Accelerat ing t he world's research.

LPV fault detection of glucose-insulin

system
Jozsef Bokor, Levente Kovacs

2006 14th Mediterranean Conference on Control and Automation

Cite this paper Downloaded from Academia.edu 

Get the citation in MLA, APA, or Chicago styles

Related papers Download a PDF Pack of t he best relat ed papers 

A novel affine qLPV model derivat ion met hod for fault diagnosis H<inf>&#x221E;</inf> perfor…
Nader Meskin

LPV design of reconfigurable and int egrat ed cont rol for road vehicles
Jozsef Bokor, Zolt án Szabó

Act ive Suspension in Int egrat ed Vehicle Cont rol

Jozsef Bokor, Zolt án Szabó
 LPV fault detection of glucose-insulin system
L. Kovács and B. Kulcsár and J. Bokor and Z. Benyó
Abstract— A new approach is proposed in the paper for un-
known glucose inlet estimation in case of diabetic patients under
insensitive care in form of a Linear Parameter Varying (LPV)
model class. The nonlinear two compartment glucose-insulin
system is rewritten in affine parameter varying form in order
to handle the dynamics with LPV Fault Detection and Isolation
(FDI) method. The aim of the closed-loop fault detection, on
the one hand, is to estimate the unknown disturbance input
of the system, on the other hand to recognize the actuator
fault. A parameter varying controller assures the performance
requirements and stability of the system.
FDI, LPV, glucose-insulin system, diabetes, Fundamental
Problem in Residual Generation
I. I NTRODUCTION
In many biomedical systems, external controller provides
the necessary input, because the human body could not
ensure it. Giving an example, diabetes is one of the most
serious disease which needs to be artificially regulated. The
newest statistics of the World Health Organization (WHO)
predate a more than 1% increase of the diabetes mellitus
from 2000 to 2025 and predestinate that 5.4% of the adult
society will suffer from, by the year 2025. The treatment of
the diabetes under insensitive care could be controlled by an
outer loop, replacing the human glucose-insulin equilibrium
system, if needed. The replacement, the outer control might
be partially or fully automatized. The self- regulation has
several strict requirements, but once it has been designed it
permits not only to facilitate the patient’s life suffering from
the disease, but also to optimize the amount of the used
dosage.
The maintenance of the glucose level, by the appropriate
choice of the insulin inlet, in a diabetic patient under inten-
sive care is currently an active research field in Biomedical
Engineering. (Van den Berghe, 2003) have shown that tight
glucose control can reduce the Intensive Care Unit (ICU)
patient mortality by 45% if the glucose level is kept less
than 6.1 mmol/L for a cardiac care population. It was shown
Partially supported by the Hungarian National Scientific Research Foun-
dation (OTKA T043111, T042990, F046726) and by the Hungarian National
Office for Research and Technology (RET-04/2004) which are gratefully
acknowledged by the authors.
L. Kovács and Z. Benyó are with the Department of Control Engineering
and Information Technology, Faculty of Electrical Engineering and Infor-
matics, Budapest University of Technology and Economics, Magyar tudósok
krt. 2., H-1117, Budapest, Hungary lkovacs@iit.bme.hu
B. Kulcsár is with the Department of Control and Transport Au-
tomation, Faculty of Transportation, Budapest University of Technol-
ogy and Economics, Bertalan L. u. 2., H-1111 Budapest, Hungary,
kulcsar@kaut.kka.bme.hu
J. Bokor is with the Computer and Automation Research Institute
Hungarian Sciences Academy H-1111 Budapest Kende u. 13-17., Hungary,
bokor@sztaki.hu
that automated control algorithms capable of tight regulation
for glucose intolerant ICU patients would reduce mortality,
as well the current burden on ICU medical resources and
time.
To design an appropriate control, an adequate model is
