Actuator and Sensor Fault Estimation of the Furuta

Pendulum System via Linear Matrix Inequalities

Eduardo Linares Miguel Amador Tonatiuh Hernández-Cortés
Dept. of Mechatronics Dept. of Mechatronics Dept. of Mechatronics
Universidad Politécnica de Pachuca Universidad Politécnica de Pachuca Universidad Politécnica de Pachuca
Zempoala, México Zempoala, México Zempoala, México

Francisco-Ronay López-Estrada Vı́ctor Estrada-Manzo

TURIX-Dynamics Diagnosis and Control Group Dept. of Mechatronics
Tecnológico Nacional de México Universidad Politécnica de Pachuca
IT Tuxtla Gutiérrez Zempoala, México
Tuxtla Gutiérrez, México victor [email protected]

Abstract—This work presents an unknown input observer observation algorithm is applied in simulation and real-time
approach for the estimation of actuator and sensor faults. The experiments on the Furuta pendulum system.
approach is based on the approximation at the origin of a given
nonlinear system. The designing conditions are obtained via the Organization: The rest of the document is organized as
Lyapunov method and are expressed in terms of linear matrix follows: in Section II, the problem to be studied is stated.
inequalities whose solvability is decidable in polynomial time. In Section III, the main results are given: conditions for the
The proposal is put at test in the Furuta pendulum system, both design of an UI observer for to estimate actuator and sensor
simulation and real-time results are discussed. faults. In Section IV, the performance of the proposal is illus-
Index Terms—Observer design, sensor fault, actuator fault,
linear matrix inequalities, nonlinear system trated via the well-known Furuta pendulum in both simulation
and real-time. Finally, Section V presents the conclusions and
I. I NTRODUCTION final remarks.
Diagnosis has become an important task in control theory, it
allows isolating and estimating faults in order to avoid physical
II. P RELIMINARIES
deterioration or malfunction of the system [1]; such faults
may occur at sensors, actuators or in the process [2], [3]. In
general, approaches for diagnosis are based on unknown input In this work the following system is considered:
(UI) observers [4]–[9]; this family of observers estimates both
the system states as well as the actuator/sensor faults. There ẋ(t) = f (x)+g(x)(u(t)+fa (t)), y(t) = h(x)+fs (t), (1)
are many methodologies for designing an UI observer: sliding
mode [10], convex [11], adaptive [12], geometric approaches where x ∈ Rn is the state vector, u ∈ Rm is the input vector,
[13]. Nevertheless, despite real systems are accurately rep- y ∈ Ro is the output vector; functions f (·) : Rn 7→ Rn ,
resented by nonlinear equations, for real-time applications g(·) : Rn 7→ Rn×m , and h(·) : Rn 7→ Ro are assumed to be
simplicity is very appreciated, so linear techniques such as sufficiently smooth in Ω ⊂ Rn ; fa (t) ∈ Rm and fs (t) ∈ Ro
linearization around an operating point come at hand [14]. are the actuator and sensor fault vectors.
Since the appearance of the book [15], linear matrix inequal- (p ) (p )
The faults are assumed to hold fa a (t) ≈ 0 and fs s (t) ≈
ities (LMIs) have been widely used by the control community,
0, where pa and ps are the derivative order of each fault vector.
even in the case of linear systems; the reason is that LMIs
This assumption allows writing the following for the actuator
are systematic, provide a solution in polynomial time, allows
faults:
including performance specifications such as speed conver-
gence and/or input/output constraints, multi-input multi-output    
0 Im 0 ··· 0 da1
systems can be easily taken into account [16]. 0 0 Im ··· 0  da2 
Contribution: Actuator and sensor faults are estimated via    
d˙a=Sa da , with Sa=0
 0 0 ··· 0, da=

 da3 , (2)

UI observers whose gain is computed by means of LMIs. The  .. .. .. .. . 
.. 
 .. 

conditions are derived by using an approximation at the origin . . . . .
of the nonlinear dynamics and the Lyapunov method. The 0 0 0 ··· 0 dapa
This work has been supported by CONACYT via scholarships for CVUs
1147641 and 1147205. where da1 (t) = fa (t), da2 (t) = f˙a (t), . . ., dapa (t) =
 (pa −1)
fa (t). As for the sensor faults we have Proof: Consider the following Lyapunov function candidate
    Va (ēa ) = ēTa Pa ēa , Pa = PaT > 0; its time-derivative along
0 Io 0 · · · 0 ds1
0 0 Io · · · the trajectories of (8) gives
 0 
 ds2 
 
d˙s=Ss ds , with Ss=0 0 0 · · ·  ds3 , (3)
0, ds= V̇a = ēTa Pa Āa − Pa L̄a C̄a + ĀTa P − C̄aT L̄Ta Pa ēa ; (10)
  

