Proc IMechE Part I: J Systems and

Control Engineering
Unknown Input Estimation Algorithms 00(0):1–9
©IMechE 2021
for a Class of LPV/Nonlinear Systems sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/ToBeAssigned

with Application to Wastewater www.sagepub.com/

SAGE
Treatment Process
Chaouche Alima1 , Zemouche Ali2 , Chaib Draa Khadidja3 ,Delattre Cédric 4 ,Ramdani
Messaoud2

Abstract
This paper addresses the problem of unknown input estimation for a class of nonlinear systems with mixed nonlinear
terms, namely Linear Parameter Varying (LPV) parts and purely Lipschitz nonlinearities. In a general context, we
consider systems with nonlinearities in the process and in the output measurements. Then, three estimation algorithms
are proposed, where each algorithm depends on the distribution of the unknown inputs in the system. In other word,
mathematically speaking, the estimation procedures depend on some necessary rank conditions ensuring the existence
of some key estimation parameters. Before introducing the unknown input estimation algorithms, a general LMI-based
design methodology is provided, as a preliminary result, to guarantee exponential convergence of the estimators by
using an H∞ −optimality criterion. To this end, a specific Lyapunov function depending on the disturbances, is used to
avoid derivatives of the disturbances. The proposed algorithms are applied to a wastewater treatment model to show
their effectiveness and performances.

Keywords
Estimation, Descriptor systems; Unknown input observers; LMIs; H∞ analysis; LPV systems; Wastewater treatment.

Introduction treatment plants require better management strategies

[chachouar eduction].Realizingthebenef itstobederivedf rommathem
Usually, controlled real systems are subject to exogenous delaysystemswithuncertainties.Similarly, in 2 , anobserverinprese
inputs, such actuator faults or disturbances affecting the performance criterion, which is based on the Lebesgue space
system state variables and the output measurements. Within L2 for obtaining local input-output stability.
this context, unknown input estimation has owned a great Since, nonlinearities in the system outputs lead to more
interest in the areas of observer-based fault diagnosis 1 , complexity in the observer design, some assumptions on the
chaotic synchronization and secure communications 2 , and system parameters such the number of unknown inputs, and
observer-based fault tolerant control 3 . nonlinear functions in the measurement equations will be
Hence, various observer design approaches have been discussed. To guarantee the observer convergence, tractable
developed with respect to some specific assumptions on LMI conditions will be formulated. Their feasibility will be
the nonlinearities of the system. One of the most popular enhanced due to a convenient use of the Young’s inequality.
methods and an important tool for output feedback control This will be possible trough the introduction of additional
of nonlinear systems is the well known high-gain observer. decision variables, which leads to additional degrees of
It is designed for systems that can be written under the freedom and, consequently, better performances in the LMI
triangular form 4 , 5 , 6 . Another method for nonlinear systems feasibility.
with Lipschitz nonlinearities is addressed using sliding The proposed estimation algorithms are applied to a
mode observers. The design of such estimators is based reduced model of activated sludge wastewater treatment
on the solution of an algebraic Riccati equation, which plant (ASM1), which was developed in the European
maybe difficult to solve 7 . Therefore, a simpler approach program COST 624 for control purposes 15 .
is the Extended Kalman Observer (EKO), which uses Notation: The following notation will be used throughout
a first order linearizion technique. However, although this paper.
the EKO is simple to design and is largely used in real
applications 8 , it is very sensitive to the initialization and its
convergence is guaranteed only locally.Dynamic simulation 1 Laboratory of Automation of Skikda (LAS), Faculty of Technology,

based on rigorous, physics-based mechanistic modeling University 20 August 1955, Skikda, Algeria
2 University of Lorraine, 186, rue de Lorraine, CRAN UMR CNRS 7039,
has become a standard tool in many engineering fields.
54400 Cosnes et Romain, France
They are quite helpful in system design, especially to 3 University of Luxembourg, Belval, Luxembourg
check system behaviour under extreme dynamic loading 4 Laboratory of Automation and Signals of Annaba (LASA), Faculty of
conditions. In order to meet the stricter wastewater effluent Engineering,University Badji-Mokhtar, Annaba, Algeria
guidelines adopted by the European Union, wastewater

Prepared using sagej.cls [Version: 2017/01/17 v1.20]

2 Proc IMechE Part I: J Systems and Control Engineering 00(0)

• ∥.∥ is the usual Euclidean norm; Assumption 1. Assume that the following condition is
• (⋆) is used for the blocks induced by symmetry; fulfilled:
• AT represents the transposed matrix of A; 
Eξ 0

• Ir represents the identity matrix of dimension r; rank = nξ + nh . (3)
Cξ Bh
• for a square matrix S, S > 0 (S < 0) means that this
matrix is positive definite (negative definite); Condition (3) implies
R  p1
+∞ p
• the notation ∥x∥Lrp = 0
∥x(t)∥ dt is the Lrp
nd + ny ≥ nξ + nh .
norm of the vector x ∈ Rr . The set Lrp is defined by
n o Under Assumption 1 we can build an observer for the
Lrp = x ∈ Rr : ∥x∥Lrp < +∞ system (1). Indeed, if (3) holds then there exist matrices Pz ∈
Rnξ ×nd , Qz ∈ Rnξ ×ny , Rz ∈ Rnh ×nd , and Sz ∈ Rnh ×ny
and then (Lrp , ∥.∥Lrp ) is called the Lebesgue space; such that
i th
z}|{
T     
• es (i) = 0, ..., 0, 1 , 0, ..., 0 ∈ Rs , s ≥ 1 is a vector Pz Qz Eξ 0 Inξ 0nξ ×nh
| {z
s components
}   = . (4)
of the canonical basis of Rs ; Rz Sz C ξ Bh 0nh ×nξ I nh

