21st International Symposium on

Mathematical Theory of Networks and Systems

July 7-11, 2014. Groningen, The Netherlands

Optimal sensor placement and scheduling of hybrid PDEs arising in

environmental and meteorological applications
Antonios Armaou Michael A. Demetriou

Abstract— We consider environmental and meteorological where Pe denotes the Péclet number, u(t, ξ) the state at spatial
applications which undergo significant parametric changes that location ξ and time t, f (u) is nonlinear term and d(ξ)w(t) the
alter the behavior of the system. For the transport process disturbance; in the latter, the spatial function d(ξ) represents
under study, which depends on the Péclet number, the behavior
changes from an advection dominated to a diffusion dominated the spatial distribution of disturbances and the function w(t)
process. For each value of the Péclet parameter, a set of optimal its temporal input. The Péclet number Pe plays an important
sensor locations and number is found and the associated state role in the quantitative and qualitative bevavior of (2). The
estimator is subsequently designed. A supervisory scheme is variation of the Péclet number renders the PDE time varying;
then utilized to schedule the sensors corresponding to the a high Péclet number represents advective-dominated process
current value of the Péclet number by first turning off all the
sensors associated with a different value of the Péclet number and a low Péclet number represents diffusive-dominated
and then activate the sensors that are optimal for the current process. This diurnal variation of the Péclet number renders
value of the Péclet number. Simulation results are included (2) a switched dynamical system with a parameter (Pe) taking
to provide further insight on parameter-dependent sensor and integer values over small set. For simplicity, it is assumed
observer scheduling for a representative 1D environmental that the parameter set Θ consists of two elements
process.
Index Terms— Distributed parameter systems; hybrid PDEs; Θ = {Pea , Ped },
optimal sensor selection; sensor scheduling; gain optimization.
a high value Pea associated with an advective-dominated
process and a low value Ped associated with a diffusive-
I. P HYSICAL MODELING AND PROBLEM DESCRIPTION dominated process.
To enable the analysis, design and optimization associated
Many environmental processes associated with weather
with the sensor location and observer design, the above PDE
prediction and monitoring are modeled by advection-
is written in an abstract form as an evolution equation in
diffusion PDEs in bounded domains, infinite domains of
a function space. Specifically, it is written as an evolution
semi-infinite domains, [1], [2], [3], [4].
equation in the Hilbert space H ⊂ L2 (Ω)
A representative PDE that captures the salient features
of such environmental processes is the diffusion-advection u̇(t) = A (Pe)u(t) + F (u(t)) + D w(t) (3)
equation restricted on the real line and given by the 1D PDE
where the parameter-dependent spatial operator A (Pe) is
with source/sink term
defined as
∂ ∂2 ∂ ¯ 1 d2 d
ū(t, z) = D 2 ū(t, z) − υ ū(t, z) + f¯(u) + d(z)w(t). (1) A (Pe)φ = φ(ξ) − φ(ξ), φ ∈ H 2 (Ω),
∂t ∂z ∂z Pe dξ 2 dξ
Many choices for the boundary conditions are considered: for Pe ∈ Θ. The space H is equipped with the standard inner
Dirichlet BCs where the concentration is specified on a product of a weighted integral
boundary and Fourier BCs at an inlet boundary when con- Z
sidering a problem defined on a semi-infinite domain ξ > 0 hφ1 , φ2 i = ρ(ξ)φ∗1 (ξ)φ2 (ξ) dξ, φ1 , φ2 ∈ H,
Ω
with inlet boundary at ξ = 0. Following [4], in the Fourier p
BCs case the flux is required to be continuous across the and with norm |φ| = hφ, φi. The spatial function ρ(ξ)
boundary. denotes the weight function of the weighted inner product.
To better present the concepts discussed in this work, we Information of the process (2), or its abstract form (3) is
focus on the non-dimensionalized PDE form made possible via a finite number of measurements obtained
by N in-domain sensing devices
∂ 1 ∂2 ∂
u(t, ξ) = u(t, ξ) − u(t, ξ) + f (u) + d(ξ)w(t) (2)
 Z 
∂t Pe ∂ξ2 ∂ξ  Ω
c1 (ξ; ξs1 )u(t, ξ) dξ

Z  
A. Armaou is with Penn State University, Dept of Chemical Engineering,
y(t) = c(ξ;ξξs )u(t, ξ) dξ = 

 .
..

