Introduction
Networks play a major role in the transmission of electricity and fluids such as gas, oil, or water. A canonical example is the power grid, a remarkable engineering realization that emerged in the past century. A natural mathematical model of these networks is a graph together with discrete conservation laws. Kirchhoff’s law and Ohm’s law yield the discrete Laplacian: the DC load flow model in the jargon of power grids. This is the model used by many grid operators. Adding inertia to the system one obtains the so-called graph wave equation. Adding driving and damping, we can model the time evolution of an electrical grid [1 , 2 ] and a fluid network [3 ] in a linearized approximation. This is a linearization of the so-called ”swing equation” used to model the electrical grid [4 , 5 ] .
The detection of disturbances on such networks is an important topic. For the grid, disturbances are electromechanical oscillations induced by faulty equipment. For other fluid network, they can be leaks, causing unexpected pressure losses. Such faults can break equipment, see [6 ] for a survey of such events for the grid. Due to the importance of the problem, many authors addressed the issue of identifying a faulty vertex in a network. Nudell and Chakrabortty [7 ] propose a graph theoretic algorithm to find forced oscillation inputs; they assume minimal information and estimate how the fault affects the Green’s function of the Laplacian. Chen et al [8 ] use an energy-based method to locate oscillation sources in power systems. Delabay et al [9 ] propose to locate a fault using a maximum likelihood approach: here the state of the network is assumed to be known everywhere, not the parameters.
In the literature, many works assume the unknown disturbances to have a given form and only parts of this (frequency, amplitude,...) are unknown and subject to the inverse problem. To address these shortfalls, six years ago, we introduced a source detection approach [10 ] for the graph wave equation with no damping, assuming distinct eigenvalues of the graph Laplacian and a disturbance on a single vertex that vanishes before reaching the final monitoring time. We defined a strategic observation set of vertices; this allows to reconstruct the state of the disturbance free system using only measurements from this set. From these observations, one can locate the source using an overdetermined linear system based on an adjoint formulation. Interestingly, the rank of the matrix in this linear system revealed graph conditions for placing observation vertices. Finally the procedure permits to reconstruct the disturbance.
Here, we revisit the problem, allow the graph laplacian to admit eigenvalues of arbitrary multiplicities, include damping in the dynamics and allow several disturbances. To monitor the network for disturbances and localizing-identifying them, one can raise the following three questions
How to select the observation set of vertices to detect the presence of a disturbance?
How to select the observation set to localize disturbances and identify their intensities ?
Can this localization and identification be done in ”real time ” ?
We also need technical conditions for which these questions can be answered.
To tackle question one, we introduce the notion of strategic observation set and show that observing such a set gives the state of the network unequivocally. This leads to requiring condition 1 where we assume a healthy time period wherein no disturbances occur. The detection is done by examining the differences between the observations of the system and it’s states when it is free of disturbances (healthy states). Once disturbances have been detected, we use these residuals to localize and identify them.
Questions three is easier to address than question two. We introduce the technical condition that disturbances occur away from the observation set. Then we show that if a submatrix of the graph Laplacian has full rank and the condition above is met, then one can detect, localize and identify disturbances in ”quasi-real time ”. We call such observation sets dominantly observable , for these, the number of vertices is larger than half the total number of network vertices. This is a strong requirement, in many cases only a limited number of sensors are available. We show that if the observation set is absorbent , then detection, localization and identification of disturbances remains possible but with a delay. We present numerical experiments illustrating these results. The article is organized as follows: section 2 presents the mathematical model and states the inverse problem, section 3 shows how to detect accidental disturbances. We present the inverse problem for dominantly absorbent observation sets and for absorbent observation sets in sections 4 and 5 respectively. Section 6 illustrates these results with numerical experiments and we conclude in section 7.
Mathematical modeling and problem statement
In the following, we consider the evolution of a miscible flow on a network modeled by a graph G ( V ) \mathcal{G}(\mathcal{V}) G ( V ) where V \mathcal{V} V is a set of vertices. We assume there are ∣ V ∣ = N |\mathcal{V}| = N ∣ V ∣ = N vertices and the graph is connected.
We consider the system of differential equations
X ¨ ( t ) + η X ˙ ( t ) − Δ X ( t ) = F ( t ) , (1) \ddot{X}(t) + \eta\dot{X}(t) - \Delta X(t) = F(t), \tag*{(1)} X ¨ ( t ) + η X ˙ ( t ) − Δ X ( t ) = F ( t ) , (1)
where X X X is the N N N dimensional vector defining the network state. Δ \Delta Δ is the N × N N \times N N × N graph Laplacian matrix defined by
Definition 2.1 The graph Laplacian matrix Δ \Delta Δ is the real, symmetric negative semidefinite matrix, such that
Δ k l = Δ l k > 0 \Delta_{kl} = \Delta_{lk} > 0 Δ k l = Δ l k > 0 if k k k and l l l are connected, 0 0 0 otherwise,
Δ k k = − ∑ l ≠ k w k l \Delta_{kk} = -\sum_{l\ne k} w_{kl} Δ k k = − ∑ l = k w k l .
Here, we choose Δ k l = 1 \Delta_{kl}=1 Δ k l = 1 for simplicity and to illustrate the influence of the topology.
The parameter η > 0 \eta> 0 η > 0 is a damping. (1) is established as a linearized model of the swing-equations for the voltage on a power grid; there x n x_n x n is associated to the phase of the voltage at vertex n n n . (1) also models the evolution of the pressure in a fluid (water or gas) network in the linear limit [11 , 3 , 12 ] .
We now recall the notions of forward and inverse problems associated to (1).
▶ Forward problem: Apart from disturbances, the right hand side term of equation (1) is F = F s o u r F = F^{sour} F = F so u r , a well known source vector that ensures a desired functioning of the network. Given a final monitoring time T > 0 T > 0 T > 0 , a source F n = 1 , … , N s o u r ∈ L 2 ( 0 , T ) F^{sour}_{n=1,\ldots,N} \in L^2(0,T) F n = 1 , … , N so u r ∈ L 2 ( 0 , T ) and two initial network state conditions X ( 0 ) = ( x 1 ( 0 ) , … , x N ( 0 ) ) T X(0) = (x_1(0), \ldots, x_N(0))^\mathsf{T} X ( 0 ) = ( x 1 ( 0 ) , … , x N ( 0 ) ) T and X ˙ ( 0 ) = ( x ˙ 1 ( 0 ) , … , x ˙ N ( 0 ) ) T \dot{X}(0) = (\dot{x}_1(0), \ldots, \dot{x}_N(0))^\mathsf{T} X ˙ ( 0 ) = ( x ˙ 1 ( 0 ) , … , x ˙ N ( 0 ) ) T , the problem (1) admits a unique solution X = ( x 1 , … , x N ) T X = (x_1, \ldots, x_N)^\mathsf{T} X = ( x 1 , … , x N ) T whose x n = 1 , … , N ∈ H 2 ( 0 , T ) x_{n=1,\ldots,N} \in H^2(0,T) x n = 1 , … , N ∈ H 2 ( 0 , T ) . This is the so-called forward problem . Thus, given a set S \mathcal{S} S of network observation vertices, we define the following observation operator:
M [ F ] : = { x n ( t ) in ( 0 , T ) , ∀ n ∈ S } . (2) M[F] := \{x_n(t) \text{ in } (0,T),\ \forall n \in\mathcal{S}\}. \tag*{(2)} M [ F ] := { x n ( t ) in ( 0 , T ) , ∀ n ∈ S } . (2)
We therefore introduce the notions of observation and non-observation sets.
Definition 2.2 The set S S S of observation vertices wherein the local network state is recorded is called observation set. The complementary of S S S , E = V − S \mathcal{E}=\mathcal{V}-S E = V − S is the set of non-observation vertices. In the following we label N S = ∣ S ∣ N_S=|S| N S = ∣ S ∣ , N E = ∣ E ∣ N_{\mathcal{E}}=|\mathcal{E}| N E = ∣ E ∣ .
▶ Inverse problem: Here, we write
F ( t ) = F s o u r ( t ) + F d i s ( t ) , ∀ t ∈ ( 0 , T ) F(t)=F^{sour}(t)+F^{dis}(t),\quad\forall t\in(0,T) F ( t ) = F so u r ( t ) + F d i s ( t ) , ∀ t ∈ ( 0 , T )
where F s o u r F^{sour} F so u r is known and F d i s F^{dis} F d i s is an unknown accidental disturbance that disrupts the healthy state of the network. In the inverse problem, we are given time records d n ( t ) d_n(t) d n ( t ) in ( 0 , T ) (0,T) ( 0 , T ) of the local state x n ( t ) x_n(t) x n ( t ) taken at observation vertices n ∈ S n\in S n ∈ S . The task is to detect-localize-identify F d i s F^{dis} F d i s from d n ( t ) d_n(t) d n ( t ) , n ∈ S n\in S n ∈ S i.e. invert the map
M [ F ] = { d n ( t ) in ( 0 , T ) , ∀ n ∈ S } . (3) M[F]=\left\{d_n(t)\ \text{in}\ (0,T),\ \forall n\in S\right\}. \tag*{(3)} M [ F ] = { d n ( t ) in ( 0 , T ) , ∀ n ∈ S } . (3)
An important question while addressing this inverse problem is: Which observation set S S S ensures the identifiability of the unknown occurring disturbances F d i s F^{dis} F d i s ? In the remainder, we show that S S S needs to be strategic , a spectral property. We will also require S S S to be absorbent i.e.
Definition 2.3 A set of vertices S S S of a graph is absorbent if each vertex of the graph is directly connected to at least one vertex of S S S .
or verify a stronger property by being dominantly absorbent that will be introduced later. See the appendix for an algorithm to find an absorbent set S S S for a given graph.
Detection of accidental disturbances
To detect unknown disturbances that may affect the network, we request the observation set S S S to be strategic . This notion is based on the spectral properties of the graph Laplacian matrix Δ \Delta Δ defined in (2.1). Here, we introduce this notion and establish that for a healthy network i.e., F d i s = 0 F^{dis}=0 F d i s = 0 , the local knowledge of its state in a strategic observation set determines in a unique manner its global state in all of its vertices. This assertion yields the main ingredient of our procedure to detect disturbances.
Strategic set of network vertices
Since the notion of a strategic observation set of vertices depends on the spectral properties of the graph Laplacian matrix Δ \Delta Δ , we employ the following notations:
For k = 1 , … , K k=1,\ldots,K k = 1 , … , K : We denote by − ω k 2 -\omega_k^2 − ω k 2 the k t h k^{\mathrm{th}} k th eigenvalue of the matrix Δ \Delta Δ , where the real numbers ω k = 1 , … , K \omega_{k=1,\ldots,K} ω k = 1 , … , K are such that 0 = ω 1 < ω 2 < ⋯ < ω K 0=\omega_1<\omega_2<\cdots<\omega_K 0 = ω 1 < ω 2 < ⋯ < ω K . Hence, the distinct eigenvalues of Δ \Delta Δ are labeled in the following decreasing order:
0 = − ω 1 2 > − ω 2 2 > ⋯ > − ω K 2 . (4) 0=-\omega_1^2>-\omega_2^2>\cdots>-\omega_K^2. \tag*{(4)} 0 = − ω 1 2 > − ω 2 2 > ⋯ > − ω K 2 . (4)
m k m_k m k is the multiplicity of − ω k 2 -\omega_k^2 − ω k 2 and M = max k ∈ { 1 , … , K } m k M=\max_{k\in\{1,\ldots,K\}}m_k M = max k ∈ { 1 , … , K } m k . Then, N = ∑ k = 1 K m k N=\sum_{k=1}^{K}m_k N = ∑ k = 1 K m k .
For k = 1 , … , K k=1,\ldots,K k = 1 , … , K : We use the indices k ℓ = 1 , … , m k k_{\ell=1,\ldots,m_k} k ℓ = 1 , … , m k to denote the normalized orthogonal eigenvectors v k ℓ v^{k\ell} v k ℓ associated to the k t h k^{\mathrm{th}} k th eigenvalue − ω k 2 -\omega_k^2 − ω k 2 of the matrix Δ \Delta Δ :
Δ v k ℓ = − ω k 2 v k ℓ , for ℓ = 1 , … , m k , where k 1 = k . (5) \Delta v^{k\ell}=-\omega_k^2v^{k\ell},\qquad\text{for }\ell=1,\ldots,m_k,\quad\text{where }k_1=k. \tag*{(5)} Δ v k ℓ = − ω k 2 v k ℓ , for ℓ = 1 , … , m k , where k 1 = k . (5)
Notice that if the graph is connected, there is a single zero eigenvalue i.e., m 1 = 1 m_1=1 m 1 = 1 [13 ] . We now introduce the notion strategic set of vertices.
Definition 3.1 A set S S S of vertices is strategic if for all eigenvalues − ω k 2 -\omega_k^2 − ω k 2 of multiplicity m k m_k m k , there exists a subset I k I_k I k of m k m_k m k vertices in S S S defining a m k × m k m_k\times m_k m k × m k invertible matrix:
A i ℓ ( k ) = v i k ℓ , for ℓ = 1 , … , m k and ∀ i ∈ I k , (6) A_{i\ell}^{(k)}=v_i^{k\ell},\qquad\text{for }\ell=1,\ldots,m_k\quad\text{and}\quad\forall i\in I_k, \tag*{(6)} A i ℓ ( k ) = v i k ℓ , for ℓ = 1 , … , m k and ∀ i ∈ I k , (6) where v k ℓ ∈ R N v^{k\ell}\in\mathbb{R}^{N} v k ℓ ∈ R N are the m k m_k m k eigenvectors associated to the eigenvalue − ω k 2 -\omega_k^2 − ω k 2 of Δ \Delta Δ .
Hence, S S S is strategic if for all eigenvalues − ω k 2 -\omega_k^2 − ω k 2 of multiplicity m k m_k m k , there exists a subset I k I_k I k of m k m_k m k row labels in S S S such that the matrix A ( k ) A^{(k)} A ( k ) whose columns are the m k m_k m k eigenvectors v k ℓ v^{k\ell} v k ℓ is of full rank. Note that a strategic set contains at least M M M vertices, where M M M is the maximal multiplicity of the eigenvalues.
This definition generalizes the one given in [10 ] for the case where all eigenvalues of the Laplacian matrix Δ \Delta Δ are of multiplicity 1. It is also a transposition for discrete systems of the result in [14 ] .
From Definition 3.1, it follows that every set containing a strategic set is also strategic . As examples:
For M = 1 M=1 M = 1 : S S S is strategic if: ∀ k ∈ { 1 , … , K = N } \forall k\in\{1,\ldots,K=N\} ∀ k ∈ { 1 , … , K = N } , ∃ I k = { i } ∈ S \exists I_k=\{i\}\in S ∃ I k = { i } ∈ S , v i k ≠ 0 v_i^k\ne0 v i k = 0 .
For M = 2 M=2 M = 2 : S S S is strategic if: ∀ k ∈ { 1 , … , K < N } \forall k\in\{1,\ldots,K<N\} ∀ k ∈ { 1 , … , K < N } , it holds:
If m k = 1 m_k=1 m k = 1 , then there exists I k = { i } ∈ S I_k=\{i\}\in S I k = { i } ∈ S such that v i k ≠ 0 v_i^k\ne0 v i k = 0 .
If m k = 2 m_k=2 m k = 2 , there exists I k = { i , j } ⊂ S I_k=\{i,j\}\subset S I k = { i , j } ⊂ S defining A ( k ) = ( v i k 1 v i k 2 v j k 1 v j k 2 ) A^{(k)}=\begin{pmatrix}v_i^{k_1}&v_i^{k_2}\\v_j^{k_1}&v_j^{k_2}\end{pmatrix} A ( k ) = ( v i k 1 v j k 1 v i k 2 v j k 2 ) invertible.
In the light of Definition 3.1, we establish the following theorem on the main property of a strategic set of vertices. This property so-called state representative of healthy networks stipulates that, in the absence of disturbances, the states of the vertices of a strategic set determine uniquely the global state of the network. This is the healthy state of the network.
Theorem 3.2 Let X 0 ∈ R N X_0 \in\mathbb{R}^N X 0 ∈ R N and X ˉ 0 ∈ R N \bar{X}_0 \in\mathbb{R}^N X ˉ 0 ∈ R N be unknowns and T 0 ∈ ( 0 , T ) T^0 \in(0,T) T 0 ∈ ( 0 , T ) . Provided S S S is a strategic set of vertices, if the solution X R ( t ) = ( x 1 R ( t ) , … , x N R ( t ) ) T X^R(t) = (x^R_1(t), \ldots, x^R_N(t))^\mathsf{T} X R ( t ) = ( x 1 R ( t ) , … , x N R ( t ) ) T of the problem:
{ X ¨ R ( t ) + η X ˙ R ( t ) − Δ X R ( t ) = 0 in ( 0 , T 0 ) , X R ( 0 ) = X 0 and X ˙ R ( 0 ) = X ˉ 0 , (7) \begin{cases}
\ddot{X}^R(t) + \eta\dot{X}^R(t) - \Delta X^R(t) = 0 & \text{in } (0,T^0),\\
X^R(0) = X_0 \quad\text{and} \quad\dot{X}^R(0) = \bar{X}_0, \tag*{(7)}
\end{cases} { X ¨ R ( t ) + η X ˙ R ( t ) − Δ X R ( t ) = 0 X R ( 0 ) = X 0 and X ˙ R ( 0 ) = X ˉ 0 , in ( 0 , T 0 ) , (7) fulfills x n R ( t ) = 0 x^R_n(t) = 0 x n R ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) \forall t \in(0,T^0) ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S \forall n \in S ∀ n ∈ S , then the initial conditions X 0 = X ˉ 0 = 0 X_0 = \bar{X}_0 = 0 X 0 = X ˉ 0 = 0 in R N \mathbb{R}^N R N .
Proof. See the appendix.
Thus, assuming the monitored network remains disturbance free during a period of time ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) , an immediate consequence of Theorem 3.2 is the following corollary:
Corollary 3.3 Let T 0 ∈ ( 0 , T ) T^0 \in(0,T) T 0 ∈ ( 0 , T ) , S S S be strategic and F s o u r F^{sour} F so u r be known in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) . The data
d n ( t ) = x n ( t ) , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S , (8) d_n(t) = x_n(t), \ \forall t \in(0,T^0), \ \forall n \in S, \tag*{(8)} d n ( t ) = x n ( t ) , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S , (8) determines uniquely the solution X ( t ) = ( x 1 ( t ) , … , x N ( t ) ) T X(t) = (x_1(t), \ldots, x_N(t))^\mathsf{T} X ( t ) = ( x 1 ( t ) , … , x N ( t ) ) T of the problem:
{ X ¨ ( t ) + η X ˙ ( t ) − Δ X ( t ) = F s o u r ( t ) in ( 0 , T 0 ) , X ( 0 ) = X 0 and X ˙ ( 0 ) = X ˉ 0 , (9) \begin{cases}
\ddot{X}(t) + \eta\dot{X}(t) - \Delta X(t) = F^{sour}(t) & \text{in } (0,T^0),\\
X(0) = X_0 \quad\text{and} \quad\dot{X}(0) = \bar{X}_0, \tag*{(9)}
\end{cases} { X ¨ ( t ) + η X ˙ ( t ) − Δ X ( t ) = F so u r ( t ) X ( 0 ) = X 0 and X ˙ ( 0 ) = X ˉ 0 , in ( 0 , T 0 ) , (9) where the two initial conditions X 0 ∈ R N X_0 \in\mathbb{R}^N X 0 ∈ R N and X ˉ 0 ∈ R N \bar{X}_0 \in\mathbb{R}^N X ˉ 0 ∈ R N are unknown.
