Reduced Order Unknown Input Observer for

Discrete Time System

Vivek Sharma B. B. Sharma R. Nath
National Institute of Technology National Institute of Technology National Institute of Technology
Hamirpur, India. Hamirpur, India. Hamirpur, India.
Email: [email protected] Email: [email protected] Email: [email protected]

Abstract—In present work, reduced order unknown input functional estimator for linear systems with unknown input is
observer is designed for a class of discrete time nonlinear systems. designed in [13].
In reduced order observer design, state estimates are developed The approach followed for the nonlinear observer design
only for those states which are not directly measured at the
output. To proceed with the design, system dynamics is split into is to use high observer gain to stabilize linear part of the
two sets, dynamics of the measured state variables and unknown dynamics while simultaneously taking care of the effect of
state variables. Then discrete observer is designed for the second nonlinearity, which is estimated in terms of its Lipschitz
set i.e. unmeasured state variables. Unknown input part of the constant. The approach for linear UIO is extended to the
system dynamics is tackled using famous decoupling approach, design of Nonlinear Unknown Input Observer (NUIO) using
because of which the state estimation error dynamics becomes
independent of the unknown input. Equivalent representation for LMI based approach [14], [15], [16], [17]. Availability of the
the nonlinear vector function is achieved using Differential Mean LMI solving software makes the difficult task of designing
Value (DMV) theorem which is likely to give less conservative NUIO easy. In [15], design of NUIO for Lipschitz nonlinear
results in comparison to Lipschitz constant based approach. systems using LMI based approach is proposed. Large value
Application of DMV theorem leads the error dynamical system of Lipschitz constant, renders the LMI infeasible. Approaches
to evolve as Linear Parameter Varying (LPV) system, which
is then analyzed using convexity principle. Finally, the Linear based on one sided Lipschitz constant [18] and Differential
Matrix Inequality (LMI) condition is achieved which is solved Mean Value Theorem (DMVT) [19] have advantages with
for feasibility to obtain observer design matrices. Reconstruction respect to conservativeness.
of unknown input is also discussed, which has application in Compared to continuous time systems, UIO design for non-
secure communication. To justify the proposed theoretical results, linear discrete time systems is less reported in literature. For
simulation results are presented for chaotic Lorenz system which
is a member of addressed class of nonlinear systems. linear discrete time system state and unknown input estimation
is discussed in [20], [21]. Some nonlinear UIO and their use in
I. I NTRODUCTION secure communication are also reported in the literature [22],
Designing observer for systems with unknown input is an [23]. Delayed output in discrete time systems is analogous
important problem in the area of observer design because to derivative of output in continuous time systems but is
of its ability to develop estimates of system states even comparatively easier to get.
in situations where driving input is completely or partially In present work design of reduced order UIO for nonlinear
unknown. More recently this class of observers i.e. Unknown discrete time system and a methodology to reconstruct un-
Input Observer (UIO) have been designed to address the known input using delayed output is proposed. DMV Theorem
problems like fault detection and isolation (FDI) [1], [2], is used to address the nonlinearity, which makes the state
chaotic systems synchronization and secure communication error dynamics evolve as LPV system, which is then analyzed
[3], [4] and lateral velocity estimation on banked roads [5]. using convexity principle. Finally, sufficient condition for the
UIO design for linear systems is discussed in [6], [7], [8], [9], existence of observer is derived in terms of LMI’s, solution of
where the used approach decouples the unknown input using which gives the observer design matrices directly. Recovery of
direct matrix calculations. Full order UIO is designed in [7] the unknown input is also achieved, and hence the proposed
using direct matrix calculation with unknown input present in observer could be used for chaotic secure communication. The
the system dynamics. Later the work is extended in [8], to a simulations results presented in the end validate the proposed
more general case where unknown input is present in output strategy.
description as well. Using similar matrix based approach, Notations used:
reduced order UIO for linear systems have been discussed in Co(Ω) means convex hull of set Ω. For Ω = {a, â} s.t.
[10], [11], [12]. The approach to design reduced order UIO in a, â ∈ Rn , Co(Ω) = {γa + (1 − γ)â, 0 ≤ γ ≤ 1}
[10] is based on generalized matrix inverse with main focus on bn (i) is canonical ith basis vector for n dimensional vector
decoupling unknown inputs or disturbances. As a result, the space i.e. it is a column vector of length n with all elements
error dynamics becomes independent of the unknown input. 0 except 1 at ith position.
Using results from the geometric theory of linear systems,

