【作者】 王鑫；

【导师】 Bernd R.Noack；

【作者基本信息】 哈尔滨工业大学 ， 能源动力（专业学位）， 2025， 博士

【摘要】 流动现象广泛存在于交通、能源和工业领域,例如汽车的阻力、机翼的升力以及燃烧过程中的混合现象等。流动建模及控制促进了相关领域流动性能的提升。随着数据驱动方法在流体力学中的迅速发展,流动建模及控制方法正在经历深刻变革。流动建模和控制问题可以被视为映射函数的回归问题。具体而言,在流动建模中需要预测动力学状态,而在闭环控制中则需要确定执行器的输出。然而,由于流动现象具有高维和非线性特性,设计有效的映射函数仍面临巨大挑战。目前,多数研究仍然依赖于数学函数对连续的流动信息进行建模和控制。例如,基于线性化理论的策略虽然表达形式简洁,但仅适用于较为简单的流动预测和优化;而机器学习模型尽管能够处理强非线性流动,却往往需要耗费大量的测试资源。聚类方法作为一种重要的数据驱动方法,能够从流动的连续特征中提取出关键的局部特征,进而支撑复杂流动的高效建模及控制。本文将连续的流动信息粗粒化为数量有限的局部离散特征,并基于这些局部离散特征之间的关联,建立了高效且适用性强的流动建模、控制与识别框架。一方面,通过聚类方法将流动中连续变化的动力学特征自动地粗粒化为若干个局部离散特征。另一方面,通过空间离散化方法,将连续分布的流场域均匀地粗粒化为离散的空间单元。首先,采用了基于聚类中心网络建模方法(Cluster-based network modeling,CNM),通过局部离散特征之间的动力学转移,实现流体动力学特征的高效重构及分析。其次,发展了基于聚类中心控制方法(Cluster-based control,CBC),通过为局部离散特征赋予控制参数,构造平滑的控制函数,实现流体动力学特征的快速优化。最后,提出了基于空间粗粒化分析方法(Spatial coarse-graining-based characterization,SCC),通过分析离散的空间单元之间的流动转移,实现流体运动学特征的有效识别。对于流动建模,本文提出了一种CNM鲁棒性的评估策略,定量评估了聚类初始化对CNM分析结果的影响。对于流动闭环控制,本文提出了多输入多输出CBC优化策略,扩展了CBC的应用范围,并通过优化器的并行化设计,将CBC的优化效率提升了约三倍。对于流动识别,SCC通过分析数量级为O(10)离散空间单元之间的流动转移关系,实现了三维大尺度环流的高效识别。这一框架在二维圆柱绕流和三维室内环流仿真装置中进行充分验证。圆柱绕流装置由三个可自由旋转的小圆柱组成,被广泛应用于流动建模与控制领域,涵盖了大多数已知的流动减阻机制。在圆柱绕流的建模中,CNM有效地捕捉并重构了周期性流动特征(0)=30,100)、准周期性流动中主要特征(0)=130)以及混沌流动的部分特征(0)=150)。整体上,CNM重构后的各簇数据的占比与原始数据误差仅在1.3%以内。此外,CNM的初始化对混沌流的分析结果影响程度有限,周期流的CNM鲁棒性最佳。在圆柱绕流的控制中,以受控下圆柱绕流的瞬时速度场作为CBC的实时反馈,并以三个小圆柱的转动作为执行器。对CBC在周期性流动(0)=30,100)和混沌流(0)=150)中进行验证,以净阻力减少作为优化目标,综合考虑圆柱的驱动功率和阻力功率。实现了圆柱绕流装置净减阻分别减少33.06%、24.15%、12.23%。此外,相比开环控制,当雷诺数0)越大时,CBC优化后的净减阻效果越明显。室内环流装置来源于哈工大-深圳技术大学无人机测试中心的简化,通过阵列式风生成器在封闭空间中生成测试风场,以满足无人机在有风条件下的飞行测试需求。室内环流装置的研究有助于分析及改进大尺度室内环流的流动特性。在室内环流的识别与控制中,SCC将室内环流粗粒化为数量有限的离散单元之间的流动转移,进而识别室内环流特征。SCC分析表明,在室内环流装置中放置等腰直角三棱柱,可以形成垂直循环流动,从而增强测试风的对称性,并提高测试区域的风速。总之,本文基于数据驱动,提出了一种高效且适用性强的流动建模、控制和识别框架。通过将连续的流动特征粗粒化为数量有限的局部离散特征,并基于离散特征之间的关联,分析或优化流动特性。该框架为流体力学的研究提供了一个新的视角,并通过圆柱绕流和室内环流装置验证了该框架的有效性。更多还原