Proof. Given F s o u r F^{sour} F so u r in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) , let X i = ( x 1 i , … , x N i ) T X^i = (x^i_1, \ldots, x^i_N)^\mathsf{T} X i = ( x 1 i , … , x N i ) T be the solution of the problem (9) with the initial conditions X i ( 0 ) = X 0 i X^i(0) = X^i_0 X i ( 0 ) = X 0 i and X ˙ i ( 0 ) = X ˉ 0 i \dot{X}^i(0) = \bar{X}^i_0 X ˙ i ( 0 ) = X ˉ 0 i , for i = 1 , 2 i = 1, 2 i = 1 , 2 . We denote d n i ( t ) = x n i ( t ) d^i_n(t) = x^i_n(t) d n i ( t ) = x n i ( t ) , ∀ t ∈ ( 0 , T 0 ) \forall t \in(0,T^0) ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S \forall n \in S ∀ n ∈ S and Y = ( y 1 , … , y N ) T Y = (y_1, \ldots, y_N)^\mathsf{T} Y = ( y 1 , … , y N ) T defined by Y = X 2 − X 1 Y = X^2 - X^1 Y = X 2 − X 1 .
Then, assuming d n 2 ( t ) = d n 1 ( t ) d^2_n(t) = d^1_n(t) d n 2 ( t ) = d n 1 ( t ) , ∀ t ∈ ( 0 , T 0 ) \forall t \in(0,T^0) ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S \forall n \in S ∀ n ∈ S , it follows that the state Y Y Y satisfies y n ( t ) = 0 y_n(t) = 0 y n ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) \forall t \in(0,T^0) ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S \forall n \in S ∀ n ∈ S and solves the problem (7) with the initial conditions Y ( 0 ) = X 0 2 − X 0 1 Y(0) = X^2_0 - X^1_0 Y ( 0 ) = X 0 2 − X 0 1 and Y ˙ ( 0 ) = X ˉ 0 2 − X ˉ 0 1 \dot{Y}(0) = \bar{X}^2_0 - \bar{X}^1_0 Y ˙ ( 0 ) = X ˉ 0 2 − X ˉ 0 1 . Theorem 3.2 implies that X 0 2 = X 0 1 X^2_0 = X^1_0 X 0 2 = X 0 1 and X ˉ 0 2 = X ˉ 0 1 \bar{X}^2_0 = \bar{X}^1_0 X ˉ 0 2 = X ˉ 0 1 . Hence, Y = 0 Y = 0 Y = 0 in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) which means that X 2 = X 1 X^2 = X^1 X 2 = X 1 in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) .
Procedure for detecting disturbances
In the remainder of this paper, we assume the observation set S S S to be strategic and the following condition regarding unknown disturbances F d i s F^{dis} F d i s to hold
(C1) : ∃ T 0 ∈ ( 0 , T ) such that F d i s ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) . (10) \text{(C1)}:\quad\exists T^0 \in(0,T) \text{ such that } F^{dis}(t) = 0,\ \forall t \in(0,T^0). \tag*{(10)} (C1) : ∃ T 0 ∈ ( 0 , T ) such that F d i s ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) . (10)
Hence, according to Corollary 3.3, the data d n ( t ) = x n ( t ) d_n(t) = x_n(t) d n ( t ) = x n ( t ) , ∀ t ∈ ( 0 , T 0 ) \forall t \in(0,T^0) ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S \forall n \in S ∀ n ∈ S determines in a unique manner the unknown network initial conditions X ( 0 ) ∈ R N X(0) \in\mathbb{R}^N X ( 0 ) ∈ R N and X ˙ ( 0 ) ∈ R N \dot{X}(0) \in\mathbb{R}^N X ˙ ( 0 ) ∈ R N involved in (9). Afterwards, from solving the problem (9) with the already determined network initial conditions and T T T instead of T 0 T^0 T 0 , we compute the healthy network state X H = ( x 1 H , … , x N H ) T X^H=(x_1^H,\ldots,x_N^H)^{\mathsf T} X H = ( x 1 H , … , x N H ) T in ( 0 , T ) (0,T) ( 0 , T ) . Notice that X H X^H X H represents the network state throughout the entire monitoring period of time ( 0 , T ) (0,T) ( 0 , T ) if it would remain out of all disturbances.
Besides, we introduce the residual variable X R = X − X H X^R=X-X^H X R = X − X H in ( 0 , T ) (0,T) ( 0 , T ) . From assuming the condition (C1) holds true, it follows that X R = ( x 1 R , … , x N R ) T X^R=(x_1^R,\ldots,x_N^R)^{\mathsf T} X R = ( x 1 R , … , x N R ) T is subject to:
{ X ¨ R ( t ) + η X ˙ R ( t ) − Δ X R ( t ) = F d i s ( t ) in ( T 0 , T ) , X R ( T 0 ) = X ˙ R ( T 0 ) = 0. (11) \begin{cases}
\ddot{X}^R(t)+\eta\dot{X}^R(t)-\Delta X^R(t)=F^{dis}(t) & \text{in }(T^0,T),\\
X^R(T^0)=\dot{X}^R(T^0)=0. \tag*{(11)}
\end{cases} { X ¨ R ( t ) + η X ˙ R ( t ) − Δ X R ( t ) = F d i s ( t ) X R ( T 0 ) = X ˙ R ( T 0 ) = 0. in ( T 0 , T ) , (11)
Since x n R = x n − x n H , ∀ n ∈ V x_n^R=x_n-x_n^H,\ \forall n\in\mathcal{V} x n R = x n − x n H , ∀ n ∈ V , the problem (11) leads in particular to
∀ T ˉ ∈ ( T 0 , T ) , F d i s = 0 in ( T 0 , T ˉ ) ⟹ x n = x n H in ( T 0 , T ˉ ) , ∀ n ∈ S . (12) \forall\bar{T}\in(T^0,T),\quad F^{dis}=0\ \text{in }(T^0,\bar{T})\ \Longrightarrow\ x_n=x_n^H\ \text{in }(T^0,\bar{T}),\ \forall n\in\mathcal{S}. \tag*{(12)} ∀ T ˉ ∈ ( T 0 , T ) , F d i s = 0 in ( T 0 , T ˉ ) ⟹ x n = x n H in ( T 0 , T ˉ ) , ∀ n ∈ S . (12)
In view of (12) , we summarize the main steps defining the procedure we developed to detect disturbances in the following algorithm:
Algorithm 1: Detection of unknown disturbances.
Assume: S \mathcal{S} S a strategic set of vertices and condition (C1) holds.
Data: T 0 ∈ ( 0 , T ) T^0\in(0,T) T 0 ∈ ( 0 , T ) , F s o u r F^{sour} F so u r in ( 0 , T ) (0,T) ( 0 , T ) , d n ∈ S d_{n\in\mathcal{S}} d n ∈ S in ( 0 , T ) (0,T) ( 0 , T ) , Δ t > 0 \Delta t>0 Δ t > 0 and a tolerance ε > 0 \varepsilon>0 ε > 0 .
Begin
Use data d n ∈ S d_{n\in\mathcal{S}} d n ∈ S in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) to determine X ( 0 ) X(0) X ( 0 ) and X ˙ ( 0 ) \dot{X}(0) X ˙ ( 0 ) of the problem (9).
Solve (9) with T T T instead of T 0 T^0 T 0 to compute the healthy state X H X^H X H in ( 0 , T ) (0,T) ( 0 , T ) .
For T ˉ = T 0 \bar{T}=T^0 T ˉ = T 0 to T T T in steps of Δ t \Delta t Δ t .
▶ Compute the residual: R = 1 N S ∑ n ∈ S ∥ d n − x n H ∥ L 2 ( T ˉ , T ˉ + Δ t ) \mathcal{R}=\frac{1}{N_{\mathcal{S}}}\sum_{n\in\mathcal{S}}\left\|d_n-x_n^H\right\|_{L^2(\bar{T},\bar{T}+\Delta t)} R = N S 1 ∑ n ∈ S d n − x n H L 2 ( T ˉ , T ˉ + Δ t ) .
▶ If R > ε \mathcal{R}>\varepsilon R > ε , then detected disturbances in ( T ˉ , T ˉ + Δ t ) (\bar{T},\bar{T}+\Delta t) ( T ˉ , T ˉ + Δ t ) . Break.
End.
The statement (12) raises a question about the effectiveness of the detection procedure: If (12) is not an equivalence, then we might have x n = x n H , ∀ n ∈ S x_n=x_n^H,\ \forall n\in\mathcal{S} x n = x n H , ∀ n ∈ S in ( T 0 , T ˉ ) (T^0,\bar{T}) ( T 0 , T ˉ ) whereas F d i s ≠ 0 F^{dis}\ne0 F d i s = 0 in ( T 0 , T ˉ ) (T^0,\bar{T}) ( T 0 , T ˉ ) . To address this issue, note first that in practice detection is more useful when unknown disturbances F d i s = ( F 1 d i s , … , F N d i s ) T F^{dis}=(F_1^{dis},\ldots,F_N^{dis})^{\mathsf T} F d i s = ( F 1 d i s , … , F N d i s ) T affect inaccessible/non-observation network vertices n ∈ E = V − S n\in\mathcal{E}=\mathcal{V}-\mathcal{S} n ∈ E = V − S . Thus, in the remainder of the article, we consider that
( C 2 ) : F n d i s = 0 in ( 0 , T ) , ∀ n ∈ S . (13) (\mathrm{C2}):\quad F_n^{dis}=0\ \text{in }(0,T),\ \forall n\in\mathcal{S}. \tag*{(13)} ( C2 ) : F n d i s = 0 in ( 0 , T ) , ∀ n ∈ S . (13)
Provided (C2) holds, selecting from the N N N equations defining the problem (11) all those associated to labels of the observation vertices n ∈ S n\in\mathcal{S} n ∈ S , we obtain: For all t ∈ ( T 0 , T ) t\in(T^0,T) t ∈ ( T 0 , T ) ,
x ¨ n R ( t ) + η x ˙ n R ( t ) − ∑ m ∈ V Δ n m x m R ( t ) = 0 , ∀ n ∈ S ⟺ ∑ m ∈ E Δ n m x m R ( t ) = d n R ( t ) , ∀ n ∈ S \ddot{x}_n^R(t)+\eta\dot{x}_n^R(t)-\sum_{m\in\mathcal{V}}\Delta_{nm}x_m^R(t)=0,\ \forall n\in\mathcal{S}
\quad\Longleftrightarrow\quad
\sum_{m\in\mathcal{E}}\Delta_{nm}x_m^R(t)=d_n^R(t),\ \forall n\in\mathcal{S} x ¨ n R ( t ) + η x ˙ n R ( t ) − m ∈ V ∑ Δ nm x m R ( t ) = 0 , ∀ n ∈ S ⟺ m ∈ E ∑ Δ nm x m R ( t ) = d n R ( t ) , ∀ n ∈ S
where d n R ( t ) = x ¨ n R ( t ) + η x ˙ n R ( t ) − ∑ m ∈ S Δ n m x m R ( t ) d_n^R(t)=\ddot{x}_n^R(t)+\eta\dot{x}_n^R(t)-\sum_{m\in\mathcal{S}}\Delta_{nm}x_m^R(t) d n R ( t ) = x ¨ n R ( t ) + η x ˙ n R ( t ) − ∑ m ∈ S Δ nm x m R ( t ) .
{#eq:14}
Let Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) be the N S × N E N_S \times N_E N S × N E submatrix obtained by selecting from the N × N N \times N N × N matrix Δ \Delta Δ the N S N_S N S rows associated to the labels of the observations vertices n ∈ S n \in\mathcal{S} n ∈ S and the N E N_E N E columns associated to the labels of the non-observation vertices m ∈ E m \in\mathcal{E} m ∈ E . Then, (14) reads
Δ ( S ; E ) X R E ( t ) = d R ( t ) , ∀ t ∈ ( T 0 , T ) , (14) \Delta(\mathcal{S};\mathcal{E})X^{R\mathcal{E}}(t)=d^R(t),\quad\forall t\in(T^0,T), \tag*{(14)} Δ ( S ; E ) X R E ( t ) = d R ( t ) , ∀ t ∈ ( T 0 , T ) , (14)
where d R ∈ R N S d^R \in\mathbb{R}^{N_S} d R ∈ R N S is defined by d n R d_n^R d n R , ∀ n ∈ S \forall n \in\mathcal{S} ∀ n ∈ S and X R E ∈ R N E X^{R\mathcal{E}} \in\mathbb{R}^{N_E} X R E ∈ R N E is given by x n R x_n^{R} x n R , ∀ n ∈ E \forall n \in\mathcal{E} ∀ n ∈ E . The effectiveness of the detection procedure relies on the properties of the matrix Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) involved in (14) . In the remainder, depending on the structure of the network, we consider the following two types of strategic observation sets. If the matrix Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) is full rank, S \mathcal{S} S is said to be dominantly absorbent . Otherwise, we will assume S \mathcal{S} S to be absorbent.
Inversion using dominantly absorbent observations
To ensure that the detection procedure is effective, we introduce the notion of dominantly absorbent set of vertices. Similarly to the strategic property that yields state representativity of healthy networks, the property dominantly absorbent ensures state representativity of contaminated networks in the sense that the assertion (12) holds as an equivalence. Thus, if x n ( t ) = x n H ( t ) x_n(t)=x_n^H(t) x n ( t ) = x n H ( t ) , ∀ t ∈ ( T 0 , T ‾ ) \forall t\in(T^0,\overline{T}) ∀ t ∈ ( T 0 , T ) for every vertex n n n of a dominantly absorbent observation set, then x n ( t ) = x n H ( t ) x_n(t)=x_n^H(t) x n ( t ) = x n H ( t ) , ∀ t ∈ ( T 0 , T ‾ ) \forall t\in(T^0,\overline{T}) ∀ t ∈ ( T 0 , T ) for all vertices n n n of the network. Hence, F d i s = 0 F^{dis}=0 F d i s = 0 in ( T 0 , T ‾ ) (T^0,\overline{T}) ( T 0 , T ) which means effectiveness of the detection procedure.
We start by introducing the notion of dominantly absorbent set of vertices. Then, we prove that in addition to making the detection effective, this new notion also ensures identifiability and leads to localize-identify all unknown detected disturbances.
Definition and practical guidelines
The notion of dominantly absorbent is defined as follows:
Definition 4.1 Let S \mathcal{S} S be a subset of N S N_S N S vertices selected from the N = N S + N E N=N_S+N_E N = N S + N E vertices defining the set V = S + E \mathcal{V}=\mathcal{S}+\mathcal{E} V = S + E of all network vertices. S \mathcal{S} S is dominantly absorbent if the N S × N E N_S \times N_E N S × N E matrix Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) introduced in (14)–(15) is of full rank.
We point out the following remarks
The properties strategic and dominantly absorbent differ in terms of their definitions and purposes. The notion of strategic depends on the spectral properties of the graph Laplacian matrix Δ \Delta Δ whereas the notion of dominantly absorbent depends on the structure of the matrix Δ \Delta Δ itself. As for purpose, the strategic property ensures uniqueness of the unknown network initial conditions for a given source term in the healthy problem (9) . The dominantly absorbent property
N S ≥ N E ⇔ N S ≥ N / 2. N_S \ge N_{\mathcal{E}} \Leftrightarrow N_S \ge N/2. N S ≥ N E ⇔ N S ≥ N /2.
If this were not the case, Δ ( S ; E ) \Delta(S;\mathcal{E}) Δ ( S ; E ) would not have full rank.
∀ m ∈ E , ∃ n ∈ S such that Δ n m ≠ 0 , (15) \forall m \in\mathcal{E},\ \exists n \in\mathcal{S}\ \text{ such that }\ \Delta_{nm} \ne0, \tag*{(15)} ∀ m ∈ E , ∃ n ∈ S such that Δ nm = 0 , (15)
Then, we establish the following result:
Proposition 4.2 Let V \mathcal{V} V be the set of all vertices defining a network whose graph Laplacian matrix is given in (2.1). All dominantly absorbent subset S ⊂ V \mathcal{S} \subset\mathcal{V} S ⊂ V is absorbent.
Proof. We prove Proposition 4.2 by contrapositive. Let S \mathcal{S} S be a subset defined by at least half of the vertices in V \mathcal{V} V . Assume that S \mathcal{S} S is not absorbent . Then, relation (16) implies:
∃ m 0 ∈ E such that ∀ n ∈ S , Δ n m 0 = 0. (16) \exists m_0 \in\mathcal{E}\ \text{ such that }\ \forall n \in\mathcal{S},\ \Delta_{nm_0}=0. \tag*{(16)} ∃ m 0 ∈ E such that ∀ n ∈ S , Δ n m 0 = 0. (16)
The assertion (17) implies that the column m 0 m_0 m 0 of the matrix Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) is null. Hence, Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) cannot be of full rank which implies that it is not dominantly absorbent . ■
In practice, selecting a dominantly absorbent observation set requires to choose N S ≥ N / 2 N_S \ge N/2 N S ≥ N /2 . This is a necessary condition for the matrix Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) to be of full rank. However, it is not sufficient. For this purpose, we recover a condition linked to the topology of the graph as in [10 ] . Indeed, for a particular graph that has a joint , we establish a result that guides the selection of the observation vertices.
We recall what is a joint vertex in a graph:
Definition 4.3 In a graph, a joint is a vertex whose removal increases the number of connected components.
Then on the selection of a dominantly absorbent set of vertices, we have:
Theorem 4.4 Consider a graph admitting a joint at a vertex k ≥ N / 2 k \ge N/2 k ≥ N /2 whose removal splits the graph vertices into two connected big and small sets. If the vertices defining S \mathcal{S} S are all taken from the big set, then the matrix Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) is not of full rank.
The proof of Theorem 4.4 is very similar to the one established in our previous work [10 ] .
To illustrate this result, we consider the 6-vertex graph admitting a joint at k = 3 k = 3 k = 3 shown in Figure 1 .
Figure 1. A graph with a joint.
The 6 × 6 6 \times6 6 × 6 graph Laplacian matrix is given by:
Δ = ( − 2 1 0 1 0 0 1 − 3 1 1 0 0 0 1 − 4 1 1 1 1 1 1 − 3 0 0 0 0 1 0 − 2 1 0 0 1 0 1 − 2 ) . (17) \Delta=
\begin{pmatrix}
-2 & 1 & 0 & 1 & 0 & 0 \\
1 & -3 & 1 & 1 & 0 & 0 \\
0 & 1 & -4 & 1 & 1 & 1 \\
1 & 1 & 1 & -3 & 0 & 0 \\
0 & 0 & 1 & 0 & -2 & 1 \\
0 & 0 & 1 & 0 & 1 & -2
\end{pmatrix}. \tag*{(17)} Δ = − 2 1 0 1 0 0 1 − 3 1 1 0 0 0 1 − 4 1 1 1 1 1 1 − 3 0 0 0 0 1 0 − 2 1 0 0 1 0 1 − 2 . (17)
The choice S = { 1 , 2 , 3 , 4 } \mathcal{S} = \{1, 2, 3, 4\} S = { 1 , 2 , 3 , 4 } and E = { 5 , 6 } \mathcal{E} = \{5, 6\} E = { 5 , 6 } is such that N S ≥ N / 2 N_{\mathcal{S}} \geq N/2 N S ≥ N /2 , however the matrix
Δ ( S ; E ) = ( 0 0 0 0 1 1 0 0 ) \Delta(\mathcal{S};\mathcal{E}) =
\begin{pmatrix}
0 & 0 \\
0 & 0 \\
1 & 1 \\
0 & 0
\end{pmatrix} Δ ( S ; E ) = 0 0 1 0 0 0 1 0
is not of full rank 2. On the other hand, ensuring that S \mathcal{S} S contains at least one vertex from each graph connected component, for example, S = { 1 , 2 , 5 } \mathcal{S} = \{1, 2, 5\} S = { 1 , 2 , 5 } and E = { 3 , 4 , 6 } \mathcal{E} = \{3, 4, 6\} E = { 3 , 4 , 6 } leads to:
Δ ( S ; E ) = ( 0 1 0 1 1 0 1 0 1 ) . (18) \Delta(\mathcal{S};\mathcal{E}) =
\begin{pmatrix}
0 & 1 & 0 \\
1 & 1 & 0 \\
1 & 0 & 1
\end{pmatrix}. \tag*{(18)} Δ ( S ; E ) = 0 1 1 1 1 0 0 0 1 . (18)
We have d e t ( Δ ( S ; E ) ) = − 1 det(\Delta(\mathcal{S};\mathcal{E})) = -1 d e t ( Δ ( S ; E )) = − 1 so that the the set S = { 1 , 2 , 5 } \mathcal{S} = \{1, 2, 5\} S = { 1 , 2 , 5 } is dominantly absorbent .
