摘要
摘要
信号识别是信息科学领域中的一个重要子领域，现代电子信息系统的发展，比如说通信、雷达等，都离不开信号处理技术的发展。各种各样的通信电子设备随着科学的不断进步以及通信技术的飞速发展被大量的使用。
深度学习是人工智能领域热潮中最火爆的一个分支，在很多应用场景上都取得了比传统方法更好的效果。因此利用深度学习进行信号识别也有了可行性，近些年来，越来越多的深度卷积神经网络被应用在信号识别领域中。尽管深度学习有很多传统方法无法比拟的优点，但是大而深的深度网络模型内存和运算量都很大，而对于嵌入式设备以及一些小型的移动设备来说无法应用。所以，在信号识别卷积神经网络性能损失合理的前提下，模型轻量化是一个很重要的研究方向。本文针对应用于信号识别的卷积神经网络模型轻量化展开研究，主要工作如下：（1）对应用于信号识别的卷积神经网络结构进行分析，将卷积神经网络中卷积层的标准卷积改为深度可分离卷积。该方法把普通卷积分解成两部分，第一部分为深度卷积，第二部分为点卷积，通过两部分的组合，从而达到对网络的轻量化作用。通过这种方法，可以有效的减小模型的大小，加快模型的前向推理速度。
（2）针对信号识别卷积神经网络卷积层中卷积核冗余导致模型太大并且前向推理时间过多的问题，本文
提出一种基于区间的卷积核剪枝方法。该方法通过提取训练好的模型的卷积层的权重参数，将卷积核的权重区间分为几个等分区间，设置每个区间的剪枝系数，对每个区间的卷积核进行剪枝，然后对模型重新训练。与传统的剪枝方法相比，该方法在损失同样精度的情况下，能够达到更大的轻量化比例。
（3）针对信号识别卷积神经网络中全连接层参数过于冗余的情况，采用奇异值分解的方法对全连接层进行轻量化。奇异值分解轻量化方法主要通过将一个大的矩阵分解成几个小矩阵相乘的形式，从而减少网络的参数量，在模型性能损失比较小的情况下达到对模型的轻量化目的。
（4）针对不同类型的信号识别卷积神经网络，合理采用上述的方法进行组合，在不同的数据集上进行详细的实验对比，在模型性能损失可以接受的情况下达到比单独的轻量化算法更好的效果。
关键词：信号识别，深度学习，模型轻量化，卷积神经网络
ABSTRACT
Signal recognition is an important sub field in the field of information science. The development of modern electronic information system, such as communication and radar, is inseparable from the development of signal processing technology. With the continuous progress of science and the rapid
development of communication technology, various kinds of communication electronic equipment are widely used.
Deep learning is one of the most popular branches in the field of artificial intelligence. It has achieved better results than traditional methods in many application scenarios. In recent years, more and more deep convolution neural networks are used in the field of signal recognition. Although deep learning has many advantages that traditional methods can't compare with, but the large and deep deep network model memory and computation are very large, and it can't be applied to embedded devices and some small mobile devices. Therefore, under the premise of reasonable performance loss of convolutional neural network, model lightweight is a very important research direction. The main work of this paper is as follows:
(1) The convolution neural network structure for signal recognition is analyzed. The convolution of convolution layer in convolution neural network is changed into deep separable convolution. In this method, the general convolution is divided into two parts, the first part is the depth convolution, the second part is the point convolution. Through the combination of the two parts, the compression of the network is achieved. This method can effectively reduce the size of the model and accelerate the forward reasoning speed of the model.
(2) Aiming at the problem that the convolution kernel redundancy in convolution neural network convolution layer of signal recognition results in too large model and too much forward reasoning time, this paper proposes an interval based convolution kernel pruning method. By extracting the weight parameters of the convolution layer of the trained model, the weight interval of the convolution kernel is divided into several equal intervals, and the pruning coefficient of each interval is set to prune the convolution kernel of the convolution layer, and then the model is retrained. Compared with the traditional method, this method can achieve a larger compression ratio when the same accuracy is lost.
(3) In this paper, singular value decomposition (SVD) is used to reduce the weight of all connection layer in convolutional neural network. SVD can reduce the number of matrix calculation by decomposing a large matrix into several small multiplication forms, and achieve the goal of model lightweight when the model performance loss is acceptable.
(4) For different types of signal recognition convolution neural network, reasonable use of the above methods for combination, in different data sets for detailed experimental comparison, in the case of acceptable model performance loss to achieve better results than a separate lightweight algorithm.
Keywords: Signal Recognition, Deep Learning, Model Lightweight, Convolutional Neural Network
目录
第一章绪论 (1)
1.1 研究背景及意义 (1)
1.2 模型轻量化方法研究现状 (2)
1.2.1 低秩分解 (2)未识别的网络
1.2.2 网络剪枝与模型量化 (3)
1.2.3 知识蒸馏 (4)
1.2.4 网络结构设计 (4)
1.3 本文的技术方案 (5)
1.4 本文的行文结构 (6)
第二章深度学习以及模型轻量化方法综述 (7)
2.1 深度学习基本概念 (7)
2.2 深度学习基本模块 (8)
2.2.1 卷积层 (8)
2.2.2 池化层 (9)
2.2.3 Dropout层 (10)
2.2.4 激活函数层 (10)
2.2.5 全连接层 (12)
2.2.6 BN层 (13)
2.3 卷积神经网络常见结构 (14)
2.3.1 LeNet5网络结构 (14)
2.3.2 AlexNet网络结构 (15)
2.3.3 VGGNet网络结构 (17)
2.3.4 GoogleNet网络结构 (17)
2.3.5 ResNet网络结构 (19)
2.4 模型轻量化方法概述 (20)
2.4.1 核的稀疏化 (21)
2.4.2 模型裁剪 (22)
2.4.3 知识精炼 (22)
2.4.4 量化 (24)
2.5 本章小结 (25)
第三章基于深度可分离卷积的信号识别模型轻量化 (26)
3.1 信号调制类型 (26)
3.2 基于一维神经网络的调制识别 (27)
3.2.1 模型设置 (27)
3.2.2 优化方法 (28)
3.2.3 损失函数 (29)
3.2.4 训练数据 (29)
3.2.5 实验与分析 (30)
3.3 基于二维神经网络的调制识别 (31)
3.3.1 模型设置 (31)
3.3.2 训练数据 (32)
3.3.3 实验与分析 (33)
3.4 深度可分离卷积轻量化方法综述 (33)
3.5 深度可分离卷积实验分析 (35)
3.6 不同调制类型的对比实验分析 (40)
3.6.1 FSK、MPSK实验 (40)
3.6.2 FSK、MPSK、APSK实验 (40)
3.7 本章小结 (41)
第四章基于区间剪枝的信号识别模型轻量化 (42)
4.1 基于区间剪枝的轻量化方法综述 (42)
4.2 区间剪枝实验分析 (43)
4.3 不同调制类型的对比实验分析 (47)
4.3.1 FSK、MPSK实验 (47)
4.3.2 FSK、MPSK、APSK实验 (47)
4.4 本章小结 (48)
第五章基于奇异值分解的信号识别模型轻量化 (49)
5.1 基于奇异值分解的轻量化方法综述 (49)
5.2 奇异值分解实验分析 (51)
5.3 三种算法的组合实验 (53)
5.3.1 Sig_model组合实验分析 (53)
5.3.2 Sig_VGG16组合实验分析 (55)
5.4 本章小结 (56)
第六章总结与展望 (57)