Академический Документы
Профессиональный Документы
Культура Документы
Abstract— The development of the artificial pancreas in the level by using a control algorithm, an insulin pump and a
Type 1 Diabetes treatment has gained strength in recent years. sensor of glucose, forming a closed control loop [7].
However, the main challenge facing this development is the
variations of the parameters in the plant, which are expected to BLOOD
change even up to an 80% of their nominal values. This INSULIN GLUCOSE
INSULIN
parametric variation presents greater difficulty when applying CONTROLLER
PUMP
controllers. This paper presents the design and evaluation of two
robust controllers, the Nonlinear PID controller and the
Dynamic Sliding Mode Controller. The simulation scenarios
consider the intake of 3 meals per day and different parametric GLUCOSE
variation. The performance analysis shows that NPID maintains SENSOR
blood glucose levels in its set-point zone better than the DSMC. Fig. 1. Artificial pancreas closed loop
Finally, the controllers are tested in the UVA/Padova metabolic
simulator obtaining the 100% of the patients in the safe range. The difficulty arises when working with biological plants,
in this case the glucose-insulin regulation cycle, since is
Keywords—Artificial pancreas, NPID, DSMC, parametric
characterized by having parameters that vary intraday and
variation, type 1 diabetes.
interpatient. The desirable result is to have a model that
describes as accurately as possible to each patient with
I. INTRODUCTION diabetes and its parameters variation throughout the day. For
The main source of energy of the human being is glucose patients with T1D, insulin sensitivity or the time of action of
and the hormone responsible for maintaining its normal level insulin are the most variable parameters since they can be
in the blood is insulin. This hormone is secreted by beta cells, affected by physical activity, stress levels, among others.
located in the islets of Langerhans of the pancreas, allowing Many of the proposed models fail to solve the aforementioned
the storage of glucose in adipose tissue, muscle and liver in problems since they are modeled based on taking samples of
the form of glycogen [1]. a few hours or days; however, the model used in the present
development [8], being one of the most current, it takes data
A person with diabetes is who has hyperglycemia or blood over 12 weeks in 8 patients, making the description of the
glucose levels above the parameters considered normal, that glucose regulation cycle very close to a real one.
is, from 70 mg/dl to 120 mg/dl, before consuming food, and
maximum 180 mg/dl after of consuming food [2]. This disease The most used control algorithms for this type of plants,
can generate cerebrovascular and cardiovascular disorders, are the PID and MPC [9]. These two types of controllers have
dysfunction in eyes, nerves, kidneys, blood vessels and heart, presented satisfactory results, however, there are small
coma and even life risk, if not treated in time [3]. In 2016 there disadvantages for them. In the case of the PID, as it is not a
were 4906 deaths in Ecuador due to Diabetes, increasing this robust controller, it does not provide the necessary reliability
number by 51% since 2007; taking into account these in case of variations of the model parameters and the MPC,
statistics, Diabetes has become the second cause of death in has a high computational cost [11]. With this in mind, in this
the country, surpassed only by ischemic diseases of the heart work different controllers are going to be designed looking for
[4]. viable control alternatives that are simple, robust and
implementable.
T1D affects 10% of patients[5], being mostly children and
adolescents, although the possibility of it occurring at any age This paper presents the design and tuning of two
is not ruled out [6]. controllers accompanied by a Feedforward controller. The
first, a Non-linear PID (NPID), whose design introduces non-
Insulin therapy is one of the most effective treatments for linear functions for the error producing variable gains
T1D and involves the delivery of insulin either by Multiple depending on the value of the functions [12]. Finally, a Sliding
Daily Injections (MDI) or Continuous Subcutaneous Insulin Mode Controller is presented; in the design of this, a filter is
Infusion (CSII), these latter one uses insulin pumps. It is added to the FOPDT model of the plant, improving stability.
important to create efficient therapies because if the insulin is Adding this particularity this becomes a Dynamic Sliding
supplied in the wrong way, hypoglycemia or low blood Mode Controller (DSMC). The advantage of this new
glucose levels can occur, which also cause serious controller is the reduction of chattering.