Identifiability-Identification of detected disturbances
In the light of Definition 4.1 , we establish the following result on the identifiability of unknown disturbances F d i s F^{\mathrm{dis}} F dis detected in ( T 0 , T ‾ ) (T^0,\overline{T}) ( T 0 , T ) affecting the source F = F s o u r F=F^{\mathrm{sour}} F = F sour of the problem (1) to become the undesired term F = F s o u r + F d i s F=F^{\mathrm{sour}}+F^{\mathrm{dis}} F = F sour + F dis in ( T 0 , T ‾ ) (T^0,\overline{T}) ( T 0 , T ) :
Theorem 4.5 Provided F d i s = ( F 1 d i s , … , F N d i s ) T F^{\mathrm{dis}}=(F_1^{\mathrm{dis}},\ldots,F_N^{\mathrm{dis}})^{\mathsf{T}} F dis = ( F 1 dis , … , F N dis ) T fulfills (C1) and (C2), let X = ( x 1 , … , x N ) T X=(x_1,\ldots,x_N)^{\mathsf{T}} X = ( x 1 , … , x N ) T be the solution of the problem (1) with F = F s o u r + F d i s F=F^{\mathrm{sour}}+F^{\mathrm{dis}} F = F sour + F dis . If the strategic set S S S defining M [ F ] M[F] M [ F ] in (2) is dominantly absorbent, then: the data
d n ( t ) = x n ( t ) , ∀ t ∈ ( 0 , T ) , ∀ n ∈ S , (19) d_n(t)=x_n(t),\ \forall t\in(0,T),\ \forall n\in S, \tag*{(19)} d n ( t ) = x n ( t ) , ∀ t ∈ ( 0 , T ) , ∀ n ∈ S , (19) determines in a unique manner all detected unknown disturbances F n d i s ∈ L 2 ( T 0 , T ) F_n^{\mathrm{dis}}\in L^2(T^0,T) F n dis ∈ L 2 ( T 0 , T ) .
Proof. Let X ( i ) = ( x 1 ( i ) , … , x N ( i ) ) T X^{(i)}=(x_1^{(i)},\ldots,x_N^{(i)})^{\mathsf{T}} X ( i ) = ( x 1 ( i ) , … , x N ( i ) ) T be the solution of the problem (1) with F ( i ) = F s o u r + F d i s ( i ) F^{(i)}=F^{\mathrm{sour}}+F^{\mathrm{dis}(i)} F ( i ) = F sour + F dis ( i ) in ( 0 , T ) (0,T) ( 0 , T ) , for i = 1 , 2 i=1,2 i = 1 , 2 . Let d n ( i ) ( t ) = x n ( i ) ( t ) , ∀ t ∈ ( 0 , T ) , ∀ n ∈ S d_n^{(i)}(t)=x_n^{(i)}(t),\forall t\in(0,T),\forall n\in S d n ( i ) ( t ) = x n ( i ) ( t ) , ∀ t ∈ ( 0 , T ) , ∀ n ∈ S and Y = ( y 1 , … , y N ) T Y=(y_1,\ldots,y_N)^{\mathsf{T}} Y = ( y 1 , … , y N ) T be the state defined by Y = X ( 2 ) − X ( 1 ) Y=X^{(2)}-X^{(1)} Y = X ( 2 ) − X ( 1 ) . Assume that
d n ( 2 ) ( t ) = d n ( 1 ) ( t ) , ∀ t ∈ ( 0 , T ) , ∀ n ∈ S . (20) d_n^{(2)}(t)=d_n^{(1)}(t),\ \forall t\in(0,T),\ \forall n\in S. \tag*{(20)} d n ( 2 ) ( t ) = d n ( 1 ) ( t ) , ∀ t ∈ ( 0 , T ) , ∀ n ∈ S . (20)
Provided F d i s ( i = 1 , 2 ) F^{\mathrm{dis}(i=1,2)} F dis ( i = 1 , 2 ) fulfill (C1), the state Y Y Y solves the problem (7) with the initial conditions Y ( 0 ) = X ( 2 ) ( 0 ) − X ( 1 ) ( 0 ) Y(0)=X^{(2)}(0)-X^{(1)}(0) Y ( 0 ) = X ( 2 ) ( 0 ) − X ( 1 ) ( 0 ) and Y ˙ ( 0 ) = X ˙ ( 2 ) ( 0 ) − X ˙ ( 1 ) ( 0 ) \dot{Y}(0)=\dot{X}^{(2)}(0)-\dot{X}^{(1)}(0) Y ˙ ( 0 ) = X ˙ ( 2 ) ( 0 ) − X ˙ ( 1 ) ( 0 ) . Since the set S S S is strategic and (21) leads to y n ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S y_n(t)=0,\ \forall t\in(0,T^0),\forall n\in S y n ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S , from applying Theorem 3.2 it follows that Y ( T 0 ) = Y ˙ ( T 0 ) = 0 Y(T^0)=\dot{Y}(T^0)=0 Y ( T 0 ) = Y ˙ ( T 0 ) = 0 . Therefore, the state Y Y Y is subject to:
{ Y ¨ ( t ) + η Y ˙ ( t ) − Δ Y ( t ) = F d i s ( 2 ) ( t ) − F d i s ( 1 ) ( t ) in ( T 0 , T ) , Y ( T 0 ) = Y ˙ ( T 0 ) = 0. (21) \begin{cases}
\ddot{Y}(t)+\eta\dot{Y}(t)-\Delta Y(t)=F^{\mathrm{dis}(2)}(t)-F^{\mathrm{dis}(1)}(t)&\text{in }(T^0,T),\\
Y(T^0)=\dot{Y}(T^0)=0. \tag*{(21)}
\end{cases} { Y ¨ ( t ) + η Y ˙ ( t ) − Δ Y ( t ) = F dis ( 2 ) ( t ) − F dis ( 1 ) ( t ) Y ( T 0 ) = Y ˙ ( T 0 ) = 0. in ( T 0 , T ) , (21)
From (21), it follows that y n ( t ) = 0 , ∀ t ∈ ( T 0 , T ) , ∀ n ∈ S y_n(t)=0,\ \forall t\in(T^0,T),\forall n\in S y n ( t ) = 0 , ∀ t ∈ ( T 0 , T ) , ∀ n ∈ S . Provided F d i s ( i = 1 , 2 ) F^{\mathrm{dis}(i=1,2)} F dis ( i = 1 , 2 ) fulfill (C2), by applying to the problem (22) similar techniques as in (14)-(15) we get
Δ ( S ; E ) Y E ( t ) = 0 , ∀ t ∈ ( T 0 , T ) , (22) \Delta(S;\mathcal{E})Y^{\mathcal{E}}(t)=0,\ \forall t\in(T^0,T), \tag*{(22)} Δ ( S ; E ) Y E ( t ) = 0 , ∀ t ∈ ( T 0 , T ) , (22)
where Y E ∈ I R N E Y^{\mathcal{E}}\in IR^{N^{\mathcal{E}}} Y E ∈ I R N E defined by y n , ∀ n ∈ E y_n,\forall n\in\mathcal{E} y n , ∀ n ∈ E . Since the set S S S is dominantly absorbent, the matrix Δ ( S ; E ) \Delta(S;\mathcal{E}) Δ ( S ; E ) is of full rank. Hence, (23) implies that Y E = 0 Y^{\mathcal{E}}=0 Y E = 0 in ( T 0 , T ) (T^0,T) ( T 0 , T ) which leads to Y = 0 Y=0 Y = 0 in ( T 0 , T ) (T^0,T) ( T 0 , T ) . From the problem (22), it follows that F d i s ( 2 ) = F d i s ( 1 ) F^{\mathrm{dis}(2)}=F^{\mathrm{dis}(1)} F dis ( 2 ) = F dis ( 1 ) in ( T 0 , T ) (T^0,T) ( T 0 , T ) . ■
To identify the unknown disturbances F d i s F^{\mathrm{dis}} F dis detected in ( T 0 , T ‾ ) (T^0,\overline{T}) ( T 0 , T ) , we recall the residual state X R = X − X H X^R=X-X^H X R = X − X H that solves the problem (11) . Since its part X R E X^{R\mathcal{E}} X R E containing the unknown local residuals x n R , ∀ n ∈ E x_n^R,\forall n\in\mathcal{E} x n R , ∀ n ∈ E solves the linear system (15), we proceed as follows:
Algorithm2: Identification using a dominantly absorbent observation set.
Assume: S S S is strategic and dominantly absorbent and conditions (C1) and (C2) hold.
Data: First detection instant T ‾ ∈ ( T 0 , T ) \overline{T}\in(T^0,T) T ∈ ( T 0 , T ) and the observations d n ∈ S d_{n\in S} d n ∈ S in ( T ‾ , T ) (\overline{T},T) ( T , T ) . Begin
From (14), use the data d n ∈ S d_n \in\mathcal{S} d n ∈ S in ( T ˉ , T ) (\bar{T},T) ( T ˉ , T ) to compute d n R d_n^R d n R , ∀ n ∈ S \forall n \in\mathcal{S} ∀ n ∈ S in ( T ˉ , T ) (\bar{T},T) ( T ˉ , T ) .
Solve the linear system (15) to determine X R ε X^{R\varepsilon} X R ε in ( T ˉ , T ) (\bar{T},T) ( T ˉ , T ) .
Using the determined X R ε X^{R\varepsilon} X R ε in ( T ˉ , T ) (\bar{T},T) ( T ˉ , T ) , reconstruct from (11) the unknown F d i s F^{dis} F d i s .
End.
Algorithm2 leads to determine F d i s F^{dis} F d i s almost in real time, depending on the sampling period Δ t \Delta t Δ t . Provided the invertible matrix extracted from Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) is well conditioned, we expect the reconstruction to be accurate.
blackInversion using absorbent observations
We address the inverse problem where the number of observation vertices is very limited smaller than half of the total network vertices. Hence, the strategic set S \mathcal{S} S is not anymore dominantly absorbent in the sense of Definition 4.1. Consequently, the matrix Δ ( S ; E ) \Delta(\mathcal{S};\mathcal{E}) Δ ( S ; E ) defining the linear system (15) can’t be of full rank which implies no guarantee about the existence/uniqueness of its solution. This raises a detection effectiveness issue since the assertion (12) might not hold as equivalence as well as an identifiability-identification issue because the solutions of (23) and (15) might be not what we expect.
In practice, disturbances usually bring the network out of its operational functioning state and always end up by affecting all of its vertices. Therefore, we believe the developed procedure to detect disturbances remains operational but it might deliver a delayed detection of the first instant disturbances. Moreover, failing the dominantly absorbent condition, we request the strategic set of observation vertices S \mathcal{S} S to be at least absorbent .
In the remainder, we assume (C1), (C2) hold and the strategic set of vertices S \mathcal{S} S to be not dominantly absorbent but rather absorbent . We develop an identification method to determine the unknown disturbances F d i s F^{dis} F d i s detected at T ˉ ∈ ( T o , T ) \bar{T} \in(T^{o},T) T ˉ ∈ ( T o , T ) . However, T ˉ \bar{T} T ˉ obtained from Algorithm1 with S \mathcal{S} S absorbent might be a delayed first instant of disturbances. Hence, we consider T ˉ k ∈ ( T ˉ , T ) \bar{T}_k \in(\bar{T},T) T ˉ k ∈ ( T ˉ , T ) and focus on identifying F d i s F^{dis} F d i s in ( T o , T ˉ k ) (T^{o},\bar{T}_k) ( T o , T ˉ k ) . For this purpose, we need an orthogonal family of L 2 ( 0 , T ˉ k ) L^2(0,\bar{T}_k) L 2 ( 0 , T ˉ k ) to expand the unknown states x m R x_m^R x m R , ∀ m ∈ E \forall m \in\mathcal{E} ∀ m ∈ E .
Orthogonal family of Legendre polynomials
We recall the Legendre orthogonal polynomials in L 2 ( − 1 , 1 ) L^2(-1,1) L 2 ( − 1 , 1 ) defined by the following recurrence formula: For all s ∈ ( − 1 , 1 ) s \in(-1,1) s ∈ ( − 1 , 1 ) ,
P ℓ + 1 ( s ) = 2 ℓ + 1 ℓ + 1 s P ℓ ( s ) − ℓ ℓ + 1 P ℓ − 1 ( s ) , ∀ ℓ ≥ 1 , P 0 ( s ) = 1 and P 1 ( s ) = s . (23) \begin{aligned}
P_{\ell+1}(s) &= \frac{2\ell+1}{\ell+1}sP_\ell(s)-\frac{\ell}{\ell+1}P_{\ell-1}(s), \quad\forall\ell\ge1,\\
P_0(s) &= 1 \quad\text{and} \quad P_1(s)=s. \tag*{(23)}
\end{aligned} P ℓ + 1 ( s ) P 0 ( s ) = ℓ + 1 2 ℓ + 1 s P ℓ ( s ) − ℓ + 1 ℓ P ℓ − 1 ( s ) , ∀ ℓ ≥ 1 , = 1 and P 1 ( s ) = s . (23)
Then, given T ‾ k ∈ ( T ‾ , T ) \overline{T}_k \in(\overline{T},T) T k ∈ ( T , T ) we employ the following change of variables:
t = T ‾ k 2 ( s + 1 ) and P ℓ ( t ) = P ℓ ( s ) = P ℓ ( 2 T ‾ k t − 1 ) , (24) t = \frac{\overline{T}_k}{2}(s+1) \qquad\text{and} \qquad\mathcal{P}_{\ell}(t) = P_{\ell}(s) = P_{\ell}\left(\frac{2}{\overline{T}_k}t-1\right), \tag*{(24)} t = 2 T k ( s + 1 ) and P ℓ ( t ) = P ℓ ( s ) = P ℓ ( T k 2 t − 1 ) , (24)
for all s ∈ ( − 1 , 1 ) s \in(-1,1) s ∈ ( − 1 , 1 ) and ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 . It follows that the sequence of Legendre’s polynomials { P ℓ } ℓ ≥ 0 \{\mathcal{P}_{\ell}\}_{\ell\ge0} { P ℓ } ℓ ≥ 0 defined by (24)–(25) forms an orthogonal basis of L 2 ( 0 , T ‾ k ) L^2(0,\overline{T}_k) L 2 ( 0 , T k ) , see [15 , 16 ] . For instance, the first four Legendre’s polynomials are given by: For all t ∈ ( 0 , T ‾ k ) t \in(0,\overline{T}_k) t ∈ ( 0 , T k ) ,
P 0 ( t ) = 1 , P 1 ( t ) = 2 T ‾ k t − 1 , \mathcal{P}_0(t)=1,\qquad\mathcal{P}_1(t)=\frac{2}{\overline{T}_k}t-1, P 0 ( t ) = 1 , P 1 ( t ) = T k 2 t − 1 ,
P 2 ( t ) = 3 2 ( 2 T ‾ k t − 1 ) 2 − 1 2 , \mathcal{P}_2(t)=\frac{3}{2}\left(\frac{2}{\overline{T}_k}t-1\right)^2-\frac{1}{2}, P 2 ( t ) = 2 3 ( T k 2 t − 1 ) 2 − 2 1 ,
P 3 ( t ) = 5 2 ( 2 T ‾ k t − 1 ) 3 − 3 2 ( 2 T ‾ k t − 1 ) . (25) \mathcal{P}_3(t)=\frac{5}{2}\left(\frac{2}{\overline{T}_k}t-1\right)^3-\frac{3}{2}\left(\frac{2}{\overline{T}_k}t-1\right). \tag*{(25)} P 3 ( t ) = 2 5 ( T k 2 t − 1 ) 3 − 2 3 ( T k 2 t − 1 ) . (25)
We set the local states x m R ∈ L 2 ( 0 , T ‾ k ) , ∀ m ∈ E x_m^R \in L^2(0,\overline{T}_k),\ \forall m \in\mathcal{E} x m R ∈ L 2 ( 0 , T k ) , ∀ m ∈ E of the residual variable X R = X − X H X^R=X-X^H X R = X − X H to:
x m R ( t ) = ∑ ℓ = 0 L x ‾ m ℓ P ℓ ( t ) , ∀ t ∈ ( 0 , T ‾ k ) , (26) x_m^R(t)=\sum_{\ell=0}^{L}\overline{x}_m^\ell\mathcal{P}_{\ell}(t),\qquad\forall t\in(0,\overline{T}_k), \tag*{(26)} x m R ( t ) = ℓ = 0 ∑ L x m ℓ P ℓ ( t ) , ∀ t ∈ ( 0 , T k ) , (26)
where x ‾ m ℓ = 1 ∥ P ℓ ∥ L 2 ( 0 , T ‾ k ) 2 ∫ 0 T ‾ k x m R ( t ) P ℓ ( t ) d t \overline{x}_m^\ell=\frac{1}{\|\mathcal{P}_{\ell}\|_{L^2(0,\overline{T}_k)}^2}\int_0^{\overline{T}_k}x_m^R(t)\mathcal{P}_{\ell}(t)\,dt x m ℓ = ∥ P ℓ ∥ L 2 ( 0 , T k ) 2 1 ∫ 0 T k x m R ( t ) P ℓ ( t ) d t and ∥ P ℓ ∥ L 2 ( 0 , T ‾ k ) 2 = T ‾ k 2 ℓ + 1 \|\mathcal{P}_{\ell}\|_{L^2(0,\overline{T}_k)}^2=\frac{\overline{T}_k}{2\ell+1} ∥ P ℓ ∥ L 2 ( 0 , T k ) 2 = 2 ℓ + 1 T k , for ℓ = 0 , … , L \ell=0,\ldots,L ℓ = 0 , … , L .