978-1-5090-2597-8/16/$31.00 2016
c IEEE 3443
 II. P ROBLEM F ORMULATION then state estimation error e(k) → 0 as k → ∞, where Z(k) =
Let us consider a class of nonlinear discrete time system as: A22 (h(k)) + EA12 (h(k)) which is defined in the proof.
Proof: To proceed with the observer design, objective
x(k + 1) = Ax(k) + g(u(k), y(k)) + B f(x(k)) + Dv(k) here is to make sate estimation error converge to zero. Using
y(k) = C x(k) (1) DMV Theorem [24] f(x) − f(x̂) can be written as
Above representation splits the dynamics into linear part with q n−p
  ∂fi (Ki (k))
A ∈ Rn×n , nonlinear part with B ∈ Rn×q and f(x(k)) ∈ Rq f(x(k)) − f(x̂(k)) = ( bq (i)bTn−p (j) )
∂wj (k)
a vector valued function, known part with g(.) ∈ Rn which i=1 j=1
is a function of y(k) and known input u(k) and unknown (w(k) − ŵ(k)) (11)
input part with v(k). Output matrix C is assumed to have
where Ki (k) ∈ Co (w(k), ŵ(k)) ∀ i = 1, . . . , q. Now,
structure [Ip 0] which is not restrictive as there always exists
using the notations
a state transformation to bring it in stated form. D ∈ Rn×r is
unknown input distribution matrix. C and D are both full rank ∂fi
hi,j (k) = (Ki (k)) (12)
matrices with p > r. Now above system dynamics can be ∂wj
⎡ ⎤
rewritten as combination of two subsystems by partitioning h1,1 (k) , . . . , h1,n−p (k)
as follows: ⎢ .. .. ⎥
h(k) = ⎣ . ,..., . ⎦ (13)
hq,1 (k) ,..., hq,n−p (k)
y(k + 1) = A11 y(k) + A12 w(k) + g1 (u(k), y(k))
Further using (9), along with definition of M from (6),
+B1 f(y(k), u(k)) + D1 v(k) (2) unknown input part from the error dynamics get decoupled
w(k + 1) = A21 y(k) + A22 w(k) + g2 (y(k), u(k)) and error dynamics (8) can be written as
+B2 f(y(k), w(k)) + D2 v(k) (3)  
B1
e(k + 1) = N e(k) + M h(k)e(k) (14)
p
where y(k) ∈ R is the measured state vector, w(k) ∈ R n−p B2
is the unknown state vector and is to be estimated. To develop Using N & M from (6), (14) gets modified as below
the estimate of unknown state vector, observer structure is
proposed as e(k + 1) = (A22 + B2 h(k)) e(k)
z(k + 1) = N z(k) + L y(k) + M B f(y, ŵ) +E (A12 + B1 h(k)) e(k) (15)
 
g1 (y,u) To simplify the above representation, let us assume following
+M (4)
g2 (y,u) relations:
ŵ = z−E y (5)
A12 (h(k)) = A12 + B1 h(k) (16)
n−p n−p
where z ∈ R , ŵ ∈ R is estimated state vector. N, L A22 (h(k)) = A22 + B2 h(k) (17)
and M are defined as given below:
So, equation (15) will become
N = A22 + EA12
L = A21 + EA11 − NE e(k + 1) = (A22 (h(k)) + EA12 (h(k))) e(k)(18)
  Say e(k + 1) = Z(k)e(k) (19)
M = E In−p (6)
where E is the design parameter and is to be selected by the To derive the conditions for the convergence of error dynamics,
designer suitably. Error in the unknown system state vector let us consider the Lyapunov function as
and estimated state vector is defined as VL (k) = eT (k)PeT (k) (20)
e(k) = ŵ(k) − w(k) (7) Applying Lyapunov stability theory, change in Lyapunov func-
Using the matrix definitions given in (6), error dynamics for tion ΔVL along error trajectories (19) is evaluated as
the proposed observer can be written as
  ΔVL (k) = VL (k + 1) − VL (k)
B1  
e(k + 1) = N e(k) + M (f(y, ŵ) − f(y,w)) = eT ZT (k)PZ(k) − P e(k)
B2
  Now the inequality condition ZT (k)PZ(k) − P < 0, if obeyed
D1
−M v(k) (8) will make ΔVL < 0 which implies error e(k) → 0 as k → ∞.
D2
This inequality condition is nonlinear matrix inequality in Z(k)
Theorem 1: If there exists a matrix E and a symmetric and is difficult to solve as such. Thus using Schur complement
positive definite matrix P such that lemma this nonlinear inequality can be can be written as LMI
E D1 = −D2 (9) (10).
 
