【ISAC】Task-Oriented Sensing, Computation, and Communication Integration for Multi-Device Edge AI

[1]-D. Wen et al., “Task-Oriented Sensing, Computation, and Communication Integration for Multi-Device Edge AI,” in IEEE Transactions on Wireless Communications, vol. 23, no. 3, pp. 2486-2502, March 2024, doi: 10.1109/TWC.2023.3303232.
[2]-D. Wen et al., “Task-Oriented Sensing, Computation, and Communication Integration for Multi-Device Edge AI,” ICC 2023 - IEEE International Conference on Communications, Rome, Italy, 2023, pp. 3608-3613, doi: 10.1109/ICC45041.2023.10279277.
[3]-P. Liu, G. Zhu, W. Jiang, W. Luo, J. Xu and S. Cui, “Vertical Federated Edge Learning With Distributed Integrated Sensing and Communication,” in IEEE Communications Letters, vol. 26, no. 9, pp. 2091-2095, Sept. 2022, doi: 10.1109/LCOMM.2022.3181612.
[4]-P. Liu et al., “Toward Ambient Intelligence: Federated Edge Learning With Task-Oriented Sensing, Computation, and Communication Integration,” in IEEE Journal of Selected Topics in Signal Processing, vol. 17, no. 1, pp. 158-172, Jan. 2023, doi: 10.1109/JSTSP.2022.3226836.
[5]-D. Wen, X. Jiao, P. Liu, G. Zhu, Y. Shi and K. Huang, “Task-Oriented Over-the-Air Computation for Multi-Device Edge AI,” in IEEE Transactions on Wireless Communications, vol. 23, no. 3, pp. 2039-2053, March 2024, doi: 10.1109/TWC.2023.3294703.
[6]-G. Li, S. Wang, J. Li, R. Wang, X. Peng and T. X. Han, “Wireless Sensing With Deep Spectrogram Network and Primitive Based Autoregressive Hybrid Channel Model,” 2021 IEEE 22nd International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), Lucca, Italy, 2021, pp. 481-485, doi: 10.1109/SPAWC51858.2021.9593198.

---

[7]-Shao J, Zhang J. Communication-computation trade-off in resource-constrained edge inference[J]. IEEE Communications Magazine, 2020, 58(12): 20-26.
[8]-F. Xi, N. Shlezinger and Y. C. Eldar, “BiLiMO: Bit-Limited MIMO Radar via Task-Based Quantization,” in IEEE Transactions on Signal Processing, vol. 69, pp. 6267-6282, 2021.
[9]-Q. Lan et al., “What is Semantic Communication? A View on Conveying Meaning in the Era of Machine Intelligence,” in Journal of Communications and Information Networks, vol. 6, no. 4, pp. 336-371, Dec. 2021.
[10]-Lan Q, Zeng Q, Popovski P, et al. Progressive feature transmission for split inference at the wireless edge[J]. arXiv preprint arXiv:2112.07244, 2021.

利用分布式ISAC为了实现多目标协作的人体姿势识别.
交换中间计算矢量而非原始的感知数据（交换联邦学习的权重而非数据）.
使用frequency-modulated continuous-wave (FMCW) signals实现ISAC.

1-Intro

Split inference（分割推理）的概念见附录，e.g.【7】设备上的模型计算成本 vs 中间特征转发到边缘服务器的通信开销.
边缘设备只提取特征，将特征给服务器，避免数据的交互，所以保护了隐私.
正是因为传输特征而非数据，本来硬件资源有限的边缘设备通信开销也减轻了.
同时繁重的推理是复杂的，这样就一下子将算力复杂的任务转移到边缘服务器端.

Split inference（分割推理）的步骤：
数据获取 $\Rightarrow$ 特征提取和量化 $\Rightarrow$ 传输
感 $\Rightarrow$ 算 $\Rightarrow$ 通
$\Downarrow$ $\Downarrow$ $\Downarrow$
协同优化，最终的性能和上面三步息息相关. 通信和感知是一对trade-off，通信和计算也是一对trade-off , 通信带宽和量化提取的特征的等级.
exploit the fact that analog signals are commonly acquired not to be reconstructed, but in order to extract
some lower-dimensional information from them 【8】. 为什么要Task-Oriented，低维 vs. 重构.

feature offload，offload这个词用的很多，【9】揭示offload的内涵就是将昂贵的计算和大量的存储甩锅给边缘服务器，本质就是一个数据上传的通信过程.
coupling，通感算的耦合.

2-系统模型

2.1-网络模型

a mobile edge server with a single-antenna access point (AP)
dual functional radarand-communication (DFRC)
edge server和K个ISAC设备都是单天线，是否太理想？
时分复用.
$T$ $\Rightarrow$ 完成推理任务的总时间.
$T_{r,k}$ $\Rightarrow$ 第$k$台设备感知的时间.
$T_{c,k}$ $\Rightarrow$ 第$k$台设备计算的时间. 常数.
$B$ $\Rightarrow$ 总带宽.
静态无线信道 $\Rightarrow$ $T$小于信道的相干时间. 其实就是准静止信道，信道在相干时间内的特性保持不变（恒定静态）.
$H_{c,k}$ $\Rightarrow$ 第$k$台设备和边缘服务器之间的信道增益.
AP可以获取全局CSI.

