were used as prediction performance indices. The k-fold cross-validation procedure was used to estimate the performance of machine learning models when making predictions on data not used during training. Grid search was applied to hyperparameter optimization. It has been shown from numerical simulation that ensemble learning methods such as random forest and extreme gradient boosting successfully predicted the chloride diffusion coefficient and artificial neural networks also provided accurate result.

• 전자공학회논문지C
• /
• 제34C권1호
• /
• pp.28-41
• /
• 1997

This paper describes a parallel implementation of a neurla network ensemble developed for object recognition on the connection machine CM-2. The implementation ensures that multiple networks are implemented simultaneously starting from different initial weights and all training samples are applied to each network by one sample per a copy of each network. When compared with a sequential implementation, this accelerates the computation speed by O(N.m.n) where N, m, and n are the network, respectively. The speedup in the computation time and the convergence characteristics of sthe modified backpropagation learning precedure were evaluated by two-dimensional object recognition problem.

• Geomechanics and Engineering
• /
• 제37권5호
• /
• pp.475-498
• /
• 2024

The prediction of the susceptibility of soil to liquefaction using a limited set of parameters, particularly when dealing with highly unbalanced databases is a challenging problem. The current study focuses on different ensemble learning classification algorithms using highly unbalanced databases of results from in-situ tests; standard penetration test (SPT), shear wave velocity (V_s) test, and cone penetration test (CPT). The input parameters for these datasets consist of earthquake intensity parameters, strong ground motion parameters, and in-situ soil testing parameters. liquefaction index serving as the binary output parameter. After a rigorous comparison with existing literature, extreme gradient boosting (XGBoost), bagging, and random forest (RF) emerge as the most efficient models for liquefaction instance classification across different datasets. Notably, for SPT and V_s-based models, XGBoost exhibits superior performance, followed by Light gradient boosting machine (LightGBM) and Bagging, while for CPT-based models, Bagging ranks highest, followed by Gradient boosting and random forest, with CPT-based models demonstrating lower G_mean(error), rendering them preferable for soil liquefaction susceptibility prediction. Key parameters influencing model performance include internal friction angle of soil (ϕ) and percentage of fines less than 75 µ (F₇₅) for SPT and V_s data and normalized average cone tip resistance (q_c) and peak horizontal ground acceleration (a_max) for CPT data. It was also observed that the addition of V_s measurement to SPT data increased the efficiency of the prediction in comparison to only SPT data. Furthermore, to enhance usability, a graphical user interface (GUI) for seamless classification operations based on provided input parameters was proposed.

• Journal of Electrical Engineering and Technology
• /
• 제12권6호
• /
• pp.2177-2186
• /
• 2017

The goal of this paper is to provide the specific forecasting steps and to explain how to design the forecasting architecture and training data sets to forecast very short-term wind power when the numerical weather prediction (NWP) is unavailable, and when the sampling periods of the wind power and training data are different. We forecast the very short-term wind power every 15 minutes starting two hours after receiving the most recent measurements up to 40 hours for a total of 38 hours, without using the NWP data but using the historical weather data. Generally, the NWP works as a predictor and can be converted to wind power forecasts through machine learning-based forecasting algorithms. Without the NWP, we can still build the predictor by shifting the historical weather data and apply the machine learning-based algorithms to the shifted weather data. In this process, the sampling intervals of the weather and wind power data are unified. To verify our approaches, we participated in the 2017 wind power forecasting competition held by the European Energy Market conference and ranked sixth. We have shown that the wind power can be accurately forecasted through the data shifting although the NWP is unavailable.

• 한국컴퓨터정보학회논문지
• /
• 제28권11호
• /
• pp.53-63
• /
• 2023

본 연구는 고령층의 치매 예방을 위한 선별검사 수단으로 자동화된 기계학습(AutoML)을 활용하여 인지기능 장애 예측모형을 개발하였다. 연구 데이터는 한국지능정보사회진흥원의 '치매 고위험군 웨어러블 라이프로그 데이터'를 활용하였다. 분석은 구글 코랩 환경에서 PyCaret 3.0.0이 사용하여 우수한 분류성능을 보여주는 5개의 모형을 선정하고 앙상블 학습을 진행하여 모형을 통합한 뒤, 최종 성능평가를 진행하였다. 연구결과, Voting Classifier, Gradient Boosting Classifier, Extreme Gradient Boosting, Light Gradient Boosting Machine, Extra Trees Classifier, Random Forest Classifier 모형 순으로 높은 예측성능을 보이는 것으로 나타났다. 특히 '수면 중 분당 평균 호흡수'와 '수면 중 분당 평균 심박수'가 가장 중요한 특성변수(feature)로 확인되었다. 본 연구의 결과는 고령층의 인지기능 장애를 보다 효과적으로 관리하고 예방하기 위한 수단으로 기계학습과 라이프로그의 활용 가능성에 대한 고려를 시사한다.

• Ecology and Resilient Infrastructure
• /
• 제10권4호
• /
• pp.107-115
• /
• 2023

고 탁도의 원수는 정수장 운영 및 수 생태 환경에 부정적인 영향을 줄 수 있어 관리가 필요한 수질 인자이며, 하천의 탁도 예측을 통해 고 탁도의 원수의 효율적 관리를 수행하기 위해 관련분야에 대한 연구가 지속되고 있다. 본 연구에서는 대표적인 앙상블 머신러닝 알고리즘 중 하나인 LightGBM (light gradient boosting machine)을 이용하여 탁도를 예측하는 다중 분류 모형을 구축하였다. 모형의 구축을 위해 입력자료를 탁도값에 따라 탁도가 낮은 경우부터 높은 경우까지 4개의 class로 구분하였으며, class 1 - 4에 속하는 자료수는 각각 945개, 763개, 95개, 25개로 분류되었다. 구축한 모형의 class 1 - 4에 대한 정밀도 (Precision) 각각 0.85, 0.71, 0.26, 0.30 재현율 (Recall)은 각각 0.82, 0.76, 0.19, 0.60로 데이터 수가 적은 소수 class에서 상대적으로 모형이 성능이 낮은 경향을 보였다. 데이터 불균형을 해소하기 위해 over-sampling알고리즘 중 SMOTE를 적용한 결과 개선된 모형의 class 1 - 4에 대한 정밀도 및 재현율은 각각 0.88, 0.71, 0.26, 0.25 및 0.79, 0.76, 0.38, 0.60으로 데이터 불균형 해소를 통해 모형의 재현율이 크게 개선되는 것을 확인할 수 있었다. 또한 데이터 구성비율이 모형성능에 미치는 영향에 대한 확인을 위하여 입력자료의 구성비를 다양하게 하고 각각의 자료로 구축된 모형의 결과를 비교하여 입력자료 구성비에 따른 모형성능의 차이를 분석하였으며, 모형 입력자료의 구성비의 적정한 산정을 통해 모형의 성능을 향상시킬 수 있음을 확인하였다.