|Alternative Title||Self-propelled swimming of a ﬂexible wing and ﬂow pattern recognition|
|Place of Conferral||北京|
|Keyword||自推进运动 流固耦合 尾迹分类 机器学习 湍流识别|
湍流和非湍流之间界面的识别是湍流研究中一个具有挑战性的课题。本研究提出了采用机器学习的方法训练检测器来识别流经圆柱的湍流区域。为了保证湍流和非湍流之间的界面与坐标系的选取无关，本文提出采用流场中的张量不变量作为输入特征来训练检测器。通过数值模拟的方式生成雷诺数为 Re=100 和 Re=3900 的圆柱绕流流场数据，用于训练和测试探测器的识别性能。为了测试不同机器学习方法的性能，分别训练了四个检测器，即基于全连接人工神经网络 (FCN) 的方法，基于极端梯度提升(XGBoost) 的方法，以及基于两组不同训练数据集的自组织映射网络 (SOM) 的方法。研究发现，有监督的学习方法(FCN 和 XGBoost 方法)在识别流动状态时的性能优于无监督学习的 SOM 方法。其中，XGBoost 检测器将 Re=100 时整个流场区域的流动状态识别为非湍流状态。FCN 检测器也正确识别了 Re=100 时流场中的绝大部分非湍流区域，除了一些远离圆柱尾迹的位置。对于 Re=3900 时的流场, FCN 和 XGBoost 方法都成功地捕获了蜿蜒的尾迹特征。通过比较 XGBoost 探测器和基于涡量模量和交叉速度脉动的检测方法，发现 XGBoost 检测器优于这些传统的检测方法，同时 XGBoost 检测器在较高雷诺数 Re=5000 时的流场中表现出了很好的鲁棒性。
Most creatures in nature are immersed in fluids, such as flying birds and swimming fish.These organisms developed different athletic skills and ability to sense the environment changes in the process of survival and evolution. For example, the high efficient self-propulsion locomotion of flying and swimming animals, aquatic organisms use the information of the surrounding flow field for navigation and motion planning, etc.The efficient propulsion of living animals and their ability to sense the environment are important inspirations for the design of artificial aircraft, underwater vehicles and signal detectors.The self-propelling motion of living animals usually involves the deformation of flexible wings or fins. The study of such problems can help us to understand the mechanism of the formation of efficient propelling of living animals and design better biomimetic propellers. Aquatic organisms can identify the difference of hydrodynamic signals in the local flow field environment by their sensory organs. The research of local flow field information recognition can provide some theoretical guidance for the corresponding bionic flow field signal detectors.In addition, there is another type of flow field recognition problem which is also the focus of people's attention, namely the identification of the interface between turbulent and non-turbulent flow.Therefore, in order to solve the problem of self-propelling motion of the bionic flexible wings, the self-propelled swimming flexible foil model is established and the influence of some key control parameters on the swimming performance are explored.For the local flow field information recognition problem, the classification of wake structures produced by self-propelled swimmers based on local measurements of flow variables is studied, and the artificial neural network recognition model is established. Finally, for the turbulence and non-turbulence interface recognition in flow field, a data-driven turbulence and non-turbulence interface recognition model is established by using a variety of machine learning methods.
The main innovative works of this paper are summarized as follows:
1. Self propulsion of a flexible foils driven by coupled plunging and pitching motions
The self propulsion performance of a flexible foil driven by coupled pitching and plunging motions at the leading edge is numerically invesigated. The influence of the driving parameters such as the bending rigidity of the flexible foil, the phase offset between plunging and pitching, the amplitudes of plunging and pitching motions on the propulsion performance are studied.It is found that with increasing rigidity, the
swimming style of the self-propelled foil gradually transits from the undulatory mode the oscillatory mode.The plunging-pitching actuation is found to be superior to the plunging-only actuation, in the sense that it prevents the decline of cruising speed at high rigidity. In addition, it has higher propulsive efficiency in a wide range of bending rigidity. It is found that the phase offset between plunging and pitching motions is a key factor affecting the propulsion performance of the flexible foil. Finally, the kinematic morphology and wake structure of the self-propelled flexible foil are compared with those of swimming animals, and the classification of the wake structure is discussed.The results of this study here provide some novel insights for the design of bionic underwater vehicles.
2. Classifying wakes produced by self-propelled foils using artificical neural networks
For the recognition of the local flow field hydrodynamic signals, the identification of self-propelled flexible foils wake structures based on measured local flow field variables is investigated. By training different artificial neural network based models, the wake structures produced by the self-propelled foils are identified by the local streamwise velocity component, the local crosswise velocity component, the local vorticity and the combination of three flow variables, respectively. It is found that the neural network model trained with two local velocity components performs well in recognizing wake types, while the neural network model trained with local vorticity value has a low classification accuracy. Finally, the combination of these three local flow field variables is used to train the artificial neural network, which can obtain a high wake recognition accuracy.The results of this study can provide some theoretical guidance for the design of environment sensing system of underwater robots.
3. Use machine learning to detect turbulent region in flow past circular cylinde
The identification of the interface between turbulence and non-turbulence is a challenging topic in turbulence research.In this study, a machine learning based method is proposed to train the detectors to recognize the turbulent regions in the flow past a circular cylinder. In order to ensure that the interface between turbulent and non-turbulent is independent of the reference coordinate system, the invariances in the flow field are used as input features to train the detectors. The training and testing data with Reynolds number of Re=100 and Re=3900 are generated by numerical simulation.To evalute the performance of different machine learning methods, four detectors were trained, namely, the fully connected neural network (FCN), the extreme gradient boosting (XGBoost), and the self-organizing map (SOM) based on two different training data sets, respectively. It is found that the performance of supervised learning method (FCN and XGBoost) is better than the unsupervised learning method (SOM) in identifying flow state. the flow state of the whole flow field is identified as non-turbulent regions by XGBoost based detector for Re=100. The FCN detector also identifies most of the flow state for Re=100 as non-turbulent regions, except for some locations away from the cylinder wakes. For the flow field at Re=3900, both FCN and XGBoost methods successfully capture the wake meandering. By comparing XGBoost detector with the detection method based on the vorticity modulus and cross-steam fluctuation intensity, it is found that XGBoost detector is superior to these conventional detection methods, and XGBoost detector shows robustness at high Renolds number flow field at Re=5000.
|李秉霖. 柔性扑翼的自主推进与流场模式识别[D]. 北京. 中国科学院大学,2020.|
|Files in This Item:|
|李秉霖硕士学位论文.pdf（7330KB）||学位论文||开放获取||CC BY-NC-SA||Application Full Text|
|Recommend this item|
|Export to Endnote|
|Similar articles in Google Scholar|
|Similar articles in Baidu academic|
|Similar articles in Bing Scholar|
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.