Localization-Identification of detected disturbances
We consider the integers 1 < I 0 < I ‾ < I 1<I_0<\overline{I}<I 1 < I 0 < I < I and the time step size Δ t = T / I \Delta t=T/I Δ t = T / I to define the discrete times t i = i Δ t t_i=i\Delta t t i = i Δ t , for i = 1 , … , I i=1,\ldots,I i = 1 , … , I with T 0 = I 0 Δ t T^0=I_0\Delta t T 0 = I 0 Δ t and T ‾ = I ‾ Δ t \overline{T}=\overline{I}\Delta t T = I Δ t . Then, we set T ‾ k = T ‾ + k Δ t \overline{T}_k=\overline{T}+k\Delta t T k = T + k Δ t , where k ∈ { 1 , … , I − I ‾ } k\in\{1,\ldots,I-\overline{I}\} k ∈ { 1 , … , I − I } with I ‾ k = I ‾ + k \overline{I}_k=\overline{I}+k I k = I + k and focus on identifying F d i s F^{dis} F d i s in ( T 0 , T ‾ k ) (T^0,\overline{T}_k) ( T 0 , T k ) . From using (27) in the linear system (14) taken at t i = I 0 + 1 , … , I ‾ k t_i=I_0+1,\ldots,\overline{I}_k t i = I 0 + 1 , … , I k , we get:
∑ m ∈ E ∑ ℓ = 0 L Δ n m P ℓ ( t i ) x ‾ m ℓ = d n R ( t i ) , ∀ n ∈ S , for i = I 0 + 1 , … , I ‾ k . (27) \sum_{m\in\mathcal{E}}\sum_{\ell=0}^{L}\Delta_{nm}\mathcal{P}_{\ell}(t_i)\overline{x}_m^\ell=d_n^R(t_i),\qquad\forall n\in\mathcal{S},\quad\text{for }i=I_0+1,\ldots,\overline{I}_k. \tag*{(27)} m ∈ E ∑ ℓ = 0 ∑ L Δ nm P ℓ ( t i ) x m ℓ = d n R ( t i ) , ∀ n ∈ S , for i = I 0 + 1 , … , I k . (27)
For the sake of clarity, we label the N S N_{\mathcal{S}} N S observation vertices of the set S = { n 1 , … , n N S } \mathcal{S}=\{n_1,\ldots,n_{N_{\mathcal{S}}}\} S = { n 1 , … , n N S } and the N E N_{\mathcal{E}} N E non-observation vertices of the set E = { m 1 , … , m N E } \mathcal{E}=\{m_1,\ldots,m_{N_{\mathcal{E}}}\} E = { m 1 , … , m N E } . Afterwards, the linear system (28) can be written under the following matrix form:
X R = D R , where X R = ( x ‾ m 1 0 ⋮ x ‾ m 1 L ⋮ ⋮ x ‾ m N E 0 ⋮ x ‾ m N E L ) ∈ R ( L + 1 ) × N E , D R = ( d n 1 R ( t I 0 + 1 ) ⋮ d n 1 R ( t I ‾ k ) ⋮ ⋮ d n N S R ( t I 0 + 1 ) ⋮ d n N S R ( t I ‾ k ) ) ∈ R I ‾ k × N S . (28) \mathcal{X}^R=\mathcal{D}^R,\qquad\text{where }\mathcal{X}^R=
\begin{pmatrix}
\overline{x}_{m_1}^{0}\\
\vdots\\
\overline{x}_{m_1}^{L}\\
\vdots\\
\vdots\\
\overline{x}_{m_{N_{\mathcal{E}}}}^{0}\\
\vdots\\
\overline{x}_{m_{N_{\mathcal{E}}}}^{L}
\end{pmatrix}
\in\mathbb{R}^{(L+1)\times N_{\mathcal{E}}},\qquad
\mathcal{D}^R=
\begin{pmatrix}
d_{n_1}^R(t_{I_0+1})\\
\vdots\\
d_{n_1}^R(t_{\overline{I}_k})\\
\vdots\\
\vdots\\
d_{n_{N_{\mathcal{S}}}}^R(t_{I_0+1})\\
\vdots\\
d_{n_{N_{\mathcal{S}}}}^R(t_{\overline{I}_k})
\end{pmatrix}
\in\mathbb{R}^{\overline{I}_k\times N_{\mathcal{S}}}. \tag*{(28)} X R = D R , where X R = x m 1 0 ⋮ x m 1 L ⋮ ⋮ x m N E 0 ⋮ x m N E L ∈ R ( L + 1 ) × N E , D R = d n 1 R ( t I 0 + 1 ) ⋮ d n 1 R ( t I k ) ⋮ ⋮ d n N S R ( t I 0 + 1 ) ⋮ d n N S R ( t I k ) ∈ R I k × N S . (28)
I ^ = I ˉ k ˉ − I 0 \hat{I}=\bar{I}_{\bar{k}}-I_0 I ^ = I ˉ k ˉ − I 0 and the ( I ^ × N S ) × ( ( L + 1 ) × N E ) (\hat{I}\times N_S)\times((L+1)\times N_{\mathcal{E}}) ( I ^ × N S ) × (( L + 1 ) × N E ) matrix C \mathcal{C} C is defined by
C = ( Δ n 1 m 1 P 0 ( t 0 + 1 ) ⋯ Δ n 1 m 1 P L ( t 0 + 1 ) ⋯ Δ n 1 m N E P 0 ( t 0 + 1 ) ⋯ Δ n 1 m N E P L ( t 0 + 1 ) ⋮ ⋮ ⋮ ⋮ Δ n 1 m 1 P 0 ( t k ˉ ) ⋯ Δ n 1 m 1 P L ( t k ˉ ) ⋯ Δ n 1 m N E P 0 ( t k ˉ ) ⋯ Δ n 1 m N E P L ( t k ˉ ) ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ Δ n N S m 1 P 0 ( t 0 + 1 ) ⋯ Δ n N S m 1 P L ( t 0 + 1 ) ⋯ Δ n N S m N E P 0 ( t 0 + 1 ) ⋯ Δ n N S m N E P L ( t 0 + 1 ) ⋮ ⋮ ⋮ ⋮ Δ n N S m 1 P 0 ( t k ˉ ) ⋯ Δ n N S m 1 P L ( t k ˉ ) ⋯ Δ n N S m N E P 0 ( t k ˉ ) ⋯ Δ n N S m N E P L ( t k ˉ ) ) (29) \mathcal{C}=\begin{pmatrix}
\Delta_{n_1m_1}\mathcal{P}_0(t_{0+1}) & \cdots& \Delta_{n_1m_1}\mathcal{P}_L(t_{0+1}) & \cdots& \Delta_{n_1m_{N_{\mathcal{E}}}}\mathcal{P}_0(t_{0+1}) & \cdots& \Delta_{n_1m_{N_{\mathcal{E}}}}\mathcal{P}_L(t_{0+1})\\
\vdots& & \vdots& & \vdots& & \vdots\\
\Delta_{n_1m_1}\mathcal{P}_0(t_{\bar{k}}) & \cdots& \Delta_{n_1m_1}\mathcal{P}_L(t_{\bar{k}}) & \cdots& \Delta_{n_1m_{N_{\mathcal{E}}}}\mathcal{P}_0(t_{\bar{k}}) & \cdots& \Delta_{n_1m_{N_{\mathcal{E}}}}\mathcal{P}_L(t_{\bar{k}})\\
\vdots& & \vdots& & \vdots& & \vdots\\
\vdots& & \vdots& & \vdots& & \vdots\\
\Delta_{n_{N_S}m_1}\mathcal{P}_0(t_{0+1}) & \cdots& \Delta_{n_{N_S}m_1}\mathcal{P}_L(t_{0+1}) & \cdots& \Delta_{n_{N_S}m_{N_{\mathcal{E}}}}\mathcal{P}_0(t_{0+1}) & \cdots& \Delta_{n_{N_S}m_{N_{\mathcal{E}}}}\mathcal{P}_L(t_{0+1})\\
\vdots& & \vdots& & \vdots& & \vdots\\
\Delta_{n_{N_S}m_1}\mathcal{P}_0(t_{\bar{k}}) & \cdots& \Delta_{n_{N_S}m_1}\mathcal{P}_L(t_{\bar{k}}) & \cdots& \Delta_{n_{N_S}m_{N_{\mathcal{E}}}}\mathcal{P}_0(t_{\bar{k}}) & \cdots& \Delta_{n_{N_S}m_{N_{\mathcal{E}}}}\mathcal{P}_L(t_{\bar{k}})
\end{pmatrix} \tag*{(29)} C = Δ n 1 m 1 P 0 ( t 0 + 1 ) ⋮ Δ n 1 m 1 P 0 ( t k ˉ ) ⋮ ⋮ Δ n N S m 1 P 0 ( t 0 + 1 ) ⋮ Δ n N S m 1 P 0 ( t k ˉ ) ⋯ ⋯ ⋯ ⋯ Δ n 1 m 1 P L ( t 0 + 1 ) ⋮ Δ n 1 m 1 P L ( t k ˉ ) ⋮ ⋮ Δ n N S m 1 P L ( t 0 + 1 ) ⋮ Δ n N S m 1 P L ( t k ˉ ) ⋯ ⋯ ⋯ ⋯ Δ n 1 m N E P 0 ( t 0 + 1 ) ⋮ Δ n 1 m N E P 0 ( t k ˉ ) ⋮ ⋮ Δ n N S m N E P 0 ( t 0 + 1 ) ⋮ Δ n N S m N E P 0 ( t k ˉ ) ⋯ ⋯ ⋯ ⋯ Δ n 1 m N E P L ( t 0 + 1 ) ⋮ Δ n 1 m N E P L ( t k ˉ ) ⋮ ⋮ Δ n N S m N E P L ( t 0 + 1 ) ⋮ Δ n N S m N E P L ( t k ˉ ) (29)
According to condition (C1), the residual states x n R = 0 , ∀ n ∈ V x_n^R=0,\forall n\in\mathcal{V} x n R = 0 , ∀ n ∈ V in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) . We determine the unknown vector x R x^R x R subject to the linear system (28) –(29) from solving:
min x R ∈ R ( L + 1 ) × N E ζ ( x R ) : = 1 2 ∥ C x R − D R ∥ 2 2 + α 2 ∑ m = 1 m N E ∥ ∑ ℓ = 0 L x ˉ m ℓ P ℓ ∥ L 2 ( 0 , T 0 ) 2 , (30) \min_{x^R\in\mathbb{R}^{(L+1)\times N_{\mathcal{E}}}}\zeta(x^R):=\frac{1}{2}\left\|\mathcal{C}x^R-\mathcal{D}^R\right\|_2^2+\frac{\alpha}{2}\sum_{m=1}^{m_{N_{\mathcal{E}}}}\left\|\sum_{\ell=0}^{L}\bar{x}_m^\ell\mathcal{P}_\ell\right\|_{L^2(0,T^0)}^2, \tag*{(30)} x R ∈ R ( L + 1 ) × N E min ζ ( x R ) := 2 1 C x R − D R 2 2 + 2 α m = 1 ∑ m N E ℓ = 0 ∑ L x ˉ m ℓ P ℓ L 2 ( 0 , T 0 ) 2 , (30)
where ∥ ⋅ ∥ 2 \|\cdot\|_2 ∥ ⋅ ∥ 2 stands for the Euclidean norm and α > 0 \alpha>0 α > 0 is a regularization coefficient [17 ] . This coefficient has to be selected as a good compromise between taking into account the measurements during the time period of disturbances ( T 0 , T ˉ k ) (T^0,\bar{T}_k) ( T 0 , T ˉ k ) and requiring the states in the non-observation vertices to match their healthy references in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) .
Solving the minimization problem (30) leads to determine an approximation in ( T 0 , T ˉ k ) (T^0,\bar{T}_k) ( T 0 , T ˉ k ) of X R X^R X R the solution of the residual problem (11) . Then, by selecting from (11) all equations associated to the non-observation vertices, it follows that each detected unknown disturbance F m d i s F_m^{dis} F m d i s occurring in a non-observation vertex m ∈ E m\in\mathcal{E} m ∈ E is subject to:
F m d i s ( t ) = x ¨ m R ( t ) + η x ˙ m R ( t ) − ∑ n = 1 N Δ m n x n R ( t ) , ∀ t ∈ ( T 0 , T ˉ k ) . (31) F_m^{dis}(t)=\ddot{x}_m^R(t)+\eta\dot{x}_m^R(t)-\sum_{n=1}^{N}\Delta_{mn}x_n^R(t),\quad\forall t\in(T^0,\bar{T}_k). \tag*{(31)} F m d i s ( t ) = x ¨ m R ( t ) + η x ˙ m R ( t ) − n = 1 ∑ N Δ mn x n R ( t ) , ∀ t ∈ ( T 0 , T ˉ k ) . (31)
Besides, for i = I 0 + 1 , … , I ˉ k i=I_0+1,\ldots,\bar{I}_k i = I 0 + 1 , … , I ˉ k we denote by φ i \varphi_i φ i the impulse response solution of:
{ φ ¨ i ( t ) − η φ ˙ i ( t ) = δ ( t − t i ) in ( t i − 1 , t i + 1 ) , φ i ( t i − 1 ) = φ i ( t i + 1 ) = 0 , (32) \begin{cases}
\ddot{\varphi}_i(t)-\eta\dot{\varphi}_i(t)=\delta(t-t_i)&\text{in }(t_{i-1},t_{i+1}),\\
\varphi_i(t_{i-1})=\varphi_i(t_{i+1})=0, \tag*{(32)}
\end{cases} { φ ¨ i ( t ) − η φ ˙ i ( t ) = δ ( t − t i ) φ i ( t i − 1 ) = φ i ( t i + 1 ) = 0 , in ( t i − 1 , t i + 1 ) , (32)
where δ ( t − t i ) \delta(t-t_i) δ ( t − t i ) is the Dirac mass at t i t_i t i . Using the Heaviside function H \mathcal{H} H , it comes that
φ i ( t ) = e η ( t − t i ) − 1 η H ( t − t i ) − e η Δ t − 1 η ( e 2 η Δ t − 1 ) ( e η ( t − t i − 1 ) − 1 ) , ∀ t ∈ ( t i − 1 , t i + 1 ) . (33) \varphi_i(t)=\frac{e^{\eta(t-t_i)}-1}{\eta}\mathcal{H}(t-t_i)-\frac{e^{\eta\Delta t}-1}{\eta(e^{2\eta\Delta t}-1)}\left(e^{\eta(t-t_{i-1})}-1\right),\quad\forall t\in(t_{i-1},t_{i+1}). \tag*{(33)} φ i ( t ) = η e η ( t − t i ) − 1 H ( t − t i ) − η ( e 2 η Δ t − 1 ) e η Δ t − 1 ( e η ( t − t i − 1 ) − 1 ) , ∀ t ∈ ( t i − 1 , t i + 1 ) . (33)
Remark 5.1 For η Δ t \eta\Delta t η Δ t small enough, using e x − 1 ≈ x e^x-1\approx x e x − 1 ≈ x it follows from (33) that the impulse response φ i \varphi_i φ i can be approximated by its affine version:
φ i ( t ) ≈ { − 1 2 ( t − t i − 1 ) , ∀ t ∈ ( t i − 1 , t i ) , 1 2 ( t − t i + 1 ) , ∀ t ∈ ( t i , t i + 1 ) . (34) \varphi_i(t)\approx\begin{cases}
-\frac{1}{2}(t-t_{i-1}),&\forall t\in(t_{i-1},t_i),\\
\frac{1}{2}(t-t_{i+1}),&\forall t\in(t_i,t_{i+1}).
\end{cases} \tag*{(34)} φ i ( t ) ≈ { − 2 1 ( t − t i − 1 ) , 2 1 ( t − t i + 1 ) , ∀ t ∈ ( t i − 1 , t i ) , ∀ t ∈ ( t i , t i + 1 ) . (34)
Graphically, (34) is the triangle of base ( t i − 1 , t i + 1 ) (t_{i-1},t_{i+1}) ( t i − 1 , t i + 1 ) and isosceles at φ i ( t i ) = − Δ t / 2 \varphi_i(t_i)=-\Delta t/2 φ i ( t i ) = − Δ t /2 .
From multiplying (31) by φ i \varphi_i φ i and integrating by parts on ( t i − 1 , t i + 1 ) (t_{i-1},t_{i+1}) ( t i − 1 , t i + 1 ) , we obtain
∫ t i − 1 t i + 1 F m d i s ( t ) φ i ( t ) d t = x m R ( t i ) − [ x m R ( t ) φ ˙ i ( t ) ] t i − 1 t i + 1 − ∑ n = 1 N Δ m n ∫ t i − 1 t i + 1 x n R ( t ) φ i ( t ) d t . (35) \int_{t_{i-1}}^{t_{i+1}} F_m^{\mathrm{dis}}(t)\varphi_i(t)\,\mathrm{d}t=x_m^R(t_i)-\left[x_m^R(t)\dot{\varphi}_i(t)\right]_{t_{i-1}}^{t_{i+1}}-\sum_{n=1}^{N}\Delta_{mn}\int_{t_{i-1}}^{t_{i+1}}x_n^R(t)\varphi_i(t)\,\mathrm{d}t. \tag*{(35)} ∫ t i − 1 t i + 1 F m dis ( t ) φ i ( t ) d t = x m R ( t i ) − [ x m R ( t ) φ ˙ i ( t ) ] t i − 1 t i + 1 − n = 1 ∑ N Δ mn ∫ t i − 1 t i + 1 x n R ( t ) φ i ( t ) d t . (35)
For the purpose of applying an appropriate numerical integration formula to determine the unknown disturbances F m d i s ( t i ) F_m^{\mathrm{dis}}(t_i) F m dis ( t i ) from (35) , we establish the following result:
Proposition 5.2 Let a < b a<b a < b be two real numbers, t i = ( a + b ) / 2 t_i=(a+b)/2 t i = ( a + b ) /2 and φ i \varphi_i φ i be the function defined in (34) with ( a , b ) (a,b) ( a , b ) instead of ( t i − 1 , t i + 1 ) (t_{i-1},t_{i+1}) ( t i − 1 , t i + 1 ) . Then,
∫ a b g ( t ) φ i ( t ) d t = g ( a ) + 10 g ( t i ) + g ( b ) 12 φ i ( t i ) b − a 2 , (36) \int_a^b g(t)\varphi_i(t)\,\mathrm{d}t=\frac{g(a)+10g(t_i)+g(b)}{12}\varphi_i(t_i)\frac{b-a}{2}, \tag*{(36)} ∫ a b g ( t ) φ i ( t ) d t = 12 g ( a ) + 10 g ( t i ) + g ( b ) φ i ( t i ) 2 b − a , (36) for all g g g a real polynomial of degree smaller or equal to 2 in ( a , b ) (a,b) ( a , b ) .
Proof. See the appendix.
From employing the numerical integration formula (36) in the assertion (35) and assuming (C1) holds, it follows that each detected unknown disturbance F m d i s F_m^{\mathrm{dis}} F m dis is subject to:
F m d i s ( t i − 1 ) + 10 F m d i s ( t i ) + F m d i s ( t i + 1 ) = 12 Δ t φ i ( t i ) d m i , for i = I 0 + 1 , … , I ˉ k , F m d i s ( t I 0 ) = 0 , (39) \begin{aligned}
F_m^{\mathrm{dis}}(t_{i-1})+10F_m^{\mathrm{dis}}(t_i)+F_m^{\mathrm{dis}}(t_{i+1})&=\frac{12}{\Delta t\varphi_i(t_i)}d_m^i,\quad\text{for }i=I_0+1,\ldots,\bar I_k,\\
F_m^{\mathrm{dis}}(t_{I_0})&=0, \tag*{(39)}
\end{aligned} F m dis ( t i − 1 ) + 10 F m dis ( t i ) + F m dis ( t i + 1 ) F m dis ( t I 0 ) = Δ t φ i ( t i ) 12 d m i , for i = I 0 + 1 , … , I ˉ k , = 0 , (39)
where t I 0 = T 0 t_{I_0}=T^0 t I 0 = T 0 and the data d m i d_m^i d m i is given by
d m i = x m R ( t i ) − [ x m R ( t ) φ ˙ i ( t ) ] t i − 1 t i + 1 − Δ t φ i ( t i ) 12 ∑ n = 1 N Δ m n ( x n R ( t i − 1 ) + 10 x n R ( t i ) + x n R ( t i + 1 ) ) . (40) d_m^i=x_m^R(t_i)-\left[x_m^R(t)\dot{\varphi}_i(t)\right]_{t_{i-1}}^{t_{i+1}}-\frac{\Delta t\varphi_i(t_i)}{12}\sum_{n=1}^{N}\Delta_{mn}\left(x_n^R(t_{i-1})+10x_n^R(t_i)+x_n^R(t_{i+1})\right). \tag*{(40)} d m i = x m R ( t i ) − [ x m R ( t ) φ ˙ i ( t ) ] t i − 1 t i + 1 − 12 Δ t φ i ( t i ) n = 1 ∑ N Δ mn ( x n R ( t i − 1 ) + 10 x n R ( t i ) + x n R ( t i + 1 ) ) . (40)
The benefit of (39) –(40) , as opposed to (31) , is that they do not involve the derivatives of the residuals x n R x_n^R x n R . Since the residuals x m R x_m^R x m R in the non-observation vertices m ∈ E m\in\mathcal{E} m ∈ E are identified from the minimization problem (30) , the determination of their derivatives might add more computational errors and lead to misinterpretations.
