Research Notes (8) Deep Learning and Its Application in WiFi Human Perception (Part 2)

论文题目：Deep Learning and Its Applications to WiFi Human Sensing: A Benchmark and A Tutorial

论文作者：Jianfei Yang, Xinyan Chen, Dazhuo Wang, Han Zou, Chris Xiaoxuan Lu, Sumei Sun, Fellow, IEEE and Lihua Xie, Fellow, IEEE

工作单位：新加坡南洋理工大学,University of Edinburgh, UK, etc

发表刊物：arxiv, pp. 1-17,2022

II. PRELIMINARIES OF WIFI SENSING

A. CSI简介

在WiFi通信中,CSI(信道状态信息)It reflects the diffraction of the wireless signal、Propagation in the physical environment after reflection and scattering,The channel characteristics of the communication link are described.For a pair of transmitter and receiver antennas,CSIMultipath phase shift and amplitude attenuation on each subcarrier are described.与RSS(接收信号强度)相比,CSIThe data has better sensing resolution,在WiFiIt can be regarded as the environment in which the signal propagates“WiFi图像”.在WiFi感知中,CSIThe recording function is provided by a dedicated tool[1]、[2]实现.CSIThe estimate can be expressed as :
$$H_i=\parallel H_i\parallel e^{j\angle H_i}$$
其中$\parallel H_i\parallel$和$\angle H_i$分别代表第$i$The magnitude and phase of the subcarriers.

B. CSI工具

CSI工具	带宽	子载波数量	设备
Intel 5300 CSI Tool	20MHz	30	Intel 5300 NIC
Atheros CSI Tool	20/40MHz	56/114	Atheros NIC
Nexmon CSI Tool	80MHz	256	智能手机&树莓派

C. CSIData transformation and cleaning

• Q:如何处理CSIData makes it applicableWiFi感知？
• A1:只使用CSIAmplitude data as input.
• A2:在基于模型的方法中,between the antennasCSIPhase difference as input.
• A3: Use processedCSIDoppler said,例如,提出BVP(Human coordinate velocity profile)to simulate the Doppler features that reflect human motion[31].

D.CSIThe impact of data on the human body

为了将CSIData is combined with deep learning models,This article summarizes what contributes to a better understanding of deep model designCSI数据属性:

• Subcarrier dimension—空间特征
• 时间维度（consecutive packets）----时间特征
• Antenna Dimensions—Resolution and channel characteristics

III. DEEP LEARNING MODELS FOR WIFI SENSING

A.用于WiFiPerceptual deep learning models（2017-2022）

B.WiFiFundamentals of commonly used deep learning models in the field of perception

C.WiFiEvaluation of commonly used deep learning models in the field of perception

• MLP：参数多,收敛速度慢,计算开销大;
• CNN：卷积层过多（>20）,The vanishing gradient problem leads to performance degradation;
• RNN：There is a vanishing gradient problem when backpropagating,无法捕获CSI的长期依赖关系;
• LSTM：克服了Vanilla RNN存在的问题;
• Transformer：参数多,训练成本高,And it is difficult to collect a large number of labeled onesCSI数据.

IV. LEARNING METHODS FOR DEEP WIFI SENSING MODELS

1. 监督学习：Traditional deep model training relies on supervised learning with large amounts of labeled data,But data collection and labeling is a realityWiFiPerceive application bottlenecks.
2. 少样本学习：Since only a small number of samples are required,因此在实际应用中,Few-shot learning helps based on WiFigesture recognition and human body recognition.
3. 迁移学习：在WiFiperceive the scene,由于CSIData is highly dependent on the environment,Therefore, cross-domain scenarios are very common.Transfer learning is used to solve cross-domain problems.
4. 无监督学习：The model can be strengthened,获得更好的泛化能力.
5. 集成学习：Can improve model classification performance,But the computational overhead also grows exponentially.

V. EMPIRICAL STUDIES OF DEEP LEARNING IN WIFI SENSING: A BENCHMARK

A. 实验设置

B. Deep Model Evaluation

C. Learning strategy assessment

VI. DISCUSSIONS AND SUMMARY

PS：Since the paper was published in axvir上的,There are inevitably some omissions in the text.比如图2The three human activities mentioned in the subheading,The actual image involves four activities,And the image source of the upper and lower lines is not given a detailed introduction.