새로운 학습 방법을 적용한 방사형 기저 함수 신경회로망 설계에 관한 연구 : 예측 모델 및 패턴분류기의 성능 개선을 중심으로

박상범 수원대학교 2022 국내박사

In this thesis, a radial basis function neural networks(RBFNNs) prediction model and pattern classifier designed with the aid of a novel learning method are proposed. The objective of this study is focused on the improvement of performance of the existing RBFNNs model by indroducing a new learning methods. In the case of the existing RBFNNs, outlier and noisy data included in dataset and the location of RBFs over the input space may be closely related to the performance. Based on these contents, two kinds of learning methods are proposed and brief explanation of the proposed learning method is enumerated as follows: 1) In the proposed prediction model and pattern classifier, the weighted FCM clustering is iteratively used for refinement of center point of each cluster. Clustering method is used to determine the center points over the input space by anaylzing the distribution of data patterns. The center points determined through the clustering method can be considered as the location of RBFs. In this study, conventional clustering method is used to initilize RBFs and then weighted FCM clustering driven with the aid of an auxiliary information is iteratively used for refinement of the locations of RBFs. Auxiliary information to be used of the weighted FCM can be obtained through sigmoid function and cross-entropy error function. Through this iterative refinement process of the location of RBFs, the performance of RBFNNs may be improved. 2) In order to trrain coefficients of connection weights between the hidden layer and output layer of the proposed prediction model and pattern classifier, margin-maximization is applied. Margin-maximization is a mechanism used to enhance the generalization ability of support vector machine(SVM) by maximizing distance from hyperplane to the nearest data pattern. Through the margin-maximization, weights each of data patterns can be obtained and then utilized to train the coefficients of connection weights. In this study, margin-maximization technique applied to least-square version of SVM(LS-SVM) is used. Since margin-maximization technique of original SVM should solve quadratic programming, lots of computational cost is consumed. Also, in regression problem, margin-maximization technique of general SVM cannot be applied. In contrast, margin-maximization technique of LS-SVM applies the least squares method to calculate weights instead of solving the quadratic programming, so computational cost is lower than original SVM. LS-SVM solves problem by implementing a linear equation, so it can apply to classification problem as well as regression problem. From the viewpoint of performance improvement through the novel learning method, the proposed prediction model and pattern classifier are evaluated by using a variety of publicly available machine learning datasets and compared with a diverse of algorithms which realized to WEKA toolkt. In addition, the Friedman test is applied for statistical analysis of the proposed prediction model and pattern classifier. Furthermore, some practical application datasets such as the actifvated sludge process datasets, Portland cement datasets, plastic wastes datasets, and partial discharge datasets are also used to evaluate the performance.

Hydrogeochemical characterization of urban groundwater in Seoul, South Korea, using self-organizing map (SOM) and fuzzy c-means (FCM) clustering

이경진 Graduate School, Korea University 2014 국내석사

In this work, combined statistical approaches were conducted for quantitative evaluation of urban groundwater quality of Seoul metro-politan city. The main purposes of this study are: 1) to classify and characterize physico-chemical properties of groundwater in Seoul, and 2) to understand geochemical evolutions of groundwater which are strongly affected by spatial distribution of environmental/anthropogenic factors. Coupled multivariate statistical methods had been applied to the groundwater chemistry data (n=343). A total of 91 prototype vectors for 13 water quality variables were derived by the self-organizing map (SOM) technique. However, the SOM result was insufficient to grasp the overall pattern of water quality. Thus, we used the fuzzy c-means (FCM) clustering algorithm to the SOM result for more effective and quantitative data interpretation. Accordingly, the prototype vectors were classified into four main hydrogeochemical groups based on their fuzzy membership values. The spatial pattern of groundwater chemistry was then examined using the ordinary cokriging on the fuzzy membership values following the additive log-ratio transformation (ALR). The result showed a distinct spatial relationship between groundwater quality and environmental/ anthropogenic factors. The physico-chemical and spatial characteristics of each groundwater group are summarized as follows: 1) Group 1 represents groundwater with low EC (median=181 μS cm-1) and high DO values (median=6.79 mg L-1), and mainly locates in the northern mountainous area of Seoul 2) Group 2 water has intermediate EC values (median=362 μS cm-1) and high pH (median=7.80), and dominantly occurs at the southern mountainous part 3) Group 3 represents water with high Eh levels (median=457.8 mV) and high nitrate concentration (median=42.7 mg L-1), and distributes ubiquitously in Seoul 4) Group 4 groundwater has the highest EC values (median=589 μS cm-1) among four groundwater groups and dominantly occurs in the center of the city.