Localization of detected disturbances
From (40)–(41), by assuming F n d i s ( t j ) = x n R ( t j ) = 0 F^{dis}_{n}(t_j)=x^{R}_{n}(t_j)=0 F n d i s ( t j ) = x n R ( t j ) = 0 , for n = 1 , … , N n=1,\ldots,N n = 1 , … , N and all j ≤ i j\le i j ≤ i whereas ∃ m 0 ∈ E \exists m_0\in\mathcal{E} ∃ m 0 ∈ E such that F m 0 d i s ( t i + 1 ) ≠ 0 F^{dis}_{m_0}(t_{i+1})\ne0 F m 0 d i s ( t i + 1 ) = 0 and F m d i s ( t i + 1 ) = 0 , ∀ m ≠ m 0 F^{dis}_{m}(t_{i+1})=0,\ \forall m\ne m_0 F m d i s ( t i + 1 ) = 0 , ∀ m = m 0 , it follows that
x m 0 R ( t i + 1 ) = C m 0 F m 0 d i s ( t i + 1 ) + ∑ n = 1 , n ≠ m 0 N Δ m 0 n x n R ( t i + 1 ) − Δ m 0 m 0 , x m R ( t i + 1 ) = C m ∑ n = 1 , n ≠ m N Δ m n x n R ( t i + 1 ) − Δ m m . (41) \begin{aligned}
x^{R}_{m_0}(t_{i+1})&=C_{m_0}\frac{F^{dis}_{m_0}(t_{i+1})+\sum_{\substack{n=1,n\ne m_0}}^{N}\Delta_{m_0n}x^{R}_{n}(t_{i+1})}{-\Delta_{m_0m_0}},\\
x^{R}_{m}(t_{i+1})&=C_m\frac{\sum_{\substack{n=1,n\ne m}}^{N}\Delta_{mn}x^{R}_{n}(t_{i+1})}{-\Delta_{mm}}. \tag*{(41)}
\end{aligned} x m 0 R ( t i + 1 ) x m R ( t i + 1 ) = C m 0 − Δ m 0 m 0 F m 0 d i s ( t i + 1 ) + ∑ n = 1 , n = m 0 N Δ m 0 n x n R ( t i + 1 ) , = C m − Δ mm ∑ n = 1 , n = m N Δ mn x n R ( t i + 1 ) . (41)
where C m = ( 1 + 12 φ ˙ i ( t i + 1 ) Δ m m φ i ( t i ) Δ t ) − 1 C_m=\left(1+\frac{12\dot{\varphi}_i(t_{i+1})}{\Delta_{mm}\varphi_i(t_i)\Delta t}\right)^{-1} C m = ( 1 + Δ mm φ i ( t i ) Δ t 12 φ ˙ i ( t i + 1 ) ) − 1 , ∀ m ∈ { 1 , … , N } \forall m\in\{1,\ldots,N\} ∀ m ∈ { 1 , … , N } . The analysis of (41) leads to:
The residual x m 0 R x^{R}_{m_0} x m 0 R in the vertex m 0 m_0 m 0 is the sum of both active and passive disturbances. The first is the actual disturbances F m 0 d i s ≠ 0 F^{dis}_{m_0}\ne0 F m 0 d i s = 0 affecting the vertex m 0 m_0 m 0 whereas the second is due to the coupling effects through the Laplacian matrix.
Since Δ m m ∈ { 0 , 1 } , ∀ n ≠ m \Delta_{mm}\in\{0,1\},\ \forall n\ne m Δ mm ∈ { 0 , 1 } , ∀ n = m and − Δ m m -\Delta_{mm} − Δ mm is the number of terms Δ m n = 1 \Delta_{mn}=1 Δ mn = 1 , the residual x m R x^{R}_{m} x m R in each vertex where F m d i s = 0 F^{dis}_{m}=0 F m d i s = 0 is defined only by the passive disturbance that is a normalized average of the residuals in all vertices directly connected to it.
Therefore, each vertex m 0 ∈ E m_0\in\mathcal{E} m 0 ∈ E hosting unknown actual disturbances is characterized by a residual x m 0 R x^{R}_{m_0} x m 0 R that, as soon as F m 0 d i s ( t ) ≠ 0 F^{dis}_{m_0}(t)\ne0 F m 0 d i s ( t ) = 0 , deviates from a common behavior followed by the residuals in all vertices m m m where F m d i s = 0 F^{dis}_{m}=0 F m d i s = 0 . That enables to localize m 0 m_0 m 0 in E \mathcal{E} E .
Identification of detected-localized disturbances
According to (40)–(41), from setting F m d i s ( t I ˉ k + 1 ) = F m d i s ( t I ˉ k ) F^{dis}_{m}(t_{\bar{I}_k+1})=F^{dis}_{m}(t_{\bar{I}_k}) F m d i s ( t I ˉ k + 1 ) = F m d i s ( t I ˉ k ) , it follows that the unknown detected disturbances F m d i s ( t i ) F^{dis}_{m}(t_i) F m d i s ( t i ) , for i = I 0 + 1 , … , I ˉ k i=I_0+1,\ldots,\bar{I}_k i = I 0 + 1 , … , I ˉ k are subject to:
\begin{pmatrix}
10 & 1 & 0 & 0 & 0 & \cdots& 0\\
1 & 10 & 1 & 0 & 0 & \cdots& 0\\
\vdots& \vdots& \vdots& \vdots& \vdots& & \vdots\\
0 & 0 & \cdots& 0 & 1 & 10 & 1\\
0 & 0 & \cdots& 0 & 0 & 1 & 11
\end{pmatrix}
\begin{pmatrix}
F^{dis}_{m}(t_{I_0+1})\\
\vdots\\
\vdots\\
\vdots\\
F^{dis}_{m}(t_{\bar{I}_k})
\end{pmatrix}
=
\frac{12}{\Delta t}
\begin{pmatrix}
d^{I_0+1}_{m}/\varphi_{I_0+1}(t_{I_0+1})\\
\vdots\\
\vdots\\
\vdots\\
d^{\bar{I}_k}_{m}/\varphi_{\bar{I}_k}(t_{\bar{I}_k})
\end{pmatrix}. \tag*{(42)}\end{pmatrix}
From solving the linear system (42) defined by a ( I ˉ k − I 0 ) × ( I ˉ k − I 0 ) (\bar{I}_k-I_0)\times(\bar{I}_k-I_0) ( I ˉ k − I 0 ) × ( I ˉ k − I 0 ) tridiagonal Toeplitz matrix, we determine the sought disturbances F m d i s ( t i ) F^{dis}_{m}(t_i) F m d i s ( t i ) , for i = I 0 + 1 , … , I ˉ k i=I_0+1,\ldots,\bar{I}_k i = I 0 + 1 , … , I ˉ k .
Algorithm3: Localization-Identification of detected disturbances
Assume: S \mathcal{S} S strategic and absorbent with conditions (C1) and (C2) hold.
Data: T 0 ∈ ( 0 , T ) T^{0} \in(0,T) T 0 ∈ ( 0 , T ) , T ‾ k ∈ ( T ‾ , T ) \overline{T}_{k} \in(\overline{T},T) T k ∈ ( T , T ) and the observations d n ∈ S d_{n\in\mathcal{S}} d n ∈ S in ( 0 , T ‾ k ) (0,\overline{T}_{k}) ( 0 , T k ) .
Begin
From (14), use d n ∈ S d_{n\in\mathcal{S}} d n ∈ S in ( T 0 , T ‾ k ) (T^{0},\overline{T}_{k}) ( T 0 , T k ) to compute d n R ( t i ) d_{n}^{R}(t_{i}) d n R ( t i ) , ∀ n ∈ S \forall n \in\mathcal{S} ∀ n ∈ S for i = I 0 + 1 , … , I ‾ k i = I_{0}+1,\ldots,\overline{I}_{k} i = I 0 + 1 , … , I k .
Solve the problem (31) to determine x m R ( t i ) x_{m}^{R}(t_{i}) x m R ( t i ) , ∀ m ∈ E \forall m \in\mathcal{E} ∀ m ∈ E for i = I 0 + 1 , … , I ‾ k i = I_{0}+1,\ldots,\overline{I}_{k} i = I 0 + 1 , … , I k .
Localize vertices hosting disturbances by analyzing the behavior of x m R x_{m}^{R} x m R , ∀ m ∈ E \forall m \in\mathcal{E} ∀ m ∈ E .
Identification: For each localized vertex m ∈ E m \in\mathcal{E} m ∈ E hosting unknown disturbances: ► From (41), compute d m i d_{m}^{i} d m i for i = I 0 + 1 , … , I ‾ k i = I_{0}+1,\ldots,\overline{I}_{k} i = I 0 + 1 , … , I k . ► Solve the linear system (43) to determine F m d i s ( t i ) F_{m}^{dis}(t_{i}) F m d i s ( t i ) , for i = I 0 + 1 , … , I ‾ k i = I_{0}+1,\ldots,\overline{I}_{k} i = I 0 + 1 , … , I k .
End.
Numerical experiments
We carry out numerical experiments on the 5–vertex graph shown in Figure 2 .
Figure 2. A graph with five vertices.
It’s graph Laplacian matrix is given by:
Δ = ( − 2 0 0 1 1 0 − 3 1 1 1 0 1 − 2 1 0 1 1 1 − 4 1 1 1 0 1 − 3 ) . \Delta=
\begin{pmatrix}
-2 & 0 & 0 & 1 & 1 \\
0 & -3 & 1 & 1 & 1 \\
0 & 1 & -2 & 1 & 0 \\
1 & 1 & 1 & -4 & 1 \\
1 & 1 & 0 & 1 & -3
\end{pmatrix}. Δ = − 2 0 0 1 1 0 − 3 1 1 1 0 1 − 2 1 0 1 1 1 − 4 1 1 1 0 1 − 3 .
Determination of the healthy network state
Provided (C1) holds, according to Corollary 3.3 the data d n ( t ) = x n ( t ) , ∀ n ∈ S d_n(t)=x_n(t),\forall n\in S d n ( t ) = x n ( t ) , ∀ n ∈ S in ( 0 , T 0 ) (0,T^0) ( 0 , T 0 ) determines uniquely the network unknown initial conditions X 0 = X ( 0 ) X_0=X(0) X 0 = X ( 0 ) and X ˙ 0 = X ˙ ( 0 ) \dot{X}_0=\dot{X}(0) X ˙ 0 = X ˙ ( 0 ) in the problem (9). We write the solution X X X of the problem (9) as follows:
X ( t ) = X 0 ( t ) + X F ( t ) , ∀ t ∈ ( 0 , T 0 ) , (43) X(t)=X^0(t)+X^F(t),\quad\forall t\in(0,T^0), \tag*{(43)} X ( t ) = X 0 ( t ) + X F ( t ) , ∀ t ∈ ( 0 , T 0 ) , (43)
where X 0 = ( x 1 0 , … , x N 0 ) T X^0=(x_1^0,\ldots,x_N^0)^\mathsf{T} X 0 = ( x 1 0 , … , x N 0 ) T and X F = ( x 1 F , … , x N F ) T X^F=(x_1^F,\ldots,x_N^F)^\mathsf{T} X F = ( x 1 F , … , x N F ) T are the solutions of:
{ X ¨ 0 ( t ) + η X ˙ 0 ( t ) − Δ X 0 ( t ) = 0 in ( 0 , T 0 ) , X 0 ( 0 ) = X 0 ∈ R N and X ˙ 0 ( 0 ) = X ˙ 0 ∈ R N , (44) \left\{
\begin{aligned}
\ddot{X}^0(t)+\eta\dot{X}^0(t)-\Delta X^0(t)&=0 &&\text{in }(0,T^0),\\
X^0(0)&=X_0\in\mathbb{R}^N &&\text{and}\quad\dot{X}^0(0)=\dot{X}_0\in\mathbb{R}^N,
\end{aligned}
\right. \tag*{(44)} { X ¨ 0 ( t ) + η X ˙ 0 ( t ) − Δ X 0 ( t ) X 0 ( 0 ) = 0 = X 0 ∈ R N in ( 0 , T 0 ) , and X ˙ 0 ( 0 ) = X ˙ 0 ∈ R N , (44)
and
{ X ¨ F ( t ) + η X ˙ F ( t ) − Δ X F ( t ) = F s o u r ( t ) in ( 0 , T 0 ) , X F ( 0 ) = 0 and X ˙ F ( 0 ) = 0. (45) \left\{
\begin{aligned}
\ddot{X}^F(t)+\eta\dot{X}^F(t)-\Delta X^F(t)&=F^{\mathsf{sour}}(t) &&\text{in }(0,T^0),\\
X^F(0)&=0 &&\text{and}\quad\dot{X}^F(0)=0.
\end{aligned}
\right. \tag*{(45)} { X ¨ F ( t ) + η X ˙ F ( t ) − Δ X F ( t ) X F ( 0 ) = F sour ( t ) = 0 in ( 0 , T 0 ) , and X ˙ F ( 0 ) = 0. (45)
We expand the unknown initial conditions of the problem (45) in the orthonormal basis of R N \mathbb{R}^N R N defined by the normalized eigenvectors v k ℓ v^{k\ell} v k ℓ of the graph Laplacian matrix Δ \Delta Δ :
X 0 = ∑ k = 1 K ∑ ℓ = 1 m k y k ℓ 0 v k ℓ and X ˙ 0 = ∑ k = 1 K ∑ ℓ = 1 m k y ˉ k ℓ 0 v k ℓ . (46) X_0=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}y_{k\ell}^0v^{k\ell}\quad\text{and}\quad\dot{X}_0=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}\bar{y}_{k\ell}^0v^{k\ell}. \tag*{(46)} X 0 = k = 1 ∑ K ℓ = 1 ∑ m k y k ℓ 0 v k ℓ and X ˙ 0 = k = 1 ∑ K ℓ = 1 ∑ m k y ˉ k ℓ 0 v k ℓ . (46)
Then, we establish the following result:
Proposition 6.1 Provided (47) holds, the solution of the problem (45) is given by
X 0 ( t ) = ∑ k = 1 K ∑ ℓ = 1 m k ( A k ( t ) y k ℓ 0 + B k ( t ) y ˉ k ℓ 0 ) v k ℓ , ∀ t ∈ ( 0 , T 0 ) , (47) X^0(t)=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}\left(A_k(t)y_{k\ell}^0+B_k(t)\bar{y}_{k\ell}^0\right)v^{k\ell},\quad\forall t\in(0,T^0), \tag*{(47)} X 0 ( t ) = k = 1 ∑ K ℓ = 1 ∑ m k ( A k ( t ) y k ℓ 0 + B k ( t ) y ˉ k ℓ 0 ) v k ℓ , ∀ t ∈ ( 0 , T 0 ) , (47) where the two functions A k A_k A k and $B_k are defined as follows:
A k ( t ) = { r ˉ k e r k t − r k e r ˉ k t r ˉ k − r k , if D k > 0 , ( 1 + η 2 t ) e − η 2 t , if D k = 0 , e − η 2 t ( cos ( − D k 2 t ) + η − D k sin ( − D k 2 t ) ) , if D k < 0 , (48) A_k(t)=
\begin{cases}
\dfrac{\bar{r}_k e^{r_k t}-r_k e^{\bar{r}_k t}}{\bar{r}_k-r_k}, & \text{if }D_k>0,\\
\left(1+\dfrac{\eta}{2}t\right)e^{-\frac{\eta}{2}t}, & \text{if }D_k=0,\\
e^{-\frac{\eta}{2}t}\left(\cos\left(\dfrac{\sqrt{-D_k}}{2}t\right)+\dfrac{\eta}{\sqrt{-D_k}}\sin\left(\dfrac{\sqrt{-D_k}}{2}t\right)\right), & \text{if }D_k<0,
\end{cases} \tag*{(48)} A k ( t ) = ⎩ ⎨ ⎧ r ˉ k − r k r ˉ k e r k t − r k e r ˉ k t , ( 1 + 2 η t ) e − 2 η t , e − 2 η t ( cos ( 2 − D k t ) + − D k η sin ( 2 − D k t ) ) , if D k > 0 , if D k = 0 , if D k < 0 , (48) B k ( t ) = { e r k t − e r ˉ k t r ˉ k − r k , if D k > 0 , t e − η 2 t , if D k = 0 , 2 e − η 2 t − D k sin ( − D k 2 t ) , if D k < 0 , (49) B_k(t)=
\begin{cases}
\dfrac{e^{r_k t}-e^{\bar{r}_k t}}{\bar{r}_k-r_k}, & \text{if }D_k>0,\\
te^{-\frac{\eta}{2}t}, & \text{if }D_k=0,\\
\dfrac{2e^{-\frac{\eta}{2}t}}{\sqrt{-D_k}}\sin\left(\dfrac{\sqrt{-D_k}}{2}t\right), & \text{if }D_k<0,
\end{cases} \tag*{(49)} B k ( t ) = ⎩ ⎨ ⎧ r ˉ k − r k e r k t − e r ˉ k t , t e − 2 η t , − D k 2 e − 2 η t sin ( 2 − D k t ) , if D k > 0 , if D k = 0 , if D k < 0 , (49) with the discriminant D k = η 2 − 4 ω k 2 D_k=\eta^2-4\omega_k^2 D k = η 2 − 4 ω k 2 and the roots r k = − η + D k 2 r_k=\dfrac{-\eta+\sqrt{D_k}}{2} r k = 2 − η + D k and r ˉ k = − η + D k 2 \bar{r}_k=\dfrac{-\eta+\sqrt{D_k}}{2} r ˉ k = 2 − η + D k .
Proof. See the appendix.
According to (47) , the identification of the unknown network initial conditions defining through (43) the state X X X solution of the problem (9) can be achieved by determining the coefficients y k ℓ 0 y_{k\ell}^{0} y k ℓ 0 and y ˉ k ℓ 0 \bar{y}_{k\ell}^{0} y ˉ k ℓ 0 in (46) . To this end, we solve the minimization problem:
min y k ℓ 0 , y ˉ k ℓ 0 J ( y k ℓ 0 , y ˉ k ℓ 0 ) : = 1 2 ∑ n ∈ S ∥ x n ( t ) − d n ( t ) ∥ L 2 ( 0 , T 0 ) 2 = 1 2 ∑ n ∈ S ∫ 0 T 0 ( ∑ k = 1 K ∑ ℓ = 1 m k ( A k ( t ) y k ℓ 0 + B k ( t ) y ˉ k ℓ 0 ) v n k ℓ + x n F ( t ) − d n ( t ) ) 2 d t , (50) \begin{aligned}
\min_{y_{k\ell}^{0},\bar{y}_{k\ell}^{0}} J(y_{k\ell}^{0},\bar{y}_{k\ell}^{0}) &:= \frac{1}{2}\sum_{n\in S}\left\|x_n(t)-d_n(t)\right\|_{L^2(0,T^0)}^2 \\
&= \frac{1}{2}\sum_{n\in S}\int_0^{T^0}\left(\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}\left(A_k(t)y_{k\ell}^{0}+B_k(t)\bar{y}_{k\ell}^{0}\right)v_n^{k\ell}+x_n^F(t)-d_n(t)\right)^2\,dt, \tag*{(50)}
\end{aligned} y k ℓ 0 , y ˉ k ℓ 0 min J ( y k ℓ 0 , y ˉ k ℓ 0 ) := 2 1 n ∈ S ∑ ∥ x n ( t ) − d n ( t ) ∥ L 2 ( 0 , T 0 ) 2 = 2 1 n ∈ S ∑ ∫ 0 T 0 ( k = 1 ∑ K ℓ = 1 ∑ m k ( A k ( t ) y k ℓ 0 + B k ( t ) y ˉ k ℓ 0 ) v n k ℓ + x n F ( t ) − d n ( t ) ) 2 d t , (50)
where from (43) , x n ( t ) = x n 0 ( t ) + x n F ( t ) x_n(t)=x_n^0(t)+x_n^F(t) x n ( t ) = x n 0 ( t ) + x n F ( t ) , ∀ n \forall n ∀ n .
We can write the gradient of the objective function J J J in (50) in the following matrix form:
∇ J ( Y ) = M Y + b . (51) \nabla J(Y)=MY+b. \tag*{(51)} ∇ J ( Y ) = M Y + b . (51)
where Y k + ℓ − 1 = y k ℓ 0 Y_{k+\ell-1}=y_{k\ell}^{0} Y k + ℓ − 1 = y k ℓ 0 , Y N + k + ℓ − 1 = y ˉ k ℓ 0 Y_{N+k+\ell-1}=\bar{y}_{k\ell}^{0} Y N + k + ℓ − 1 = y ˉ k ℓ 0 , k = 1 , … , K k=1,\ldots,K k = 1 , … , K , ℓ = 1 , … , m k \ell=1,\ldots,m_k ℓ = 1 , … , m k . The expressions of the matrix M M M and the right hand side b b b are given in the Appendix with more details.