consequences [5]. With this in mind, the artificial pancreas
(AP) is an excellent alternative for the treatment of T1D since In the results obtained after the simulation, in nominal
it automates the whole process, controlling the blood glucose conditions and in the presence of parametric variations, blood
𝑈𝐷 = ∫ 𝑘𝐷 𝑠𝑖𝑔𝑛(𝑆(𝑡)) ()
Fig. 2. NPID graphical interpretation The final control law () is the sum of these two
components, but it must be taken into account that the design
of this controller is done in deviation variables, so to
Finally, the control law is: implement it in plants with initial conditions, the following
1 considerations must be taken:
𝑈𝑁𝑃𝐼𝐷 = 𝑘𝑝 [𝑓𝑎𝑙(𝑒𝑝 , 𝛼𝑝 , 𝛿) + 𝑓𝑎𝑙(𝑒𝑖 , 𝛼𝑖 , 𝛿)
𝑇𝑖 () ̅ ± 𝑈(𝑡)
𝑚(𝑡) = 𝑚 ()
+ 𝑇𝑑 𝑓𝑎𝑙(𝑒𝑑 , 𝛼𝑑 , 𝛿)]
− (𝑡)
𝑋𝑚 − (𝑡)
= 𝑋𝑚 − 𝐺̅ ()
Given that the base of this controller is a PID, to find the 𝜏𝜆 1
𝑘𝑝 , 𝑇𝑖 and 𝑇𝑑 values, Dahlin tuning is used [14]. 𝑚(𝑡) = 𝑚
̅ ± ∫( 𝑒(𝑡) + (𝑋 − (𝑡) − 𝐺̅ )
𝐾𝑡𝑓 𝐾𝑡𝑓 𝑚
𝜏𝑝 ()
𝑘𝑝 = () 1
2 𝐾𝑝 𝑡0𝑝 − 𝑈𝑐 (𝑡)) 𝑑𝑡 + ∫ 𝑘𝐷 𝑠𝑖𝑔𝑛(𝑆(𝑡))
𝑡𝑓
𝑇𝑖 = 𝜏𝑝 ()
With their respective tuning constants, which are obtained
𝑇𝑑 = 0.5 𝑡0𝑝 () according to [16]:
B. DSMC 0.51 𝜏 0.76
𝐾𝐷 = 𝑐 [ ] ()
The FOPDT model of the plant is obtained and a filter is |𝐾| 𝑡0
added in the numerator and denominator, where the value of
the pole and the zero of this filter is 𝑡𝑓 > 1: 𝜏 + 𝑡0
𝜆=𝑐 ()
𝜏𝑡0
Where, 𝐾, 𝜏 y 𝑡0 are the values of the parameters of the 𝜶𝒊 0.8
FOPDT model of the plant and 𝑐 is a tuning constant [16]. 𝜶𝒅 0.5
For the design of the controller, the diagram shown in Fig.
3, that part of the Smith predictor, should be considered. The DSMC uses tuning constants shown in the TABLE III.
TABLE III. DSMC TUNING PARAMETERS
R (s) E(s) DSMC U (s) Y (s)
SETPOINT +
CONTROLLER
PLANT
--
Parameter Value
Xm-(s) +
Xm(s) +
Gp-(s) Gp+(s) + -- 𝒄 100
I. Cond. 𝒕𝒇 1000
+ I. Cond.
+
𝝀 1.1021
X(s) Em(s)
+
+ 𝒌𝑫 5.3537
Fig. 3. DSMC diagram Finally, the 𝐾𝑑 parameter is obtained for the feedforward
controller ().
C. Feedforward
The food intake is taken into account as the disturbance of 𝐺𝑓𝑓 = b ∗ 1.4 ()
the plant.
Based on the equations proposed by Marchetti, in [17], the 𝑏 = 4.2 ()
information of two transfer functions is used. The first, 𝐺𝑃
(), is the FOPDT model of the plant, which represents the
effect of the controlled variable 𝑌(𝑠) against a change of the Where 𝑏 is a re-tuning parameter to improve the controller
manipulated variable 𝑈(𝑠); and the second, 𝐺𝑑 (), is the performance.
FOPDT model of the perturbation that represents the effect of
The feedforward controller is applied to the NPID and the
the disturbance 𝐷(𝑠), in the controlled variable 𝑌(𝑠) . DSMC.
Consequently, 𝐺𝑓𝑓 is the feedforward controller, like this:
For the simulation, the consumption of food distributed in
𝑌(𝑠) 𝑒 −𝑡0𝑑 𝑠 three daily meals with a duration of 10 minutes each was
𝐺𝑑 (𝑠) = = 𝐾𝑑 () considered, which had the following amounts of
𝐷(𝑠) 𝜏𝑑 𝑠 + 1
carbohydrates; breakfast at 8:00 am with 40 grams of
𝐾𝑑 carbohydrates, lunch at 1:00 pm with 60 grams and finally
𝐺𝑓𝑓 = − ()
𝐾𝑝 dinner at 7:00 pm with 30 grams of carbohydrates.
After applying the controllers to the Hovorka model [8],
IV. RESULTS with nominal values, the results shown in Fig. 4 were
obtained, and the controller effort observed in Fig. 5 y Fig. 6.
A. Matlab simulation
In the TABLE I, the FOPDT model parameters are shown,
as shown in Section III.
TABLE I. PLANT FOPDT MODEL PARAMETERS
Parameter Value
𝝉𝒑 (min) 298.975
𝒕𝟎𝒑 (min) 130.28
𝑲𝒑 (mg⁄dl per U/h) -321.355
Fig. 4 NPID and DSMC glucose response
As explained in the previous section, using these Where it can be observed, that for the NPID, the glucose
parameters and the equations (), () y (), we obtain the always remains within the normoglycemia levels, being its
tuning of a PID; after this, the values of the NPID parameters maximum value 170.2 (mg/dl) and the minimum 70 (mg/dl).
are assigned. On the other hand, the DSMC presents hypoglycemia at
TABLE II. NPID PARAMETERS minute 1143, descending from normoglycemia 7.7 (mg/dl).
Parameter Value
𝐤𝐩 -0.00357
𝐓𝐢 298.975
𝐓𝐝 64.14
𝜹 20
𝜶𝒑 1.2
Fig. 5 NPID and DSMC control effort When observing the results of the glucose levels after
applying scenario 2, the NPID shows hyperglycemia with a
value of 184 (mg/dl) in the second meal, while the DSMC
presents hyperglycemia in the second meal reaching 216.6
(mg/dl) and in the third meal the glucose level reaches the
upper limit.
B. UVA/Padova simulation
Below are the graphs of the 11 adult patients for both the
NPID (Fig. 11) and the DSMC (Fig. 12). Being able to observe
that there are no hypoglycemias in none of the cases.