순차적 클러스터링기법을 이용한 송전 계통의 지역별 그룹핑

황호윤 건국대학교 대학원 2009 국내석사

A Korean electric power market is forming recently a market of period of transition state to change to a price system by a region at single price systems. According to change of paradigms, don't provide price signal of LMP(Locational marginal price) and no price differentiation in a local area by provide same price signal to a local area. For a similar price signal to local consumer, a regional clustering is important that can transfer a equal sign to the bus which has each different price on the basis of bus. If you take LMP and local information this method into consideration at the same time, and you give weight on neither one side, buses on a borderline can move on arbitrary. For solving this problem, in this paper will propose a clustering method of relative fluctuation rate.

퍼지 클러스터링 기반 신경회로망과 SVM 패턴 분류기 설계에 관한 연구 : 검은색 폐플라스틱 분류를 중심으로

배종수 수원대학교 2017 국내석사

Lately, the amount of waste plastics including black plastics is getting more and more increasing. According as lots of plastics are widely used in various industrial fields. Under these circumstances, necessity for recycling of limited useful resources is getting more and more important gradually and research related to plastic sorting system is being largely required for plastic recycling. Plastic sorting system constructed currently by Near Infrared Ray(NIR) is being exploited to classify colored plastics besides black plastic. However, the classification of black plastics still remains a challenging issue, because of the absorption of infrared rays of NIR spectrometer for black plastics. Design methodology to identify black plastics in introduced. ATR FT-IR, Raman, and LIBS spectroscopies are used to carry out qualitative as well as quantitative analysis and also comparative studies for black plastics. For ATR FT-IR spectrometer, the spectra data of black plastics can be measured through the contact of interval gap between the spectrometer and plastic. Its measurement speed is faster compared to NIR spectrometer. ATR FT-IR spectrometer which is the contact type of interval gap, has difficulty in the on-line application. As the contactless type of interval gap, Raman spectrometer can measure the samples quickly, but its ensuing effect leads to the difficulty of data extraction due to lots of noises as well as the difficulty of application to on-line system. Therefore, LIBS spectrometer which is the contactless type, is used to effectively extract spectra data being applied in the on-line system. But, whenever the spectra data are measured in the same sample through spectrometer, the position of peak points of the characteristic spectra data are partially changed or shifted. Design methodology which takes into consideration for the changed or shifted spectra data are introduce in this study. The design method of determining input variables corresponding to data peak points based on the chemical characteristic lead to more reasonable and effective technique for improving the performance of FRBFNN and SVM classifiers. Moreover, in order to improve the identification performance, intelligent computing algorithms such as Principal Component Analysis(PCA), Fuzzy Transform(FT), Fuzzy Radial Basis Function Neural Networks(FRBFNN), Support vector machine classifiers(SVM) and Particle Swarm Optimization(PSO) are considered to analyze and classify some types of black plastics. In the preprocessing step for classifying some black plastics, the characteristic peak points are extracted and region corresponding to each characteristic peak point is taken into consideration. Here, as the preprocessing techniques, PCA and Fuzzy Transform algorithms are used for the dimension reduction of data. And FRBFNN and SVM are exploited as intelligent classifiers. FRBFNN classifier is considered as the powerful tool with the synthesis technologies of fuzzy theory and neural networks for the identification of black plastics. SVM classifier is used for comparative studies with FRBFNN classifier. In conclusion, the design methodology related to preprocessing techniques based FRBFNN classifier is demonstrated as competitive and preferred network architecture, as well as superb performance.