【Abstract】 The performance of transport,energy and industrial systems may strongly depend on flow properties.Examples are drag of cars,lift of airplane,and mixing in combus-tion.Optimizing flow properties is enabled by modeling and control technologies.In recent decades,data-driven machine-learning approaches have significantly contributed to effective modeling and optimization.Flow modeling and control for multiple-input multiple-output plants can be formu-lated as regression problems.In modeling,the data allows to identify a dynamics,i.e.a mapping from the current to the future state.In control,the feedback law from sensor sig-nals to actuation command need to be optimized for performance.Both regression prob-lems are challenged by the inherent high-dimensional,nonlinear dynamics of turbulence.Most approaches concern the continuous flow information by mathematical functions for modeling and control.For example,strategies based on linearization theory offer elegant formulations but are practical only for predicting and optimizing simple flow;machine learning,as a black-box model,is able to be applied to strongly nonlinear flows but re-quires large test budgets.Clustering,as a typical data-driven method,can extract key local states from the continuous flow information.The key local states facilitate a compromise between linear and strongly nonlinear dynamics,enabling analytical modeling and effi-cient feedback control for computational fluid dynamics within limited budgets.This study coarse-grains continuous time-dependent or space-dependent flow into a limited number of local states or subdomains.And,efficient flow modeling and control strategies are developed by leveraging the associations between these discrete states or subdomains.One approach is clustering,which coarse-grains time-dependent flow char-acteristics automatically into cluster centroids as discrete states.Another coarse-gaining method separates the flow mathematically into subdomains.First,a cluster-based net-work modeling(CNM)method is introduced to predict flow dynamics by capturing the transfers among the discrete states.Second,a cluster-based control(CBC)method is de-veloped to optimize flow dynamics by assigning control parameters to the discrete states.Finally,a spatial coarse-graining-based characterization(SCC)approach is proposed to identify flow kinematics through transitions among the subdomains.For flow modeling,a robustness evaluation strategy for CNM is proposed to quantitatively assess the impact of clustering initialization on CNM results.For flow control,a multi-input multi-output(MIMO)CBC is established to extend the application domain,and a parallel optimizer is introduced to achieving approximately a three-fold acceleration in optimization pro-cess.For flow identification,SCC enables both quantitative and qualitative identification of large-scale three-dimensional circulations by discrete subdomains of the order O(10).This framework is exemplified by a two-dimensional bluff-body benchmark and an engineering investigation of fan array wind generator in a drone testing lab.The academic benchmark is the fluidic pinball composed of three equal parallel rotating cylinders in am-bient flow.This configuration encapsulates most known drag reduction mechanisms,and is widely used in the field of flow modeling and control.CNM effectively captures and reconstructs periodic flows(0)=30,100),the dominant dynamics of quasi-periodic flow(0)=130),and partial dynamics of chaotic flow(0)=150).Overall,the proportions of each cluster in the data reconstructed by CNM deviate from the original data by no more than 1.3%.Initialization of CNM has limited influence at chaotic flow,while CNM exhibits the greatest robustness for periodic flows.For control of the fluidic pinball,the instantaneous state of the controlled flow serves as real-time feedback for CBC,while the rotations of the three small cylinders act as actuator.The enhanced CBC framework was then validated under periodic flows(0)=30,100)and chaotic flow(0)=150),with net drag reduction as the optimization objective.This objective incorporated considerations of both the actuation power and the drag power of the cylinders.The results demonstrated net drag reductions of 33.06%,24.15%,and 12.23%for the respective flow conditions.Moreover,as the Reynolds number increases,the net drag reduction achieved by CBC op-timization becomes more pronounced compared to open-loop control.The engineering scenario involves the recirculation derived from a simplification of the HIT-SZTU Aerial city lab,which generates a test wind field in an enclosed room by fan-array wind generator to satisfy the needs of drone test.Here,an analysis of the recirculation helps to identify and improve the test conditions.For identification and control of the recirculation,SCC coarse-grains the room flow into flow transfers among a limited number of discrete sub-domains,thereby identifying the characteristics of the recirculation.SCC is particularly revealing for changes of the recirculation by a triangular prism.This prism symmetrizes the flow by establishing a vertical circulation loop and increases the velocity in the test region.In summary,this thesis proposes an efficient data-driven framework for flow con-trol,modeling,and identification.By coarse-graining continuous flow characteristics into a limited number of locally discrete clusters or subdomains,it effectively analyzes and manipulates the flow characteristics.The framework provides a new perspective on the study of fluid mechanics,exemplified by the fluidic pinball and the recirculation in the aerial city lab.更多还原