Equation (4) means that

Problem Formulation and Preliminary
Results 
 Pz Eξ + Qz Cξ = Inξ

Rz Eξ + Sz Cξ = 0nh ×nξ

This section is devoted to some preliminary results we will (5)
use for the unknown input estimation. This result requires the 
 Qz Bh = 0nξ ×nh
Sz Bh = Inh .

following assumption.

 z }|˙ { The matrices Pz , Qz , Rz , Sz are computed by
 E ξ(t) = A ρ(t)
ξ(t) + B ρ(t)
f ξ(t), u(t)

ξ ξ
f

 + Gg ρ(t) g y(t),
u(t) + Dξ ω(t)    ⊤  !−1  ⊤
y(t) = Cξ ξ(t) + Bh h ξ(t), u(t) + Dy ω(t) Pz Qz Eξ 0 Eξ 0 Eξ 0

Rz Sz
= Cξ Bh Cξ Bh Cξ Bh
.
(1)
nξ nu ny
where ξ ∈ R , u ∈ R , and y ∈ R are respectively the (6)
state vector of the system, the known control input vector, the
 
Pz Qz
output measurements vector. The vector ω ∈ Rnω represents In other word, is the Moore-Penrose pseudo-
Rz Sz
the unknown disturbances affecting the system dynamics and 
Eξ 0

the measured variables. The matrices Eξ ∈ Rnd ×nξ , Dξ ∈ inverse of the full column rank matrix .
C ξ Bh
Rnd ×nω , Cξ ∈ Rny ×nξ , Bh ∈ Rny ×nh , and Dy ∈ Rny
×nω Consider the following nonlinear filter:
are known and constant. The affine matrix Aξ ρ(t) , is
expressed under the form 
ż(t) =



Az ρ(t) z(t) + Ay ρ(t) yη1 (t)

ˆ





s


 +Bz ρ(t) f ξ(t), u(t) + Gz ρ(t) g y(t), u(t)

X  ˆ
ξ(t) = z(t) + Qz yη2 (t)
Aξ ρ(t) = A0 + ρj (t)Aj
yηi (t) = y(t) − ηi (t), i = 1, 2
j=1

ˆ


η1 (t) = Bh h ξ(t), u(t)




with η2 (t) = 0Rny .
(7)
ρj,min (t) ≤ ρj (t) ≤ ρj,max (t)
By taking into account the equalities (5), the dynamics of
which means that the parameter ρ(t) belongs to a bounded ˜ ≜ ξ(t)
the estimation error ξ(t) ˆ − ξ(t) can be rewritten as
convex set for which the set of 2s vertices can be defined by: follows:

Vρ = ρ(t) ∈ Rs : ρj (t) ∈ {ρj,min (t), ρj,max (t)}


˜ − Qz Dy ω(t) = z(t) − Pz Eξ ξ(t).
ξ(t) (8)

The functions Therefore, we can write

nξ nu nf
f :R ×R −→ R ˙ {
z }|
ζ̇(t) = ż(t) − Pz Eξ ξ(t) (9a)
and
h : Rnξ × Rnu −→ Rnh
˜ − Qz Dy ω(t).
ζ(t) = ξ(t) (9b)
are globally Lipschitz* with respect to ξ, uniformly on u,
with Lipschitz constants γf and
γh , respectively. Assume
also that the couple Eξ , Cξ satisfies the following
condition:   ∗ If f and h are only locally Lipschitz, we may consider their saturated
Eξ versions f s and hs , respectively, on an invariant compact set X in which
rank = nξ . (2) f s and hs satisfy the global Lipschitz property 16 .
Cξ

Prepared using sagej.cls

Alima 3

Using the dynamics (1) and (7) we get the following detailed Unknown Input Estimation Algorithms
structure of (9): This section is devoted to unknown input estimation for
nonlinear systems with nonlinear outputs. The aim consists
 
˜ + Az ρ(t) − A + LCξ ξ



ζ̇(t) = Az ρ(t) ξ(t)
 to use the results of the previous section to estimate, in the


 H∞ sense, unknown inputs occurring in the system. The
+ Bz ρ(t) − Pz Bf ρ(t) f (ξ, u)
 class of systems concerned by this study is described by the


following nonlinear equations:


