基于深度压缩技术的Hybrid像素探测器读出系统原型机研制

The new Hybrid pixel detector not only features high position resolution, but also provides information of time and energy, which could be widely used in high energy physics experiments and other fields. Ultra-high collision rates of colliders will generate massive data. This project is conceived to be used to develop scalable readout system prototype machine, which takes and compresses data from pixel microchips, and transmits to computing center. Besides, we will lay emphasis on developing deep compression algorithm through neural networks; convolutional neural networks are used to compress information of position and energy, and Long Short-Term Memory (LSTM) networks are used to handle time information with relatively long delay. This project will also discuss how to shrink the parameter storage and computation consumption, “deeply compress” the compression algorithm itself and ease the implementation of algorithms in ASIC and FPGA chips.

新型Hybrid像素探测器不仅具备高精度位置分辨率，还可提供时间分辨和能量分辨信息，将广泛应用于高能物理实验等领域。在对撞机的超高碰撞率下，大量像素通道将产生海量数据。本项目拟研制用于Hybrid像素探测器的可扩展型读出系统原型机，读取并压缩像素芯片数据，传输至计算中心。重点研究基于神经网络的深度压缩算法，利用卷积神经网络的模式识别能力压缩位置和能量信息，利用LSTM长短期记忆网络处理时间序列中延迟相对较长的时间信息。本项目还将研究如何降低神经网络的参数存储和计算消耗，对深度压缩算法本身进行“深度压缩”，降低算法直接在ASIC或FPGA中实现的难度。

像素探测器具有位置分辨率高且兼具幅度分辨和时间分辨的特点，被广泛应用于高能物理实验。本项目完成了针对像素探测器脉冲信号特征提取的研究。首先我们使用深度神经网络进行脉冲信号的时间和幅度特征信号提取，并针对信号实时处理的需求设计完成了神经网络加速运算芯片，实现约1.9ms的单次推理延时。而后我们完成了对深度神经网络进行特征提取的固有局限性和应用潜力分析，并针对像素探测器的特点研究了低采样率、低量化精度、低计算精度下的神经网络性能。最后，我们完成了像素芯片Topmetal-M2的设计，研发完成的基于像素芯片的CXPD立方星原型机已经正样交付，预计于2023年发射用于空间X射线偏振探测。