 (4)
University Park, PA 16802-4400, USA, [email protected]. The Ω

author gratefully acknowledges financial support from NSF, grant CBET
 Z 
 
12-64902. cN (ξ; ξsN )u(t, ξ) dξ
M. A. Demetriou is with Worcester Polytechnic Institute, Dept of Me- Ω
chanical Engineering, Worcester, MA 01609, USA, [email protected].
The author gratefully acknowledges financial support from the AFOSR, The output vector y(t) denotes the vector of measured
grant FA9550-12-1-0114. variables and the N dimensional vector of spatial functions

ISBN: 978-90-367-6321-9 124

MTNS 2014
Groningen, The Netherlands

c(ξ;ξξs ) denotes the sensor shape model. The N-dimensional Following our earlier work [5], one can obtain a transfer
vector ξ s ∈ (Ω)N denotes the vector of sensor locations or function model for the i th eigenmode, via the use of Laplace
sensor centroids. transforms with Xi (s) = L[xi (t)]
Typical choice for the spatial distribution of the sensing 1 hφi , di
devices ci (ξ; ξsi ) is the Dirac delta function ci (ξ; ξsi ) = Xi (s) = xi (0) + W (s), i = 1, . . . ∞,
s − (λi + k) s − (λi + k)
δ(ξ − ξsi ) which provides the state information at the spatial
location ξsi , where λi denotes the i th eigenvalue of the operator A (Pe).
Z In a similar fashion, one can obtain the expression for the
yi (t) = δ(ξ − ξsi )u(t, ξ) dξ = u(t, ξsi ), i = 1 . . . , N. Laplace transform of the measured output (6)
Ω
∞ Z 
We let ξ s = (ξs1 , ξs2 , . . . , ξsN ) denote the vector of the sensor Y (s;ξξs ) = ∑ c(ξ;ξξs )φi (ξ) dξ Xi (s)
locations and explicitly parameterize the output vector by ξ s i=1 Ω
in order to emphasize the dependence on sensor location  ∞ Z  
 Z  ∑ δ(ξ − ξs1 )φi (ξ) dξ Xi (s)

y1 (t; ξs1 )
 δ(ξ − ξs1 )u(t, ξ) dξ 
 i=1 Ω 

 Ω   
   
= 
 .. 
y(t;ξξs ) =  .
.
 .
.

.
=  (5)
   
 .   .  
 Z 


Z   ∞ 
yN (t; ξsN ) δ(ξ − ξsN )φi (ξ) dξ Xi (s)
 
δ(ξ − ξ )u(t, ξ) dξ
sN ∑ Ω
Ω i=1

The above can be written in terms of the output operator,

To present the output in a compact form and to facilitate
parameterized by both the sensor location ξ s and the Péclet
the computation of an associated spatial norm, we define the
number
N-dimensional output matrix
y(t;ξξs ) = C (ξξs ; Pe)u(t). (6) Z
We can now state the design objective: Ci (ξξs ) = c(ξ;ξξs )φi (ξ) dξ, i = 1, . . . , ∞
Ω
Problem statement: Given the parameter-dependent hybrid
which in view of the specific sensor ci (ξ; ξsi ) = δ(ξ − ξsi )
system (3) with location-parameterized measurements (5),
takes the form
1) Find the optimal sensor location that corresponds to  
each switch mode of the parameter-dependent process φi (ξs1 )
 
(3); in other words, for each value of the parameter Ci (ξξs ) =  ..
 i = 1, . . . ∞.
 
Pe ∈ Θ = {Pea , Ped }, find the corresponding set of  . 
sensor locations that render the pair (A (Pe), C (ξξs ; Pe)) φi (ξsN )
approximately observable and optimize an appropriate
Similarly, we define the disturbance distribution matrix,
measure of observability and filter performance
where Di represents the projection of the spatial distribution
2) For each value of Pe ∈ Θ, design the optimal filter as-
of the disturbances d(ξ) with respect to the basis function
sociated with the pair (A (Pe), C (ξξs ; Pe)) and schedule
set {ϕi (ξ)}i=1 , i.e.
the set of sensors that correspond to that value of Pe.
II. S PATIAL DISCRETIZATION AND FINITE DIMENSIONAL
Z
REPRESENTATION
Di = ρ(ξ)φ∗i (ξ)d(ξ) dξ = hφi , di, i = 1, . . . ∞.
Ω
An appropriate approximation that plays a central role The above enable the Laplace transform of the N-
in defining the optimization measures is based on spectral dimensional output vector expressed as
expansion, where the state admits the expansion ∞ ∞
xi (0) Di
∞ Y (s;ξξs ) = ∑ Ci (ξξs ) + ∑ Ci (ξξs ) W (s)
u(t, ξ) = ∑ φi (ξ)xi (t). i=1 s − (λi + k) i=1 s − (λi + k)
i=1    
φi (ξs1 )xi (0) φi (ξs1 )hφi , di
The spatial functions φi (ξ), are the eigenfunctions of the  s − (λi + k)   s − (λi + k) 
   
