电磁量能器脉冲信号的幅度与时间特征实时提取算法研究

In present electromagnetic calorimeters, ADC samples are transmitted from front-end electronics to the computing nodes. The computing nodes calculate the amplitude and arrival time of the pulse through fitting methods. To meet the requirements of increased number of the event rate and number of channels, the designers need to increase the bandwidth of the high-speed data links and the scale of the computing clusters. This project will study amplitude and timing feature extraction algorithms in real time based on Field Programming Gate Array (FPGA), and the real-time performance and resource utilization rate. In the FPGA, the filters such as Moving Average filter will be used to calibrate the baseline shift caused by perturbation; functions such as Weight Integrators will be used to calculate signal peak-amplitude precisely; methods such as multi-points Linear Interpolation and Constant-Fraction Discrimination will be used to measure the arrival time of the pulse precisely; processing methods such as Moving Window Deconvolution will be used to process the pile-up events. These algorithms move the feature extraction work from the computing nodes to the front-end electronics, only the pulse amplitude and timing information need be transferred from the front-end electronics to the computing nodes. In this way, the system real-time performance will be improved greatly, and the data will be compressed significantly. Therefore, this method can greatly ease the the highly dependancy on the high-speed data link of the system.

在现有电磁量能器中,大量ADC采样数据需要从前端电子学传输到计算机节点,然后再通过拟合算法提取信号的幅度与时间信息。当探测器的事件率和通道数目提高后,人们必须相应地提高数据链路的带宽和计算机集群的规模。本项目将研究基于可编程逻辑器件(FPGA)的实时幅度时间特征提取算法以及该算法在FPGA中实现后的时效性和资源利用率。该算法将在FPGA内利用滑动平均等滤波器对因微扰引起的基线偏移进行矫正;利用权重积分等函数精确计算信号的峰值幅度;利用多点线性内插法和恒比甄别等方法精确测量信号的到达时间;利用多级滑动窗口反卷积运算处理堆积脉冲。该算法将信号的特征提取工作从计算机节点转移到前端电子学中进行,将显著提高系统实时性,大幅度压缩数据量,减轻系统发展对高速数据链路和大规模计算机集群的高度依赖。

本项目的目标是研究脉冲信号特征提取算法。首先，研制了包括高精度高采样率ADC卡和kintex-7 FPGA控制板的实验室数据获取系统HPDAQ，用于开展不同脉冲信号特征提取算法的性能比较研究。其次，深入研究了传统的最小二乘曲线拟合算法在脉冲信号特征提取问题上的应用，构建了曲线拟合算法的一阶近似方程，定量分析了长期漂移、短期漂移和随机噪声，并进行了仿真研究。最后，通过曲线拟合与神经网络的对比研究，论证了该问题的深度学习算法潜力，提出了一种基于深度学习的时间信息提取算法，研究了针对该问题的神经网络结构。仿真结果表明，在非理想情况下，专用神经网络结构可以有效地抑制噪声，提高时间分辨率。最后用ALICE实验中电磁量能器前端电子学板和HPDAQ系统进行实验，在实验条件下，该神经网络的性能比曲线拟合好20%以上。