+ Gz ρ(t) − Pz Gg ρ(t) g(y, u)
  (






+ Eω − LDω ω + Bz ρ(t) δft − Ay Bh ρ(t) δht

ẋ = Ax ρ(t) x + Aµ
ρ(t) µ
+ Bσ ρ(t) σ x, µ, u
+ Gϱ ρ(t) ϱ y, u +
Dx ω
(10a) y = Cx x + Cµ µ + Bς ς x, µ, u + D̄y ω
(25)
˜ − Qz Dy ω(t), δft = f (ξ,
ˆ u) − f (ξ, u) (10b) where x(t) ∈ Rnx is the system state and µ(t) ∈ Rnµ is
ζ(t) = ξ(t)
the unknown input vector. The nonlinearities σ(.) and ς(.)
satisfy the global Lipschitz assumption as in the previous
ˆ u) − h(ξ, u), A = Pz Aξ ,
δht = h(ξ, (10c) section. For sake of brevity, we omit the LPV parameter ρ(t)
in the sequel.
We will present some schemes to estimate simultaneously
L = Ay − Az Qz , Eω = −Pz Dξ , Dω = −Dy . (10d) the system state x(t) and the unknown input µ(t). In
each estimation scheme, we rewrite system (25) under the
H∞ criterion descriptor form (1) according to specific assumptions on
This subsection is devoted to some definitions related to the parameters of (25), namely the
matrices Aµ , Cµ , Bς , Cx ,
the H∞ performance criterion. The aim consists in finding and the nonlinear function ς x, u . There are several ways
η1 (t), η2 (t) and the matrices Az , Ay , Bz , Gz , and Qz so that to organize the presentation of all the possible scenarios
˜ satisfies the criterion:
the estimation error ξ(t) due to different assumptions required in each scenario.
q Nevertheless, for practical considerations we will investigate
˜ nξ ≤ µ∞ ∥ω∥2 nω + ν∞ ∥∥ξ˜0 ∥2
∥ξ∥ (11) three general cases summarized as follows:
L 2 L 2

where µ∞ > 0 is the disturbance attenuation level and ν∞ > a) rank(Cµ ) = nµ ;

√ b) Cµ = 0;
0 is to be determined. In fact, µ∞ is the disturbance gain
˜ c) 0 < rank(Cµ ) < nµ .
from ω to ξ.
Usually we use Lyapunov functions to get checkable The organization above is interesting in the sense that
conditions guaranteeing (11). In the LMI framework, we take it depends on the matrix Cµ which intervenes in the
˜ such that
a quadratic Lyapunov function V (ξ), output measurements. Indeed, the different scenarios we
will provide strongly depend on the nature and the number
dV ˜ ˜ 2 − µ∞ ∥ω∥2 ≤ 0.
ϑ(t) ≜ (ξ) + ∥ξ∥ (12) of sensors needed to be able to estimate simultaneously
dt the system states and the unknown inputs. This plays a
The objective is to develop some LMI conditions under crucial role because from practical point of view, some
which the inequality (12) holds. In section , a LMI sensors may be extremely expensive or unavailable at any
technique will be proposed, where it is valid under specific cost. This organization offers, for a large number of users
assumptions. and a wide field of applications, the possibility to estimate
unknown variables according to their sensors availability and
LMI design procedure accessibility.
In the following, we will provide LMI conditions ensuring
Theorem 1. Assume that there exist symmetric positive def-
i i i i the asymptotic estimation of the states and the unknown
inite matrices P ∈ Rnξ ×nξ , Sij ∈ Rnf ×nf , Mij ∈ Rnh ×nh
inputs. In each case, after transformation of the model (25),
and a matrix R ∈ Rny ×nξ such that the following convex
we use the LMI design technique given in Section II.
optimization problem is solvable:

min(µ∞ ) subject to (14) (13) First case: rank(Cµ ) = nµ

System (25) can be rewritten under the form (1) with

Az = A − LCξ , Ay = L + Az Qz (24a)    
Eξ = Inx 0nx ×nµ , Aξ = Ax Aµ , (26a)
 
Bz = Pz Bf , Gz = Pz Gg (24b)   x
Cξ = Cx Cµ , ξ = , (26b)
µ
Then, the H∞ criterion (11) is satisfied with ν∞ = λmax (P).
Hence, the observer parameter Lj is computed by
Bf = Bσ , Gg = Gϱ , Bh = Bς ,
−1
Lj = P RTj . Dξ = Dx , Dy = D̄y , (26c)

Proof. The proof is omitted.

Prepared using sagej.cls

4 Proc IMechE Part I: J Systems and Control Engineering 00(0)

 Ps Σf Σh

PEω − RT0 Dω − Ω1 − ρj RTj Dω

(1, 1) j=1 z }| { z }| {



 Σf1 ... Σfm Σh1 ... Σhq 
 (⋆)
 − Ω2 + ΩT2 − µ∞ Iq 

 ≤0 (14)
 

 (⋆) −Λf S 0 

 
(⋆) (⋆) −Λh M
with
s
X  
AT0 P + PA0 − CξT R0 − RT0 Cξ + ρj ATj P + PAj − CξT Rj − RTj Cξ (15)
j=1
h    i h    i
Σfi = Ni,1
f f
P, Si1 . . . Ni,n i P, Sinif , Σhi = Ni,1
h h
Rj , Mi1 . . . Ni,ni Rj , Minih (16)
f h

