1 Introduction
2 Preliminaries and Problem Formulation
2.1 System Description
betweenness centrality
2.2 Anomaly Detector Dynamics
2.3 Problem Formulation
3 Secure Scheme Design for MASs
3.1 Independent Anomaly Detection
Theorem 1
離散時間代數黎卡提系統
Algorithm 1: Independent Anomaly Detection
3.2 Betweenness Centrality Based Cooperative Anomaly Detector Design
Theorem 2
3.3 Feasibility Analysis of the Cooperative Detector
3.4 Anomaly Identification
4 Experiments on a Multivehicle Platform
Ref
compensated exogenous
x i ( k + 1 ) = A x i ( k ) + B u i ( k ) + B d d i ( k ) + B f f i ( k ) y i ( k ) = C x i ( k ) + D d d i ( k ) + D f f i ( k ) (1) \begin{aligned} x_i(k+1) &= A x_i(k) + B u_i(k) + &B_d d_i(k) + B_f f_i(k) \\ y_i(k) &= C x_i(k) + &D_d d_i(k) + D_f f_i(k) \end{aligned}\tag{1} xi(k+1)yi(k)=Axi(k)+Bui(k)+=Cxi(k)+Bddi(k)+Bffi(k)Dddi(k)+Dffi(k)(1)
其中,
x i ( k ) x_i(k) xi(k) 表示狀态;
u i ( k ) u_i(k) ui(k) 表示控制輸入;
y i ( k ) y_i(k) yi(k) 表示測量輸出向量;
d i d_i di 表示未知外界幹擾信号,上界 ∥ d 2 ∥ 2 ≤ Δ d \| d_2 \|_2 \le \Delta_d ∥d2∥2≤Δd;
f i f_i fi 表示故障。
cardinality
方程(1)的控制法則為
u i ( k ) = P i [ y i ( k ) , y j ( k ) ] , j ∈ N i (2) u_i(k) = \mathcal{P}_i[y_i(k), y_j(k)], j\in\mathcal{N}_i \tag{2} ui(k)=Pi[yi(k),yj(k)],j∈Ni(2)
其中, P i \mathcal{P}_i Pi 表示控制目标。即, ∀ i , j ∈ V N , P i = P j \forall i,j\in\mathcal{V}_N, \mathcal{P}_i=\mathcal{P}_j ∀i,j∈VN,Pi=Pj。
vulnerability
控制輸入向量 u i u_i ui 依賴于它的鄰居節點的輸出資訊決定。
是以,定義 y i j ( k ) y_i^j(k) yij(k) 表示 i i i 從 j j j 這裡獲得的輸出資訊。那麼,襲擊模型可以表示為
y i j ( k ) = y i ( k ) + b i j ( k ) (3) y_i^j(k) = y_i(k) + b_i^j(k) \tag{3} yij(k)=yi(k)+bij(k)(3)
這裡,
b i j ( k ) b_i^j(k) bij(k) 表示由于 false-data-injection 襲擊導緻的 corruption 資訊,并且有上界 ∥ b i j ∥ 2 ≤ Δ b \| b_i^j \|_2 \le \Delta_b ∥bij∥2≤Δb( Δ b \Delta_b Δb 是已知 的上界)。
η i , j \eta_{i,j} ηi,j 表示邊 e ( i , j ) e{(i,j)} e(i,j) 的 betweenness centrality。它有如下形式
η i , j = ∑ p = 1 N ∑ q = 1 , q > p N g p → j → q g p → q , p , q ∈ V N , ( i , j ) ∈ E i (4) \eta_{i,j} = \sum_{p=1}^{N} \sum_{q=1, q>p}^{N} \frac{g_{p \rightarrow j \rightarrow q}}{g_{p \rightarrow q}}, \quad p,q\in \mathcal{V}_N, (i,j)\in\mathcal{E}_i \tag{4} ηi,j=p=1∑Nq=1,q>p∑Ngp→qgp→j→q,p,q∈VN,(i,j)∈Ei(4)