FT-IR 분광법을 이용한 플라스틱 재질에 대한 최적화된 뉴로-퍼지 패턴 분류기 설계

송찬석 수원대학교 2016 국내석사

In this thesis, the new method is proposed for classifying plastic materials based on Neuro-Fuzzy algorithm. In the automatic separation process of plastics, some types of plastics such as PP, PS and PET are classified by using near infrared spectroscopy. However, it is difficult to separate black plastics due to its properties of absorbing Near Infrared Rays(NIR). In this study, in order to overcome the drawback of the existing method to classify black plastics, FT-IR with ATR is used. Transmittance spectrum data is obtained through Fourier transform infrared spectroscopy. The obtained transmittance spectrum data is preprocessed by the proposed two methods and then the preprocessed data is used as inputs to Neuro-fuzzy networks classifier. In the preprocessing step for classifying black plastics for some materials, the characteristic of materials is analysed. In this study, two preprocessing methods are considered. The first method is to extract characteristic peak points from spectrum data. The second one is to extract region based on each characteristic point. The data extracted by the preprocessing methods is used as the input values of RBFNN pattern classifier. RBFNNs consist of three modules such as condition, consequence and aggregation phase. FCM clustering is used as the activation function of hidden layer in the condition phase. In the consequence phase, the coefficients of polynomial function are estimated by using least square estimation. In the aggregation phase, the final output of RBFNN classifier is calculated by using fuzzy inference. The optimal parameters of RBFNN classifier are obtained by using differential evolution. The differential evolution is well known as a kind of various genetic algorithms. Finally, the experimental results of the preprocessing methods are compared and analyzed from the viewpoint of classification performance.

Machine Learning of Breast DCE-MRI in Assessing Background Parenchymal Enhancement for Cancer Risk Assessment

Douglas, Lindsay Nicole The University of Chicago ProQuest Dissertations & 2023 해외박사(DDOD)

To enhance breast cancer screening practices, artificial intelligence (AI) systems have been developed to aid radiologists in a variety of tasks. Machine learning (ML) techniques for computer-aided diagnosis are based on human-engineered or deep learning methods, and they depend on accurate segmentation for useful feature extraction. As the use of dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) has increased in breast imaging, particularly for high-risk screening, the potential for AI to provide significant clinical benefit has grown. There is need for a deeper understanding of how breast MRI can be used for diagnosis and risk assessment in order to develop robust, generalizable AI systems for quantifying clinically valuable breast characteristics.This dissertation presents novel methods for computerized assessment of background parenchymal enhancement (BPE), a known risk factor for breast cancer, from breast DCE-MRI. In Chapter 1, we introduce the background of breast cancer screening with a focus on AI applications to motivate the subsequent chapters. In Chapter 2, we investigate segmentation techniques for lesions and breast regions. In Chapter 3, we develop an ML technique for computer BPE scoring that includes electronic lesion removal. In Chapter 4, we perform an independent evaluation of the BPE scoring algorithm applied to high-risk patients. Ultimately, the results of this work have the potential to encourage future incorporation of quantitative image analysis into the clinical workflow for radiologists and therefore improve patient care.Segmentation of lesions and breasts: Methods for segmentation of breast lesions and breasts from DCE-MRI were investigated using a dataset of patients diagnosed with cancerous or benign mass- or nonmass-enhancing lesions. Lesion segmentation performances of U-Net convolutional neural networks were compared to the fuzzy c-means (FCM) clustering algorithm and to radiologist delineations. Breast segmentation was performed on post-contrast subtraction maximum intensity projection images. Results suggest that using a 2D U-Net on post-contrast subtraction DCE-MRIs is feasible and could be an effective alternative to FCM or 3D U-Net for lesion segmentation.Computerized assessment of BPE: An automatic computer BPE scoring method that includes electronic lesion removal was developed using a dataset of DCE-MRIs that had radiologist BPE ratings available from prior clinical review. Qualitative, radiologist-reported BPE ratings and quantitative, computer BPE scores were evaluated for different breast regions, and the effect of varying image types and magnet strengths was investigated. A statistically significant correlation was found between the radiologist and computer BPEs. Results demonstrated promising performances of the computerized method for classifying BPE levels across various viewing projections and DCE timepoints.BPE scoring on a high-risk dataset: The role of BPE in predicting breast cancer was explored for a dataset of high-risk screening DCE-MRIs. An independent validation of the BPE scoring algorithm reproduced findings from the initial dataset on an independent dataset. In addition, results found a statistically significant difference between the computer BPE scores of patients that developed cancer and those of non-cancer patients with low BPE. Future investigations involving enriched datasets would expand the understanding of the role that computer BPE scores can have in predicting cancer.