Writing ∇ J = 0 \nabla J=0 ∇ J = 0 , we get the linear system M Y = − b MY=-b M Y = − b whose solution gives the network initial condition as follows:
X 0 = ∑ k = 1 K ∑ ℓ = 1 m k Y k + ℓ − 1 v k ℓ and X ˉ 0 = ∑ k = 1 K ∑ ℓ = 1 m k Y N + k + ℓ − 1 v ˉ k ℓ . (52) X_0=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}Y_{k+\ell-1}v^{k\ell}
\qquad\text{and}\qquad
\bar{X}_0=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}Y_{N+k+\ell-1}\bar{v}^{k\ell}. \tag*{(52)} X 0 = k = 1 ∑ K ℓ = 1 ∑ m k Y k + ℓ − 1 v k ℓ and X ˉ 0 = k = 1 ∑ K ℓ = 1 ∑ m k Y N + k + ℓ − 1 v ˉ k ℓ . (52)
Afterwards, from solving the forward problem (9) with the initial conditions obtained in (52) and T T T instead of T 0 T^0 T 0 , we determine the healthy network state X H X^H X H in ( 0 , T ) (0,T) ( 0 , T ) .
Inverse problem for dominantly absorbent vertices
We consider the network shown in Figure 2 . From the network shown in Figure 2 , we select the strategic and dominantly absorbent set of observation vertices S = { 1 , 4 , 5 } S=\{1,4,5\} S = { 1 , 4 , 5 } . That leaves the set of non-observation vertices E = { 2 , 3 } E=\{2,3\} E = { 2 , 3 } . Then, we set F s o u r ( t ) = 0 F^{sour}(t)=0 F so u r ( t ) = 0 and generate synthetic measurements from disturbances located in the vertex m = 2 m=2 m = 2 of intensity
F d i s ( t ) = sin π ( t − T 0 ) T , t > T 0 , 0 , otherwise F^{dis}(t)=\sin\frac{\pi(t-T_0)}{T},\quad t>T_0,\quad0,\ \text{otherwise} F d i s ( t ) = sin T π ( t − T 0 ) , t > T 0 , 0 , otherwise
where T = 100 T=100 T = 100 , T 0 = 10 T_0=10 T 0 = 10 , η = 0.1 \eta=0.1 η = 0.1 , Δ t = 0.1 \Delta t=0.1 Δ t = 0.1 . The initial condition of the system is X 0 = X ˉ 0 = 0 X_0=\bar{X}_0=0 X 0 = X ˉ 0 = 0 .
Figure 3 shows the reconstructed disturbances in vertices 2 and 3 (left panel) and the reconstructed and exact disturbances in vertices 2 and 3 (right panel). In the left panel, the reconstructed disturbance at vertex 3 is zero as expected. The disturbance at vertex 2 is non zero and is accurately reconstructed as seen in the right panel of Figure 3 . Indeed, the reconstructed disturbances are non-null only in the vertex m = 2 m = 2 m = 2 origin of these disturbances. Moreover, the reconstruction is accurate because the unmeasured residuals in the vertices E \mathcal{E} E are determined by solving the well posed linear system (15), where the full rank matrix is
Figure 3. Left panel: reconstructed disturbances in vertices 2 and 3, Right panel: reconstructed and exact disturbances in vertex 2.
Δ ( S ; E ) = ( 0 0 1 1 1 0 ) . \Delta(\mathcal{S};\mathcal{E}) =
\begin{pmatrix}
0 & 0 \\
1 & 1 \\
1 & 0
\end{pmatrix}. Δ ( S ; E ) = 0 1 1 0 1 0 .
Inverse problem in absorbent observation vertices
Now assume that the observation set is S = { 1 , 4 } \mathcal{S} = \{1, 4\} S = { 1 , 4 } , it is not anymore dominantly absorbent but only absorbent. The non observed vertices are E = { 2 , 3 , 5 } \mathcal{E} = \{2, 3, 5\} E = { 2 , 3 , 5 } . The reconstruction of the residuals x m R ( t ) x_m^R(t) x m R ( t ) , m = 2 , 3 , 5 m = 2, 3, 5 m = 2 , 3 , 5 is done using the minimization problem (31) with a regularization factor α = 10 − 2 \alpha= 10^{-2} α = 1 0 − 2 .
Figure 4 shows the time evolution of the residuals x m R ( t ) x_m^R(t) x m R ( t ) for a source term in vertices 2,3 or 5. These numerical results back our interpretations of the assertions (41) . Indeed, these results show that the residual in the vertex origin of disturbances is the first to become significantly non-zero and is clearly larger than the residuals in the other vertices wherein there are no disturbances. These elements define our procedure to localize disturbances using a non-dominantly absorbent observation set S S S . Notice that overtime, disturbances end up contaminating all the vertices of the network so that residuals cannot be used anymore to find the initial location and intensities of the detected disturbances. Therefore, we reconstruct residuals and disturbances in the interval ( T 0 , T k ) (T_0,T_k) ( T 0 , T k ) , where T k T_k T k is reasonably larger than the first instant of detection of disturbances.
Figure 4. Reconstructed Residuals x n R ( t ) x_n^R(t) x n R ( t ) for disturbances located in vertex: 2(left), 3(middle) and 5(right).
Figure 5 shows the reconstruction of the disturbance F m d i s ( t ) F_m^{dis}(t) F m d i s ( t ) once the source node has been found.
Figure 5. Reconstruction of disturbances F m d i s ( t ) F_m^{dis}(t) F m d i s ( t ) located in vertex: 2(left), 3(middle) and 5(right).
Note that F m d i s ( t ) F_m^{dis}(t) F m d i s ( t ) is better approximated for a source in vertices 3 and 5 than in vertex
Two factors could amplify the error on the residual in a vertex m m m of E \mathcal{E} E : the total number of network vertices directly connected to the vertex m m m and how many of these are not observed. Since the residual x m R x_m^{R} x m R is the major contributor in the computed data from (41), this perspective might explain the errors on the reconstructed disturbances F m d i s F_m^{dis} F m d i s presented in Figure 5 . Indeed, the smallest error is on F 3 d i s F_3^{dis} F 3 d i s since from Figure 2 vertex m = 3 m=3 m = 3 is directly connected to 2 vertices with only one of them being not observed. As expected, the largest error is for F 2 d i s F_2^{dis} F 2 d i s , because vertex m = 2 m=2 m = 2 is directly connected to 3 vertices, 2 of them being not observed. The error on F 5 d i s F_5^{dis} F 5 d i s is smaller than the one for F 2 d i s F_2^{dis} F 2 d i s since only 1 of the 3 directly connected vertices to m = 5 m=5 m = 5 is not observed. On the other hand, it is larger than the error on F 3 d i s F_3^{dis} F 3 d i s since m = 5 m=5 m = 5 is directly connected to more vertices than m = 3 m=3 m = 3 .
Conclusion
In the present study, we addressed the inverse source problem of detection-localization-identification of unknown disturbances that might accidentally affect transmission networks after an initial healthy period of time. This study is a follow up on our previous work where we broaden the spectrum of applications to cover more general networks defined by graph Laplacian matrices admitting eigenvalues of arbitrary multiplicities. We also allow a damping coefficient on the transmission process and consider that disturbances might affect multiple non-accessible vertices.
We generalized the notion of strategic set of observation vertices and proved that recording the network state in such a set enables to uniquely identify its initial state during the healthy period of time. From this we determine the global healthy network state. This answers the question of the introduction on how to choose the observation set to detect the presence of a disturbance. A strategic observation set is enough to detect the presence of a disturbance; we can use algorithm 1: detection of disturbances.
Then, we introduced the notion of dominantly absorbent set of network vertices. We proved that if the observation set is strategic and dominantly absorbent, then these observations determine in a unique manner the states of the network in the non-observation vertices since these solve a well posed linear system. This enables to efficiently detect and uniquely identify the unknown disturbances in quasi-real-time. This answers the third question of the introduction: a strategic and dominantly absorbent observation set allows quasi-real-time localization and identification of a disturbance. In practice, this can be done using Algorithm 2.
In practice, the number of observation vertices is limited and the observation set is not dominantly absorbent. Then the linear system becomes ill-posed. To identify the network states in the non-observation vertices, we need the observation set to be at least absorbent. To obtain these states, we introduce a least squares problem augmented by an appropriate regularization term. The deviation of these states from their healthy values enables to localize the unknown disturbances and identify their intensities. This answers question two of the introduction: we use in practice Algorithm 3.
Appendix:
An algorithm to find an absorbent set of a graph
A possible algorithm to find an absorbing set is to find a covering tree of the graph using for example Kruskal’s algorithm, see [18 ] . Then one decimates the vertices of largest degree until the remaining set is absorbent.
Algorithm: Absorbing set Choose: S \mathcal{S} S empty set
Begin
Find a covering tree of the graph using Kruskal’s algorithm. Save the vertices in the set T \mathcal{T} T .
Find maximal degree d max d_{\max} d m a x of the vertices in T \mathcal{T} T .
For d = d max d=d_{\max} d = d m a x to 2 Select all vertices in T \mathcal{T} T of degree ≥ d \ge d ≥ d Include them in S \mathcal{S} S If S \mathcal{S} S is not absorbent then Break end For
End.
Proof of Theorem 3.2
Let X 0 ∈ R N X_0 \in\mathbb{R}^{N} X 0 ∈ R N , X ˉ 0 ∈ R N \bar{X}_0 \in\mathbb{R}^{N} X ˉ 0 ∈ R N and X R = ( x 1 R , … , x N R ) T X^{R}=(x_1^{R},\ldots,x_N^{R})^{\mathsf{T}} X R = ( x 1 R , … , x N R ) T be the solution of the problem:
{ X ¨ R ( t ) + η X ˙ R ( t ) − Δ X R ( t ) = 0 in ( 0 , T 0 ) , X R ( 0 ) = X 0 and X ˙ R ( 0 ) = X ˉ 0 . (53) \begin{cases}
\ddot{X}^{R}(t)+\eta\dot{X}^{R}(t)-\Delta X^{R}(t)=0 & \text{in }(0,T^{0}),\\
X^{R}(0)=X_0 & \text{and}\quad\dot{X}^{R}(0)=\bar{X}_0. \tag*{(53)}
\end{cases} { X ¨ R ( t ) + η X ˙ R ( t ) − Δ X R ( t ) = 0 X R ( 0 ) = X 0 in ( 0 , T 0 ) , and X ˙ R ( 0 ) = X ˉ 0 . (53)
From expanding the solution X R X^{R} X R of the problem (54) in the orthonormal basis of R N \mathbb{R}^{N} R N defined by the normalized eigenvectors of the graph Laplacian matrix Δ \Delta Δ , we write
X R ( t ) = ∑ k = 1 K ∑ ℓ = 1 m k y k ℓ ( t ) v k ℓ , ∀ t , where y k ℓ ( t ) = ⟨ X R ( t ) , v k ℓ ⟩ . (54) X^{R}(t)=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}y_{k\ell}(t)v^{k\ell},\quad\forall t,\qquad\text{where } y_{k\ell}(t)=\langle X^{R}(t),v^{k\ell}\rangle. \tag*{(54)} X R ( t ) = k = 1 ∑ K ℓ = 1 ∑ m k y k ℓ ( t ) v k ℓ , ∀ t , where y k ℓ ( t ) = ⟨ X R ( t ) , v k ℓ ⟩ . (54)
Here, ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ denotes the inner product in R N \mathbb{R}^{N} R N . Moreover, according to (5), the functions y k ℓ y_{k\ell} y k ℓ in (55) solve the following second order differential equation:
{ y ¨ k ℓ ( t ) + η y ˙ k ℓ ( t ) + ω k 2 y k ℓ ( t ) = 0 in ( 0 , T 0 ) , y k ℓ ( 0 ) = ⟨ X 0 , v k ℓ ⟩ and y ˙ k ℓ ( 0 ) = ⟨ X ˉ 0 , v k ℓ ⟩ . (55) \begin{cases}
\ddot{y}_{k\ell}(t)+\eta\dot{y}_{k\ell}(t)+\omega_k^{2}y_{k\ell}(t)=0 & \text{in }(0,T^{0}),\\
y_{k\ell}(0)=\langle X_0,v^{k\ell}\rangle& \text{and}\quad\dot{y}_{k\ell}(0)=\langle\bar{X}_0,v^{k\ell}\rangle. \tag*{(55)}
\end{cases} { y ¨ k ℓ ( t ) + η y ˙ k ℓ ( t ) + ω k 2 y k ℓ ( t ) = 0 y k ℓ ( 0 ) = ⟨ X 0 , v k ℓ ⟩ in ( 0 , T 0 ) , and y ˙ k ℓ ( 0 ) = ⟨ X ˉ 0 , v k ℓ ⟩ . (55)
The discriminant of the characteristic equation associated to (56) is D k = η 2 − 4 ω k 2 D_k=\eta^2-4\omega_k^2 D k = η 2 − 4 ω k 2 . Depending on the sign of D k D_k D k , the general form of the real-valued solutions of (56) reads
y k ℓ ( t ) = e − η 2 t { λ k ℓ e D k 2 t + μ k ℓ e − D k 2 t , if D k > 0 , λ k ℓ + μ k ℓ t , if D k = 0 , λ k ℓ cos ( − D k 2 t ) + μ k ℓ sin ( − D k 2 t ) , if D k < 0 , (56) y_{k\ell}(t)=e^{-\frac{\eta}{2}t}\begin{cases}
\lambda_{k\ell}e^{\frac{\sqrt{D_k}}{2}t}+\mu_{k\ell}e^{-\frac{\sqrt{D_k}}{2}t}, & \text{if }D_k>0,\\
\lambda_{k\ell}+\mu_{k\ell}t, & \text{if }D_k=0,\\
\lambda_{k\ell}\cos\left(\frac{\sqrt{-D_k}}{2}t\right)+\mu_{k\ell}\sin\left(\frac{\sqrt{-D_k}}{2}t\right), & \text{if }D_k<0,
\end{cases} \tag*{(56)} y k ℓ ( t ) = e − 2 η t ⎩ ⎨ ⎧ λ k ℓ e 2 D k t + μ k ℓ e − 2 D k t , λ k ℓ + μ k ℓ t , λ k ℓ cos ( 2 − D k t ) + μ k ℓ sin ( 2 − D k t ) , if D k > 0 , if D k = 0 , if D k < 0 , (56)
where the real coefficients λ k ℓ \lambda_{k\ell} λ k ℓ and μ k ℓ \mu_{k\ell} μ k ℓ to be determined from the initial conditions y k ℓ ( 0 ) y_{k\ell}(0) y k ℓ ( 0 ) and y ˙ k ℓ ( 0 ) \dot{y}_{k\ell}(0) y ˙ k ℓ ( 0 ) . Assuming S \mathcal{S} S to be a strategic set of vertices, suppose that
x n R ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S . (57) x_n^R(t)=0,\quad\forall t\in(0,T^0),\ \forall n\in\mathcal{S}. \tag*{(57)} x n R ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S . (57)
Then, in view of (55)-(58), it follows that
x n R ( t ) = ∑ k = 1 K ∑ ℓ = 1 m k y k ℓ ( t ) v n k ℓ = 0 , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S . (58) x_n^R(t)=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}y_{k\ell}(t)v_n^{k\ell}=0,\quad\forall t\in(0,T^0),\ \forall n\in\mathcal{S}. \tag*{(58)} x n R ( t ) = k = 1 ∑ K ℓ = 1 ∑ m k y k ℓ ( t ) v n k ℓ = 0 , ∀ t ∈ ( 0 , T 0 ) , ∀ n ∈ S . (58)
Besides for arbitrary values η > 0 \eta>0 η > 0 and ω k 2 ≥ 0 \omega_k^2\geq0 ω k 2 ≥ 0 , we notice that according to (4) and since ω 1 = 0 \omega_1=0 ω 1 = 0 , the set of discriminants fulfills
η 2 − 4 ω K 2 = D K < D K − 1 < ⋯ < D 2 < D 1 = η 2 . (59) \eta^2-4\omega_K^2=D_K<D_{K-1}<\cdots<D_2<D_1=\eta^2. \tag*{(59)} η 2 − 4 ω K 2 = D K < D K − 1 < ⋯ < D 2 < D 1 = η 2 . (59)
To establish the proof for the most general case of the three possible forms defining the functions y k ℓ y_{k\ell} y k ℓ in (57) are involved in (59), we assume that
∃ k ˉ ∈ { 2 , … , K − 1 } such that D k ˉ = 0. (60) \exists\bar{k}\in\{2,\ldots,K-1\}\quad\text{such that}\quad D_{\bar{k}}=0. \tag*{(60)} ∃ k ˉ ∈ { 2 , … , K − 1 } such that D k ˉ = 0. (60)
According to (57)-(61), it follows from (59) by analytic continuation that: ∀ n ∈ S \forall n\in\mathcal{S} ∀ n ∈ S ,
∑ k = 1 k ˉ − 1 ∑ ℓ = 1 m k ( λ k ℓ e D k 2 t + μ k ℓ e − D k 2 t ) v n k ℓ + ∑ ℓ = 1 m k ˉ ( λ k ˉ ℓ + μ k ˉ ℓ t ) v n k ˉ ℓ + ∑ k = k ˉ + 1 K ∑ ℓ = 1 m k ( λ k ℓ cos ( − D k 2 t ) + μ k ℓ sin ( − D k 2 t ) ) v n k ℓ = 0 , ∀ t ∈ R . (61) \sum_{k=1}^{\bar{k}-1}\sum_{\ell=1}^{m_k}\left(\lambda_{k\ell}e^{\frac{\sqrt{D_k}}{2}t}+\mu_{k\ell}e^{-\frac{\sqrt{D_k}}{2}t}\right)v_n^{k\ell}
+\sum_{\ell=1}^{m_{\bar{k}}}\left(\lambda_{\bar{k}\ell}+\mu_{\bar{k}\ell}t\right)v_n^{\bar{k}\ell}
+\sum_{k=\bar{k}+1}^{K}\sum_{\ell=1}^{m_k}\left(\lambda_{k\ell}\cos\left(\frac{\sqrt{-D_k}}{2}t\right)+\mu_{k\ell}\sin\left(\frac{\sqrt{-D_k}}{2}t\right)\right)v_n^{k\ell}=0,\quad\forall t\in\mathbb{R}. \tag*{(61)} k = 1 ∑ k ˉ − 1 ℓ = 1 ∑ m k ( λ k ℓ e 2 D k t + μ k ℓ e − 2 D k t ) v n k ℓ + ℓ = 1 ∑ m k ˉ ( λ k ˉ ℓ + μ k ˉ ℓ t ) v n k ˉ ℓ + k = k ˉ + 1 ∑ K ℓ = 1 ∑ m k ( λ k ℓ cos ( 2 − D k t ) + μ k ℓ sin ( 2 − D k t ) ) v n k ℓ = 0 , ∀ t ∈ R . (61)
Notice that: 1. If k ˉ \bar{k} k ˉ doesn’t exist and D K > 0 D_K>0 D K > 0 , then the Left Hand Side (LHS) of (62) is defined only by its first term with K K K instead of k ˉ − 1 \bar{k}-1 k ˉ − 1 . 2. If k ˉ \bar{k} k ˉ doesn’t exist and D K < 0 D_K<0 D K < 0 , then the LHS of (62) is defined only by its first and third terms with k ^ \hat{k} k ^ instead of k ˉ − 1 \bar{k}-1 k ˉ − 1 and k ^ + 1 \hat{k}+1 k ^ + 1 instead of k ˉ + 1 \bar{k}+1 k ˉ + 1 , where k ^ \hat{k} k ^ is such that D k ^ > 0 D_{\hat{k}}>0 D k ^ > 0 whereas D k ^ + 1 < 0 D_{\hat{k}+1}<0 D k ^ + 1 < 0 . 3. If k ˉ = K \bar{k}=K k ˉ = K , then D K = 0 D_K=0 D K = 0 and the LHS of (62) is defined only by its two first terms.
