制作了论文的思维导图，用于学习和交流
Over the air deep learning based radio signal classification
1. 论文针对问题
1.1. 利用CV领域的深度学习网络+OTA、合成数据实现调制信号的分类识别
2. 问题的解决方法
2.1. 思路
- 利用最新深度学习的深度网络resnet来训练模型，进行调制信号的识别
2.2. 数据集
2.2.1. 标签
- Normal Classes
OOK, 4ASK, BPSK, QPSK, 8PSK, 16QAM, AM-SSB-SC, AM-DSB-SC, FM, GMSK, OQPSK
- Difficult Classes
OOK, 4ASK, 8ASK, BPSK, QPSK, 8PSK, 16PSK, 32PSK, 16APSK, 32APSK, 64APSK, 128APSK, 16QAM, 32QAM, 64QAM, 128QAM, 256QAM, AM-SSB-WC, AM-SSB-SC, AM-DSB-WC, AM-DSB-SC, FM, GMSK, OQPSK
2.2.2. 生成方法
several simulated wireless channels generated from the model
over the air (OTA) transmission channel of clean signals with no synthetic channel impairments
channel initialization of variables
2.3. 模型
Baseline Method
- 思路
leverages the list of higher order moments and other aggregate signal behavior statistics given in table
- 方法
2.1. baseline模型 ：XGBoost
数据 特征：1024样本统计特征
数据 降维 ：1024∗2维→28维
效果：outperforms a single decision tree or support vector machine (SVM) significantly on the task
2.2. Convolutional Neural Network
思路：CNN在CV中的运用地非常好
模型： VGG
滤波器：最小 3x3
池化层：最小 2x2
效果： This represents a simple DL CNN design approach which can be readily trained and deployed to effectively accomplish many small radio signal classification tasks
优势： 无需手动提取特征（do not perform any expert feature extraction or other pre-processing on the raw radio signal , instead allowing the network to learn raw time-series features directly on the high dimension data）
2.3. Residual Neural Network
思路： 更深的网络模型+残差网络特性，提供更好的性能
模型： ResNet
residual unit
stack of residual units
调整：全连接层激活函数 the scaled exponential linear unit (SELU)，a slight improvement over conventional ReLU performance，not ReLU
对比：
RestNet：236,344 参数
CNN/VGG：257,099 参数
3. SENSING PERFORMANCE ANALYSIS
A. Classification on Low Order Modulations
高SNR： VGG/CNN和ResNet差不多，ResNet 比 baseline 获得大概 5dB 优势
ResNet: 99.8%, VGG: 98.3%, Baseline: 94.6%
B. Classification under AWGN conditions
数据： N = 239, 616 examples
模型： L = 6 residual stacks
效果： the best performance at both high and low SNRs on the difficult dataset by a margin of 2-6 dB in improved sensitivity for equivalent classification accuracy.
C. Classification under Impairments
效果：
- ResNet
ResNet performance improves under LO offset rather than degrading.
At high SNR performance ranges from around 80% in the best case down to about 59% in the worst case.
- Baseline
in the best case at high SNR this method obtains about 61% accuracy while in the worst case it degrades to around 45% accuracy
D. Classifier performance by depth
L=5：121 layers, 229k trainable parameters
L = 0： 25 layers and 2.1M trainable parameters
E. Classification performance by modulation type
10dB SNR 可以让所有调制类型都达到80%以上的正确率
融合矩阵中error最大的调制类型
high order phase shift keying (PSK) (16/32-PSK)
high order quadrature amplitude modulation (QAM) (64/128/256-QAM)
AM modes (confusing with-carrier (WC) and suppressed-carrier (SC)
high order QAM and PSK can be extremely difficult to tell apart through any approach
F. Classifier Training Size Requirements
样本数量：
4-8k样本： 模型正确率为随机
100M样本左右： 提升5-20%
200万的所有数据训练： 单个 NVIDIA V100 GPU （125Tera-FLOPS) 需要花费大约16小时
从100M提升到200M：并没有看到非常大的性能提升；在高SNR时，正确率大约都为95%
单个样本序列长度
G. Over the air performance
利用USRP装置，生成了1.44M的数据集
利用NVIDIA V100 GPU训练了大约14小时
所有样本在SNR为10dB，测试集正确率为95.6%
H. Transfer learning over-the-air performance
思路： 利用迁移学习把训练好的模型做资源整合利用；freeze 网络参数， 再次训练时只更新租后的3层全连接层
方法： 利用ResNet在1.2M合成数据中训练，然后在OTA样本中测试
no fine-tuning：24类中，正确率为64% - 80%
fine-tuning：24类中，正确率在84%-96%
