基于卷积神经网络的量能器触发信号自动产生方法研究

Compared to TPC gas detector, particles pass through the calorimeter very quickly. Calorimeter can provide fast photon and jet trigger for global system. The traditional algorithms such as sliding window algorithm, Grid Cluster and Region of Interest trigger algorithm, identify the signal and background events by comparing the window energy and trigger threshold. These algorithms use "safe" trigger threshold to achieve high efficiency. However, it also considers some high-energy background events as signal and increase false trigger rate. This project intends to use the most advanced industry field deep image recognition learning technology, study new calorimeter trigger algorithm based on convolutional neural network. In this algorithm, high precise raw data input to the data layer of convolutional neural network, through automatic denoising, convolution, linear correction, normalized layer, pool layer and softmax hidden layer prior to the conduction process, calculate the probability of photon and jet, and generate trigger signal. According to the characteristics of the calorimeter data, this project will study triggering algorithm framework, including network depth, feature layer depth, convolution kernel size, and reverse conduction network structure, and optimises trigger performance, reduce false trigger rate.

相对于TPC等气体探测器，粒子在量能器中停留的时间极短，量能器可为全局系统提供快速光子和喷注触发。传统的滑动窗算法、Grid Cluster、Region of Interest等触发算法通过比较窗口能量与触发门限来甄别信号和背景事件，并利用“安全”的触发门限获得较高的触发效率。然而，随之被误判为信号的背景事件也越多，伪触发率也越高。本项目拟借助工业界图像识别领域最先进的深度学习技术，研究基于卷积神经网络的新型量能器触发算法。在该算法中，高精度原始数据连接到卷积神经网络的数据层，经过自动去噪层、卷积层、线性矫正层、归一化层、池化层和softmax等隐层的前向传导处理，计算出该事件为光子与喷注的概率，并以此产生触发信号。根据量能器的数据特点，设计触发算法的构架，包括网络深度、特征层深度、卷积核大小和反向传导网络结构，优化触发性能，降低伪触发率。

本项目的目标是研究基于卷积神经网络的量能器触发信号自动产生方法。首先，研究了将量能器触发问题转化为对背景事件和目标事件的二分类问题的约束条件，根据不同的物理实验目标，定义了分类器的全召回率模式（稀有事件）和最优阈值判定模式（大概率事件）和对应的性能指标量化标准。其次，研究了基于传统支持向量机的经典二分类方法，揭示了该方法在空间关联能力上的不足；研究了基于多尺度卷积核的分类器对具有大动态范围的量能器准高斯脉冲信号采样数据的敏感性，利用大尺度卷积捕获大信号，利用小尺度卷积捕获弱信号。然后，研究了神经网络的网络架构轻量化方法，利用深度可分离卷积加速卷积运算，通过通道混洗弥补分组卷积导致的空间弱关联问题。最后，设计和构建了低中高三种不同层级的FPGA电子学研制平台，用于对分类触发算法的加速研究和性能验证。通过实验发现，改进后的分类算法在100%召回率下，伪触发率比原分类算法降低了6%。