spatial operator, and are used as a basis for the space H ∞   ∞  
= ∑ .. .
..
+ ∑  W (s).
   
in order to express the system state, u, as a function of sys- .
i=1 
  i=1  
tem eigenmodes xi (t). Please note that these eigenfunctions   
depend on the parameter Pe.
 φi (ξsN )xi (0)   φi (ξsN )hφi , di 
To analyze the evolution equation (3), the linear com- s − (λi + k) s − (λi + k)
ponent of F (u) ≈ k u in (3) is retained, resulting in the In more compact form, it is written as
following infinite dimensional ODE in H ∞ ∞

ẋ(t)= A (Pe)x(t) + kx(t) + D (Pe)w(t) Y (s;ξξs ) = ∑ Gi (s;ξξs )xi (0) + ∑ Hi (s;ξξs )W (s)
i=1 i=1 (8)
(7)
y(t;ξξs )= C (ξξs ; Pe) x(t). = G(s;ξξs , Pe)x(0) + H(s;ξξs , Pe)W (s),

125
MTNS 2014
Groningen, The Netherlands

where Gi (s;ξξs ) or Hi (s;ξξs ) denote the transfer function be- Using the expansion and taking Laplace transforms of (10),
tween the measurement vector or the disturbance and the i th we arrive at an expression similar to that for the process state
eigenmode, respectively. Equation (8) can be used to provide ∞
Ci (ξξs )ei (0) ∞ Ci (ξξs )Di
the optimal sensor locations using open-loop techniques; E (s;ξξs ) = ∑ +∑ W (s)
i=1 s − µi (ξ s ) i=1 s − µi (ξ s )
ξ ξ
essentially the optimal sensor locations are found so that the ∞ ∞
system becomes “more observable”. Appropriate measures (11)
= ∑ Ji (s;ξξs )e(0) + ∑ Ki (s;ξξs )W (s)
of enhanced observability may include the observability i=1 i=1
Gramian, the spatial H2 and H∞ norms. = J(s;ξξs , Pe)e(0) + K(s;ξξs , Pe)W (s).
The above open loop measures may be used to obtain the
Similar to the definition of λi , the location-parameterized
set of admissible sensor locations that can subsequently be
µi (ξξs ) denotes the eigenvalues of the (parameter-dependent)
used to design optimal state estimators. A state estimator for
closed loop operator A (Pe) + kI − L (ξξs , Pe)C (ξξs , Pe).
(7), parameterized by the sensor locations, is given by
In view of (11), one can now quantify the first design ob-
 