−P ZT (k)P Now to design observer, it is required to solve (9) and (10)
< 0 (10) simultaneously, which doesn’t have a direct solution. So it is
P Z(k) −P

3444 2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference
proposed to proceed by, finding all possible solutions of E 60
x3

by solving equality (9) in terms of arbitrary variable S using 50

x̂3

generalized inverse. Further, replace the value of E in LMI 40

(10) and then solve it for new variable S and P > 0. General
solution to ED1 = −D2 is
30

x3 , & x̂3
20

E = −D2 D+ +
1 + S(Ip − D1 D1 ) (21) 10

where D+
1 = (DT1 D1 )−1 DT1is the generalized inverse of 0

matrix D1 and S is an arbitrary matrix of dimension (n−p)×p. −10

0 1000 2000 3000 4000 5000 6000

By multiplying D1 on both sides of (21), one can get the Sample

equation ED1 = −D2 because D+ 1 D1 = Ip . Therefore, Fig. 1. Convergence of system and observer state
solution given by (21) is indeed a general solution of (9).
Let us assume that U = −D2 D+ +
1 and V = Ip − D1 D1 for
notational convenience, then general solution for E in (21) is III. N UMERICAL S IMULATION
rewritten as:
E = U + SV (22) To validate the proposed methodology, let us consider
This value of E is replaced in (10) and the LMI (10) gets discrete time Lorenz chaotic system with dynamics as
redefined. Moreover for effective observer design, it is required
that the LMI (10) be feasible for all possible values of h(k) x(k + 1) = A x(k) + g(u,y) + B f(x) + D v(k)
defined in (13). Since this LMI is affine in h(k), hence using y(k) = Cx(k) (26)
convexity principles, it is sufficient to solve LMI at all the ⎡ ⎤
vertices VHq,(n−p) , of the convex hull encapsulating all possible 1 − 10T 10T 0
values of h(k). where A = ⎣ 28T 1−T 0 ⎦;
8
Thus the redefined LMI to be solved for new set of variables ⎡ 0
⎤ 0 1 − 3 T
0 0  
i.e. P = PT > 0 and Ps = P S, for the design of the observer x1 (k)x3 (k)
is B = T ⎣ −1 0 ⎦ f(x) =
  x1 (k)x2 (k)
0 1  
−P ZT1 (α)  T
< 0 (23) 1 0 0
Z1 (α) −P D = 1 1 1 C = for simula-
0 1 0
∀α ∈ VHq,(n−p) tion unknown input v(k) and sampling period T is taken
where Z1 (α) = PA22 (α) + PUA12 (α) + Ps VA12 (α). as 0.1sin(2πT k) and 0.001 respectively. Depending on the
Design Algorithm: structure of C the above system can be expressed in the form 
1) Partition the matrices A, B, C, D, state vector x to write (2) with matrices defined as A11 = 1 − 10T 10T
the system in the form given by (2).   28T 1−T
+ + 0    
2) Compute U = −D2 D1 and V = Ip − D1 D1 . A12 = A21 = 0 0 A22 = 1 − 83 T
3) Find all the vertices α such that α ∈ VHq,(n−p) by  0   
0   1
running simulations. B1 = B2 = 0 D1 =
4) Solve LMIs in (23) for Ps and symmetric matrix P > 0. −1   1
−1   0
5) Calculate S using S = P Ps and E from (22). D2 = 1 g1 (u,y) = g2 (u,y) = y1 y2
6) Using E, all the gain matrices of the proposed observer 0
can be computed by using (6). f(y,w) = y1 w1 . Now by solving LMI condition (23)
observer
 matrices are found  as follows
 P = 1.764 × 105,
Remark 1: Unknown input recovery has its applications
S =  −0.5076  0.4914  , E = −0.9995  −0.0005
in varied fields like secure communication, parameter iden-
N = 0.9973 , L = 0.0073 −0.01 
tification and FDI (Fault detection and identification). For
M = −0.9995 −0.0005 1.0000 . For simulation pur-
the system (1) - observer (4) it has been proved above that,
pose, initial condition of the system is taken as x(0) =
observer designed with feasible solutions of LMI will estimate  T
1 2 1 . As per the definition of output matrix, only
the system states i.e. ŵ → w. Thus rewriting system (1) in
third state needs to be estimated. Thus the observer is designed
terms of state estimates as
for the estimation of the third state. Initial  condition of the
x̂(k + 1) = Ax̂(k) + g(u,y) + B f(x̂) + Dv(k) (24) observer is taken as x̂(0) = −3.99 . Fig (1) shows the
  convergence of the observer and system third state, for which
y
where x̂ = . Unknown input v(k) can be recovered the observer is designed. State estimation error converging to
ŵ
from above as zero is shown in Fig (2). Further, as discussed in Remark (1),
T −1 T it is possible to recover unknown input using equation (25).
v(k) = (D D) D [x̂(k + 1) − Ax̂(k) − g(u,y) − B f(x̂)] Fig (3) shows the unknown input v(t) and the recovered input.
(25)

