http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
유연 다물체 동역학 시스템의 실시간 해석을 위한 DNN 기반 메타모델링 기법
한성지(Seongji Han),최희선(Hee-Sun Choi),최주환(Juhwan Choi),최진환(Jin Hwan Choi),김진균(Jin-Gyun Kim) 대한기계학회 2021 大韓機械學會論文集A Vol.45 No.10
본 연구에서는 유연 다물체 동역학 시스템에 포함된 유연체의 절점 위치, 응력 및 변형률을 지정된 설계 변수 입력을 통해 실시간으로 얻기 위한 심층신경망 기반 메타모델링 기법을 제시하였다. 유연체의 거동을 학습시키기 위해 적절한 형태의 입력 및 출력 데이터 구조를 구성하였으며, 수많은 유연체 절점 자유도로 인한 데이터 증가로 발생하는 학습 시간을 단축하기 위한 효율적인 훈련 알고리즘을 개발하였다. 알고리즘은 확률적 경사 하강(SGD: stochastic gradient descent) 단계와 오차 보정 단계로 구성되며, 전체 데이터의 일부만을 사용하여 학습된다. 확률적 경사 하강 단계에서 심층신경망은 거친 데이터 모음(coarse data set)을 사용하여 순차적/반복적으로 학습되며, 이 과정에서 심층신경망 모델의 개선이 충분하지 않은 경우 오차 보정 단계를 통해 정확도를 추가적으로 향상시킬 수 있다. 개발된 알고리즘의 효율성 및 정확성, 실시간성을 수치 예제를 통해 확인하였다. In this study, a deep neural network-based metamodeling technique is proposed to obtain the positions, stresses, and strains of the nodes of a flexible body in real-time using the input of designated design variables for the flexible multibody dynamics systems. To train the responses of a flexible body, an appropriate type of input and output data structure is constructed, and an efficient training algorithm is developed to reduce the computational resources caused by data increase owing to the numerous degrees of freedom of nodes in the flexible body. The algorithm includes a stochastic gradient descent (SGD) step and an error correction step, and the deep neural network model is trained using only a part of the entire data. In the SGD step, the model is trained sequentially and iteratively using a coarse data set. As the iterations proceed, when the improvement of the model is insufficient, the accuracy is further improved by an error correction step. The efficiency, accuracy, and the possibility of real-time simulation of the proposed algorithm are presented through a numerical example.
해저 채굴용 해양 로봇 해석을 위한 입자법 및 다물체동역학 기반 연성해석
김영광(Young Kwang Kim),한성지(SeongJi Han),한종부(Jong-Boo Han),여태경(Tae-Kyeong Yeu),김진균(Jin-Gyun Kim) 대한기계학회 2021 대한기계학회 춘추학술대회 Vol.2021 No.11
This research presents a co-simulation method of deep-sea excavation operations for Cyber Physics Operation System(CPOS) implementation of marine robots. Multi-body dynamics (MBD) is used to describe the deep-sea excavator, and discrete element method (DEM) is employed to modeling and co-simulation of the excavation operation under seabed. In DEM, the reference bonding force between particles are employed to well describe characteristics of the seabed. Finally, the co-simulation is established by using RecurDyn and EDEM software environments. To validate the system, we compared the results with the seafloor sulfide deposition mining simulation.
동역학 시스템을 위한 DNN 메타모델 기반의 시스템 보정 네트워크
장성일(Sungil Jang),최주환(Juhwan Choi),한성지(Seongji Han),김진균(Jin-Gyun Kim),최진환(Jin Hwan Choi) 대한기계학회 2023 대한기계학회 춘추학술대회 Vol.2023 No.11
Modeling and simulation of dynamic systems are widely used in mechanical system design and control. System identification, a process of correlation with experimental data or target data, is essential for the reliability of the implemented model. The process is possible when the relationship between various variables that affect the responses of the system can be understood. Since various assumptions and approximations are considered in the modeling process of a nonlinear system such as a multibody dynamics system, it is difficult to directly understand the relationship between variables. Therefore, a surrogate model is needed to understand the system. In this study, we propose a system identification framework using deep neural networks. This framework consists of two processes: System metamodel and System identification network. System metamodel that is trained by the nonlinear relationship between the changing output responses to the input variables of the system using a deep neural network is introduced. The data required for training the neural network is obtained from the numerical model of the system. The loss minimization process for initializing a new neural network, which is a system identification network to find an input that produces a target output is proceeded.