  PB H   T   RT B F + Ps ρ RT B F   T 
Hi Sik Fi Mik
 
0 h ik j=1 j j h ik
f
Ni,k P, Sik = ¯⊤ z ik + , Ni,kh
Rj , Mik = −
D y PBz Hik 0 D¯⊤ y RT0 Bh Fik 0
! (17)
  2 2
Λf = block-diag Λf1 , . . . , Λfnf , Λfi = block-diag I i ,..., I i (18)
bi1 nf binif nf

nif times
  z }| {
S = block-diag S1 , . . . , Snf , Si = block-diag Si1 , . . . , Sinif (19)
!
  2 2
Λ = block-diag Λh1 , ..., Λhnh , Λhi = block-diag
h
Inih , . . . , I i (20)
di1 dinih nh

h ni times
  z }| {
M = block-diag M1 , . . . , Mnh , Mi = block-diag Mi1 , . . . , Minih (21)
s
X s
X
Ω1 = AT0 PQz Dy − CξT R0 Qz Dy + ρj ATj PQz Dy − ρj CξT Rj Qz Dy (22)
j=1 j=1
s
X
Ω2 = ETω PQz Dy − DTω R0 Qz Dy − ρj DTω Rj Qz Dy (23)
j=1

f ξ, u = σ x, µ, u , h ξ, u = ς x, µ, u , g y, u = ϱ y, u .able to use the LMI design method presented in Section , we
(26d) need to find a convenient transformation of the system (25)
into the form (1) for which the rank condition (2) is fulfilled.
Since the matrix Cµ is full column rank then it follows that
There are some methods in the literature to deal
the condition (2) holds. Indeed,
with this issue but it still remains open and some new
 
Eξ

Inx 0nx ×nµ
 ideas and improvements are possible even if this may
rank   = rank   require additional assumptions. These methods are generally
Cξ Cx Cµ established for systems where all the nonlinearities are
independent of µ(t). What we propose in this section
= nx + nµ = nξ . (27) is inspired from the basic work in 17 developed for
linear systems with unknown inputs. The generalization to
Hence the LMI design technique given in Section can be
nonlinear systems with nonlinear outputs, in the presence
applied according to additional assumptions.Indeed, under
of disturbances in both the dynamics of the system and
Assumption 1, the observer-based filter (7) provides a
the measurements, is not obvious. However, to use the
simultaneous estimation of the state x(t) and the unknown
results reported in Section , we will introduce specific linear
input µ(t) in the H∞ sense defined in (11) provided
transformations, inspired from 17 , to get a new system under
that the filter parameters are given by solving the convex
the form (1).
optimization problem (13) in Theorem 1 according to the
notations (57). As often required in all unknown input estimation
procedures summarized in the three previous items, we need
Second case: Cµ = 0 the following assumption on the nonlinearities σ(., .) and
ς(., .):
In this case, the transformation (57) is not convenient
because the rank condition (2) (or (27)) does not hold. To be

Prepared using sagej.cls

Alima 5

Assumption 2. Assume that the nonlinearities σ(., .) and Finally, we can apply the design procedure of Section
ς(., .) satisfy the following conditions: “Problem Formulation and Preliminary Results”. It follows

that under Assumption 1 and Assumption 3, the observer-
dσ x, µ, u based filter (7) provides an estimation of the state ξ(t) ≜
≡ 0, (28)
dµ S−1 x(t) in the H∞ sense defined in (11) provided that

the filter parameters are given by solving the convex
dς x, µ, u
≡ 0. (29) optimization problem (13) in Theorem 1 according to the
dµ notations (33). The system state x(t) is estimated by
Assumption 2 means that the unknown input µ(t) appears
ˆ
x̂(t) = Sξ(t) (36)
linearly in the system (25). Without loss of generality, the
matrix Aµ is defined as full column rank. Indeed, if Aµ is
which satisfies the H∞ inequality
not of column rank, then there always exist a full column
rank matrix Āµ̄ and a vector µ̄ such that Aµ µ(t) = Āµ̄ µ̄(t).
q
∥x̃∥Ln2 x ≤ µx ∥ω∥2Lnω + νx ∥x̃0 ∥2 (37)
From Assumption 2 we can write 2

σ x, µ, u ≡ σ̸µ x, u and ς x, µ, u ≡ ς̸µ x, u . (30) where x̃(t) = x(t) − x̂(t) and

Estimation of the system state: Since Aµ is full column µx = µ∞ ∥S∥2 , νx = ν∞ ∥S∥2 ∥S−1 ∥2 .
rank, then there is a matrix T ∈ Rnx ×(nx −nµ ) such that
Estimation of the unknown input: Now that the state of
the system is estimated, it remains to estimate the unknown
 
S = T Aµ (31)
input µ(t), which was not estimated simultaneously with
is nonsingular and x(t). To do that, we will exploit x̂(t), but we need to use its
  ˙
derivative x̂(t), which may be calculated numerically online.
0(nx −nµ )×nµ
S−1 Aµ = . (32) Introducing the notations
Inµ
ξ2 = 0nµ ×(nx −nµ ) Inµ ξ ≜ S2 ξ, A2,ξ = S2 S−1 Ax S
 
Then using the transformation ξ(t) ≜ S−1 x(t), we get a
ξ-system under the descriptor form (1) with the following B2,f = S2 S−1 Bσ , G2,h = S2 S−1 Gϱ , D2,ξ = S2 S−1 Dx ,
convenient parameters:
then from the transformation (31)-(32), we get
Eξ = Inx −nµ 0(nx −nµ )×nµ , Aξ = Eξ S−1 Ax S, (33a)
 
µ(t) = ξ˙2 − A2,ξ ξ − B2,f σ̸µ Sξ, u


Cξ = Cx T Cx Aµ , Bf = Eξ S−1 Bσ ,
 
(33b)