 .. .. .. . . ..   .. 
. . . . .  . 
then, once substituted N̄a = Pa L̄a , V̇a < 0 holds if and only
0 0 0 ··· 0 dsps if the LMIs (9) hold too. □
where ds1 (t) = fs (t), ds2 (t) = f˙s (t), . . ., dsps (t) = B. Estimation of sensor faults
(p −1)
fs s (t).
In order to design UI observer for (4), with fa (t) = 0 and
Further developments are based on the linearization of the
(3), let us write the following augmented system:
previous system, that is,
x̄˙ s (t) = Ās x̄s (t) + B̄s u(t), (11)
ẋ(t) = Ax(t)+B(u(t)+fa (t)), y(t) = Cx(t)+fs (t), (4)
where A ∈ Rn×n , B ∈ Rn×m , C ∈ Ro×n are matrices with with
     
constant entries obtained from linearizing (1) at the origin. x A 0 B  
x̄s= , Ās= , B̄s= , C̄s= C [Io 0o×o(ps −1) ] .
In order to estimate the state x(t) and the faults fa (t) and ds 0 Ss 0
fs (t) an UI observer will be employed [1]. The following
Thus, an UI observer for the previous augmented system is
section presents conditions in the form of LMIs for the design
of a Thau-Luenberger UI observer. ˆ˙ s (t) = Ās x̄
x̄ ˆs (t) + B̄s u(t) + L̄s (y − ŷ), (12)
III. M AIN R ESULTS ˆs = [x̂T dˆTs ]T , (Ās , B̄s , C̄s ) are defined as in (11); while
with x̄
The estimation of both the actuator and sensor faults is car- L̄s = [LTP s LTIs ]T , LP s ∈ Ro×n , LIa ∈ Ro×(n+ps o) is the
ried out via a proportional-integral observer. First, an actuator observer gain to be designed such that the observation error
PI observer is design via LMIs and secondly for sensor faults. ēs is asymptotically stable. By defining the observation error
as
A. Estimation of actuator faults ˆ
   
x̄ − x̄ e
ēs = = x . (13)
In order to design UI observer for (4) with fs (t) = 0 and ds − dˆs efs
(2), let us write the following augmented system:
Its dynamics is
x̄˙ a (t) = Āa x̄a (t) + B̄a u(t), (5)

ē˙ s (t) = Ās − L̄s C̄s es (t). (14)
with
       The following result states LMI conditions for such task.
x A B Im 0m×m(pa −1) B Theorem 2: The UI observer (12) asymptotically estimates
x̄a= , Āa= , B̄a= ,
da 0 Sa 0 the augmented system (11) if there exist Ps = PsT ∈
and C̄a = [C 0o×(n+pa m) ]. Thus, an UI observer for the R(n+ps o)×(n+ps o) and N̄s ∈ Ro×(n+ps o) such that
previous augmented system is
Ps > 0, Ps Ās − N̄s C̄s + ĀTs Ps − C̄sT N̄sT < 0 (15)
ˆ˙ a (t) = Āa x̄
x̄ ˆa (t) + B̄a u(t) + L̄a (y − ŷ), (6)
is satisfied. In this case, the observer gain is L̄s = Ps−1 N̄s .
ˆa = [x̂T dˆTa ]T , (Āa , B̄a , C̄a ) are defined as in (5); while
with x̄ Proof: Consider the following Lyapunov function candidate
L̄a = [LTP a LTIa ]T , LP a ∈ Ro×n , LIa ∈ Ro×(n+pa m) is the Vs (ēs ) = ēTs Ps ēs , Ps = PsT > 0; its time-derivative along the
observer gain to be designed such that the observation error trajectories of (14) gives
ēa is asymptotically stable. By defining the observation error
V̇s = ēTs Ps Ās − Ps L̄s C̄s + ĀTs P − C̄sT L̄Ts Ps ēs ; (16)


as
ˆ
   
x̄ − x̄ e
ēa = = x , (7) then, once substituted N̄s = Ps L̄s , V̇s < 0 holds if and only
da − dˆa efa
if the LMIs (15) hold too. □
whose dynamics is Remark 1: One of the advantages of the LMI framework

