Hi there, I'm Qingyi Zhou!
An undergraduate student.
Department of Electrical Engineering
School of Electronics Engineering and Computer Science
Peking University
Email: zhouqingyi@pku.edu.cn
Contact: No. 5 Yiheyuan Road Beijing, 100871, People's Republic of China
Projects
Light detection and Ranging (LiDAR) system
Collaborators: Zhongwei Tan, Chuanchuan Yang.
In this project we analyze LiDAR systems on autonomous cars theoretically. We aim to provides guidance for choosing laser emitters and receivers when designing LiDAR systems. First we derive the Cramer-Rao lower bound of distance estimation. Based on that, we also calculate the minimum optical output power.
Two different ranging techniques, namely time-of-flight (ToF) and Quadrature Phase Detection (QPD) are compared in detail, revealing the reason why ToF is preferred when designing LiDAR systems. The influence of receiver (an APD) is also discussed.
Qingyi Zhou, Zhongwei Tan, Chuanchuan Yang*, “Theoretical Limit Evaluation of Ranging
Accuracy and Power for LiDAR Systems in Autonomous Cars”,
Opt. Eng., 57(9), pp. 096104 (2018)
Application of machine learning algorithms in Optical Communication
Collaborators: Anzhong Liang, Kang Hu, Xiaolong Zheng, Fukui Tian, Zhongwei Tan, Chuanchuan Yang, Fan Zhang, Feng Yang, Xin Qin.
In this project we intend to test multiple machine learning algorithms and see whether they can be utilized to improve the performance of optical communication systems.
Our platform contains an optical transmission system based on 850-nm VCSEL and OM4-MMF. VCSEL based short range optical interconnect links will continue to be the most widely deployed optical link in data center networks because of low power consumption, low cost and wider availability.
We've already tested several machine learning algorithms (including support vector machine and recurrent neural networks) by utilizing these algorithms to equalize and classify severely distorted optical signals.
We're currently testing algorithms based on probability graph models (PGM), including CRF, HMM. We're also trying to combine these algorithms with conventional equalization techniques (FFE, Volterra series, and MLSE).
Qingyi Zhou, Chuanchuan Yang*, Anzhong Liang, Xiaolong Zheng, Zhangyuan Chen,
“Low-Complexity Recurrent Neural Network For High Speed Optical Fiber Transmission”,
Opt. Commun., 441, pp.121-126 (2019)
Feature-map compression for high-speed video transmission
Collaborators: Xiaolong Zheng, Chuanchuan Yang, Feng Yang.
Compression.
Inverse design of Nano-photonic Devices
Collaborators: Jonathan A. Fan, Jiaqi Jiang.
Researchers have been using topology optimization (TO) to design irregularly shaped nanophotonic devices. In this project, our main goal is to accelerate the TO process for 2D/3D deflectors.
First we tested multiple optimizers, including vanilla gradient descent, momentum, Adam, AdaGrad and RMSprop. RMSprop is significantly better compared with other optimizers, since it produces good deflection efficiency in about 100 iterations, taking 60% less time compared with vanilla gradient descent.
Second, an adaptive meshgrid is implemented by using less pixels at the beginning stage of TO and gradually increase the number of pixels during the process. For TM wave deflectors, using adaptive meshgrid can save up to 50% time.
At last, we focus on 2D aperiodic deflectors, whose design involves both topology optimization and boundary optimization (BO). By adding an L-2 regularizer on variation of refractive index distribution, we've proved that we only need few BO iterations. L-1 regularizer (inspired by TV-based image processing), on the other hand, give poor results.
By accelerating TO, hopefully we can reduce the time it takes to build a data base, which contains deflectors of different shapes. On one hand, the data base will enable us to construct meta-lens with deflector units (different units at different radius). On the other hand, other researchers will gain access to these devices, which might promote development in this field.