http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
시간 합성곱 신경망을 이용한 베어링 성능 저하 예측 방법 개발
류동흠(Dong Heum Ryu),이용빈(Yongbin Lee),최동훈(Dong-Hoon Choi) 대한기계학회 2021 대한기계학회 춘추학술대회 Vol.2021 No.11
Performance degradation of a rolling bearing is a major factor for machine failure, which makes its prediction and diagnosis an important topic in engineering. Various studies have been carried out to predict this performance degradation, where recent developments in deep-learning led many researchers to tackle this problem using recurrent neural network (RNN) architectures such as long-short-term-memory (LSTM). However, RNN architectures consists of intrinsic problems such as the ‘requirement of substantial memory resource’ or ‘gradient vanishing’. Hence, this study proposes using a temporal convolutional network (TCN) for the bearing performance-degradation prediction. TCN is a convolutional neural network (CNN) that can predict time-series data using dilated causal convolutions. It has recently gained popularity in time-series prediction due to its capability to overcome issues caused by RNN architectures. A bearing dataset provided by NASA has been used to verify the performance of TCN, and the results were compared to other conventional deep-learning techniques.
대용량 데이터를 다루기 위한 합성우도함수 기반의 새로운 Gaussian process regression 모델 개발
류동흠(Dong Heum Ryu),이용빈(Yongbin Lee),최동훈(Dong-Hoon Choi) 대한기계학회 2018 대한기계학회 춘추학술대회 Vol.2018 No.12
Gaussian process (GP) model, a well-known surrogate model, are frequently used in various engineering fields due to its predictive performance. However, since GP model requires large numbers of matrix inversion for optimizing userdefined parameters, constructing the model becomes difficult as the size of data increases. This drawback limits the use of GP model for accumulated large data. In this study, a novel Gaussian process regression model, which can handle large data, is proposed. The proposed method is based on a main idea of a composite likelihood function where three key techniques -1) data splitting method, 2) user-defined parameter optimization, 3) fuzzy inference- are introduced to improve performance of the model. In order to verify the performance, benchmark problems with various tendencies were used. The results were compared with existing machine learning methods where the proposed method proved to have better performance.
다층 퍼셉트론의 네트워크 아키텍쳐 결정을 위한 빅데이터 기반 효율적인 탐색 기법
류동흠(Dong Heum Ryu),이용빈(Yongbin Lee),최동훈(Dong-Hoon Choi) 대한기계학회 2020 대한기계학회 춘추학술대회 Vol.2020 No.12
A multilayer perceptron (MLP) is a deep learning model commonly used in various fields of engineering for building a regression model as a substitute for a nonlinear system. However, selection of MLP network architecture, which much influences the performance of an MLP, remains a challenge. In this study, an efficient search algorithm based on big data is proposed for selecting the appropriate network architecture. The proposed algorithm extracts 9 suitable candidates of network architectures and selects the structure with the best predictive performance for a new data. The candidates are extracted from a big data of approximately 1.4 million MLP network architectures optimized for benchmark regression problems with various tendencies. Various benchmark regression problems were used in order to verify the performance of the proposed algorithm and the results were compared to other existing algorithms.
Proper Orthogonal Decomposition을 활용한 극저온 수직 평판의 열전달 예측
류익현(Ikhyun Ryu),류동흠(Dongheum Ryu),이용빈(Yongbin Lee) 대한기계학회 2023 대한기계학회 춘추학술대회 Vol.2023 No.11
In this research, heat transfer prediction of cryogenic vertical flat plate subject to frosting was conducted using Proper Orthogonal Decomposition (POD). Experiments were conducted using a heat flux sensor from varying the air temperature, air humidity, air velocity, and wall temperature. With POD, one of the popular Reduced Order Modeling (ROM) techniques, dominant behaviors of the heat flux variation were extracted to generate a predictive model together with Polynomial Regression (PR). To compare with the previous work of the correlation equation built from the common heat transfer characteristics, two performance metrics were calculated: Mean Absolute Error (MAE) and the relative error of the total heat transfer prediction. As a result, the predictive model built using POD showed better prediction performance compared to the correlation equations, showing great potential in predicting time-series outputs of the heat transfer of cryogenic vertical flat plate under forced convection.
박준혁(Joon-Hyuk Park),류동흠(Dongheum Ryu),이용빈(Yongbin Lee) 대한기계학회 2023 대한기계학회 춘추학술대회 Vol.2023 No.11
Latin Hypercube Design (LHD) is one of the most widely used design of experiment methods in computer simulation, not involving random errors. However, because of randomness in LHD, its space-filling property is not robust. Although optimal LHD (OLHD) has been developed to resolve the robustness issue, the increase in the size of samples makes it practically impossible to find the optimal solution, due to rapid increases in computational time and the number of design variables for optimization problems of OLHD generation. Translational Propagation Latin Hypercube Design (TPLHD), creating near-optimal LHD based on rules without optimization, handled these shortcomings. TPLHD, however, has poor space-filling property compared to OLHD. Therefore, this paper proposed a method to efficiently create a large set of samples with space-filling property comparable to OLHD, through optimizing seed design of TPLHD instead of entire experimental points. Its performance was proved effective by comparing it with TPLHD and OLHD.
생성모델로 생성한 인공 실험점을 활용한 순차적 샘플링 기법
김동건(Dong-Keon Kim),류동흠(Dong Heum Ryu),이용빈(Yongbin Lee) 대한기계학회 2022 대한기계학회 춘추학술대회 Vol.2022 No.11
The existing sequential sampling method has a drawback that convergence is very poor when the number of initial experimental points is small. To handle this issue, we present a method to improve the performance of sequential sampling by increasing the number of initial experimental points using artificial experimental points created by a generative model designed with an artificial neural network. A generative model suitable for tabular data (TVAE) is used for generating the artificial experimental points and Expected Improvement (EI) is applied for additional sampling. Finally, the proposed algorithm is verified by comparing its performance with the existing EI technique.
생성모델로 생성한 인공 실험점을 활용한 순차적 샘플링 기법
김동건(Dong-Keon Kim),류동흠(Dong Heum Ryu),이용빈(Yongbin Lee) 대한기계학회 2022 대한기계학회 춘추학술대회 Vol.2022 No.11
The existing sequential sampling method has a drawback that convergence is very poor when the number of initial experimental points is small. To handle this issue, we present a method to improve the performance of sequential sampling by increasing the number of initial experimental points using artificial experimental points created by a generative model designed with an artificial neural network. A generative model suitable for tabular data (TVAE) is used for generating the artificial experimental points and Expected Improvement (EI) is applied for additional sampling. Finally, the proposed algorithm is verified by comparing its performance with the existing EI technique.