is that control specifications can be naturally added as LMIs
ē˙ a (t) = Āa − L̄a C̄a ea (t). (8)
too. For example, the speed convergence of the proposed UI
The following result states LMI conditions for such task. observers can be increased if instead of (9), the following LMI
Theorem 1: The UI observer (6) asymptotically estimates is used:
the augmented system (5) if there exist Pa = PaT ∈
Pa > 0, Pa Āa − N̄a C̄a + ĀTa Pa − C̄aT N̄aT +2αPa < 0, (17)
R(n+pa m)×(n+pa m) and N̄a ∈ Ro×(n+pa m) such that
where α > 0 is the speed rate, also viewed, as the index
Pa > 0, Pa Āa − N̄a C̄a + ĀTa Pa − C̄aT N̄aT < 0 (9)
for exponential stability of the error system (8); the previous
is satisfied. In this case, the observer gain is L̄a = Pa−1 N̄a . inequality has been obtained from V̇a ≤ −2αVa [17], [18]. A
similar LMI set for the sensor-fault case can be achieved, this occur. Initial conditions are set as x(0) = 0 and x̂(0) = [30 −
is to say, that in instead of (15) we can check 30 30 − 30]T . In what follows the estimation of actuator and
sensor faults is carried out by means of theorems 1 and 2,
Ps > 0, Ps Ās − N̄s C̄s + ĀTs Ps − C̄sT N̄sT +2βPs < 0, (18)
respectively. It will become clear that the larger the value of
where β > 0 being the speed convergence rate. Decay rate pa and ps is, the order of the considered derivatives for the
LMIs are generally use when facing real-time implementa- actuator and sensor faults, the better the estimation results.
tions.
TABLE I
IV. I MPLEMENTATION ON THE F URUTRA P ENDULUM PARAMETERS FOR THE F URUTA P ENDULUM

Let us consider the Furuta pendulum, it is an under actuated Variable Description Value
system that consists on two couple beams: the horizontal one Mp Pendulum mass 0.027Kg
Lp Large pendulum 0.153m;
rotates with a DC motor while the vertical one is joint to the r Large arm 0.0826m
first one so it freely rotates, an scheme is shown in Figure Jp Pendulum inertia 0.000177m4
1. Choosing as state variables x = [θ, α, θ̇, α̇]T , a state-space Jarm Arm inertia 0.00018374m4
kt Armor constant 0.0333
representation in the form of (1) with fa = 0, fs = 0, Km Motor constant 0.0333

x3
 Rm Motor resistance 8.7 ohm
g Gravity 9.81 sm2
 x4 
f (x) =  p1 cos2 x2 +p2 cos x2 sin x2 −p3 ,
 
 δ 
2 2 2
(p4 cos x2 +p5 sin(2x2 ))(Mp Lp sin x2 +Mp r +Jarm )
A. Actuator-fault case
δ
In this part, we have considered a sinusoidal signal as the

0
0   actuator fault, that is, fa (t) = 4 sin(0.75t). LMIs in Theorem
x
 
g(x) =  Kt Mp L2p +Kt Jp  , and h(x) = 1 (3)
1 have run for pa = 1 (f˙a (t) ≈ 0) and pa = 3 (fa (t) ≈ 0)
 
Rm δ
x2
but including a decay rate (see Remark 1). For the first case,
 
−Kt Lp Mp r cos x2
Rm δ they are found feasible with α = 11.5 and
where p1 = gL2p Mp2 r, p2 = 0.5Mp2 L3p r, p3 =
 
60.7411 7.7292 −3.0681 −0.4409 3.0804
(Kt Km )(Mp L2p +Jp )/Rm , p4 = −gLp Mp , p5 = −0.5Mp L2p ,  7.7292 45.6267 −0.4830 −2.1052 1.3048 
and δ = Jp Jarm + Mp2 L4p sin2 x2 + Mp2 L2p r2 + Jarm Mp L2p +
 
Pa = −3.0681 −0.4830 0.2572
 0.0459 −0.3298 ,
Jp Mp r2 + Jp L2p Mp sin2 x2 − L2p Mp2 r2 cos2 x2 . The parameter −0.4409 −2.1052 0.0459 0.1547 −0.1318
values are given in Table I. 3.0804 1.3048 −0.3298 −0.1318 0.5377
A linearization at the origin produces a linear system of the  
form (4) with matrices 42.4 13.5
     −3.9 32.6 
0 0 1 0 0  
0 0 0 1 L̄a = 1073.8 571.8