g p g_{p} gp 表示從智能體 p p p 到 q q q 的最短路徑數量,
g p → j → q g_{p \rightarrow j \rightarrow q} gp→j→q 表示這些路徑經過邊 e ( i , j ) e{(i,j)} e(i,j) 的數量。
Ref: A centrality-based security game for multi-hop networks
四個假設:
LTI 系統(1)穩定;
狀态空間( A , B f , C , D f A, B_f, C, D_f A,Bf,C,Df)是最小實作;
協作範圍 r r r 和最大通信距離 R R R 滿足 2 r ≤ R 2r \le R 2r≤R;
異常檢測器不會被襲擊者破壞 (tampered)。
基于觀測器 (observer-based) 的異常檢測器:
x ^ i j ( k + 1 ) = A x ^ i j ( k ) + B u i ( k ) + L ( y i j ( k ) − y ^ i j ( k ) ) y ^ i j ( k ) = C x ^ i j ( k ) r i j ( k ) = V ( y i j ( k ) − y ^ i j ( k ) ) (5) \begin{aligned} \hat{x}_i^j(k+1) &=A \hat{x}_i^j(k) + Bu_i(k) + L(y_i^j(k) - \hat{y}_i^j(k)) \\ \hat{y}^j_i(k) &=C \hat{x}^j_i(k) \\ r^j_i(k) &=V (y_i^j(k) - \hat{y}_i^j(k)) \end{aligned}\tag{5} x^ij(k+1)y^ij(k)rij(k)=Ax^ij(k)+Bui(k)+L(yij(k)−y^ij(k))=Cx^ij(k)=V(yij(k)−y^ij(k))(5)
x ^ i j ( k ) \hat{x}_i^j(k) x^ij(k) 表示估算狀态 (estimated state);
y ^ i j ( k ) \hat{y}_i^j(k) y^ij(k) 表示估算輸出 (estimated output);
u i ( k ) u_i(k) ui(k) 表示由通信網絡産生的控制輸入。
r ^ i j ( k ) \hat{r}_i^j(k) r^ij(k) 表示殘差信号 (residual signal) 且滿足:
{ ∥ r i j ( k ) ∥ 2 ≥ J t h i , ⇒ 有 異 常 ∥ r i j ( k ) ∥ 2 < J t h i , ⇒ 無 異 常 (6) \left\{\begin{aligned} \| r^j_i(k) \|_2 \ge J^i_{th}, \Rightarrow 有異常 \\ \| r^j_i(k) \|_2 < J^i_{th}, \Rightarrow 無異常 \end{aligned}\right.\tag{6} {∥rij(k)∥2≥Jthi,⇒有異常∥rij(k)∥2<Jthi,⇒無異常(6)
J t h j J^j_{th} Jthj 表示門檻值函數 (threshold function)。
殘差信号關于網絡攻擊,實體故障和誤差的 Z 變換表示為:
r i j ( z ) = V [ G b ( z ) b i j ( z ) + G f ( z ) f i ( z ) + G d ( z ) d i ( z ) ] G b ( z ) = I − C ( z I − A + L C ) − 1 L (7) \begin{aligned} & r^j_i(z) = V [G_b(z) b_i^j(z) + G_f(z) f_i(z) + G_d(z)d_i(z)] \\ & G_b(z) = I - C(zI-A+LC)^{-1}L \end{aligned}\tag{7} rij(z)=V[Gb(z)bij(z)+Gf(z)fi(z)+Gd(z)di(z)]Gb(z)=I−C(zI−A+LC)−1L(7)
為了同時檢測和驗者異常,需要解決以下兩個問題:
如何檢測是否存在 b i j ( k ) b_i^j(k) bij(k) 和 f i ( k ) f_i(k) fi(k),在 d i ( k ) d_i(k) di(k) 的影響下?
如何區分 b i j ( k ) b_i^j(k) bij(k) 和 f i ( k ) f_i(k) fi(k) 在檢測到的時候?