2.2-雷达感知与特征计算

2.2.1-感知信号

感知信号是 $\Rightarrow$ 线性频率上升（linear frequency up-ramp）线性调频（chirp）. 其频率随时间线性增加. 频率随时间变化，因此它没有一个固定的周期，但我们可以分析其瞬时频率.
$s(t)=\cos \left(2\pi \left(1000t+\frac{2000t^2}{2}\right)\right),0\le t\le 2$
$s^{\prime }(t)=-2000\pi (1+2t)\sin \left(2\pi \left(1000t+1000t^2\right)\right)$

• 初始频率 $f_0=1$ kHz
• 频率变化率 ( k = 2 ) kHz/s
• 振幅 ( A = 1 )
• 持续时间 ( T = 2 ) 秒
  为了便于展示，频率下降100倍
  
  回波信号the echo signal $\Rightarrow$ $r_k(t)=u_k(t)+{\sum }_{j=1}^J{v}_{k,j}(t)+n_r(t)$
  理想的回波信号desired $\Rightarrow$ $u_k(t)=H_{r,k}(t)s_k(t-\tau )$，$H_{r,k}(t)$是反射系数，往返路劲损耗
  同一个目标从其他第$j$个反射径返回的信号 $\Rightarrow$ $v_{k,j}(t)=C_{r,k,j}(t)s_k(t-{\tau }_j)$, 同样由反射系数和时延构成.

2.2.2-感知信号的处理

每台ISAC设备接收的回波采样成复数矢量.
奇异值分解减轻杂波，提取有用信息.
第$k$台ISAC设备获取的第$n_k$个特征元素 $\Rightarrow$ ${\bar{r}}_k(n_k)={\bar{u}}_k(n_k)+{\sum }_{j=1}^J{\bar{v}}_{k,j}(n_k)+{\bar{n}}_r(n_k)$. ${\bar{u}}_k(n_k)$是理想的真实的特征(标签)，其实和上述回波信号是对应的.
第$k$台ISAC设备获取的第$n_k$个特征元素,雷达发射功率归一化 $\Rightarrow$ $\hat{x}(n_k)=\frac{r_k(n)}{\sqrt{P_{r,k}}}=x(n_k)+c_{r,k}(n_k)+\frac{n_r(n_k)}{\sqrt{P_{r,k}}}$
根据中心极限定理，$c_{r,k}(n_k)$服从高斯分布 (反射径$J$足够多)， $c_{r,k}(n_k)\sim \mathcal{N}(0,{\sigma }_{c,k}^2)$.
此外，$n_r(n_k)/\sqrt{P_{r,k}}\sim \mathcal{N}\left(0,{\sigma }_r^2/P_{r,k}\right)$
中心极限定理：无论原始数据的分布是什么，当样本容量足够大时，样本均值的分布将趋近于正态分布.

第$k$台ISAC设备生成的特征子集 $\Rightarrow$ x ^ k = { x ^ ( n k ) , 1 ≤ n k ≤ N k } \hat{\mathbf{x}}_k=\{\hat{x}(n_k), 1 \leq n_k \leq N_k\} x^k​={ x^(nk​),1≤nk​≤Nk​}. $N_k$是总的特征数量. 假设不同ISAC设备生成的特征子集是独立的 (因为ISAC设备部署的很稀疏，并且各自的感知区域是不重叠的).

2.2.3-量化模型

对${\hat{\mathbf{x}}}_k$使用线性量化器量化，比如对${\hat{\mathbf{x}}}_k$中第$n_k$个元素量化，量化后是$\Rightarrow$$z(n_k)=\sqrt{Q_k}\hat{x}(n_k)+d_k$，由量化增益$\sqrt{Q_k}$和量化噪声$d_k\sim \mathcal{N}(0,{\delta }_k^2)$组成.

接收端需要对量化的版本进行恢复 $\Rightarrow$ $\tilde{x}(n_k)=z(n_k)/\sqrt{Q_k}=\hat{x}(n_k)+d_k/\sqrt{Q_k}$. 量化增益$\sqrt{Q_k}$越大，量化噪声相对就小些
每一个$\hat{x}(n_k)$都要量化编码传输，所以前面有$N_k$ $\Rightarrow$ $I({\tilde{\mathbf{x}}}_k;{\hat{\mathbf{x}}}_k)=N_k{\log }_2(1+Q_k/{\delta }_k^2),\forall k$

2.2.4-判别增益 Discriminant Gain

symmetric divergence

x ^ ( n k ) = r k ( n ) P r , k = x ( n k ) + c r , k ( n k ) + n r