感知噪声分布 e k ∼ N ( 0 , ϵ k 2 I ) \mathrm{e}_k\sim\mathcal{N}\left(0,\epsilon_k^2\mathrm{I}\right) ek∼N(0,ϵk2I)
d k \mathrm{d}_k dk
感知噪声经过PCA d k ∼ N ( 0 , ϵ k 2 I ) \mathrm{d}_k\sim\mathcal{N}\left(0,\epsilon_k^2\mathbf{I}\right) dk∼N(0,ϵk2I)
F ℓ ( x ) \mathcal{F}_\ell(\mathbf{x}) Fℓ(x)
第 ℓ \ell ℓ类别特征的分布 F ℓ ( x ) = N ( μ ℓ , Σ ) , 1 ≤ ℓ ≤ L \mathcal{F}_\ell(\mathbf{x})=\mathcal{N}\left(\boldsymbol{\mu}_\ell,\boldsymbol{\Sigma}\right),\quad1\leq\ell\leq L Fℓ(x)=N(μℓ,Σ),1≤ℓ≤L
n \mathbf{n} n
OAC过程的AWGN n ∼ N ( 0 , δ 0 2 I ) \mathbf{n}\sim\mathcal{N}\left(0,\delta_0^2\mathbf{I}\right) n∼N(0,δ02I)
x m i x_{m_i} xmi
每个特征元素的分布 x m i ∼ 1 L ∑ ℓ = 1 L N ( μ ℓ , m i , σ m i 2 ) x_{m_i}\sim\frac1L\sum_{\ell=1}^L\mathcal{N}\left(\mu_{\ell,m_i},\sigma_{m_i}^2\right) xmi∼L1∑ℓ=1LN(μℓ,mi,σmi2)
d k , m i d_{k,m_i} dk,mi
每个噪声元素的分布 d k , m i ∼ N ( 0 , ϵ k 2 ) d_{k,m_i}\sim\mathcal{N}\left(0,\epsilon_k^2\right) dk,mi∼N(0,ϵk2)
x k , m i x_{k,m_i} xk,mi
本地元素 x k , m i ∼ 1 L ∑ ℓ = 1 L N ( μ ℓ , m i , σ m i 2 + ϵ k 2 ) x_{k,m_i}\sim\frac1L\sum_{\ell=1}^L\mathcal{N}\left(\mu_{\ell,m_i},\sigma_{m_i}^2+\epsilon_k^2\right) xk,mi∼L1∑ℓ=1LN(μℓ,mi,σmi2+ϵk2)
x ^ m i \hat{x}_{m_i} x^mi
接收元素分布 x ^ m i ∼ 1 L N ( μ ^ ℓ , m i , σ ^ m i 2 ) \hat{x}_{m_i}\sim\frac1L\mathcal{N}\left(\hat{\mu}_{\ell,m_i},\hat{\sigma}_{m_i}^2\right) x^mi∼L1N(μ^ℓ,mi,σ^mi2)
解释 x ^ m i \hat{x}_{m_i} x^mi
{ μ ^ ℓ , m i = ∑ k = 1 K c k μ ℓ , m i , σ ^ m i 2 = σ m i 2 ( ∑ k = 1 K c k ) 2 + ∑ k = 1 K c k 2 ϵ k 2 + δ 0 2 2 ( f 1 T f 1 + f 2 f 2 ) \begin{cases}\hat{\mu}_{\ell,m_i}=\sum_{k=1}^Kc_k\mu_{\ell,m_i},\\\hat{\sigma}_{m_i}^2=\sigma_{m_i}^2\left(\sum_{k=1}^Kc_k\right)^2+\sum_{k=1}^Kc_k^2\epsilon_k^2+\frac{\delta_0^2}2\left(\mathbf{f}_1^T\mathbf{f}_1+\mathbf{f}_2\mathbf{f}_2\right)\end{cases} ⎩⎨⎧μ^ℓ,mi=∑k=1Kckμℓ,mi,σ^mi2=σmi2(∑k=1Kck)2+∑k=1Kck2ϵk2+2δ02(f1Tf1+f2f2)
A
C
A
C
A
C
符号
描述
z k \mathbf{z}_k zk
传感器采集的信号 z k = z + e k \textcolor{red} {\mathbf{z}_{k}=\mathbf{z}+\mathbf{e}_{k}} zk=z+ek z = [ z 1 , z 2 , … , z S ] T \textcolor{blue}{\mathbf{z}=[z_1,z_2,\ldots,z_S]^T} z=[z1,z2,…,zS]T z k = [ z k , 1 , z k , 2 , … , z k , S ] T \textcolor{green}{\mathbf{z}_{k}=[z_{k,1},z_{k,2},\ldots,z_{k,S}]^{T}} zk=[zk,1,zk,2,…,zk,S]T e k ∼ N ( 0 , ϵ k 2 I ) \mathbf{e}_k\sim\mathcal{N}\left(0,\epsilon_k^2\mathbf{I}\right) ek∼N(0,ϵk2I)
x k \mathbf{x}_k xk
PCA特征提取后的信号分量 x = U T z = [ x 1 , x 2 , … , x M ] T \mathbf{x}=\mathbf{U}^T\mathbf{z}=[x_1,x_2,\ldots,x_M]^T x=UTz=[x1,x2,…,xM]T,前 M 个最大的奇异值对应的主成分
d k \mathrm{d}_k dk
PCA特征提取后的感知分量 d k = U T e k \mathrm{d}_k=\mathrm{U}^T\mathrm{e}_k dk=UTek
F ( x ) \mathcal{F}(\mathbf{x}) F(x)
PCA特征提取后的信号分布 F ( x ) = 1 L ∑ ℓ = 1 L F ℓ ( x ) \mathcal{F}(\mathbf{x})=\frac1L\sum_{\ell=1}^L\mathcal{F}_\ell(\mathbf{x}) F(x)=L1∑ℓ=1LFℓ(x)