2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference 3445
 [3] M.Boutayeb, M.Darouach, and H.Rafaralahy, “Generalized state-space
5
observers for chaotic synchronization and secure communication,” IEEE
4.5 Trans. Circuits Syst. I, vol. 49, no. 3, pp. 345–349, 2002.
4
[4] B. Sharma and I. Kar, “Observer-based synchronization scheme for a
class of chaotic systems using contraction theory,” Nonlinear Dynamics,
3.5 vol. 63, no. 3, pp. 429–445, 2011.
3 [5] L. Imsland, T. A. Johansen, H. F. Grip, and T. I. Fossen, “On non-linear
unknown input observers–applied to lateral vehicle velocity estimation
e3

2.5
on banked roads,” International Journal of Control, vol. 80, no. 11, pp.
2 1741–1750, 2007.
1.5
[6] F.Yang and R.W.Wilde, “Observers for linear systems with unknown
inputs,” IEEE Trans. Autom. Control, vol. 33, no. 7, pp. 677–681, 1988.
1
[7] M.Darouach, M.Zasadzinski, and S.J.Xu, “Full-order observers for linear
0.5 systems with unknown inputs,” IEEE Trans. Autom. Control, vol. 39,
no. 3, pp. 606–609, 1994.
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 [8] M. Darouach, “Complements to full order observer design for linear
Sample systems with unknown inputs,” Applied Mathematics Letters, vol. 22,
no. 7, pp. 1107 – 1111, 2009.
Fig. 2. State estimation error for estimated state x3 [9] C.C.Tsui, “A new design approach to unknown input observers,” IEEE
Trans. Autom. Control, vol. 41, no. 3, pp. 464–468, 1996.
[10] P.Kudva, N.Viswanadham, and A.Ramakrishna, “Observers for linear
0.1 systems with unknown inputs,” IEEE Trans. Autom. Control, vol. 25,
0.08
Recovered Signal
Unknown Input no. 1, pp. 113–115, 1980.
[11] M.Darouach, M.Zasadzinski, and M.Hayar, “Reduced-order observer
Unknown input, Recovered signal