necessary. In the last 50 years, a variety of models for the
interaction between glucose and insulin have been suggested.
However, the simplest model proved to be the minimal
model of (Bergman et al., 1979), but its shortcoming is its
big sensitivity to variance in the parameters and that the
plasma insulin concentration must be known as a function
of time. Therefore, extensions of this minimal model have
been proposed.
Strategies have been designed and applied to the problem
(Fischer and Teo, 1989; Juhász, 1993; Benyó et al., 1998)
and (Hovorka et al., 2004; Lin et al., 2004; Kovács et al.,
2005) were trying to capture the changes in patient dynamics
of the glucose-insulin interaction, particularly with respect
to insulin sensitivity. On the other hand, biomedical control
applications are safety related problems, thus the recognition
of the malfunction or the disturbance could be capital to
preserve human life.
The goal of the Fault Detection and Isolation (FDI) is
well-known; detect the malfunction of the safety specific
system and take any measures in order to avoid catastrophes,
damages, injuries or simply the system breakdown. There
exist several alternative ways for detection and isolation of
failures. Using the approaches like multiple model, general-
ized likelihood methods (Bokor et al., 1989), state observer
or parameter estimation, the failure appearance in system can
be detected.
The design of the detection filter is a model-based method-
ology to recognize abrupted behavior of the system. The idea
behind the model-based filter design is the appearance of
the unknown inputs of the dynamic in the error system. The
palette of the detection filter design is large (Massoumnia,
1986; White and Speyer, 1987; Bassville, 1998).
The robust detection filter is one of the most exciting re-
search area in FDI techniques, (Edelmayer et al., 1994; Edel-
mayer et al., 1997) or (Stoustrup and Grimble, 1997; Balas et
al., 2002; Szászi et al., 2002; Hammouri et al., 1999; Kulcsár
and Bokor, 2004).
The aim of the paper is double. First, the nonlinear
biomedical model, the two-compartment Bergman model is
transformed in input affine and linear parameter varying
mode in order to formulate the control and fault accom-
modation in LPV strategy. Secondly, the fault identification
problem is used to unknown input detection to provide a
useful information for supervisory and higher control level.
 The layout of the paper is the following. After the intro- The LPV system is said to be quasi (qLPV) when a part
ductory section, the insulin-glucose nonlinear system is set of the parameters ρ(t) coincides with the states x(t) (Szászi
up. The third part describes shortly the LPV FDI system et al., 2005).
design. Before the conclusion a numerical example is given A modified minimal model for the insulin-glucose inter-
to demonstrate the unknown disturbance detection algorithm. action in human body can be written in the following two-
compartment nonlinear mode (Benyó et al., 1998):
II. QALPV AND NONLINEAR GLUCOSE - INSULIN
MODELING ẋ1 (t) = p1 x1 (t) + p2 d(t)
(9)
The section briefly describes the applied nonlinear ẋ2 (t) = (p3 − x1 (t))x2 (t) + u(t) + p4
Bergman realization and formulates the problem in quasi where the parameters have been identified by (Juhász,
Affine Linear Parameter Varying (qALPV) model. The mod- 1993):
ified nonlinear and minimal model is set for Type I diabetes
patients under intensive care. p1 = −0.021151
Linear Parameter Varying systems are class of nonlinear p2 = 0.092551
(10)
systems where the parameter could be an arbitrary time p3 = −0.014188
varying, piecewise-continuous and vector valued function p4 = 0.077947
denoted by ρ(t) defined on a compact set P. In order to The terms d(t) and u(t) are the exogenous glucose
evaluate the system, the parameter trajectory is requested to [g/100ml] and insulin [µ U/ml] inlets, x1 (t) and x2 (t) stand
be known either by measurement or by computation. for the concentration of glucose in the plasma and for the
A formal definition of the parameter varying systems is concentration of the insulin remote from plasma. In our case
given below. x1 (t) and x2 (t) represent both the states and the output of the
Definition 1: For a compact subset P ⊂ Rs , the parameter system, as the dynamical performances of the measurement
variation set FP denotes the set of all piecewise continuous and actuator devices are considerably faster than the system
function mapping R+ (time) into P with a finite number of itself.
discontinuities in any interval. The compact set P ⊂ Rs , Firstly, the nonlinear system is transformed into quasi
along with continuous functions A : Rs 7→ Rn×n , B : Rs 7→ input-affine LPV by the appropriate choice of the parameters.
Rn×nu , C : Rs 7→ Rny ×n and D : Rs 7→ Rny ×nu represent Let us define two scheduled variable on the parameter set by
an nth order LPV system whose dynamics evolve as ρ(t), given by:
ẋ(t) = A(ρ)x(t) + B(ρ)u(t) (1) · ¸ · ¸
y(t) = C(ρ)x(t) + D(ρ)u(t) (2) ρ1 (t) x1
ρ(t) = = 1 . (11)
ρ2 (t) x1
with ρ(t) ∈ FP , (Wu et al., 2000).
Several methods exist to create an LPV description. In the Introducing these variables, (9). can be rewritten in the
paper, affine parameter dependent model is used which can following input affine form:
formulated in general parameter dependent form such as (1)-
(2). The affine dependency with d(ρ(t)) = N in the system ẋ(t) = A(ρ(t))x(t) + B2 u(t) + B1 d(t)
(12)
means: y(t) = Cx(t)
A(ρ) = A0 + ρ1 A1 + ρ2 A2 + ... + ρN AN (3) where x(t) represents the states, u(t) the insulin control
B(ρ) = B0 + ρ1 B1 + ρ2 B2 + ... + ρN BN (4) input, d(t) the glucose disturbance, y(t) is the measured
output.
C(ρ) = C0 + ρ1 C1 + ρ2 C2 + ... + ρN CN (5)
The parameter matrices are given by:
D(ρ) = D0 + ρ1 D1 + ρ2 D2 + ... + ρN DN . (6)
A(ρ(t)) = A0 + A1 ρ1 (t) + A2 ρ2 (t) =
The affine LPV system can be written as: · ¸ · ¸ · ¸
p1 0 0 0 0 0
( N
) + ρ (t) + ρ (t) (13)
X 0 p3 0 −1 1 p4 0 2
Σ(t) := Σ0 + ρi Σi : ρi ∈ [ρi ρi ], ρ̇i ∈ [ρ̇i ρ̇i ] (7) · ¸ · ¸ · ¸
i=1 p2 0 1 0
B1 = , B2 = , C= . (14)
0 1 0 1
with i = 0, 1, ..., N and
· ¸ For instance, no actuator and no sensor dynamics are taken
Ai Bi into consideration, they are supposed to be neglected because
Σi = . (8)
Ci Di of the slow glucose-insulin dynamics.
In (7)., the parameters are varying between known minimal Once the x1 (t) glucose concentrate is measured, the
and maximal bounds (respectively the limits of theirs rates parameters are known. The extremal values of the ρ(t)’s
are known). Parameter dependent models are very useful and determine the parameter polytope with the minimal and
could even describe uncertainty (parameter uncertainty). It maximal point of the vertex. If the bounds of the parameter
is rewarding to create affine LPV systems, since control and rates are known, it could be implemented into the design
FDI synthesis becomes easier. process.
 III. PARAMETER VARYING RESIDUAL GENERATION
In this section, the nonlinear unknown glucose and insulin
failure detection problem is presented. The useful tool of the
detection can be rooted in the geometric based parameter
varying Fundamental Problem in Residual Generation, FPRG
(Balas et al., 2002).
The general block scheme of FPRG can be seen in Fig.
1. The so-called unknown input strategy can be recognized,
since the residual generator works from the insulin control
input and from the measured system outputs and does not use
the disturbance or failure if any. The goal of the LPV FPRG
u ✲
✲ r✲
1
RG1 ✲ the fault detection filter, namely the residual. Note that this
✲ geometric based parameter dependent fault detection filter is
✲ r✲
2 solution is the minimization of the effect of the false alarm
y RG2 ✲
✲ ✲
Fig. 1. Fundamental Problem in Residual Generation scheme