Success strategies for Smart Rehabilitation Device clusters : Applying exploratory analysis and FCM

Sung-jin Kim 대전대학교 일반대학원 2019 국내박사

현재 우리나라는 전 세계에서 가장 빠른 속도로 초고령사회로 진입하고 있다. 통계청 인구추계에 따르면 국내 1970년 전체인구 중 65세 이상 노령인구 비율 3.1%에서, 2017년 13.8%, 2040년부터는 전체인구 30%를 초과하는 고령사회가 될 전망이다. 한국의 기대수명은 지난 46년간 약 20여 년 증가하여 현재 선진국 수준에 도달했지만, 건강수명은 2015년 73세로 기대수명과 격차가 약 9년 정도 발생하였다. 이처럼 의료 기술의 발전으로 인간 수명이 연장되었으나 고령화로 인해 뇌졸중, 치매 등 만성질환 또한 증가하고 있다. 고령화의 진전으로 다양한 사회문제가 발생할 전망이지만, 다른 한편으로 재활산업이라는 새로운 산업이 태동하여 신성장동력으로 성장할 수 있다. 외국의 경우 국민소득 1만 달러 도달 후 본격적으로 재활 의료기기 시장이 형성되는데, 미국과 일본은 각각 1970대 말, 1985년부터 활성화되었다. 이에 비교해 한국은 2006년 (국민소득이 1만6천 달러)을 기점으로 재활 의료기기 시장이 형성되기 시작하였으나, 아직은 시장규모가 상대적으로 작은 상황이다. 재활산업이란 이러한 고령자, 장애인 등의 일상생활을 지원해 주며 삶의 질과 생산성을 높여주는 제품 및 서비스를 제공하는 산업을 뜻한다. 초고령사회에 진입함에 따라 재활 의료산업 또한 성장할 것으로 전망되지만, 아직 국내 재활 기기 업체는 대부분 영세한 관계로 연구개발 및 혁신제품 개발이 어려운 상황이다. 이에 국내업체의 역량을 높이기 위해 산학병연 간 협업 네트워크를 구성하여 기업을 육성시킬 수 있는 재활 의료산업 클러스터 조성이 필요할 것으로 생각한다. 본 논문에서는 재활 의료기기산업 성공을 위하여 크게 두 가지 연구기법을 적용하였다. 먼저 탐색적 연구기법을 적용하여 문헌 연구를 통하여 바이오 클러스터의 특정 및 주요 요인에 대해 분석하였으며, 국내외 의료기기 클러스터 사례분석 비교를 통해 관련 성공 요인들을 도출하였다. 도출한 요인들을 바탕으로 국내 의료기기업체의 애로사항 및 클러스터 조성이 필요사항에 대한 설문조사를 진행하였다. 이를 바탕으로 의료기기 전반적인 실태 및 필요사항에 대해 분석할 수 있으며, 추가적인 통계적 분석을 통해 유의한 결과를 끌어낼 수 있었다. 다음으로 관련 전문가들 의견을 통해 재활 의료산업에 영향을 미치는 요인들의 인과관계를 정리하며, 요인별 가중치 또한 산출하여 이를 기반으로 퍼지인식도(Fuzzy Cognitive Map, FCM)를 작성하였다. 