− G2,g ϱ y, u + D2,ξ ω. (38)
Gg = Eξ S−1 Gϱ , Bh = Eξ S−1 Bς , (33c)
−1 −1
Then, we propose to estimate the unknown input by:
Dξ = Eξ S Dx , Dy = Eξ S D̄y , (33d)


˙
µ̂(t) = ξˆ2 − A2,ξ ξˆ − B2,f σ̸µ̂ Sξ,
ˆ u − G2,g ϱ y, u (39)



f ξ, u = σ̸µ Sξ, u ,



h ξ, u = ς̸µ Sξ, u ,


Now we can state two propositions related to the
g y, u = ϱ y, u , (33e) estimation of the unknown input µ(t). To simplify, we first
Thanks to the transformation (31)-(32) the state vector ξ is present the results in the output disturbance-free case, i.e:
decoupled from the unknown input, which vanishes from the D̄y = 0. Later, we will provide new bounds for D̄y ̸= 0,
system (1) corresponding to (33). If the rank condition (2) where the derivative of the disturbance will appear in the
is fulfilled, then it is possible to construct an H∞ filter formulas.
providing an estimation of the vector ξ. As we can see, the
difference between this case and the case where rank(Cµ ) = Proposition 1. Under Assumption 1 and Assumption 3, if the
nµ is that now condition (2) is not a direct consequence convex optimization problem (13) in Theorem 1 according
of (33a)-(33b). It depends on the two matrices Aµ and Cx . to the notations (33) is solvable, then the observer-based
Therefore, the following assumption is required: filter (7) and the unknown input estimation (39) guarantee
the following L2 bound:
Assumption 3. The following condition is fulfilled: q
∥µ − µ̂∥Lnµ ≤ λω (ϵ)∥ω∥2Lnω + λ0 (ϵ)∥∥ξ˜0 ∥2 (40)
2
rank (Cx Aµ ) = rank (Aµ ) ≜ nµ . (34) 2

where
Assumption 3 leads to
   


Eξ Inx −nµ 0(nx −nµ )×nµ λω (ϵ) = S2 A − LCξ − A2,ξ + γf S2 Bz − B2,f
rank   ≜ rank  
Cξ Cx T Cx Aµ !2
 1
+ γh S2 LBh 1+ µ∞


= rank Inx −nµ + rank (Cx Aµ ) ϵ
= nx − nµ + rank (Aµ )
2
= nx . (35) + 1+ϵ S2 Eω + D2,ξ , (41a)

Prepared using sagej.cls

6 Proc IMechE Part I: J Systems and Control Engineering 00(0)

Then, as in the previous section, after transforming the

λ0 (ϵ) = S2 A − LCξ − A2,ξ + γf S2 Bz − B2,f system (25), the dynamics of the augmented vector
!2  −1 
 1 M x(t)
+ γh S2 LBh 1+ ν∞ . (41b) ξ(t) ≜ (50)
ϵ µ1 (t)

can be written under the descriptor form (1) with the

Proof. The proof is straightforward, then it is omitted.
following parameters:
The parameter ϵ > 0 comes from the use of the Young’s  
inequality. The criterion (40) holds ∀ϵ > 0. Then, the Eξ = Inx −nµ2 0(nx −nµ2 )×nµ2 0(nx −nµ2 )×nµ1 ,
bounds (41) can be minimized, but for simplicity, we can   (51a)
choose ϵ = 1. Kξ = Inx −nµ2 0(nx −nµ2 )×nµ2 , (51b)
Aξ = Kξ M−1 Ax M Aµ1 ,
 
(51c)
Remark 1. In the disturbance free case, i.e. ω(.) ≡ 0, we
Cξ = Cx N Cx Aµ2 Cµ1 , Bf = Kξ S−1 Bσ , (51d)
 
have the asymptotic convergence of the estimated unknown
input:
lim µ(t) − µ̂(t) = 0 Gg = Kξ S−1 Gϱ , Bh = Kξ S−1 Bς , (51e)
t→+∞
Dξ = Kξ S−1 Dx , Dy = Kξ S−1 D̄y , (51f)
because we have dV ˜ < 0, ∀ξ˜ ̸= 0.
(ξ)






dt f ξ, u = σ̸µ Sξ, u , h ξ, u = ς̸µ Sξ, u , g y, u = ϱ y, u .
(51g)
Third case: 0 < rank(Cµ ) < nµ
In this section, we combine the two previous cases, namely Now let us introduce the following assumption.
the case rank(Cµ ) = nµ and rank(Cµ ) = 0.
In such a situation, 0 < rank(Cµ ) < nµ , there exist a Assumption 5. The following condition is fulfilled:
matrix Cµ1 ∈ Rny ×nµ1 and a variable µ1 (t) ∈ Rnµ1 such
that rank (Cx Aµ2 ) = rank (Aµ2 ) ≜ nµ2 . (52)


Cµ µ(t) = Cµ1 µ1 (t) and rank Cµ1 = nµ1 . (42) Under Assumption 5 and the fact that Cµ1 is full column
rank, it is easy to show that
The nonlinear functions may depend on the vector µ1 .
 