For p = 1 , … , K p=1,\ldots,K p = 1 , … , K , let A ( p ) A^{(p)} A ( p ) be the N S × m p N_{\mathcal{S}}\times m_p N S × m p matrix whose columns are the m p m_p m p eigenvectors of Δ \Delta Δ associated to − ω p 2 -\omega_p^2 − ω p 2 and rows are the N S N_{\mathcal{S}} N S components of these vectors corresponding to the vertex labels defining S \mathcal{S} S . Since S \mathcal{S} S is strategic, it follows from Definition 3.1 that the matrix A ( p ) A^{(p)} A ( p ) is of full rank. Let Λ ( p ) = ( λ p 1 , … , λ p m p ) T \Lambda^{(p)}=(\lambda_{p1},\ldots,\lambda_{pm_p})^{\mathsf{T}} Λ ( p ) = ( λ p 1 , … , λ p m p ) T and ζ ( p ) = ( μ p 1 , … , μ p m p ) T \zeta^{(p)}=(\mu_{p1},\ldots,\mu_{pm_p})^{\mathsf{T}} ζ ( p ) = ( μ p 1 , … , μ p m p ) T .
▶ Proving that λ k ℓ = μ k ℓ = 0 \lambda_{k\ell}=\mu_{k\ell}=0 λ k ℓ = μ k ℓ = 0 , for k = 1 , … , k ˉ − 1 k=1,\ldots,\bar{k}-1 k = 1 , … , k ˉ − 1 and ℓ = 1 , … , m k \ell=1,\ldots,m_k ℓ = 1 , … , m k :
Using D p = 1 D_{p=1} D p = 1 : Since D 1 > D k = 2 , … , K D_1>D_{k=2,\ldots,K} D 1 > D k = 2 , … , K , from multiplying (61) by e − D 1 2 t e^{-\frac{\sqrt{D_1}}{2}t} e − 2 D 1 t and setting the limit when t t t tends to + ∞ +\infty + ∞ , then by e D 1 2 t e^{\frac{\sqrt{D_1}}{2}t} e 2 D 1 t and setting the limit when t t t goes to − ∞ -\infty − ∞ , we get
∀ n ∈ S , ∑ ℓ = 1 m 1 λ 1 ℓ v n 1 ℓ = 0 and ∑ ℓ = 1 m 1 μ 1 ℓ v n 1 ℓ = 0. (62) \forall n\in S,\qquad\sum_{\ell=1}^{m_1}\lambda_{1\ell}v_n^{1\ell}=0\qquad\text{and}\qquad\sum_{\ell=1}^{m_1}\mu_{1\ell}v_n^{1\ell}=0. \tag*{(62)} ∀ n ∈ S , ℓ = 1 ∑ m 1 λ 1 ℓ v n 1 ℓ = 0 and ℓ = 1 ∑ m 1 μ 1 ℓ v n 1 ℓ = 0. (62)
Moreover, (62) can be rewritten under the following matrix form:
A ( 1 ) Λ ( 1 ) = 0 and A ( 1 ) ζ ( 1 ) = 0 ⟹ λ 1 ℓ = 1 , … , m 1 = μ 1 ℓ = 1 , … , m 1 = 0. (63) A^{(1)}\Lambda^{(1)}=0\qquad\text{and}\qquad A^{(1)}\zeta^{(1)}=0\qquad\Longrightarrow\qquad\lambda_{1\ell=1,\ldots,m_1}=\mu_{1\ell=1,\ldots,m_1}=0. \tag*{(63)} A ( 1 ) Λ ( 1 ) = 0 and A ( 1 ) ζ ( 1 ) = 0 ⟹ λ 1 ℓ = 1 , … , m 1 = μ 1 ℓ = 1 , … , m 1 = 0. (63)
The implication in (63) is obtained since the matrix A ( 1 ) A^{(1)} A ( 1 ) is of full rank.
Using D p = 2 , … , k ˉ − 1 D_{p=2,\ldots,\bar{k}-1} D p = 2 , … , k ˉ − 1 : By taking into account of (63) in (61) and iterating the same process for p = 2 , … , k ˉ − 1 p=2,\ldots,\bar{k}-1 p = 2 , … , k ˉ − 1 : We multiply (61) by e − D p 2 t e^{-\frac{\sqrt{D_p}}{2}t} e − 2 D p t and set the limit when t t t tends to + ∞ +\infty + ∞ , then by e D p 2 t e^{\frac{\sqrt{D_p}}{2}t} e 2 D p t and set the limit when t t t goes to − ∞ -\infty − ∞ , we obtain: For p = 2 , … , k ˉ − 1 p=2,\ldots,\bar{k}-1 p = 2 , … , k ˉ − 1 ,
A ( p ) Λ ( p ) = 0 and A ( p ) ζ ( p ) = 0 ⟹ λ p ℓ = 1 , … , m p = μ p ℓ = 1 , … , m p = 0. (64) A^{(p)}\Lambda^{(p)}=0\qquad\text{and}\qquad A^{(p)}\zeta^{(p)}=0\qquad\Longrightarrow\qquad\lambda_{p\ell=1,\ldots,m_p}=\mu_{p\ell=1,\ldots,m_p}=0. \tag*{(64)} A ( p ) Λ ( p ) = 0 and A ( p ) ζ ( p ) = 0 ⟹ λ p ℓ = 1 , … , m p = μ p ℓ = 1 , … , m p = 0. (64)
Proving that λ k ˉ ℓ = μ k ˉ ℓ = 0 \lambda_{\bar{k}\ell}=\mu_{\bar{k}\ell}=0 λ k ˉ ℓ = μ k ˉ ℓ = 0 , for ℓ = 1 , … , m k ˉ \ell=1,\ldots,m_{\bar{k}} ℓ = 1 , … , m k ˉ : In view of (63) -(64) , the first double sum in (61) vanishes. We compute the integral over ( 0 , T ) (0,T) ( 0 , T ) of the updated equation (61) , where T > 0 T>0 T > 0 . From dividing the obtained result, firstly by T 2 T^2 T 2 and setting the limit when T T T tends to + ∞ +\infty + ∞ , then by T T T and resetting again the limit when T T T goes to + ∞ +\infty + ∞ , we find
∀ n ∈ S , ∑ ℓ = 1 m k ˉ μ k ˉ ℓ v n k ˉ ℓ = 0 and ∑ ℓ = 1 m k ˉ λ k ˉ ℓ v n k ˉ ℓ = 0 ⟹ λ k ˉ ℓ = 1 , … , m k ˉ = μ k ˉ ℓ = 1 , … , m k ˉ = 0. (65) \forall n\in S,\qquad\sum_{\ell=1}^{m_{\bar{k}}}\mu_{\bar{k}\ell}v_n^{\bar{k}\ell}=0\qquad\text{and}\qquad\sum_{\ell=1}^{m_{\bar{k}}}\lambda_{\bar{k}\ell}v_n^{\bar{k}\ell}=0\qquad\Longrightarrow\qquad\lambda_{\bar{k}\ell=1,\ldots,m_{\bar{k}}}=\mu_{\bar{k}\ell=1,\ldots,m_{\bar{k}}}=0. \tag*{(65)} ∀ n ∈ S , ℓ = 1 ∑ m k ˉ μ k ˉ ℓ v n k ˉ ℓ = 0 and ℓ = 1 ∑ m k ˉ λ k ˉ ℓ v n k ˉ ℓ = 0 ⟹ λ k ˉ ℓ = 1 , … , m k ˉ = μ k ˉ ℓ = 1 , … , m k ˉ = 0. (65)
The implication in (65) is obtained since the matrix A ( k ˉ ) A^{(\bar{k})} A ( k ˉ ) is of full rank.
Proving that λ k ℓ = μ k ℓ = 0 \lambda_{k\ell}=\mu_{k\ell}=0 λ k ℓ = μ k ℓ = 0 , for k = k ˉ + 1 , … , K k=\bar{k}+1,\ldots,K k = k ˉ + 1 , … , K and ℓ = 1 , … , m k \ell=1,\ldots,m_k ℓ = 1 , … , m k : According to (64) -(65) , it follows that (61) is now reduced to:
∀ n ∈ S , ∑ k = k ˉ + 1 K ∑ ℓ = 1 m k ( λ k ℓ cos ( − D k 2 t ) + μ k ℓ sin ( − D k 2 t ) ) v n k ℓ = 0 , ∀ t ∈ R . (66) \forall n\in S,\qquad\sum_{k=\bar{k}+1}^{K}\sum_{\ell=1}^{m_k}\left(\lambda_{k\ell}\cos\left(\frac{\sqrt{-D_k}}{2}t\right)+\mu_{k\ell}\sin\left(\frac{\sqrt{-D_k}}{2}t\right)\right)v_n^{k\ell}=0,\qquad\forall t\in\mathbb{R}. \tag*{(66)} ∀ n ∈ S , k = k ˉ + 1 ∑ K ℓ = 1 ∑ m k ( λ k ℓ cos ( 2 − D k t ) + μ k ℓ sin ( 2 − D k t ) ) v n k ℓ = 0 , ∀ t ∈ R . (66)
For p = k ˉ + 1 , … , K p=\bar{k}+1,\ldots,K p = k ˉ + 1 , … , K : We multiply (66) by cos ( − D p 2 t ) \cos\left(\frac{\sqrt{-D_p}}{2}t\right) cos ( 2 − D p t ) and compute its integral over ( 0 , T ) (0,T) ( 0 , T ) . From dividing the result by T T T and setting the limit when T T T tends to + ∞ +\infty + ∞ , we get
∀ n ∈ S , ∑ ℓ = 1 m p λ p ℓ v n p ℓ = 0 ⟹ λ p ℓ = 1 , … , m p = 0. (67) \forall n\in S,\qquad\sum_{\ell=1}^{m_p}\lambda_{p\ell}v_n^{p\ell}=0\qquad\Longrightarrow\qquad\lambda_{p\ell=1,\ldots,m_p}=0. \tag*{(67)} ∀ n ∈ S , ℓ = 1 ∑ m p λ p ℓ v n p ℓ = 0 ⟹ λ p ℓ = 1 , … , m p = 0. (67)
Afterwards, by taking into account the result (67) in (66) and then, computing the derivative of the updated equation, it follows that
∀ n ∈ S , ∑ k = k ˉ + 1 K ∑ ℓ = 1 m k − D k 2 μ k ℓ cos ( − D k 2 t ) v n k ℓ = 0 , ∀ t ∈ R . (68) \forall n\in S,\qquad\sum_{k=\bar{k}+1}^{K}\sum_{\ell=1}^{m_k}\frac{\sqrt{-D_k}}{2}\mu_{k\ell}\cos\left(\frac{\sqrt{-D_k}}{2}t\right)v_n^{k\ell}=0,\qquad\forall t\in\mathbb{R}. \tag*{(68)} ∀ n ∈ S , k = k ˉ + 1 ∑ K ℓ = 1 ∑ m k 2 − D k μ k ℓ cos ( 2 − D k t ) v n k ℓ = 0 , ∀ t ∈ R . (68)
Then, reapplying the same techniques used in (66) –(67) , we obtain: For p = k ˉ + 1 , … , K p=\bar{k}+1,\ldots,K p = k ˉ + 1 , … , K ,
∀ n ∈ S , ∑ ℓ = 1 m p μ p ℓ v n p ℓ = 0 ⟹ μ p ℓ = 1 , … , m p = 0. (69) \forall n\in S,\qquad\sum_{\ell=1}^{m_p}\mu_{p\ell}v_n^{p\ell}=0 \qquad\Longrightarrow\qquad\mu_{p\ell=1,\ldots,m_p}=0. \tag*{(69)} ∀ n ∈ S , ℓ = 1 ∑ m p μ p ℓ v n p ℓ = 0 ⟹ μ p ℓ = 1 , … , m p = 0. (69)
The implications in (67) and in (69) are obtained since the matrix A ( p ) A^{(p)} A ( p ) is of full rank.
In view of (63) –(64) , (65) , (67) and (69) , it follows that λ k ℓ = 1 , … , m k = μ k ℓ = 1 , … , m k = 0 \lambda_{k\ell=1,\ldots,m_k}=\mu_{k\ell=1,\ldots,m_k}=0 λ k ℓ = 1 , … , m k = μ k ℓ = 1 , … , m k = 0 , for k = 1 , … , K k=1,\ldots,K k = 1 , … , K . Therefore, (54) –(56) implies that the solution of the problem (53) is X R ( t ) = 0 , ∀ t X^R(t)=0,\ \forall t X R ( t ) = 0 , ∀ t which leads to X R ( 0 ) = X R ( 0 ) = 0 X^R(0)=X^R(0)=0 X R ( 0 ) = X R ( 0 ) = 0 .
Proof of Propostion 5.2
Let a < b a<b a < b be two real numbers, t i = ( a + b ) / 2 t_i=(a+b)/2 t i = ( a + b ) /2 and φ i \varphi_i φ i be the function introduced in (??) with ( a , b ) (a,b) ( a , b ) instead of ( t i − 1 , t i + 1 ) (t_{i-1},t_{i+1}) ( t i − 1 , t i + 1 ) . Let α , β , γ \alpha,\beta,\gamma α , β , γ be three real numbers and g ( t ) = α t 2 + β t + γ , ∀ t g(t)=\alpha t^2+\beta t+\gamma,\ \forall t g ( t ) = α t 2 + β t + γ , ∀ t .
▶ ∫ a t i g ( t ) φ i ( t ) d t = − 1 2 ∫ a t i ( α t 3 + [ β − α a ] t 2 + [ γ − β a ] t − γ a ) d t , = − 1 2 ( α 4 ( t i 4 − a 4 ) + 1 3 [ β − α a ] ( t i 3 − a 3 ) + 1 2 [ γ − β a ] ( t i 2 − a 2 ) − γ a ( t i − a ) ) . \begin{aligned}
\blacktriangleright\int_a^{t_i}g(t)\varphi_i(t)\,dt
=-\frac{1}{2}\int_a^{t_i}\left(\alpha t^3+[\beta-\alpha a]t^2+[\gamma-\beta a]t-\gamma a\right)\,dt, \\
&=-\frac{1}{2}\left(\frac{\alpha}{4}(t_i^4-a^4)+\frac{1}{3}[\beta-\alpha a](t_i^3-a^3)+\frac{1}{2}[\gamma-\beta a](t_i^2-a^2)-\gamma a(t_i-a)\right).
\end{aligned} ▶ ∫ a t i g ( t ) φ i ( t ) d t = − 2 1 ∫ a t i ( α t 3 + [ β − α a ] t 2 + [ γ − β a ] t − γ a ) d t , = − 2 1 ( 4 α ( t i 4 − a 4 ) + 3 1 [ β − α a ] ( t i 3 − a 3 ) + 2 1 [ γ − β a ] ( t i 2 − a 2 ) − γ a ( t i − a ) ) .
▶ ∫ t i b g ( t ) φ i ( t ) d t = 1 2 ∫ t i b ( α t 3 + [ β − b α ] t 2 + [ γ − β b ] t − γ b ) d t , = 1 2 ( α 4 ( b 4 − t i 4 ) + 1 3 [ β − b α ] ( b 3 − t i 3 ) + 1 2 [ γ − β b ] ( b 2 − t i 2 ) − γ b ( b − t i ) ) . \begin{aligned}
\blacktriangleright\int_{t_i}^{b}g(t)\varphi_i(t)\,dt
=\frac{1}{2}\int_{t_i}^{b}\left(\alpha t^3+[\beta-b\alpha]t^2+[\gamma-\beta b]t-\gamma b\right)\,dt, \\
&=\frac{1}{2}\left(\frac{\alpha}{4}(b^4-t_i^4)+\frac{1}{3}[\beta-b\alpha](b^3-t_i^3)+\frac{1}{2}[\gamma-\beta b](b^2-t_i^2)-\gamma b(b-t_i)\right).
\end{aligned} ▶ ∫ t i b g ( t ) φ i ( t ) d t = 2 1 ∫ t i b ( α t 3 + [ β − b α ] t 2 + [ γ − β b ] t − γ b ) d t , = 2 1 ( 4 α ( b 4 − t i 4 ) + 3 1 [ β − b α ] ( b 3 − t i 3 ) + 2 1 [ γ − β b ] ( b 2 − t i 2 ) − γ b ( b − t i ) ) .
Since t i − a = b − t i t_i-a=b-t_i t i − a = b − t i , from employing Chasles rule it follows that
∫ a b g ( t ) φ i ( t ) d t = − b − t i 2 ( α 12 ( b 3 − a 3 ) + [ α 12 t i + β 6 ] ( b 2 − a 2 ) + [ α 12 t i 2 + β 6 t i + γ 2 ] ( b − a ) ) . (70) \int_a^b g(t)\varphi_i(t)\,dt
=-\frac{b-t_i}{2}\left(\frac{\alpha}{12}(b^3-a^3)+\left[\frac{\alpha}{12}t_i+\frac{\beta}{6}\right](b^2-a^2)+\left[\frac{\alpha}{12}t_i^2+\frac{\beta}{6}t_i+\frac{\gamma}{2}\right](b-a)\right). \tag*{(70)} ∫ a b g ( t ) φ i ( t ) d t = − 2 b − t i ( 12 α ( b 3 − a 3 ) + [ 12 α t i + 6 β ] ( b 2 − a 2 ) + [ 12 α t i 2 + 6 β t i + 2 γ ] ( b − a ) ) . (70)
Besides, t i = ( a + b ) / 2 t_i=(a+b)/2 t i = ( a + b ) /2 implies that a b = 2 t i 2 − ( b 2 + a 2 ) / 2 ab=2t_i^2-(b^2+a^2)/2 ab = 2 t i 2 − ( b 2 + a 2 ) /2 . Therefore, we get
b 3 − a 3 = ( b − a ) ( b 2 + a b + a 2 ) = b − a 2 ( b 2 + 4 t i 2 + a 2 ) and b 2 − a 2 = 2 ( b − a ) t i . (71) b^3-a^3=(b-a)(b^2+ab+a^2)=\frac{b-a}{2}(b^2+4t_i^2+a^2)\qquad\text{and}\qquad b^2-a^2=2(b-a)t_i. \tag*{(71)} b 3 − a 3 = ( b − a ) ( b 2 + ab + a 2 ) = 2 b − a ( b 2 + 4 t i 2 + a 2 ) and b 2 − a 2 = 2 ( b − a ) t i . (71)
Thus, from using (71) in (70) and since 2 β t i = β ( a + b ) 2\beta t_i=\beta(a+b) 2 β t i = β ( a + b ) , we obtain
∫ a b g ( t ) φ i ( t ) d t = − b − t i 2 b − a 2 ( 10 12 α t i 2 + β t i + α 12 ( b 2 + a 2 ) + γ ) , = − b − t i 2 b − a 2 ( 10 12 α t i 2 + 10 12 β t i + β 12 ( a + b ) + α 12 ( b 2 + a 2 ) + γ ) , = − b − t i 2 b − a 2 ( 10 12 g ( t i ) + 1 12 g ( b ) + 1 12 g ( a ) ) . (72) \begin{aligned}
\int_a^b g(t)\varphi_i(t)\,dt
=-\frac{b-t_i}{2}\frac{b-a}{2}\left(\frac{10}{12}\alpha t_i^2+\beta t_i+\frac{\alpha}{12}(b^2+a^2)+\gamma\right), \\
&=-\frac{b-t_i}{2}\frac{b-a}{2}\left(\frac{10}{12}\alpha t_i^2+\frac{10}{12}\beta t_i+\frac{\beta}{12}(a+b)+\frac{\alpha}{12}(b^2+a^2)+\gamma\right), \\
&=-\frac{b-t_i}{2}\frac{b-a}{2}\left(\frac{10}{12}g(t_i)+\frac{1}{12}g(b)+\frac{1}{12}g(a)\right). \tag*{(72)}
\end{aligned} ∫ a b g ( t ) φ i ( t ) d t = − 2 b − t i 2 b − a ( 12 10 α t i 2 + β t i + 12 α ( b 2 + a 2 ) + γ ) , = − 2 b − t i 2 b − a ( 12 10 α t i 2 + 12 10 β t i + 12 β ( a + b ) + 12 α ( b 2 + a 2 ) + γ ) , = − 2 b − t i 2 b − a ( 12 10 g ( t i ) + 12 1 g ( b ) + 12 1 g ( a ) ) . (72)
Since t i − a = b − t i t_i-a=b-t_i t i − a = b − t i , from (??) it follows that φ i ( t i ) = − ( b − t i ) / 2 \varphi_i(t_i)=-(b-t_i)/2 φ i ( t i ) = − ( b − t i ) /2 . Hence, the last equality in (72) yields the result announced in (36) .