is to design parameter dependent residuals which are sensi-
tive only to the appropriate unknown input events. In (9).,
respectively in its qALPV form, the only exogenous noise
direction is the disturbance, but let us suppose the presence
of an occasionally actuator failure acting through the control
input direction. The failure direction is denoted by L and
modeled as a linear additive term in the input affine LPV
model. Let us suppose a constant fault direction. However,
this is not the only way because parameter dependent failures
or disturbances could be treated as well.
L = B2 .
On the other hand, the problem is a nonlinear disturbance
decoupling and estimation problem.
Using the LPV FPRG two residuals will be designed. The
first residual generator, sensitive only to the d(t) disturbance,
is influenced only by the glucose intake. The second residual
is dedicated only to the possible actuator malfunction, where
the fault signature is denoted by m(t).
The existence of the residual generator, the unknown input
observer is allied to the observability and input detectability
criteria. The (C, A(ρ)) pair is observable and C is epic (i.e.
it has full row rank). The solvability condition, the input
observability needs to be verified by:
S2∗ ∩ L = 0
where S2∗ is the minimal unobservability subspace of image
B1 (B1 ) and L is the image Im(L).
Henceforward, the check of the following condition is
indispensable for the existance of the residual generator
sensitivity only for the actuator fault:
S1∗ ∩ B1 = 0
The theory behind the solvability condition is to look for
a minimal subset, where one of the failure to be decoupled
can be caught.
If (16)-(17). are held one can design FDI filters in the
form of a parameter dependent observer:
ẇ(t) = F (ρ)w(t) − E(ρ)y(t) + G(ρ)u(t) (18)
r(t) = M w(t) − Hy(t) (19)
where H will be the solution of Ker HC = Ker C + S ∗ ,
and M is the unique solution of MP=HC with P : X →
X /S(ρ). F (ρ) can be given as F (ρ) = A0 (ρ) + D1 (ρ)M ,
E = P D(ρ) and D = D0 (ρ) + P −r D1 (ρ)H, finally G =
P B(ρ) and P −r is the right inverse of the projection matrix
(Massoumnia, 1986; Balas et al., 2002). r(t) is the output of
not necessary stable.
If the input obsevability can not be met, the most obvious
on residual output by disturbance attenuation on residual
(Kulcsár and Bokor, 2004).
IV. Q UESTION OF STABILITY
The stability requirements of the insulin control and of the
unknown input detection filter are treated in the sequel.
For the first slight, the two problems, namely the stability
of the closed loop and of the FDI filter, seem to be different,
but in case of the two compartment qALPV model the same
stabilizing method could be applied. The connection between
them is only the duality and the P transformation described
in the previous part.
Both the insulin inlet and the detection filter uses param-
eter dependent gain. The quadratically stabilizing parameter
dependent Lyapunov criteria is written:
(15) Acl (ρ)T Q + QAcl (ρ) = (A(ρ) + B2 K(ρ))T Q +
+ Q(A(ρ) + B2 K(ρ)) < 0.
where Q is the solution and K(ρ) is the state feedback term.
Acl (ρ) is the closed loop parameter varying matrix.
Secondly, one can assure the stability of the LPV FDI
system with quadratically stabilizing D1 (ρ) if F (ρ) =
A0 (ρ) + D1 (ρ)M . In this case the gain is given by:
D1 (ρ) = D10 + ρ1 D11 + ρ2 D12
The quadratically stable parameter varying gain can be
obtained by
F (ρ)T Q̄ + Q̄F (ρ) = (A0 (ρ) + D1 (ρ)M )T Q̄ +
(16) + Q̄(A0 (ρ) + D1 (ρ)M ) < 0
with Q̄ = Q̄T > 0 solution for all ρ. If one introduces
a new variable Z(ρ) = D1 (ρ)Q̄ the only thing to get the
appropriate solution for stable gain is to solve (in case of
affine parameter model) the following inequalities at each
corner point of the polytropic:
(17) A0 (ρ)Q̄ + Q̄A0 (ρ) + M T Z(ρ)T + Z(ρ)M < 0
 Control input signal
Note that parameter independent gain D1 (ρ) can assure the 1.4