Then Assumption 2 is replaced by the following one: Eξ
rank   = nx + nµ1 ≜ nξ .
Assumption 4. Assume that the nonlinearities σ(., .) and
Cξ
ς(., .) satisfy the following conditions:


Hence, we can apply the results of Section to deduce an
∂σ x, µ, u ∂σ x, µ, u estimation of the unknown input µ(t).
≡ , (43)
∂µ ∂µ1 From (47), the unknown input µ(t) is estimated in the H∞


sense as follows:
∂ς x, µ, u ∂ς x, µ, u
≡ . (44)
∂µ ∂µ1 ˙
µ̂(t) = Tµ1 µ̂1 (t) + Tµ2 ξˆµ2 − TµA2 ξˆ
From Assumption 3 and since each of the nµ unknown ˆ u − T G ϱ y, u
−TµB2 σ̸µ̂ Mµ1 ξ,



(53)
µ2
inputs should appear in system equations (25), then there
exist matrices Aµ2 ∈ Rnx ×nµ2 , Tµ1 ∈ Rnµ ×nµ1 , Tµ2 ∈ where
Rnµ ×nµ2 , and a variable µ2 (t) ∈ Rnµ2 such that
Mµ2

Aµ µ(t) = Aµ1 µ1 (t) + Aµ2 µ2 (t) z }| {

(45)

ξˆµ2 ≜ 0nµ2 ×(nx −nµ2 ) Inµ2 ξ,ˆ (54)
rank Aµ2 = nµ2 , nµ = nµ1 + nµ2 (46) −1 −1
TµA2 ≜ Mµ2 M Ax M, TµB2≜ Mµ2 M Bf , (55)
µ(t) = Tµ1 µ1 (t) + Tµ2 µ2 (t). (47)  
M 0
TµG2 ≜ Mµ2 M−1 Gg , Mµ1 ≜ . (56)
Notice that equations (45)-(47) are not possible only if at 0 Inµ1
least one of the unknown inputs in the vector µ(t) does not
appear in the system (25). However, such a case is useless Remark 2. To avoid repetition, the detailed mathematical
because there is no unknown input to estimate. calculations, which we can get by following exactly the same
Since Aµ2 is full column rank, then there is a matrix steps as in the previous sections, are omitted. In addition,
N ∈ Rnx ×(nx −nµ1 ) such that Proposition 1 still remains valid with convenient λω and λ0
corresponding to the new vector ξ(t) defined in (50) and the
new matrices defined in (54)-(56).
 
M = N Aµ2 (48)

is nonsingular and
  Remark 3. Since the matrix Aµ2 in (45)-(46) is not unique,
0(nx −nµ2 )×nµ2 then we have the possibility to choose the one that satisfies
M−1 Aµ2 = . (49)
Inµ2 Assumption 5.

Prepared using sagej.cls

Alima 7

Application to Wastewater Treatment where ρ(t) = Din is the influent flow rate and
Process  ⊤
x = XDCO SN O SN H , SN D SO ,
In order to illustrate effectiveness of the proposed nonlinear
estimation algorithms, we will apply it to a reduced model of  ⊤
u = θ2 , Kla, XDCOmoy , SN Dmoy ,
activated sludge wastewater treatment plant, the ASM1. The
model, which contains five state variables, was developed in and  in ⊤
in
the European program COST 624 for control purposes. For µ = SN O SN H d .
a detailed study of the model, we refer the reader to 15 . The The matrices Dx and Dy represent the distribution of the
five state variables of the model are given and defined below: disturbances added in the process and in the sensors. We
• XDCO : biodegradable substrate; assume that a uniformly distributed random noise is injected
• SN O : nitrate concentration; in the system according to matrices Dx and Dy . The
• SN H : ammonia concentration; nonlinear functions of the system are given by
• SN D : soluble biodegradable organic nitrogen concen-
XDCO SO
tration; σ1 ξ, u = θ1 (58)
220 + XDCO 0.2 + SO
• SO : dissolved oxygen concentration.
The state variables available for measurement are given in the
XDCO 0.2 SN O
 ⊤ σ2 ξ, u = θ1 ηN O,g (59)
vector y = SN O SN H SO . In the the ASM1 model, 220 + XDCO 0.2 + SO 0.2 + SN O
there are two variables which are usually used as unknown
quantities to be estimated 15 , namely the inlet concentrations

SN H SO
σ3 ξ, u = θ3 (60)
SNin in
O , and SN H .
SN H + 1 SO + 0.4
Since the unknown input estimation algorithm given in
XDCO SO ηN O,h 0.2 SN O
σ4 ξ, u = θ5
the previous Section, the third case where 0 < rank(Cµ ) < XDCO + 258 0.2 + SO 0.2 + SO 0.5 + SN O
nµ , is the more general one, then it is sufficient to provide (61)
simulations for this case only. To this end, in addition to the After calculating the partial derivatives of the nonlinearities,
model unknown inputs SN in in we find that all the lower and upper bounds aij and bij are
O and SN H , we consider that the
SO sensor is affected by a fault d(t) given by: equal to 0 and 0.186, respectively.
We have rank(Cµ ) = 1. Then we will write

10 sin(4πt) if 4(day) < d(t) < 7(day) ⊤
d(t) ≜ 
Cµ µ = Cµ1 µ1 , with Cµ1 = 0 0 1 , µ1 = d(t),
0 otherwise.