問題 1 的優化結果可以通過設計一下優化檢測器來産生,使
influence of disturbances anomaly sensitivity → min . (8) \frac{\text{influence of disturbances}}{\text{anomaly sensitivity}} \rightarrow \min. \tag{8} anomaly sensitivityinfluence of disturbances→min.(8)
幹擾的影響 異常敏感性 → min . (8) \frac{\text{幹擾的影響}}{\text{異常敏感性}} \rightarrow \min. \tag{8} 異常敏感性幹擾的影響→min.(8)
問題 2,由于襲擊信号 b i j ( k ) b_i^j(k) bij(k) 可以任意被襲擊者設計,甚至于可以完全等價于故障信号 f i ( k ) f_i(k) fi(k)。舉個例子,讓 b i j ( k ) b_i^j(k) bij(k) 滿足:
ξ ( k + 1 ) = A ξ ( k ) + B f ζ ( k ) b i ( k ) = C ξ ( k ) + D f ζ ( k ) \xi(k+1) = A \xi(k) + B_f \zeta(k) \\ b_i(k) = C \xi(k) + D_f \zeta(k) ξ(k+1)=Aξ(k)+Bfζ(k)bi(k)=Cξ(k)+Dfζ(k)
ξ ( k ) \xi(k) ξ(k) 是輔助狀态向量 (auxiliary state vector);
ζ \zeta ζ 是随機信号;
那麼,(7)中的殘差信号 r i j ( z ) r_i^j(z) rij(z) 可以表示為:
r i j ( k ) = V [ G ζ ( z ) ζ ( k ) + G f ( z ) f i ( z ) + G d ( z ) d i ( z ) ] r_i^j(k) = V [G_\zeta (z) \zeta(k) + G_f(z)f_i(z) + G_d(z)d_i(z)] rij(k)=V[Gζ(z)ζ(k)+Gf(z)fi(z)+Gd(z)di(z)]
這裡, G ζ ( z ) = G f ( z ) G_\zeta(z) = G_f(z) Gζ(z)=Gf(z)。
重新考慮異常檢測器(5)。
x i ( k + 1 ) = A x i ( k ) + B u i ( k ) + B d d i ( k ) + B f f i ( k ) y i ( k ) = C x i ( k ) + D d d i ( k ) + D f f i ( k ) (1) \begin{aligned} \quad\ \ x_i(k+1) &= A x_i(k) + B u_i(k) + &B_d d_i(k) + B_f f_i(k) \\ y_i(k) &= C x_i(k) + &D_d d_i(k) + D_f f_i(k) \end{aligned}\tag{1} xi(k+1)yi(k)=Axi(k)+Bui(k)+=Cxi(k)+Bddi(k)+Bffi(k)Dddi(k)+Dffi(k)(1)
假設幹擾信号 d i ( k ) ≡ 0 d_i(k)\equiv 0 di(k)≡0。
實體故障 f i ( k ) ≠ 0 f_i(k) \ne 0 fi(k)=0 且襲擊信号 b i j ( k ) = 0 b_i^j(k) = 0 bij(k)=0,産生 ∥ r i j ( k ) ∥ 2 ≠ 0 , ∀ j ∈ N i \| r^j_i(k) \|_2 \ne 0, \forall j \in \mathcal{N}_i ∥rij(k)∥2=0,∀j∈Ni;
實體故障 f i ( k ) = 0 f_i(k) = 0 fi(k)=0 且襲擊信号 b i j ( k ) ≠ 0 b_i^j(k) \ne 0 bij(k)=0,産生 ∥ r i j ( k ) ∥ 2 ≠ 0 \| r^j_i(k) \|_2 \ne 0 ∥rij(k)∥2=0 和 ∥ r i j ( k ) ∥ 2 ≠ 0 , ∀ l ∈ N i , l ≠ j , ∥ r i l ( k ) ∥ 2 = 0 \| r^j_i(k) \|_2 \ne 0, \forall l \in \mathcal{N}_i, l \ne j, \| r^l_i(k) \|_2 = 0 ∥rij(k)∥2=0,∀l∈Ni,l=j,∥ril(k)∥2=0。
是以,組織所有屬于 N i \mathcal{N}_i Ni 的智能體一起合作監視 i i i ,實體故障信号
f i ( k ) f_i(k) fi(k) 和襲擊信号 b i j ( k ) b_i^j(k) bij(k) 将會表現出完全不同的特點,可以區分它們。
具體來說,将 f i ( k ) f_i(k) fi(k) 作為待檢測的故障, b i j ( k ) b_i^j(k) bij(k) 作為待抑制的外部幹擾,通過在 N i \mathcal{N}_i Ni 中設計一個協作檢測器,将故障與幹擾、攻擊區分開來,這樣