;
, B =  0 ,
 
A= 0 76.2884 −0.5689  149.4 670.1
0  17.084
0 82.2655 −0.239941 0 7.2054 484.7 301.4
  (3)
1 0 0 0 For the case when fa (t) ≈ 0, LMIs are feasible with α =
C= .
0 1 0 0 9.45 and the observer gain
 
All the experiments have been performed with a control law 0.0009 0.0004
K = [−7.2523 83.261 − 2.9707 11.1655], this gain helps to 0.0001 0.0005
 
keep the system oscillating around the origin when the faults 0.0456 0.0288
L̄a = 1 × 105 
 
0.0152 0.0166 .

0.0639 0.0428
 
0.7338 0.4953
3.2380 2.1958
Real-time results show that the larger the value of pa , the
better the fault estimation is; this can be seen from Figure
2 (first order) and 3 (third order). The reconstruction of the
state signals x3 and x4 , corresponding to the speeds of both
(3)
beams, is shown in figures 4 and 5 when fa (t) ≈ 0 is under
consideration. It is important to stress that there is mismatched
between the real and the estimated signals, because we have
design a lineal, but simple-to-implement, observer for a system
Fig. 1. Scheme of the Furuta pendulum system whose dynamics is fast.
 Fa X3
Fa vs Fag Fag X3 VS X3g X3g
8 250
6 200
150
4
100

Angle (deg)
2
Voltage (V)

50
0 0
-2 -50
-100
-4
-150
-6 -200
-8 -250
10.0 11.0 12.0 13.0 14.0 15.0 16.0 17.0 18.0 19.0 20.0 10.0 11.0 12.0 13.0 14.0 15.0 16.0 17.0 18.0 19.0 20.0
Time (s) Time (s)

(3)
Fig. 2. Real-time estimation of the actuator fault when f˙a (t) ≈ 0. Fig. 4. Real-time estimation of x3 when fa (t) ≈ 0

Fa X4
Fa vs Fag Fag X4g
X4 VS X4g
12
50
10 40
8 30
6 20
Voltage (V)

Angle (deg)
10
2 0
0 -10
-2 -20
-4 -30
-6 -40
-8 -50
10.0 11.0 12.0 13.0 14.0 15.0 16.0 17.0 18.0 19.0 20.0 -60
Time (s) 10.0 11.0 12.0 13.0 14.0 15.0 16.0 17.0 18.0 19.0 20.0
Time (s)

(3)
Fig. 3. Real-time estimation of the actuator fault when fa (t) ≈ 0. (3)
Fig. 5. Real-time estimation of x4 when fa (t) ≈ 0

B. Sensor-fault case The considered sensor fault is fs (t) = 1.2sgn(sin(0.75t));

Similar as before, LMIs in Theorem 2 are tested for f˙s ≈ 0 real-time results are plotted in figures 6 and 7, the first one
(3)
and fs ≈ 0, they also included a decay rate (see Remark 1). shows the estimation of the fault under the assumption f˙s (t) ≈
They have been found feasible with β = 15 0 while the second one uses a third order approximation.
 
3.2530 −0.3736 −0.1188 0.0066 2.5220 V. C ONCLUSIONS
−0.3736 3.4222 0.0037 −0.1206 2.6512 
  It has been presented an unknown input observer for the es-
Ps=−0.1188 0.0037
 0.0067 0.0004 −0.0866  timation of actuator and sensor faults of the Furuta pendulum.
 0.0066 −0.1206 0.0004 0.0066 −0.0841 Conditions in the form of LMIs have been derived by means
2.5220 2.6512 −0.0866 −0.0841 4.6995 of an approximation of the nonlinear dynamics combined with
  the Lyapunov method. Experimental results illustrate how the
0.2250 0.2673
 0.1776 0.3106 
Fs
and L̄s = 1 × 103 
 
 1.6065 1.2933 
 Fs vs Fsg Fsg
 0.7302 2.2348  40
−0.1705 −0.2475 30

for the first order derivative; while for the third order, with 20
10
Voltage (V)

β = 9 we have
  0
0.0261 0.2808 -10
 0.0232 0.2861  -20
 
 0.2235 2.3678  -30
L̄s = 1 × 104 
 
 0.1993 2.5442 . -40
10.0 11.0 12.0 13.0 14.0 15.0 16.0 17.0 18.0 19.0 20.0
−0.0225 −0.2721 Time (s)
 