stability of the LPV filter as well.

1.2

V. U NKNOWN INPUT DETECTOR DESIGN FOR Q ALPV 1

MODEL OF GLUCOSE - INSULIN SYSTEM
0.8
The section interprets a possible simulation setup of faulty,

Insulin (µ U/ml)
disturbance corrupted closed loop insulin system along a 0.6

specified trajectory. The goal of the numerical example is to

0.4
show detection of the unmeasured input signal under closed
loop glucose-insulin control. Based on the validity domain 0.2

of the minimal two compartment nonlinear Bergman model,

0
the extremal values of the parameters can be easily evaluated
such as: −0.2
0 20 40 60 80 100 120 140 160 180

Time (min.)
Real and estimated glucose intake
0.05
Real glucose intake
Estimated glucose intake
Fig. 4. The control input change
0.04

0.03 effect of these inputs can be separated, i.e. the solvability

Glucose (g/100 ml)

condition for LPV FDI is held.

0.02 Glucose and actuator malfunction is taken into account
during the 180 min simulation. The glucose input variation
0.01
(Kovács et al., 2005) can be seen in Fig. 2, where one plots
the estimated intake respectively. Since the detection is based
0
on the responses of the system and on the control input one
concludes a slight delay between them. The actuator failure
−0.01
0 20 40 60 80 100 120 140 160 180
Glucose output Insulin output
Time (min.) 0.06 7.5

0.05
Fig. 2. The disturbance and the detected disturbance 7

0.04

Insulin (µ U/100 ml)

Glucose (g/100 ml)

6.5
Detected and real actuator fault
1.2 0.03

1 0.02

5.5
0.8 0.01
Insulin (µ U/ml)

0.6 0 5
0 50 100 150 200 0 50 100 150 200

Time (min.) Time (min.)

0.4

0.2 Fig. 5. The outputs of the system

0
and its estimation is shown on the Fig. 3. The isolation and
Detected actuator fault
Real actuator fault detection of the actuator fault can be stated at one hour.
−0.2
0 20 40 60 80 100 120 140 160 180
The actuator fault is a step function where one supposes the
Time (min.)
malfunction of the control input with an additive 1 µU/ml
magnitude. The control input and measured output (Fig. 4-5)
Fig. 3. The actuator failure and its recognition is perturbed by the glucose and the actuator fault since the
closed loop stability is assured by a quadratically stabilizing
·
ρ1 ρ1
¸ ·
x x1
¸ parameter dependent gain (Section IV).
G= = 11 1 (20)
ρ2 ρ2 x1 x1 VI. C ONCLUSION
where G denotes the parameter margins and the under-, Biomedical control applications are safety oriented sys-
respectively overlined values are the maximal and minimal tems. The paper presents the closed, glucose-insulin loop
terms, where ρ1 = 10−3 and ρ1 = 10−1 µU/ml. During the qALPV fault detection problem, where not only a possible
simulation setup one would like to reconstruct the unknown actuator failure is recognized, but also the unmeasured glu-
inputs, the glucose disturbance and the actuator fault. The cose intake is estimated. The nonlinear problem was trans-
formed into linear, but parameter dependent form. Further Wu, F., K. M., Grigoriadis and A. Packard (2000). Anti-windup controller
robustness analysis and synthesis will be elaborated on an design using linear parameter-varying control methods. International
Journal of Control 73(12), 1104–1114.
extended dynamic by the reconfiguration of the controller.