0.06

0.04
design for descriptor systems with unknown inputs,” IEEE Trans. Autom.
0.02
Control, vol. 41, no. 7, pp. 1068–1072, 1996.
0
[12] H. Trinh, T. D. Tran, and T. Fernando, “Disturbance decoupled observers
−0.02
for systems with unknown inputs,” IEEE Trans. Autom. Control, vol. 53,
no. 10, pp. 2397–2402, Nov 2008.
−0.04
[13] S.P.Bhattacharyya, “Observer design for linear systems with unknown
−0.06
inputs,” IEEE Trans. Autom. Control, vol. AC-23, no. 3, pp. 483–484,
−0.08
1978.
−0.1
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 [14] D. Koenig, “Observer design for unknown input nonlinear descriptor
Sample
systems via convex optimization,” IEEE Trans. Autom. Control, vol. 51,
no. 6, pp. 1047–1052, June 2006.
Fig. 3. Unknown input v(t) and recovered unknown input [15] W.Chen and M.Saif, “Unknown input observer design for a class of
nonlinear systems: an lmi approach,” in Proc. IEEE ACC. Minneapolis,
Minnesota, USA: IEEE, 2006, pp. 834–838.
[16] H.Dimassi and A.Loria, “Adaptive unknown-input observers-based syn-
IV. C ONCLUSION chronization of chaotic systems for telecommunication,” IEEE Trans.
Circuits Syst. I, vol. 58, no. 4, pp. 800–812, 2011.
In present work, reduced order unknown input observer [17] L. Hassan, A. Zemouche, and M. Boutayeb, “Robust unknown input
(ROUIO) is designed for nonlinear discrete time system. observers for nonlinear time-delay systems,” SIAM Journal on Control
With the proposed observer structure, state estimation error and Optimization, vol. 51, no. 4, pp. 2735–2752, 2013.
[18] W. Zhang, H. Su, F. Zhu, and G. M. Azar, “Unknown input observer
evolves independent of the unknown input. Thus the designed design for one-sided lipschitz nonlinear systems,” Nonlinear Dynamics,
observer is immune to unknown input and is able to esti- vol. 79, no. 2, pp. 1469–1479, 2015.
mate the unmeasurable states of the system. Using equivalent [19] V. Sharma, V. Agrawal, B. B. Sharma, and R. Nath, “Unknown input
nonlinear observer design for continuous and discrete time systems with
representation for nonlinearity in the system dynamics using input recovery scheme,” Nonlinear Dynamics, vol. 85, no. 1, pp. 645–
DMV theorem, error dynamics is molded as LPV system 658, 2016.
and hence becomes tractable for treatment using convexity [20] M. E. Valcher, “State observers for discrete-time linear systems with
unknown inputs,” IEEE Transactions on Automatic Control, vol. 44,
principle. Finally observer matrix gains are obtained directly no. 2, pp. 397–401, 1999.
as a solution of resultant LMI condition obtained using the [21] T. Floquet and J.-P. Barbot, “State and unknown input estimation for
proposed approach. Proposed results are validated on discrete linear discrete-time systems,” Automatica, vol. 42, no. 11, pp. 1883–
1889, 2006.
time Lorenz system which belongs to the proposed class of [22] I. Belmouhoub, M. Djemai, and J. Barbot, “Cryptography by discrete-
discrete time systems. time hyperchaotic systems,” in IEEE Conference on Decision and
Control, 2003, pp. 1902–1907.
[23] G. Millerioux and J. Daafouz, “Unknown input observers for message-
ACKNOWLEDGMENT embedded chaos synchronization of discrete-time systems,” Interna-
tional Journal of Bifurcation and Chaos, vol. 14, no. 04, pp. 1357–1368,
The work is funded by TEQIP-II Project under MHRD, 2004.
Govt. of India. [24] A. Zemouche, M. Boutayeb, and G. I. Bara, “Observer design for
nonlinear systems: An approach based on the differential mean value
R EFERENCES theorem.” in 44th IEEE Conference on Decision and Control. IEEE,
2005, pp. 6353–6358.
[1] W. Chen and M. Saif, “Fault detection and isolation based on novel
unknown input observer design,” in American Control Conference, June
2006, pp. 5129–5134.
[2] K. C. Veluvolu and Y. C. Soh, “Fault reconstruction and state estimation
with sliding mode observers for lipschitz non-linear systems,” IET
control theory & applications, vol. 5, no. 11, pp. 1255–1263, 2011.

3446 2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference