http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
이종 데이터 학습 및 다중 시점 예측을 이용한 생산 수율 예측 기법
김경휘(Kyung Hwi Kim),손동연(Dong Yeon Son),임채환(Chehwan Lim),추호석(Ho Seok Choo),김재협(Jae Hyup Kim) 대한전자공학회 2024 대한전자공학회 학술대회 Vol.2024 No.6
In this paper, we propose a yield prediction method using design data and processing data. Key factors that affect the yield include the products unique characteristics (e.g. design) and their impact on the manufacturing process. Because the data of the two factors are static data and dynamic (time series) data, respectively, multi-modal learning is necessary. Additionally, it is an environment in which perfect control cannot be achieved during the production process, and if production continues, the yield will inevitably show average and variance. Therefore, the yield prediction problem should not be considered as a single numerical prediction, but rather as a problem of predicting production results at various points in time (i.e. multi-horizon) when production continues in the future, and using these to predict distribution. (i.e. multi-horizon forecasting) So, the proposed method was composed of learning multi-modal data and multi-horizon forecasting, and ultimately predict the distribution of the yield.
사전 클러스터링과 다중열 합성곱 신경망을 이용한 한글필기체 인식 연구
천원정(Won Jung Chun),김경휘(Kyung Hwi Kim),조현종(Hyun Jong Cho),김형중(Hyung Joong Kim),김재협(Jae Hyup Kim) 대한전자공학회 2018 대한전자공학회 학술대회 Vol.2018 No.11
In the case of Hangul handwriting, unlike the form consisting of a single structure of alphabets or numbers, it is composed of a combination of a initial/medial/final consonants. So, the character similarity is very high in image data, and there is over 2000 characters. In this paper, we propose a method to improve the recognition performance by reducing the complexity of the problem of Hangeul handwriting recognition. We propose a structure that recognizes more than 2000 characters at the same level by grouping high similarity characters through pre-clustering and assigning a separate deep learning model to each cluster. Experimental results, we show about 65% recognition performance for 2350 characters that can be combined, and about 79% for commonly used 1052 characters.