R EFERENCES
Balas, G. J., J. Bokor and Z. Szabo (2002). Failure detection for LPV sys-
tems - a geometric approach. Proc. of American Control Conference,
Alaska pp. 4421–4426.
Bassville, M. (1998). Detecting changes in signals and systems-a survey.
Automatica 3(24), 309–326.
Benyó, Z., B. Paláncz, Cs. Juhász and P. Várady (1998). Design of glucose
control via symbolic computation. Proc. of the 20th Ann. Int. Conf. of
the IEEE Engineering in Medicine and Biology Society, Hong Kong
20(6), 3116–3119.
Bergman, B.N., Y.Z. Ider, C.R. Bowden and C. Cobelli (1979). Quantitive
estimation of insulin sensitivity. American Journal of Physiology
236, 667–677.
Bokor, J., P. Gaspar, M. Tanyi and A. Edelmayer (1989). Change detection
in signals and systems: experiences using the multiple model and
the generalized likelihood ratio approach. Proc. of Adv. Information
Processing in Aut. Contr. 2, 13–18.
Edelmayer, A., J. Bokor and L. Keviczky (1994). An H∞ filtering approach
to robust detection of failures in dynamical systems. Proc. 33rd IEEE
Conf. Dec.Cont. pp. 3037–3039.
Edelmayer, A., J. Bokor, F. Szigeti and L. Keviczky (1997). Robust detection
filter design in the presence of time-varying system perturbation.
Automatica 33, 471–475.
Fischer, M.E. and K.L. Teo (1989). Optimal insulin infusion resulting from a
mathematical model of blood glucose dynamics. IEEE Trans. Biomed.
Eng. 36(3), 479–486.
Hammouri, H., M Kinneart and E.H. El Yaagoubi (1999). Observed based
approach to fault detection and isolation for nonlinear systems. IEEE
Trans. Automat. Contr. 10(44), 1879–84.
Hernjak, N. and F. J. Doyle (2005). Glucose control design using nonlin-
earity assessment techniques. AIChE Journal.
Hovorka, R., V. Canonico, L. J. Chassin, U. Haueter, M. Massi-Benedetti,
M. Orsini Federici, T. R. Pieber, H.C. Schaller, L. Schaupp, T. Vering
and M.E. Wilinska (2004). Nonlinear model predictive control of
glucose concentration in subjects with type 1 diabetes. Physiological
measurement 25, 905–920.
Juhász, Cs. (1993). Medical Application of Adaptive Control, Supporting
Insulin-Therapy in case of Diabetes Mellitus. PhD dissertation, Bu-
dapest University of Technology and Economics, Electrical Engineer-
ing Program. Budapest.
Kovács, L., B. Paláncz and Z. Benyó (2005). Classical and modern control
strategies in glucose-insulin stabilization. Proc. 16th IFAC World
Congress, Prague, Czech Republic. Electronic Proceedings #04165.
Kulcsár, B. and J. Bokor (2004). Robust LPV detection filter design using
geometric approach. Proc. 12th IEEE Mediterranean Conference on
Control and Automation. Electronic Proceedings #1123.
Lin, J., J.G. Chase, G.M. Shaw, C.V. Doran, C.E. Hann, M.B. Robertson,
P.M. Browne, T. Lotz, G.C. Wake and B. Broughton (2004). Adaptive
bolus-based set-point regulation of hyperglycemia in critical care.
Proc. of 26th Ann. Int. Conf. of IEEE Eng. in Biomedicine Soc., San
Francisco, USA pp. 3463–3466.
Massoumnia, M.A. (1986). A geometric approach to the synthesis of failure
detection filters,. IEEE Trans. Automat. Contr. AC-31(9), 839–846.
Parker, R.S., F.J. Doyle III, J.H. Ward and N.A. Peppas (2000). Robust
H∞ glucose control in diabetes using a physiological model. AIChE
Journal.
Stoustrup, J. and M. J. Grimble (1997). Integration control and fault
diagnosis: A separation result. Proc. of IFAC Symposium on Fault
Detection, Hull pp. 341–46.
Szászi, I., A. Marcos, G. J. Balas and J. Bokor (2005). LPV detection
filter design for a Boeing 747-100/200 aircraft. Journal of Guidance,
Control and Dynamics 28(3), 439–445.
Szászi, I., B. Kulcsár, G.J. Balas and J. Bokor (2002). Design of FDI filter
for an aircraft control system. Proc. of American Control Conference,
Alaska pp. 4232–4237.
Van den Berghe, G. (2003). Insulin therapy for the critically ill patient.
Clinical Cornerstone 5(2), 56–63.
White, J.E. and J.L. Speyer (1987). Detection filter design: spectral theory
and algorithms. IEEE Trans. Automat. Contr. AC-32, 593–603.