influence of disturbances and attacks influence of faults → min . (9) \frac{\text{influence of disturbances and attacks}}{\text{influence of faults}} \rightarrow \min . \tag{9} influence of faultsinfluence of disturbances and attacks→min.(9)
幹擾和襲擊的影響 故障的影響 → min . (9) \frac{\text{幹擾和襲擊的影響}}{\text{故障的影響}} \rightarrow \min . \tag{9} 故障的影響幹擾和襲擊的影響→min.(9)
獨立異常檢測器設計用來解決 Problem 1 。
min L , V J 1 ( L , V ) = min L , V ∥ V G d ( z ) ∥ ∞ ∥ V G f ( z ) ∥ ∞ (10a) \min_{L,V} J_1(L,V) = \min_{L,V} \frac{\|VG_d(z)\|_\infty}{\|VG_f(z)\|_\infty} \tag{10a} L,VminJ1(L,V)=L,Vmin∥VGf(z)∥∞∥VGd(z)∥∞(10a)
min L , V J 2 ( L , V ) = min L , V ∥ V G d ( z ) ∥ ∞ ∥ V G b ( z ) ∥ ∞ (10b) \min_{L,V} J_2(L,V) = \min_{L,V} \frac{\|VG_d(z)\|_\infty}{\|VG_b(z)\|_\infty} \tag{10b} L,VminJ2(L,V)=L,Vmin∥VGb(z)∥∞∥VGd(z)∥∞(10b)
給出檢測器(5)并假設(1-4)均滿足,那麼
L o p t = − L o T , V o p t = V o (11) L_{opt} = -L_o^T,\quad V_{opt} = V_o \tag{11} Lopt=−LoT,Vopt=Vo(11)
V o V_o Vo 是列滿秩矩陣 H H H 的左逆, H H H 滿足 H H T = C X C T + D d D d T HH^T = CXC^T + D_dD_d^T HHT=CXCT+DdDdT。
并且 ( X , L o X,L_o X,Lo) 是下述離散時間代數黎卡提系統的穩定解
[ A X A T − X + B d B d T A X C T + B d D d T C X A T + D d B d T C X C T + D d D d T ] [ I L 0 ] = 0 (12) \left[\begin{matrix} AXA^T - X + B_dB_d^T & AXC^T + B_dD_d^T \\ CXA^T + D_dB_d^T & CXC^T + D_dD_d^T \end{matrix}\right] \left[\begin{matrix} I \\ L_0 \end{matrix}\right] = 0 \tag{12} [AXAT−X+BdBdTCXAT+DdBdTAXCT+BdDdTCXCT+DdDdT][IL0]=0(12)
離散時間代數黎卡提系統 (discrete-time algebraic Riccati system)
在(7)中的殘差信号 r i j ( k ) r_i^j(k) rij(k),可以通過使用 L = L o p t L=L_{opt} L=Lopt、 V = V o p t V=V_{opt} V=Vopt 來産生。
然後是一個合适的門檻值函數 J t h j J^j_{th} Jthj 來決定是否異常。
令 J t h j = sup b i j = 0 , f i = 0 ∥ r i j ( z ) ∥ 2 J^j_{th}=\sup_{b_i^j=0, f_i=0} \|r_i^j(z)\|_2 Jthj=supbij=0,fi=0∥rij(z)∥2
J t h j = sup d i ∥ V G d ( z ) d i ( z ) ∥ 2 = ∥ V G d ( z ) ∥ ∞ Δ d J^j_{th} = \sup_{d_i}\| VG_d(z)d_i(z)\|_2 = \|VG_d(z)\|_\infty \Delta_d Jthj=disup∥VGd(z)di(z)∥2=∥VGd(z)∥∞Δd
Input: 智能體 i i i 的輸入 u i ( k ) u_i(k) ui(k) 和輸出 y i j ( k ) y_i^j(k) yij(k);
Output: 智能體 j j j 對 i i i 獨立檢測器結果。
基于定理1計算 L o p t L_{opt} Lopt、 V o p t V_{opt} Vopt;
通過通信網絡得到 u i ( k ) u_i(k) ui(k)、 y i j ( k ) y_i^j(k) yij(k);
通過(7)獲得殘差信号 r i j ( k ) r_i^j(k) rij(k),其中 L = L o p t L=L_{opt} L=Lopt、 V = V o p t V=V_{opt} V=Vopt;
通過(6)評價殘差信号 r i j ( k ) r_i^j(k) rij(k)(是否異常?);
k = k + 1 k = k+1 k=k+1 并傳回步驟2。
自我了解:
最優參數 L o p t L_{opt} Lopt、 V o p t V_{opt} Vopt 為系統矩陣的最優,是以全局僅存在一個最優。這一點也可以通過觀察算法執行後傳回了第2步,而沒有再次重新計算 L o p t L_{opt} Lopt、 V o p t V_{opt} Vopt 再次佐證(如果最優不是全局最優,那麼每檢測一次就要重新計算最優)。
協作異常檢測器用來檢測所有屬于 N i N_i Ni 的智能體,以解決 Problem2。
x ^ i j ( k + 1 ) = A x ^ i j ( k ) + B u i ( k ) + L [ y i j ( k ) + ∑ l = i 1 i c i w l ( y i l ( k ) − y i j ( k ) ) − y ^ i j ( k ) ] y ^ i j ( k ) = C x ^ i j ( k ) r ˉ i j ( k ) = V [ y i j ( k ) + ∑ l = i 1 i c i w l ( y i l ( k ) − y i j ( k ) ) − y ^ i j ( k ) ] (16) \begin{aligned} \hat{x}_i^j(k+1) =& A \hat{x}_i^j(k) + Bu_i(k) \\ &+L[y_i^j(k) + \sum_{l=i_1}^{i_{c_i}} \red{w_l}(y_i^l(k)-y_i^j(k)) - \hat{y}_i^j(k)] \\ \hat{y}_i^j(k) =& C\hat{x}_i^j(k) \\ \bar{r}_i^j(k) =& V[y_i^j(k) + \sum_{l=i_1}^{i_{c_i}} \red{w_l} (y_i^{l}(k) - y_i^j(k)) - \hat{y}_i^j(k)] \end{aligned}\tag{16} x^ij(k+1)=y^ij(k)=rˉij(k)=Ax^ij(k)+Bui(k)+L[yij(k)+l=i1∑iciwl(yil(k)−yij(k))−y^ij(k)]Cx^ij(k)V[yij(k)+l=i1∑iciwl(yil(k)−yij(k))−y^ij(k)](16)
令 w l T = [ w i 1 w i 2 ⋯ w i c i ] w_l^T = [w_{i_1} \quad w_{i_2} \quad \cdots w_{i_{c_i}}] wlT=[wi1wi2⋯wici],
b N i ( k ) = c o l ( b i i 1 ( k ) , b i i 2 ( k ) , ⋯ , b i i c i ( k ) ) b_{N_i}(k) = col(b_i^{i_1}(k), b_i^{i_2}(k), \cdots, b_i^{i_{c_i}}(k)) bNi(k)=col(bii1(k),bii2(k),⋯,biici(k))。
(16)變成如下形式:
x ^ i j ( k + 1 ) = A x ^ i j ( k ) + B u i ( k ) + L [ y i ( k ) − y ^ i j ( k ) + ( w T Γ ⊗ I n b ) b N i ( k ) ] y ^ i j ( k ) = C x ^ i j ( k ) r ˉ i j ( k ) = V [ y i ( k ) − y ^ i j ( k ) + ( w T Γ ⊗ I n b ) b N i ( k ) ] (17) \begin{aligned} \hat{x}_i^j(k+1) =& A \hat{x}_i^j(k) + Bu_i(k) \\ &+L[y_i(k) - \hat{y}_i^j(k) + (w^T \Gamma \otimes I_{n_b}) b_{\mathcal{N}_i}(k)] \\ \hat{y}_i^j(k) =& C\hat{x}_i^j(k) \\ \bar{r}_i^j(k) =& V[y_i(k) - \hat{y}_i^j(k) + (w^T \Gamma \otimes I_{n_b}) b_{\mathcal{N}_i}(k)] \end{aligned}\tag{17} x^ij(k+1)=y^ij(k)=rˉij(k)=Ax^ij(k)+Bui(k)+L[yi(k)−y^ij(k)+(wTΓ⊗Inb)bNi(k)]Cx^ij(k)V[yi(k)−y^ij(k)+(wTΓ⊗Inb)bNi(k)](17)
這裡, Γ = d i a g { γ i 1 , γ i 2 , ⋯ , γ i c i } \Gamma = diag\{\gamma_{i_1}, \gamma_{i_2}, \cdots, \gamma_{i_{c_i}}\} Γ=diag{γi1,γi2,⋯,γici} 是二進制對角矩陣。
r l = { 1 , edge ( i , l ) is under attack 0 , no attack exists (18) r_l = \left\{\begin{aligned} &1,\quad \text{edge}\ (i, l)\ \text{is under attack} \\ &0, \quad \text{no attack exists} \end{aligned}\right.\tag{18} rl={1,edge (i,l) is under attack0,no attack exists(18)