−0.1321 −1.8427
−0.3737 −6.9996 Fig. 6. Real-time estimation of the sensor fault when f˙s (t) ≈ 0.
 Fs [17] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear matrix
Fs vs Fsg Fsg inequalities in system and control theory. SIAM, 1994, vol. 15.
80 [18] K. Tanaka and H. Wang, Fuzzy Control Systems Design and Analysis:
60
A linear matrix inequality approach. New York: John Wiley & Sons,
2001.
40
20
Voltage (V)

0
-20
-40
-60
-80
10.0 11.0 12.0 13.0 14.0 15.0 16.0 17.0 18.0 19.0 20.0
Time (s)

(3)
Fig. 7. Real-time estimation of the sensor fault when fs (t) ≈ 0.

proposal performs in real-time. A nonlinear generalization of

the UI observes is left as future work.
R EFERENCES
[1] M. Blanke, M. Kinnaert, J. Lunze, M. Staroswiecki, and J. Schröder,
Diagnosis and fault-tolerant control. Springer, 2006, vol. 2.
[2] X. Zhang, M. M. Polycarpou, and T. Parisini, “A robust detection and
isolation scheme for abrupt and incipient faults in nonlinear systems,”
IEEE Transactions on Automatic Control, vol. 47, no. 4, pp. 576–593,
2002.
[3] N. Bedioui, R. Houimli, and M. Besbes, “Simultaneous sensor and
actuator fault estimation for continuous-time polytopic LPV system,”
International Journal of Systems Science, vol. 50, no. 6, pp. 1290–1302,
2019.
[4] Y. Xiong and M. Saif, “Unknown disturbance inputs estimation based
on a state functional observer design,” Automatica, vol. 39, no. 8, pp.
1389–1398, 2003.
[5] B. Marx, D. Koenig, and J. Ragot, “Design of observers for Takagi-
Sugeno descriptor systems with unknown inputs and application to fault
diagnosis,” IET Control Theory & Applications, vol. 1, no. 5, pp. 1487–
1495, 2007.
[6] R. Orjuela, B. Marx, J. Ragot, and D. Maquin, “On the simultaneous
state and unknown input estimation of complex systems via a multiple
model strategy,” IET Control Theory & Applications, vol. 3, no. 7, pp.
877–890, 2009.
[7] D. Ichalal and S. Mammar, “On unknown input observers for LPV
systems,” IEEE Transactions on Industrial Electronics, vol. 62, no. 9,
pp. 5870–5880, 2015.
[8] D. Rotondo, M. Witczak, V. Puig, F. Nejjari, and M. Pazera, “Robust
unknown input observer for state and fault estimation in discrete-
time Takagi–Sugeno systems,” International Journal of Systems Science,
vol. 47, no. 14, pp. 3409–3424, 2016.
[9] B. Marx, D. Ichalal, J. Ragot, D. Maquin, and S. Mammar, “Unknown
input observer for LPV systems,” Automatica, vol. 100, pp. 67–74, 2019.
[10] C. Edwards and C. P. Tan, “A comparison of sliding mode and unknown
input observers for fault reconstruction,” European Journal of control,
vol. 12, no. 3, pp. 245–260, 2006.
[11] D. Quintana, V. Estrada-Manzo, and M. Bernal, “Fault detection and
isolation via a novel convex optimization scheme,” IEEE Latin America
Transactions, vol. 17, no. 07, pp. 1096–1101, 2019.
[12] D. Wang and K.-Y. Lum, “Adaptive unknown input observer approach
for aircraft actuator fault detection and isolation,” International Journal
of Adaptive Control and Signal Processing, vol. 21, no. 1, pp. 31–48,
2007.
[13] W. Baumann and W. Rugh, “Feedback control of nonlinear systems
by extended linearization,” IEEE Transactions on Automatic Control,
vol. 31, no. 1, pp. 40–46, 1986.
[14] H. Kailath, Linear Systems. New York, USA: Prentice Hall, 1980.
[15] S. Boyd, L. E. Ghaoui, E. Feron, and V. Belakrishnan, Linear Matrix
Inequalities in System and Control Theory. Philadelphia, USA: SIAM:
Studies In Applied Mathematics, 1994, vol. 15.
[16] C. Scherer, Linear Matrix Inequalities in Control Theory. Delf, The
Netherlands: Delf University, 2004.