˙ (t) = A (Pe) + kI − L (ξξ , Pe)C (ξξ , Pe) xb(t)
xb jective presented in the previous section. The sensor location
s s
(9) optimization can now be stated as that of finding the optimal
+L (ξs , Pe)y(t; ξs ), locations ξ s for each value Pe ∈ Θ so that
where xb(t, ξ) is the estimate of the state x(t) and L (ξξs , Pe) • it maximizes the spatial H2 norm of G(s;ξ ξs , Pe) in (8),
denotes the location-parametrized filter gain. The design of i.e. it maximizes the value of information,
the filter gain is not explicitly defined, it may be a Kalman • it minimizes the spatial H2 norm of K(s;ξ ξs , Pe) in (11),
filter design, or simply based on Luenberger observer design. i.e. it enhances the estimator performance.
The state error e(t) = x(t) − xb(t) associated with the above Please note that the optimization of the spatial H2 norm of
estimator is given by G(s;ξξs , Pe) is considered an “open-loop” optimization since
  it only improves the observability of the sensor locations. The
ė(t) = A (Pe)+kI − L (ξξs , Pe)C (ξξs , Pe) e(t)+ D w(t). (10) minimization of the spatial H2 norm of K(s;ξξs , Pe) chooses
from the admissible set of sensor locations that have already
The state error (10) can be used to find the optimal sensor enhanced observability and then improves the performance
locations; it can be used to find the locations ξ s that minimize of the state estimator.
the effects of the disturbance w(t) on the state error and also The next two sections will summarize the sensor place-
to minimize an energy cost associated with the state error. ment scheme and the sensor scheduling scheme.
The sensor location optimization for the state error must
III. O PTIMAL SENSOR PLACEMENT SCHEME
also take into account the hybrid nature of the process (7). An
optimal sensor location for a given value of the parameter Pe Once the state error is represented in (11) explicitly in
may not be an optimal value for a different value. This aspect terms of the initial conditions and the disturbances, the sensor
must be considered in the sensor optimization problem. placement can be presented. The location optimization can
In order to quantify the optimization objectives, we pro- be posed as a constrained nonlinear optimization problem.
ceed with an assumption similar to the one made about the To do so, we must first define the objective function and
expansion of the process state. We assume that the spatially then present the optimization formulation.
distributed state estimation error e(t, ξ) can be expanded as A. Observability measures
∞ Following our earlier fundamental work [5], we provide a
e(t, ξ) = ∑ ϕi (ξ)ei (t), quantitative metric for observability, which is based on the
i=1
spatial H2 norms first presented in [6]. The spatial H2 norm
where ϕi (ξ) is the ith eigenfunction of the parameter- at sensor location ξ s of G(s, ξ;ξξs ) in (8) is defined as
dependent closed loop operator A (Pe) + kI − 1 ∞
Z Z
L (ξξs , Pe)C (ξξs , Pe). However, to obtain expressions similar kGk2H2 = ρ(ξ)tr {G∗ ( jω, ξ;ξξs )G( jω, ξ;ξξs )} dξ dω
2π −∞ Ω
to the sensor sensitivities to specific sensor locations, we
consider the output estimation error and represents a measure of sensor sensitivity placed at
location ξ s over the entire spatial domain in an average sense.
∞ Z 
e(t;ξξs ) = ∑ c(ξ;ξξs )ϕi (ξ) dξ ei (t) Using the modal expansion for the process state, the spatial
i=1 Ω norm can be expressed in terms of the system modes
∞ Z ∞
We also define the vector of the (closed-loop) modes at each kGk2H = ∑ tr {G∗i ( jω;ξξs )Gi ( jω;ξξs )} dξ
2
of the N sensor locations i=1 −∞ (12)
∞

ϕi (ξs1 )
 = ∑ kGi (s; ξs )k22 .
i=1
 
.. This then prompts the simplification involving the
Ci (ξξs ) =   i = 1, . . . ∞, modal
 
 .  Ci (ξξs )
norms fi (ξξs , Pe) = kGi (s;ξξs , Pe)k2 =
.
ϕi (ξsN ) s − (λi + k) 2

126
MTNS 2014
Groningen, The Netherlands

The modal norm fi (ξξs , Pe) represents a measure of sensor leading to sensor locations with attributes:
network sensitivity placed at location ξ s to the i th system • maximize the network sensitivity to dominant system
mode. Each fi is an N-dimensional vector. dynamics
The modal norm in (12) is an infinite dimensional vector. • minimize the effects of disturbances on observer perfor-
Since the higher modes become progressively more stable, mance
then one can truncate the infinite series expansion and only
IV. O PTIMAL SENSOR SCHEDULING SCHEME
consider a finite number of modes. Using the above, we have
∞ ∞ Once the optimal set os sensors (equiv. sensor centroids
kG(s, ξ;ξξs , Pe)k2H = ∑ kGi (s;ξξs , Pe)k22 = ∑ fi2 (ξξs , Pe) ξs ) for each value of the parameter Pe ∈ Θ, one can proceed
2
i=1 i=1 with the scheduling policy of both the sensors and the
M ∞ estimator gains. We denote the output matrix associated with
= ∑ fi2 (ξξs , Pe) + ∑ fi2 (ξξs , Pe) the optimal values of the sensor centroids for a given value
i=1 i=M+1 of the parameter Pek ∈ Θ by
2 2
=H (ξξs , Pe) + S (ξξs , Pe), C(ξks ; Pek ), Pek ∈ Θ
(13)
M and denote the optimal filter gain associated with the pair
where H2 (ξξs , Pe) = ∑ fi2 (ξξs , Pe) denotes the truncation (A (Pek ),C(ξks ; Pek ) by
i=1
of the H2 spatial norm to the first M slow modes and L(ξks ; Pek ), Pek ∈ Θ.
∞
2
S (ξs , Pe) = ∑ fi2 (ξξs , Pe) denotes the “spillover” of the Then, for each value of Pek ∈ Θ we implement the following
i=M+1 estimator that utilizes the current set of sensors defined by
higher modes’ dynamics on the sensor accuracy. With only
C(ξks ; Pek ) via
open-loop measures, the sensor location optimization prob-
˙ (t; ξk ) = A (Pek ) − L(ξk ; Pek )C(ξk ; Pek ) xb(t; ξk )