殘差信号 r N i ( z ) r_{\mathcal{N}_i}(z) rNi(z) 的 Z 變換如下
r N i ( z ) = V N i [ G ˉ b ( z ) b N i ( z ) + G ˉ f ( z ) f i ( z ) + G ˉ d ( z ) d i ( z ) ] G ˉ d ( z ) = 1 c i ⊗ G d ( z ) , G ˉ f ( z ) = 1 c i ⊗ G f ( z ) G ˉ b ( z ) = ( W Γ ) ⊗ G b ( z ) (20) \begin{aligned} r_{\mathcal{N}_i}(z) &= V_{\mathcal{N}_i} [\bar{G}_b(z) b_{\mathcal{N}_i}(z) + \bar{G}_f(z) f_i(z) + \bar{G}_d(z) d_i(z)] \\ \bar{G}_d(z) &= 1_{c_i} \otimes G_d(z), \bar{G}_f(z) = 1_{c_i} \otimes G_f(z) \\ \bar{G}_b(z) &= (W\Gamma) \otimes G_b(z) \end{aligned}\tag{20} rNi(z)Gˉd(z)Gˉb(z)=VNi[Gˉb(z)bNi(z)+Gˉf(z)fi(z)+Gˉd(z)di(z)]=1ci⊗Gd(z),Gˉf(z)=1ci⊗Gf(z)=(WΓ)⊗Gb(z)(20)
優化問題(9)表達如下
min L , V J 3 ( L , V ) = min L , V ∥ V N i G ˉ d ( z ) ∥ ∞ ∥ V N i G ˉ f ( z ) ∥ ∞ (21a) \min_{L,V} J_3(L,V) = \min_{L,V} \frac{\|V_{\mathcal{N}_i} \bar{G}_d(z)\|_\infty}{\|V_{\mathcal{N}_i} \bar{G}_f(z)\|_\infty} \tag{21a} L,VminJ3(L,V)=L,Vmin∥VNiGˉf(z)∥∞∥VNiGˉd(z)∥∞(21a)
min L , V J 4 ( L , V , W ) = min L , V , W ∥ V N i G ˉ b ( z ) ∥ ∞ ∥ V N i G ˉ f ( z ) ∥ ∞ s . t . W = 1 c i ⊗ w T , w > 0 , w T 1 c i = 1. (21b) \min_{L,V} J_4(L,V,W) = \min_{L,V,W} \frac{\|V_{\mathcal{N}_i} \bar{G}_b(z)\|_\infty}{\|V_{\mathcal{N}_i} \bar{G}_f(z)\|_\infty} \\ s.t.\ W=1_{c_i} \otimes w^T, w>0, w^T 1_{c_i} = 1. \tag{21b} L,VminJ4(L,V,W)=L,V,Wmin∥VNiGˉf(z)∥∞∥VNiGˉb(z)∥∞s.t. W=1ci⊗wT,w>0,wT1ci=1.(21b)
下一步就是找出最優參數 L = L o p t L=L_{opt} L=Lopt, V = V o p t V=V_{opt} V=Vopt, W = W o p t W=W_{opt} W=Wopt 來同時解決(21a)和(21b)。
但是利用定理1解出來的最優參數,(21a)和(21b)并不相同。是以需要一個合适的妥協。
令 σ ˉ ( ⋅ ) \bar{\sigma}(\cdot) σˉ(⋅) 表示最大奇異值, ρ ( ⋅ ) \rho(\cdot) ρ(⋅) 表示譜半徑 (spectral radius)。
由于 ∥ V N i G ˉ b ( z ) ∥ ∞ = sup θ ∈ [ 0 , π ] ρ [ ( W Γ ⊗ ( V G b ( e j θ ) ) T W Γ ⊗ ( V G b ( e j θ ) ) ) ] = sup θ ∈ [ 0 , π ] ρ ( Γ T W T W Γ ) ρ [ ( V G b ( e j θ ) ) T ( V G b ( e j θ ) ) ] = σ ˉ ( W Γ ) sup θ ∈ [ 0 , π ] σ ˉ [ V G b ( e j θ ) ] \begin{aligned} & {\|V_{\mathcal{N}_i} \bar{G}_b(z)\|_\infty} \\ &= \sup_{\theta\in[0,\pi]} \sqrt{\rho [(W\Gamma \otimes (VG_b(e^{j\theta}))^T W\Gamma \otimes (VG_b(e^{j\theta})))]} \\ &= \sup_{\theta\in[0,\pi]} \sqrt{\rho (\Gamma^T W^T W\Gamma)} \sqrt{\rho[(VG_b(e^{j\theta}))^T (VG_b(e^{j\theta}))]} \\ &= \bar{\sigma}(W\Gamma) \sup_{\theta\in[0,\pi]} \bar{\sigma}[VG_b(e^{j\theta})] \end{aligned} ∥VNiGˉb(z)∥∞=θ∈[0,π]supρ[(WΓ⊗(VGb(ejθ))TWΓ⊗(VGb(ejθ)))] =θ∈[0,π]supρ(ΓTWTWΓ) ρ[(VGb(ejθ))T(VGb(ejθ))] =σˉ(WΓ)θ∈[0,π]supσˉ[VGb(ejθ)]
