http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
딥러닝 기반 화력발전 보일러 출구 NOx 농도 선행 예측 모델
조현빈(Hyunbin Jo),강동협(Donghyup Kang),박성민(Seongmin Park),이종욱(Jongwuk Lee),류창국(Kang Y. Huh) 한국연소학회 2022 KOSCOSYMPOSIUM논문집 Vol.2022 No.11
In this study, we developed a deep learning model to forecast the NOx and oxygen concentration, and gas temperature at the boiler exit of a coal-fired power plant. The target boiler is a 500 MWe tangential firing boiler, which is one of 20 units often referred to as standard coal power plant. From the database of the power plant, 73 raw items of operation data with one-minute frequency were collected for a period of approximately 5 months. Through the feature selection procedure, the raw data items were condensed into 19 features which include coal feeder throughput to burners, air flow rate, and burner tilt. The features were then used to establish two types of data segments: segment #1 for current operation status and segment #2 for recent histories measured at the boiler exit. Considering the large fluctuations, the histories of the recent values at the boiler exit values were averaged over 5 min. After evaluating different prediction models with respect to the nature of the data segments, suitable models were applied in the form of ensemble model to forecast the boiler exit values 1 min in advance. When compared to measured data, the prediction quality was sufficiently high with a mean square error of 0.0123 for NOx emission.
분류기 불일치 기반 비감독 도메인 적응 기법을 통한 회전체 고장 진단 기술
이진욱(Jinwook Lee),김명연(Myungyon Kim),고진욱(Jin Uk Ko),신용진(Yongjin Shin),윤병동(Byeng D. Youn) 대한기계학회 2020 대한기계학회 춘추학술대회 Vol.2020 No.12
Recently, deep learning-based fault diagnosis has been studied in many fields because of its superiority compared to existing data-driven method. However, it requires assumption that training and test data are drawn from same distribution which is generally not satisfied in real industry field. In order to address this problem, called domain shift, alleviating discrepancy between marginal distributions is widely utilized in existing methods. While performance is enhanced, there is limitation that these methods cannot consider the relation between target samples and decision boundaries. This can cause performance degradation compared to not using domain adaptation, called negative transfer. To solve this problem, we propose to maximize the discrepancy between two classifiers’ probability outputs to detect target samples that are far from the source samples. A feature extractor learns to generate target features near the source samples to minimize the discrepancy. Two bearing datasets are utilized to validate the effectiveness of proposed method. The results show that proposed method outperforms existing marginal distribution matching method.
프레넬 렌즈 성형에서 tip radius가 광학 성능에 미치는 영향
심종명(Jongmyeong Shim),김호관(Hokwan Kim),김중억(Joongeok Kim),명호(Ho Myung),김명연(Myungyon Kim),강신일(Shinill Kang) 한국소성가공학회 2012 한국소성가공학회 학술대회 논문집 Vol.2012 No.10
This paper describes the effect of tip radius on optical performance for fabricating Fresnel lens. We firstly designed Fresnel lens for uniform light distribution by backward design method and energy compensation method. To verify the effect of tip radius, simulation was conducted varying lens tip radius. Designed Fresnel lenses were replicated by UV imprinting and injection molding process using metal mold fabricated by direct machining. Lens profile fabricated by each process and optical performance with each fabricated lens were measured. We compared the results with simulation data, and analyzed the effect of tip radius.