In a compact form, by adding d(t) as unknown input, with rank(Cµ1 ) = nµ1 = 1. Hence, we can get (45)-(47)
the ASM1 model is described under the form (25) with the with
following parameters 15 :  
0

0 0

  0 ρ(t) 0 
−0.7954ρ(t) 0 0 0 0    
Aµ1 = 0 , Aµ2 =  0
   ρ(t)
, (62)
 0 −ρ(t) 0 0 0 0  0
  0 
Ax =  0 0 −ρ(t) 0 0 (57a) 1 0 0
 0 0 0 −ρ(t) 0
0 0 0 0 1
     in 
    0 1 0 SN O
0 0 0 1 0 0 0 Tµ1 = 0 , Tµ2 = 0 1 , µ2 =  . (63)
ρ(t) 0 0 0 0 0 0 in
    1 0 0 SN H
Aµ =  0
 ρ(t) 0 , Gg = 0
  0 0 0 ,
 0 0 0 0 0 0 0 It is obvious that rank(Cx Aµ2 ) = rank(Aµ2 ) = nµ2 = 2.
0 0 0 0 SOsat 1 0 Then Assumption 5 holds, which means that we can apply
(57b) the estimation algorithm (53) to estimate the whole unknown
input µ, after estimating the vector ξ(t) given by
 
−1.49 −1.49 0 0
 −1 
M x(t)
 0 0.17 4.16 0  ξ(t) ≜
  µ1 (t)
Bσ =  −0.08 −0.08 −4.16 0 
 (57c)
 0 0 0 1.0  where M comes from (48) with
0 0 −19.04 −10.00  
1 0 0
    0 0 0
0 1 0 0 0 0 0 0  
Cx = 0 0 1 0 0 , Cµ = 0 0 0 , (57d) N= 0 0 0
.
0 0 1
0 0 0 0 1 0 0 1
  0 1 0
 ⊤ 0.1
Dx = 0.2 0.2 1.1 0.1 0 , Dy = 0.2 , (57e) By applying the proposed method on the descriptor
0.3 system obtained by using the transformations (48)-(51), the

Prepared using sagej.cls

8 Proc IMechE Part I: J Systems and Control Engineering 00(0)

corresponding matrices Pz and Qz are computed as

    SND azote concentration SND (g/l)
1 0 0 0 0 0 2
0 1 0  0 0 0 ξ4
1
ξˆ4
   
0 0 1  0 0 0
Pz = 
 , Qz = 0.0125
   (64) 0
0 0 0  0 0
 0 2 4 6 8 10 12 14
0 0 0  0 0.0125 0 time (days)
0 −1 0 0 0 1 SO oxygen concentration SO (g/l)
10
By solving the convex optimization problem (13) under the 5
ξ5
LMI constraints (14), obtained the following observer gains: ξˆ5
0
0 2 4 6 8 10 12 14
 
−40, 08513 −38, 7799 460, 04244
time (days)
 42, 55344 41, 33500 −236, 45936

−36, 41289 −35, 17368 441, 99827 
 actuator fault (unknown input) (g/l)
10
L0 =   (65)
 0, 11180 0, 11097 6, 47413  0
ξ6
 0, 11038 0, 11637 6, 27554  ξˆ6
−42, 49719 −41, 27906 237, 50855 -10
0 2 4 6 8 10 12 14
time (days)
 
0, 78717 0, 76231 −7, 12587
−0, 535440 −0, 51698 9, 81997 
 
 0, 93714 0, 90707 −6, 58848 Figure 2. Behavior of ξ4 , ξ5 , the sensor fault ξ6 = µ1 (t), and
L1 =   (66)
 0, 01046 0, 01014 0, 08815 
 their estimates.
 0, 01008 0, 00984 0, 08719 
0, 53594 0, 51750 −9, 74496
and the optimal value of the disturbance attenuation level
State estimation error of XDCO
is µ∞ = 1.7008. The simulation results over a horizon of 10
0
T = 14days are depicted in Figures 1 and 2. -10
0 2 4 6 8 10 12 14
State estimation error of SNO
0.05
organic concentration XDCO (g/l) 0
200 -0.05
0 2 4 6 8 10 12 14
ξ1
100 State estimation error of SNH
ξˆ1 0.05
0
0 -0.05
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
time (days) State estimation error of SND
0.1
nitrate/nitrites azote concentration SNO (g/l) 0
20 -0.1
0 2 4 6 8 10 12 14
ξ2
10 State estimation error of SO
ξˆ2 10
0 0
0 2 4 6 8 10 12 14 -10
0 2 4 6 8 10 12 14
time (days)
ammoniac azote concentration SNH (g/l)
20
ξ3
Figure 3. Estimation errors of the system states.
10
ξˆ3
0
0 2 4 6 8 10 12 14 Conclusion
time (days)
In this paper, we have proposed three unknown input
estimation algorithms for a class of systems with Lipschitz
Figure 1. Behavior of ξ1 , ξ2 and ξ3 and their estimates. nonlinearities in both of the process dynamics and the output
measurements. The presence of LPV terms in the process is
The simulation results show the efficiency of the proposed also considered and motivated by the main application of the
observer design method. Indeed, the estimation of the system established theoretical results, namely diagnosis and control
states and the sensor fault d(t) is accurate as exhibit for a biological wastewater treatment model. Indeed, three
the behavior of the estimation errors in Figures 3 and 4, cases of study are proposed to tackle all the possibilities
respectively. The estimation results for the system unknown of the presence of unknown inputs in the overall system.
in in
inputs SN O and SN H , i.e. µ2 (t), are depicted in Figure 5, Using a Lyapunov approach, a general LMI-based synthesis
and the related estimation errors are shown in Figure 6. technique is proposed to provide a numerical computation
Overall, the simulation results show the effectiveness and method of the unknown input observer parameters, which
performances of the proposed design methods despite the ensure an H∞ −optimality criterion. The proposed method
presence of the disturbances. is successfully applied to an activated sludge wastewater