那麼
min J 4 = min L , V , W \min J_4 = \min_{L,V,W} \frac{}{} minJ4=L,V,Wmin
給出檢測器(19)并假設條件 1-4 都滿足,那麼
L = L o p t L=L_{opt} L=Lopt, V = V o p t V=V_{opt} V=Vopt,解決(21a),這裡 L o p t L_{opt} Lopt 和 V o p t V_{opt} Vopt 由(11)産生;
W = W o p t W = W_{opt} W=Wopt 解決(24),且 W o p t W_{opt} Wopt 滿足
W o p t = 1 c i ⊗ w o p t T w o p t T = [ w o p t , i 1 , w o p t , i 2 , ⋯ , w o p t , i c i ] w o p t , l = ∏ p = i 1 i c i γ ^ p 2 γ ^ l 2 ∑ q = i 1 i c i ∏ p = i 1 , p ≠ q i c i γ ^ p 2 , ∀ l = i 1 , ⋯ , i c i (25) W_{opt} = 1_{c_i} \otimes w_{opt}^T \\ w_{opt}^T = [w_{opt,i_1}, w_{opt,i_2}, \cdots, w_{opt,i_{c_i}}] \\ w_{opt,l} = \frac{\prod_{p=i_1}^{i_{c_i}} \hat{\gamma}_p^2}{\hat{\gamma}_l^2 \sum_{q=i_1}^{i_{c_i}} \prod_{p=i_1, p\ne q}^{i_{c_i}} \hat{\gamma}_p^2}, \quad \forall l=i_1,\cdots,i_{c_i} \tag{25} Wopt=1ci⊗woptTwoptT=[wopt,i1,wopt,i2,⋯,wopt,ici]wopt,l=γ^l2∑q=i1ici∏p=i1,p=qiciγ^p2∏p=i1iciγ^p2,∀l=i1,⋯,ici(25)
拉格朗日乘數 (Lagrange multipliers)
儲存仿真相關參數
傳遞函數陣
準備計算H無窮範數
系統傳遞函數 MIMO
[21] J. R. Riehl and M. Cao, “A centrality-based security game for multi-hop networks,” IEEE Trans. Control Netw. Syst., vol. 5, no. 4, pp. 1507–1516, Dec. 2018.
[22] S. Shiliang, W. Ting, Y. Chen, L. Xiaofan, and Z. Hai, “Distributed fault detection and isolation for flocking in a multi-robot system with imperfect communication,” Int. J. Adv. Robot. Syst., vol. 11, no. 6, pp. 86–93, 2014.
[23] S. Ding, T. Jeinsch, P. Frank, and E. Ding, “A unified approach to the optimization of fault detection systems,” Int. J. Adaptive Control Signal Process., vol. 14, no. 7, pp. 725–745, 2000.
[24] P. Zhang, S. X. Ding, G. Z. Wang, and D. H. Zhou, “Fault detection of linear discrete-time periodic systems,” IEEE Trans. Autom. Control, vol. 50, no. 2, pp. 239–244, Feb. 2005.
[25] Y. Li, H. Fang, J. Chen, and C. Yu, “Distributed cooperative fault detection for multiagent systems: A mixed h ∞/h 2optimization approach,” IEEE Trans. Ind. Electron., vol. 65, no. 8, pp. 6468–6477, Aug. 2018.
[26] D. B. Rawat and C. Bajracharya, “Detection of false data injection attacks in smart grid communication systems,” IEEE Signal Process. Lett., vol. 22, no. 10, pp. 1652–1656, Oct. 2015.