Proof of Proposition 6.1
We expand the solution X 0 X^0 X 0 of the problem (44) in the orthonormal basis defined by the family { v k ℓ , k = 1 , … , K and ℓ = 1 , … , m k } \{v^{k\ell},\ k=1,\ldots,K\text{ and }\ell=1,\ldots,m_k\} { v k ℓ , k = 1 , … , K and ℓ = 1 , … , m k } of the N N N normalized eigenvectors associated to the graph Laplacian matrix Δ \Delta Δ :
X 0 ( t ) = ∑ k = 1 K ∑ ℓ = 1 m k y k ℓ ( t ) v k ℓ , ∀ t ∈ ( 0 , T 0 ) , (73) X^0(t)=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}y_{k\ell}(t)v^{k\ell},\quad\forall t\in(0,T^0), \tag*{(73)} X 0 ( t ) = k = 1 ∑ K ℓ = 1 ∑ m k y k ℓ ( t ) v k ℓ , ∀ t ∈ ( 0 , T 0 ) , (73)
where y k ℓ ( t ) = ⟨ X 0 ( t ) , v k ℓ ⟩ y_{k\ell}(t)=\langle X^0(t),v_{k\ell}\rangle y k ℓ ( t ) = ⟨ X 0 ( t ) , v k ℓ ⟩ with ⟨ , ⟩ \langle,\rangle ⟨ , ⟩ designates the inner product in R N \mathbb{R}^N R N . From the problem (44) , it follows in view of (5) and (46) that the functions y k ℓ y_{k\ell} y k ℓ solve:
{ y ¨ k ℓ ( t ) + η y ˙ k ℓ ( t ) + ω k 2 y k ℓ ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) , y k ℓ ( 0 ) = y k ℓ 0 and y ˙ k ℓ ( 0 ) = y ˉ k ℓ 0 . (74) \begin{cases}
\ddot{y}_{k\ell}(t)+\eta\dot{y}_{k\ell}(t)+\omega_k^2y_{k\ell}(t)=0,\quad\forall t\in(0,T^0),\\
y_{k\ell}(0)=y^0_{k\ell}\quad\text{and}\quad\dot{y}_{k\ell}(0)=\bar{y}^0_{k\ell}. \tag*{(74)}
\end{cases} { y ¨ k ℓ ( t ) + η y ˙ k ℓ ( t ) + ω k 2 y k ℓ ( t ) = 0 , ∀ t ∈ ( 0 , T 0 ) , y k ℓ ( 0 ) = y k ℓ 0 and y ˙ k ℓ ( 0 ) = y ˉ k ℓ 0 . (74)
The discriminant of the characteristic equation associated to (74) is D k = η 2 − 4 ω k 2 D_k=\eta^2-4\omega_k^2 D k = η 2 − 4 ω k 2 . Therefore, we distinguish the following three cases defining the solutions of (74) :
y k ℓ ( t ) = e − η 2 t { λ k ℓ e D k 2 t + μ k ℓ e − D k 2 t , if D k > 0 , λ k ℓ + μ k ℓ t , if D k = 0 , λ k ℓ cos ( − D k 2 t ) + μ k ℓ sin ( − D k 2 t ) , if D k < 0 , (75) y_{k\ell}(t)=e^{-\frac{\eta}{2}t}
\begin{cases}
\lambda_{k\ell}e^{\frac{\sqrt{D_k}}{2}t}+\mu_{k\ell}e^{-\frac{\sqrt{D_k}}{2}t}, & \text{if }D_k>0,\\
\lambda_{k\ell}+\mu_{k\ell}t, & \text{if }D_k=0,\\
\lambda_{k\ell}\cos\left(\frac{\sqrt{-D_k}}{2}t\right)+\mu_{k\ell}\sin\left(\frac{\sqrt{-D_k}}{2}t\right), & \text{if }D_k<0,
\end{cases} \tag*{(75)} y k ℓ ( t ) = e − 2 η t ⎩ ⎨ ⎧ λ k ℓ e 2 D k t + μ k ℓ e − 2 D k t , λ k ℓ + μ k ℓ t , λ k ℓ cos ( 2 − D k t ) + μ k ℓ sin ( 2 − D k t ) , if D k > 0 , if D k = 0 , if D k < 0 , (75)
where the real coefficients λ k ℓ \lambda_{k\ell} λ k ℓ and μ k ℓ \mu_{k\ell} μ k ℓ to be determined from the initial conditions y k ℓ ( 0 ) y_{k\ell}(0) y k ℓ ( 0 ) and y ˙ k ℓ ( 0 ) \dot{y}_{k\ell}(0) y ˙ k ℓ ( 0 ) . Then, depending on the sign of the discriminant D k D_k D k , we obtain:
Case 1. If D k > 0 D_k>0 D k > 0 , we denote by: r k = − η + D k 2 r_k=\frac{-\eta+\sqrt{D_k}}{2} r k = 2 − η + D k and r ˉ k = − η + D k 2 \bar{r}_k=-\frac{\eta+\sqrt{D_k}}{2} r ˉ k = − 2 η + D k . From (75) , we get
\begin{cases}
\lambda_{k\ell}+\mu_{k\ell}=y_{k\ell}(0),\\
\lambda_{k\ell}r_k+\mu_{k\ell}\bar{r}_k=\dot{y}_{k\ell}(0).
\end{cases}
\quad\Longrightarrow\quad
\lambda_{k\ell}=\frac{\bar{r}_k}{\bar{r}_k-r_k}y_{k\ell}(0)-\frac{1}{\bar{r}_k-r_k}\dot{y}_{k\ell}(0),
\quad
\mu_{k\ell}=-\frac{r_k}{\bar{r}_k-r_k}y_{k\ell}(0)+\frac{1}{\bar{r}_k-r_k}\dot{y}_{k\ell}(0). \tag*{(76)}\end{cases}
Hence, in this case, the solution y k ℓ y_{k\ell} y k ℓ of (74) is then given by:
y k ℓ ( t ) = r ˉ k e r k t − r k e r ˉ k t r ˉ k − r k y k ℓ ( 0 ) + e r ˉ k t − e r k t r ˉ k − r k y ˙ k ℓ ( 0 ) . (77) y_{k\ell}(t)=\frac{\bar{r}_ke^{r_kt}-r_ke^{\bar{r}_kt}}{\bar{r}_k-r_k}y_{k\ell}(0)+\frac{e^{\bar{r}_kt}-e^{r_kt}}{\bar{r}_k-r_k}\dot{y}_{k\ell}(0). \tag*{(77)} y k ℓ ( t ) = r ˉ k − r k r ˉ k e r k t − r k e r ˉ k t y k ℓ ( 0 ) + r ˉ k − r k e r ˉ k t − e r k t y ˙ k ℓ ( 0 ) . (77)
Case 2. If D k = 0 D_k=0 D k = 0 , then from (75) , it follows that the solution y k ℓ y_{k\ell} y k ℓ of (74) is defined by:
y k ℓ ( t ) = ( 1 + η 2 t ) e − η 2 t y k ℓ ( 0 ) + t e − η 2 t y ˙ k ℓ ( 0 ) . (78) y_{k\ell}(t)=\left(1+\frac{\eta}{2}t\right)e^{-\frac{\eta}{2}t}y_{k\ell}(0)+te^{-\frac{\eta}{2}t}\dot{y}_{k\ell}(0). \tag*{(78)} y k ℓ ( t ) = ( 1 + 2 η t ) e − 2 η t y k ℓ ( 0 ) + t e − 2 η t y ˙ k ℓ ( 0 ) . (78)
Case 3. If D k < 0 D_k<0 D k < 0 , then (75) leads to:
y k ℓ ( t ) = e − η 2 t ( cos ( − D k 2 t ) + η − D k sin ( − D k 2 t ) ) y k ℓ ( 0 ) + 2 e − η 2 t − D k sin ( − D k 2 t ) y ˙ k ℓ ( 0 ) . (79) y_{k\ell}(t)=e^{-\frac{\eta}{2}t}\left(\cos\left(\frac{\sqrt{-D_k}}{2}t\right)+\frac{\eta}{\sqrt{-D_k}}\sin\left(\frac{\sqrt{-D_k}}{2}t\right)\right)y_{k\ell}(0)+\frac{2e^{-\frac{\eta}{2}t}}{\sqrt{-D_k}}\sin\left(\frac{\sqrt{-D_k}}{2}t\right)\dot{y}_{k\ell}(0). \tag*{(79)} y k ℓ ( t ) = e − 2 η t ( cos ( 2 − D k t ) + − D k η sin ( 2 − D k t ) ) y k ℓ ( 0 ) + − D k 2 e − 2 η t sin ( 2 − D k t ) y ˙ k ℓ ( 0 ) . (79)
Setting y k ℓ ( 0 ) = y k ℓ 0 y_{k\ell}(0)=y^0_{k\ell} y k ℓ ( 0 ) = y k ℓ 0 and y ˙ k ℓ ( 0 ) = y ˉ k ℓ 0 \dot{y}_{k\ell}(0)=\bar{y}^0_{k\ell} y ˙ k ℓ ( 0 ) = y ˉ k ℓ 0 in (77) –(79) leads to the result in Proposition 6.1.
Details on the gradient ∇ J \nabla J ∇ J (52)
We verify that the first order partial derivatives of the objective function J J J in (51) are given by: Let ν k ℓ p q = ∑ n ∈ S ν n p q ν n k ℓ \nu_{k\ell}^{pq}=\sum_{n\in\mathcal{S}}\nu_n^{pq}\nu_n^{k\ell} ν k ℓ pq = ∑ n ∈ S ν n pq ν n k ℓ ,
∂ J ( y k ℓ 0 , y ˉ k ℓ 0 ) ∂ y p q 0 = ∑ k = 1 K ∑ ℓ = 1 m k ( ν k ℓ p q ∫ 0 T 0 A p ( t ) A k ( t ) d t y k ℓ 0 + ν k ℓ p q ∫ 0 T 0 A p ( t ) B k ( t ) d t y ˉ k ℓ 0 ) + ∑ n ∈ S ν n p q ∫ 0 T 0 A p ( t ) ( x n F ( t ) − d n ( t ) ) d t . (80) \frac{\partial J(y_{k\ell}^{0},\bar{y}_{k\ell}^{0})}{\partial y_{pq}^{0}}=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}\left(\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{A}_{p}(t)\mathcal{A}_{k}(t)\,\mathrm{d}t\,y_{k\ell}^{0}+\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{A}_{p}(t)\mathcal{B}_{k}(t)\,\mathrm{d}t\,\bar{y}_{k\ell}^{0}\right)+\sum_{n\in\mathcal{S}}\nu_{n}^{pq}\int_{0}^{T^{0}}\mathcal{A}_{p}(t)\left(x_{n}^{F}(t)-d_{n}(t)\right)\,\mathrm{d}t. \tag*{(80)} ∂ y pq 0 ∂ J ( y k ℓ 0 , y ˉ k ℓ 0 ) = k = 1 ∑ K ℓ = 1 ∑ m k ( ν k ℓ pq ∫ 0 T 0 A p ( t ) A k ( t ) d t y k ℓ 0 + ν k ℓ pq ∫ 0 T 0 A p ( t ) B k ( t ) d t y ˉ k ℓ 0 ) + n ∈ S ∑ ν n pq ∫ 0 T 0 A p ( t ) ( x n F ( t ) − d n ( t ) ) d t . (80)
∂ J ( y k ℓ 0 , y ˉ k ℓ 0 ) ∂ y ˉ p q 0 = ∑ k = 1 K ∑ ℓ = 1 m k ( ν k ℓ p q ∫ 0 T 0 B p ( t ) A k ( t ) d t y k ℓ 0 + ν k ℓ p q ∫ 0 T 0 B p ( t ) B k ( t ) d t y ˉ k ℓ 0 ) + ∑ n ∈ S ν n p q ∫ 0 T 0 B p ( t ) ( x n F ( t ) − d n ( t ) ) d t . (81) \frac{\partial J(y_{k\ell}^{0},\bar{y}_{k\ell}^{0})}{\partial\bar{y}_{pq}^{0}}=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}\left(\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{B}_{p}(t)\mathcal{A}_{k}(t)\,\mathrm{d}t\,y_{k\ell}^{0}+\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{B}_{p}(t)\mathcal{B}_{k}(t)\,\mathrm{d}t\,\bar{y}_{k\ell}^{0}\right)+\sum_{n\in\mathcal{S}}\nu_{n}^{pq}\int_{0}^{T^{0}}\mathcal{B}_{p}(t)\left(x_{n}^{F}(t)-d_{n}(t)\right)\,\mathrm{d}t. \tag*{(81)} ∂ y ˉ pq 0 ∂ J ( y k ℓ 0 , y ˉ k ℓ 0 ) = k = 1 ∑ K ℓ = 1 ∑ m k ( ν k ℓ pq ∫ 0 T 0 B p ( t ) A k ( t ) d t y k ℓ 0 + ν k ℓ pq ∫ 0 T 0 B p ( t ) B k ( t ) d t y ˉ k ℓ 0 ) + n ∈ S ∑ ν n pq ∫ 0 T 0 B p ( t ) ( x n F ( t ) − d n ( t ) ) d t . (81)
To express (81)-(82) under a matrix form, we introduce the 2 N × 2 N 2N\times2N 2 N × 2 N matrix M M M and the two vectors b , Y b,Y b , Y of R 2 N \mathbb{R}^{2N} R 2 N defined by: For
p = 1 , … , K and k = 1 , … , K q = 1 , … , m p ℓ = 1 , … , m k . \begin{aligned}
p&=1,\ldots,K && \text{and} && k=1,\ldots,K\\
q&=1,\ldots,m_p && && \ell=1,\ldots,m_k.
\end{aligned} p q = 1 , … , K = 1 , … , m p and k = 1 , … , K ℓ = 1 , … , m k .
M p + q − 1 , k + ℓ − 1 = ν k ℓ p q ∫ 0 T 0 A p ( t ) A k ( t ) d t , M p + q − 1 , N + k + ℓ − 1 = ν k ℓ p q ∫ 0 T 0 A p ( t ) B k ( t ) d t , M N + p + q − 1 , k + ℓ − 1 = ν k ℓ p q ∫ 0 T 0 B p ( t ) A k ( t ) d t , M N + p + q − 1 , N + k + ℓ − 1 = ν k ℓ p q ∫ 0 T 0 B p ( t ) B k ( t ) d t . (82) \begin{aligned}
M_{p+q-1,k+\ell-1}&=\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{A}_{p}(t)\mathcal{A}_{k}(t)\,\mathrm{d}t, &
M_{p+q-1,N+k+\ell-1}&=\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{A}_{p}(t)\mathcal{B}_{k}(t)\,\mathrm{d}t,\\
M_{N+p+q-1,k+\ell-1}&=\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{B}_{p}(t)\mathcal{A}_{k}(t)\,\mathrm{d}t, &
M_{N+p+q-1,N+k+\ell-1}&=\nu_{k\ell}^{pq}\int_{0}^{T^{0}}\mathcal{B}_{p}(t)\mathcal{B}_{k}(t)\,\mathrm{d}t. \tag*{(82)}
\end{aligned} M p + q − 1 , k + ℓ − 1 M N + p + q − 1 , k + ℓ − 1 = ν k ℓ pq ∫ 0 T 0 A p ( t ) A k ( t ) d t , = ν k ℓ pq ∫ 0 T 0 B p ( t ) A k ( t ) d t , M p + q − 1 , N + k + ℓ − 1 M N + p + q − 1 , N + k + ℓ − 1 = ν k ℓ pq ∫ 0 T 0 A p ( t ) B k ( t ) d t , = ν k ℓ pq ∫ 0 T 0 B p ( t ) B k ( t ) d t . (82)
b p + q − 1 = ∑ n ∈ S ν n p q ∫ 0 T 0 A p ( t ) ( x n F ( t ) − d n ( t ) ) d t , b N + p + q − 1 = ∑ n ∈ S ν n p q ∫ 0 T 0 B p ( t ) ( x n F ( t ) − d n ( t ) ) d t and Y k + ℓ − 1 = y k ℓ 0 , Y N + k + ℓ − 1 = y ˉ k ℓ 0 . (83) \begin{aligned}
b_{p+q-1}&=\sum_{n\in\mathcal{S}}\nu_{n}^{pq}\int_{0}^{T^{0}}\mathcal{A}_{p}(t)\left(x_{n}^{F}(t)-d_{n}(t)\right)\,\mathrm{d}t,\\
b_{N+p+q-1}&=\sum_{n\in\mathcal{S}}\nu_{n}^{pq}\int_{0}^{T^{0}}\mathcal{B}_{p}(t)\left(x_{n}^{F}(t)-d_{n}(t)\right)\,\mathrm{d}t
\end{aligned}
\qquad\text{and}\qquad
\begin{aligned}
Y_{k+\ell-1}&=y_{k\ell}^{0},\\
Y_{N+k+\ell-1}&=\bar{y}_{k\ell}^{0}. \tag*{(83)}
\end{aligned} b p + q − 1 b N + p + q − 1 = n ∈ S ∑ ν n pq ∫ 0 T 0 A p ( t ) ( x n F ( t ) − d n ( t ) ) d t , = n ∈ S ∑ ν n pq ∫ 0 T 0 B p ( t ) ( x n F ( t ) − d n ( t ) ) d t and Y k + ℓ − 1 Y N + k + ℓ − 1 = y k ℓ 0 , = y ˉ k ℓ 0 . (83)
According to (83)-(84), the first order partial derivatives in (81) and (82) are given by:
∂ J ( Y ) ∂ Y k + ℓ − 1 = ∑ k = 1 K ∑ ℓ = 1 m k ( M p + q − 1 , k + ℓ − 1 Y k + ℓ − 1 + M p + q − 1 , N + k + ℓ − 1 Y N + k + ℓ − 1 ) + b p + q − 1 , (84) \frac{\partial J(Y)}{\partial Y_{k+\ell-1}}=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}\left(M_{p+q-1,k+\ell-1}Y_{k+\ell-1}+M_{p+q-1,N+k+\ell-1}Y_{N+k+\ell-1}\right)+b_{p+q-1}, \tag*{(84)} ∂ Y k + ℓ − 1 ∂ J ( Y ) = k = 1 ∑ K ℓ = 1 ∑ m k ( M p + q − 1 , k + ℓ − 1 Y k + ℓ − 1 + M p + q − 1 , N + k + ℓ − 1 Y N + k + ℓ − 1 ) + b p + q − 1 , (84)
∂ J ( Y ) ∂ Y N + k + ℓ − 1 = ∑ k = 1 K ∑ ℓ = 1 m k ( M N + p + q − 1 , k + ℓ − 1 Y k + ℓ − 1 + M N + p + q − 1 , N + k + ℓ − 1 Y N + k + ℓ − 1 ) + b N + p + q − 1 . \frac{\partial J(Y)}{\partial Y_{N+k+\ell-1}}=\sum_{k=1}^{K}\sum_{\ell=1}^{m_k}\left(M_{N+p+q-1,k+\ell-1}Y_{k+\ell-1}+M_{N+p+q-1,N+k+\ell-1}Y_{N+k+\ell-1}\right)+b_{N+p+q-1}. ∂ Y N + k + ℓ − 1 ∂ J ( Y ) = k = 1 ∑ K ℓ = 1 ∑ m k ( M N + p + q − 1 , k + ℓ − 1 Y k + ℓ − 1 + M N + p + q − 1 , N + k + ℓ − 1 Y N + k + ℓ − 1 ) + b N + p + q − 1 .
Then one can write
∇ J ( Y ) = M Y + b . (85) \nabla J(Y)=MY+b. \tag*{(85)} ∇ J ( Y ) = M Y + b . (85)