Prepared using sagej.cls

Alima 9

State estimation error of unknown input d(t) References

6

1. Yaz EE, Azemi A. Actuator fault detection and isolation in

5
nonlinear systems using LMIs and LMEs. In: Proceedings of
4
the 1998 American Control Conference. ACC (IEEE Cat. No.
98CH36207). vol. 3. IEEE; 1998. p. 1590–1594.
3 2. Zemouche A, Boutayeb M. Sobolev norms-based state
estimation and input recovery for a class of nonlinear systems.
2
design and experimental results. IEEE Transactions on Signal
1
Processing. 2008;57(3):1021–1029.
3. Teh PS, Trinh H. Design of unknown input functional
0 observers for nonlinear systems with application to fault
diagnosis. Journal of Process Control. 2013;23(8):1169 – 1184.
-1
0 2 4 6 8 10 12 14 4. Gauthier JP, Hammouri H, Othman S. A simple observer
for nonlinear systems, applications to bioreactors. IEEE
Transactions on Automatic Control. 1992;37(6):875–880.
Figure 4. Estimation error of unknown input d(t).
5. Zemouche A, Zhang F, Mazenc F, Rajamani R. High-Gain
Nonlinear Observer with Lower Tuning Parameter. IEEE
nitrate concentration (unknown input) SNOin (g/l) Transactions on Automatic Control. 2019;64(8):3194–3209.
20
ξ7
6. Khalil H. Cascade high-gain observers in output feedback
10 ξˆ7 control. Automatica. 2017;80:110–118.
7. Alessandri A. Design of sliding-mode observers and filters for
0
nonlinear dynamic systems. In: Proceedings of the 39th IEEE
-10 Conference on Decision and Control (Cat. No. 00CH37187).
0 2 4 6 8 10 12 14
time (days) vol. 3. IEEE; 2000. p. 2593–2598.
20
ammoniac azote concentration (unknown input) SNHin (g/l) 8. Darouach M, Zasadzinski M, Onana AB, Nowakowski S.
ξ8 Kalman filtering with unknown inputs via optimal state
ξˆ8
10 estimation of singular systems. International journal of systems
0
science. 1995;26(10):2015–2028.
9. Hou M, Muller P. Design of observers for linear systems
-10
0 2 4 6 8 10 12 14
with unknown inputs. IEEE Transactions on automatic control.
time (days) 1992;37(6):871–875.
10. Darouach M, Zasadzinski M, Xu SJ. Full-order observers for
in in
Figure 5. Behavior of ξ7 and ξ8 (SN O and SN H ) and their linear systems with unknown inputs. IEEE transactions on
estimates. automatic control. 1994;39(3):606–609.
11. Yang F, Wilde RW. Observers for linear systems with
unknown inputs. IEEE transactions on automatic control.
State estimation errors of SNOin
10 1988;33(7):677–681.
5 12. Fernando T, MacDougall S, Sreeram V, Trinh H. Exis-
tence conditions for unknown input functional observers.
0
International Journal of Control. 2013;86(1):22–28. Avail-
-5
able from: https://doi.org/10.1080/00207179.
-10
0 2 4 6 8 10 12 14
2012.715802.
State estimation errors of SNHin
13. Phanomchoeng G, Rajamani R. Real-Time Estimation of
10
Rollover Index for Tripped Rollovers With a Novel Unknown
5 Input Nonlinear Observer. IEEE/ASME Transactions on
0 Mechatronics. 2014 April;19(2):743–754.
-5
14. Hassan L, Zemouche A, Boutayeb M. Robust unknown input
observers for nonlinear time-delay systems. SIAM Journal on
-10
0 2 4 6 8 10 12 14 Control and Optimization. 2013;51(4):2735–2752.
15. Boulkroune B, Halabi S, Zemouche A. H− /H∞ fault
Figure 6. Estimation errors of unknown inputs µ2 (t). detection filter for a class of nonlinear descriptor systems.
International Journal of Control. 2013 Feb;86(2):253–262.
16. Shim H, Son YI, Seo JH. Semi-global observer for multi-
output nonlinear systems. Systems and Control Letters.
treatment process model (benchmark COST 624), where an 2001;43(4):233–244.
actuator fault is estimated as an unknown input. 17. Hou M, Muller PC. Design of observers for linear systems with
As future work, we aim to apply these preliminary results unknown inputs. IEEE Transactions on Automatic Control.
to develop a fault tolerant control strategy, which will 1992;37(6):871–875.
provide a generic method for reference trajectory tracking
in biological wastewater treatment models. In addition, we
plan to consider parameter uncertainties in the process.

Prepared using sagej.cls