【作者】 赵洋；

【导师】 王剑；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2022， 硕士

【摘要】 无线通信技术的高速发展正使万物互联成为可能。在越来越多终端设备接入无线通信网络的同时,无线电频谱资源也越来越紧张。由于无线电波的开放性,无线通信网络对干扰相对更加敏感。为了更有效地利用有限的无线频谱资源,管理部门引入了认知无线电（Cognitive Radio,CR）技术来动态地共享紧张的频谱资源,但这也给频谱资源的非法盗用者提供了可乘之机。为了避免无线频谱资源被非法利用,需要对其使用情况进行监测,对给定传输频带内的无线电信号的合法性进行确认,因此无线电信号识别技术成为近年来研究的重点之一。由于无线电信号识别不仅仅是简单地判断传输信道内是否有通信信号存在,而且要对通信信号的基本特征进行分析和确认,传统的基于信号统计特征分析的无线电信号识别方法通常难以达到性能要求。随着深度学习,神经网络尤其是面向图像识别的卷积神经网络的发展,借鉴卷积神经网络在目标识别领域的优秀成果,提供端到端的无线电信号智能识别算法成为主流的研究方向。矛盾的是,卷积神经网络结构复杂,运算量极大,边缘终端设备的算力往往不足以部署完整的卷积神经网络,因此需要对网络进行轻量化设计。传统的卷积神经网络轻量化设计方法主要包括探索新的网络运算架构,或在现有运算架构的缩减和性能劣化之间找到折中。面对这样的问题,本文从一种区别于二进制编码系统的数值表征方式——随机计算系统出发,从系统功耗、运算延迟和系统资源消耗的角度,设计并优化了一种基于随机计算的无线电信号识别神经网络,并最终将其部署到FPGA测试平台上完成验证。本文的主要工作如下:·在随机计算原理方面,在对确定性随机计算原理分析的基础上,实现了确定性随机计算序列构造和高精度加法器。在计算误差明显优于传统随机计算的同时,将随机计算的计算延迟从2²ⁱ²缩减到2ⁱ;·在无线电信号识别网络算法优化方面,本文应用了一种自下而上的网络结构搜索算法。在此基础上,对搜索得到的卷积神经网络进行剪枝和量化以压缩其运算量和存储量。在不影响识别率的情况下,对网络整体的压缩比约为5%左右;·在无线电信号识别网络硬件实现方面,本文设计了基于确定性随机计算原理的乘加并行结构与稀疏的网络存储结构;·完成了软硬件系统的设计与搭建,在Vivado开发环境中完成了硬件代码的编写,在Modelsim平台上完成了硬件代码全功能的仿真测试,最终在VC707FPGA测试平台上实现了整个系统,完成了系统硬件性能分析,在卷积神经网络识别率为87.5%的条件下,网络整体计算量压缩至5%左右,系统整体功耗为2.45W左右。更多还原

【Abstract】 The rapid development of wireless communication technology is making the inter-connection of everything possible.While more and more terminal devices are connected to wireless communication networks,radio spectrum resources are getting tighter and tighter.Due to the open nature of radio waves,wireless communication networks are relatively more sensitive to interference.In order to use the limited wireless spectrum resources more effectively,the management has introduced Cognitive Radio（CR）technology to dynamically share the tight spectrum resources,but this also provides opportunities for illegal users of spectrum resources.In order to avoid illegal use of radio spectrum re-sources,it is necessary to monitor its usage and confirm the legitimacy of radio signals in a given transmission band,so radio signal identification technology has become one of the research focuses in recent years.Since radio signal recognition is not only simply to determine whether a communication signal exists in the transmission channel,but also to analyze and confirm the basic characteristics of the communication signal,traditional radio signal recognition methods based on statistical signal feature analysis are usually difficult to achieve the performance requirements.With the development of deep learn-ing,neural networks,especially convolutional neural networks for image recognition,it has become a mainstream research direction to provide end-to-end radio signal intelligent recognition algorithms by drawing on the excellent results of convolutional neural net-works in the field of target recognition.Paradoxically,convolutional neural networks are complex in structure and extremely large in computation,and the computing power of edge terminal devices is often insufficient to deploy complete convolutional neural networks,so lightweight design of the network is required.Traditional approaches to lightweight design of convolutional neural networks mainly include exploring new network comput-ing architectures or finding a compromise between the reduction and performance degra-dation of existing computing architectures.Facing such a problem,this paper designs and optimizes a stochastic computation-based neural network for radio signal recognition from the perspective of system power consumption,computational latency and system re-source consumption,starting from a numerical representation that is different from the bi-nary coding system,the stochastic computation system,and finally deploys it to an FPGA testbed The verification is completed.The main work of this paper is as follows.·In the aspect of random calculation principle,based on the analysis of the principle of deterministic random calculation,a deterministic random calculation sequence structure and a high-precision adder are realized.Reduced from 2²ⁱ²to 2ⁱ.·In the aspect of network algorithm optimization for radio signal recognition,a bottom-up network structure search algorithm is applied in this paper.Based on that,the convolutional neural network obtained by the algorithm is pruned and quantized to compress its computation and storage.Without affecting the recognition rate,the overall compression ratio of the network is about 5%.·In terms of hardware implementation of radio signal recognition network,this paper designs a multiplication-add parallel structure and a sparse network storage structure based on the principle of deterministic random computing.·Completed the design and construction of the software and hardware system,the hardware code was written in Vivado development environment,the hardware code was simulated and tested on the Modelsim platform,and the system was finally implemented on the VC707 FPGA test platform.Under the condition that the recog-nition rate of convolutional neural network is 87.5%,the overall computation of the network is compressed to about 5%,and the overall power consumption of the system is about 2.45W.更多还原