작성된 퍼지 인식도를 바탕으로 시나리오별 추론 및 시뮬레이션하여 재활의료산업 육성을 위한 전략을 제시하였다. 퍼지 인식도를 통해 실시한 다양한 시뮬레이션 결과들이 향후 스마트재활 기기산업 클러스터 육성에 도움이 될 것으로 기대된다. 퍼지인식도맵 및 재활 의료기기 클러스터 성공 요인 분석을 위한 탐색적 결과의 시사점은 다음과 같다. 첫째, 이론적 배경을 통해 클러스터의 정의에 대한 탐색적 연구를 시도하였으며, 최종적으로 클러스터는 같은 산업에 속한 기관들이 지리적 인접성에 위치하면 차별화 및 비용 절감 효과에 따라 다른 지역에 입지한 기업들보다 경쟁 우위에 설 수 있다는 것을 주요 특징으로 도출할 수 있었다. 두 번째, 국내외 주요 의료기기 클러스터 사례 비교 분석을 통하여 앞서 이론적 배경에서 도출한 의료 관련 클러스터 주요 요인 가운데 우수한 연구기반 및 전문인력, 기반 등도 중요하지만, 그에 못지않게 자생적으로 발생하는 산학병연 네트워킹이 필요하며, 이를 이루기 위해 선도적 기업 또는 기관의 리더십이 중요하다는 것을 알 수 있었다. 세 번째, 클러스터 성공 요인을 바탕으로 한 설문조사를 통해 실제 의료기기업체들은 이론적 성공 요인보다 일차적으로 기업의 수익/비용에 영향을 미치는 부지나 물류, 인력 등에 따라 입지를 먼저 선택하지만, 궁극적으로 연구개발이나 마케팅에 대한 지원이 필요하다는 점을 도출하였다. 네 번째, 설문 결과를 바탕으로 통계적 분석을 한 결과, 국가 메디컬 클러스터와 인접한 기업들이 입지 사회적 요인에는 상대적인 만족도를 느끼지만, 물리적 환경에 대한 만족도는 상대적으로 낮으므로 이에 대한 보완이 필요하다는 것을 탐색할 수 있었다. 다섯 번째, 앞서 도출된 주요 입지요인을 바탕으로 FCM 시뮬레이션을 시행하여 긍정 시나리오와 부정 시나리오에 대해 시뮬레이션을 할 수 있었으며, 이를 통하여 클러스터의 주요 요인들, 즉 정부와 지자체의 제도적 지원 요인, 산학병연의 협력 등의 사회적 요인, 또 기반 시설 등의 물리적 요인의 중요성을 도출할 수 있었다. 본 연구의 의의는 기존 의료클러스터 성공 요인에 대한 이론적 배경을 실제 국내외 의료기기 클러스터의 사례와 비교 분석하여 공통점과 차이를 분석하였으며, 더 나아가 이를 바탕으로 실제 의료기기업체들 대상 설문조사를 진행하여 현장의 애로사항 및 클러스터에 대한 필요사항을 도출할 수 있었다는 것이다. 향후 이 논문의 시사점을 활용한 클러스터 관련 정책이 개발 및 적용될 것을 기대한다.

SAGE와 automatic clusterung 알고리즘을 이용한 SCM 연구

구자호 서울대학교 대학원 2007 국내석사