lem should maximize the value of information over the xb s s s s
dominant modes, i.e. maximize H2 (ξξs , Pe) over a set of (16)
k k k
+L(ξs ; Pek )C(ξs ; Pek )y(t; xis ).
admissible locations, for each value Pe ∈ Θ. Additionally,
it should minimize the effects of the disturbance on the
measurements. The later can be included in the minimization Algorithm 1 Sensor scheduling for process monitoring
of the disturbance effects on the state estimation error (11). 1: measure physical parameter Pek ∈ Θ
B. Filter performance measures 2: switch to optimal sensors ξks associated with current
Similar to the expansion (14), we can expand (11) in Pek ∈ Θ
terms of the contributions of the individual modes of the 3: output process measurements y(t; ξks )
closed-loop system. Using the orthogonality assumption of 4: repeat for next value of Pe` ∈ Θ
the closed loop modes and the fact that the higher mode
of the closed loop system are also becoming progressively
more stable, then in a similar fashion to (14), we have the Algorithm 2 Estimation gain scheduling for state estimation
truncated expressions for the spatial H2 norm 1: read physical parameter Pek ∈ Θ
∞
2: obtain process measurements with current sensor set ξks
kK(s;ξξs , Pe)k2
H2 = ∑ kKi (s;ξξs , Pe)k22
i=1
3: compute filter gain L(ξks ; Pek ) corresponding to current
optimal sensor locations
M ∞
4: implement state estimator (17)
= ∑ kKi (s;ξξs , Pe)k22 + ∑ kKi (s;ξξs , Pe)k22
i=1 i=M+1 5: repeat for next value of Pe` ∈ Θ
2 2
=L (ξξs , Pe) + M (ξξs , Pe),
(14) R EFERENCES
M
2 [1] J. H. Seinfeld and S. N. Pandis, Atmospheric Chemistry and Physics:
where L (ξξs , Pe) = ∑ kKi (s;ξξs , Pe)k22 denotes the truncation From Air Pollution to Climate Change. New York: Wiley-Interscience,
i=1 1997.
of the H2 spatial norm to the first slow M modes and [2] J. H. Seinfeld, Atmospheric Chemistry and Physics of Air Pollution.
∞
New York: John Wiley & Sons, Inc., 1995.
M2 (ξs , Pe) = ∑ kKi (s;ξξs , Pe)k22 denotes the “spillover” of [3] E. Kalnay, Atmospheric Modeling, Data Assimilation and Predictability.
i=M+1 40 West 20th Street, New York, NY: Cambridge University Press, 2002.
the higher modes’ dynamics on the sensor accuracy. [4] J. D. Logan, Transport modeling in hydrogeochemical systems, ser.
Interdisciplinary Applied Mathematics. New York: Springer-Verlag,
C. Design of sensor network 2001, vol. 15, geophysics and Planetary Sciences.
Using both open-loop (enhanced observability) and close- [5] A. Armaou and M. A. Demetriou, “Optimal actuator/sensor placement
for linear parabolic pdes using spatial H2 norm,” Chemical Engineering
loop (performance) measures, an optimization problem can Science, vol. 61, no. 22, pp. 7351–7367, 2006.
be formulated as [6] D. Halim and S. R. Moheimani, “An optimization approach to optimal
placement of collocated piezoelectric actuators and sensors on a thin
ξ∗s = arg max H2 (ξξs ) − ωL L2 (ξξs )

ξs ∈Ωs
(15) plate,” Mechatronics, vol. 13, pp. 27–47, 2003.

127