【作者】 陈曦；

【导师】 梁生；

【作者基本信息】 北京交通大学 ， 光学工程， 2024， 硕士

【摘要】 随着汽车智能化程度的提升,传统仪表盘逐渐难以满足多维度的人机交互需求。作为第三代车载抬头显示系统,增强现实抬头显示(Augmented Reality HeadUp Display,AR-HUD)能够实现真实世界和虚拟世界的信息集成,有望成为未来车载人机交互的发展方向。在AR-HUD的光学组合器中,表面浮雕光栅(Surface Relief Grating,SRG)以其灵活的设计与紧凑的结构成为一种主流技术方案。本文面向SRG针对多种光栅结构展开研究,基于机器学习实现了一种高效率的光栅性能预测方法。本文的具体研究内容如下:(1)针对光栅衍射效率提升难度大,光栅设计优化过程复杂等难题,基于有限元法(Finite Element Method,FEM)构建了目前应用相对广泛的倾斜光栅、镀膜光栅、夹层光栅、双柱光栅以及U形光栅共5种光栅的物理模型,研究了其衍射效率,通过分析光栅结构对衍射效率的影响机理设计出了一种混合光栅,其衍射效率与传统倾斜光栅相比提升了40%以上。(2)针对AR-HUD的大视场需求,通过转折光栅实现了波导内光线的方位偏转,完成光栅波导的二维扩瞳。同时本文构建出了由耦入光栅,转折光栅和耦出光栅组成的光栅波导全光路模型,通过射线追踪获取光强分布情况,基于衍射效率与衍射次数对出射光均匀度进行优化,实现了入射FOV 15°×15°,耦出扩瞳100mm×60 mm的光栅波导一体化设计。此外,本文还对复色光入射光栅进行了探索,对单光栅中红光、绿光和蓝光的衍射效率与出射角度差异问题进行了分析。(3)针对传统光栅设计方法在处理复杂系统以及进行多参数优化时效率较低的问题,搭建神经网络并利用光栅数据集进行训练,得到了8种光栅的高精度神经网络模型。通过神经网络模型进行光栅性能预测,对8种光栅的结构进行了优化,经有限元法数值仿真验证,本文设计的混合光栅在0°～14°入射角范围内的平均衍射效率达94.21%,均匀度达94.94%,与目前的衍射光栅研究工作相比具有优势。证明了本文提出的机器学习设计方法能够高效率、高精度地实现光栅结构设计。(4)针对机器学习设计方法的软件固化问题,基于Tkinter模块以及Py Installer完成了神经网络模型编译器代码的部署,经过图形用户界面设计以及光栅性能预测功能完善,实现了光栅性能预测软件可行性与有效性的验证。本文提出的面向AR-HUD的人工智能研究方法,实现了高效率的光栅波导结构优化设计,通过软件工具的开发为研究成果的实际应用奠定基础,能够进一步为基于人工智能的光学设计提供理论指导与技术参考。更多还原

【Abstract】 With the advancement of vehicle intelligence,traditional dashboards are increasingly unable to meet the multidimensional human-machine interaction needs.As the third generation of vehicle head up display system,Augmented Reality Head Up Display(AR-HUD)can integrate information from the real and virtual world and is expected to become a future direction for vehicle human-machine interaction.In the optical combiner of AR-HUD,Surface Relief Grating(SRG)with its flexible design and compact structure has become a mainstream technological solution.This thesis conducts research on various grating structures aimed at SRG and has implemented an efficient grating performance prediction method based on machine learning.The specific research content of this thesis is as follows:(1)Faced with challenges such as the difficulty in improving the diffraction efficiency of gratings and the complexity of the grating design optimization process,physical models of five types of gratings including tilted grating,coated grating,interlayer grating,double-column grating,and U-shaped grating,which are currently widely used,were built using the Finite Element Method(FEM).Their diffraction efficiencies were studied,and by analyzing the mechanism by which the grating structure affects diffraction efficiency,a hybrid grating was designed,improving diffraction efficiency by over 40% compared to traditional tilted gratings.(2)Addressing the large field of view requirements of AR-HUD,azimuthal deflection of light within the waveguide was achieved through turning grating,accomplishing two-dimensional pupil expansion of the grating waveguide.Additionally,an integrated model of the grating waveguide consisting of in-coupling grating,turning grating,and out-coupling grating was constructed.Light intensity distribution was obtained through ray tracing,and the uniformity of the emergent light was optimized based on diffraction efficiency and the number of diffractions,realized an integrated design of grating waveguide with incident FOV of 15°×15° and coupled out pupil expansion of 100 mm×60 mm.Furthermore,the thesis explored polychromatic light incidence grating and analyzed the differences in diffraction efficiency and exit angles between red,green,and blue light in a single grating.(3)Addressing the inefficiency of traditional grating design methods in handling complex systems and multi-parameter optimization,a high-precision neural network model for eight types of gratings was obtained by constructing neural networks and training them with grating dataset.Grating performance predictions were made using the neural network model,and the structures of eight types of gratings were optimized.Verified through finite element numerical simulation,the hybrid grating designed in this thesis achieved an average diffraction efficiency of 94.21% and a uniformity of 94.94%within an incident angle range of 0° to 14°,demonstrating advantages over current diffraction grating research.The machine learning design method proposed in this thesis was proven to efficiently and accurately implement grating structure design.(4)Addressing the software solidification issues of machine learning design methods,the neural network model compiler code was deployed using the Tkinter module and Py Installer.After designing the graphical user interface and perfecting the grating performance prediction function,the feasibility and effectiveness of the grating performance prediction software were validated.The artificial intelligence research method for AR-HUD proposed in this thesis achieves efficient optimization design of the grating waveguide structure.Through the development of software tools,the foundation is laid for the practical application of research results,further providing theoretical guidance and technical reference for AIbased optical design.更多还原