《Over the Air Deep Learning Based Radio Signal Classification》思維導圖

制作了論文的思維導圖，用于學習和交流

2018-Over the Air Deep Learning Based Radio Signal Classification.jpg

Over the air deep learning based radio signal classification

1. 論文針對問題

1.1. 利用CV領域的深度學習網(wǎng)絡+OTA借杰、合成數(shù)據(jù)實現(xiàn)調(diào)制信號的分類識別

2. 問題的解決方法

2.1. 思路

利用最新深度學習的深度網(wǎng)絡resnet來訓練模型搁拙，進行調(diào)制信號的識別

2.2. 數(shù)據(jù)集

2.2.1. 標簽

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

數(shù)據(jù) 特征：1024樣本統(tǒng)計特征

數(shù)據(jù) 降維：1024?2維→28維

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中的運用地非常好

模型： VGG

濾波器：最小 3x3

池化層：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

優(yōu)勢：無需手動提取特征（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：更深的網(wǎng)絡模型+殘差網(wǎng)絡特性奔脐，提供更好的性能

模型： ResNet

residual unit

stack of residual units

調(diào)整：全連接層激活函數(shù) the scaled exponential linear unit (SELU)，a slight improvement over conventional ReLU performance，not ReLU

對比：

RestNet：236,344 參數(shù)
CNN/VGG：257,099 參數(shù)

3. SENSING PERFORMANCE ANALYSIS

A. Classification on Low Order Modulations

高SNR： VGG/CNN和ResNet差不多滞谢，ResNet 比 baseline 獲得大概 5dB 優(yōu)勢
ResNet: 99.8%, VGG: 98.3%, Baseline: 94.6%

B. Classification under AWGN conditions

數(shù)據(jù)： 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 可以讓所有調(diào)制類型都達到80%以上的正確率

融合矩陣中error最大的調(diào)制類型

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