Available online at www.sciencedirect.

com

Journal of the Franklin Institute 357 (2020) 5143–5172

www.elsevier.com/locate/jfranklin

PID controller parameters adjustment using a single

memory neuron
Samir Ladjouzi a,c,∗, Said Grouni b,c
a Department of Electrical engineering, Faculty of Sciences and Applied Sciences, University of Bouira, Rue Frères
Boussendalah 10000 Bouira, Algeria
b Industrial Maintenance Department, Faculty of Sciences Engineering, University UMBB of Boumerdes, Av. de

l’Indépendance 35000 Boumerdes, Algeria

c Applied Automation Laboratory, Faculty of Oil and Chemistry, University of Boumerdes, Av.de l’Indépendance,

35000 Boumerdes, Algeria

Received 3 February 2018; received in revised form 13 February 2020; accepted 18 February 2020
Available online 3 March 2020

Abstract
The work presented in this paper presents the use of a single memory neuron to find optimal gains
for a PID controller. The adopted strategy with the principal equations is discussed. The efficiency of
the proposed method is shown for two usual problems which frequently occur in the industry: a single
tank and a boiler and heat exchanger applications. A comparison with the Ziegler–Nichols method is
presented in order to prove the effectiveness of our approach.
© 2020 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.

1. Introduction

One of most used techniques in the feedback control field is the Proportional, Integral
and Derivative (PID) controller developed in a formal law by Minorsky a Russian American
engineer in 1922. The basic idea of PID is to calculate an error value between a desired

∗ Corresponding author at: Department of Electrical engineering, University of Bouira, Faculty of Sciences and

Applied Sciences, Rue Frères Boussendalah 10000, Bouira, Algeria.

E-mail addresses: [email protected], [email protected] (S. Ladjouzi), s.grouni@univ-
boumerdes.dz (S. Grouni).

https://doi.org/10.1016/j.jfranklin.2020.02.027
0016-0032/© 2020 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.
5144 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

reference and the output process, using an error with three terms: proportional, integral, and
derivative (with the following gains Kp , Ki and Kd , respectively) to produce a control signal
which is applied to the system. Several applications have used the PID controller as an
efficient tool to obtain a good control law. In [1] a PID was used to control a refrigeration
cycle based on a CO2 fluid. Authors in [2] have proposed a PID to control the ball position
in a ball and beam system.
An efficient method in [3], based on the use of the particle swarm optimization technique
associated with a PID controller, was proposed by adapting the operating point of a proton
exchange membrane fuel cell to reach the maximum power. In industrial processes, PID con-
trollers are present in more than 90% of the control loops [4]. Despite its various capabilities,
the persistent problem in a PID controller is how to find optimal values for the three gains
cited above (i.e., proportional gain Kp , integral gain Ki and derivative gain Kd ). Many articles
have been proposed solutions to this problem, like the famous Ziegler–Nichols tuning method
[5]. In [6], the pole placement technique was used to adjust PID parameters and a combination
of pattern recognition with expert systems was applied in [7] on an industrial PID controller.
Recently with the appearance of evolutionary, bio-inspired and intelligent algorithms research
on PID tuning has advanced.
Among papers introducing these algorithms in PID parameter adaptation, we can cite:
Particle Swarm Optimization algorithm [8,9], Genetic Algorithm [10,11], Fuzzy logic [12],
BAT algorithm [13] and others.
Artificial Neural Networks have also been used for the PID tuning to compute the optimal
parameters of the PID by using a single neuron [14–17].
The objective of this paper is to apply a Single Memory Neuron (SMN) to adapt parameters
of the PID. This memory neuron consists of a single neuron with a single delay, which makes
the output history sensitive. The adaptive PID obtained by this single neuron is applied to
control two common industrial processes: a single tank and a heat exchanger.
This paper is arranged as follows: Section 2 describes the memory neuron model. In
Section 3, details are given on the PID tuning procedure. A short overview on the Ziegler–
Nichols method is presented in Section 4. Simulations, results and interpretations are exposed
in Section 5. Section 6 presents the conclusions.

2. Memory neuron model

In this section, we begin by describing the memory neuron model used in this work. This
specific neuron was proposed by Poddar and Unnikrishnan [18], and it was used later in
many applications [19–21]. The term “memory” is used because the memory unit contains
past information from itself and from the output neuron.
The memory neuron is a simple neuron connected with a unit time memory. The neuron
receives external inputs with the delayed output memory unit. The memory unit contains past
information from itself and from the output neuron. Each connection has its proper weight
as depicted in Fig. 1.
In the previous figure, the general representation of the memory neuron is given in part
(a), while in part (b) we have shown the details of the weighted links between the simple
neuron and its memory unit.
We use the following notations related to our work:

xi (k) : the ith input of the neuron at time k (1 ≤ i ≤ n)

 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5145

Fig. 1. The memory neuron model. (a) General scheme. (b) Detailed architecture with weighted links.

s(k) : the output of the memory neuron at time k.

v(k) : the output of the memory unit at time k.
wi (k) : the weight of the connection from the ith input xi (k) to the input of the neuron at
time k.
α(k) : the weight of the connection from the output s(k) to the input of the memory unit
at time k.
β(k) : the weight of the connection from the output of the memory unit to the input of
the neuron at time k
f : the activation function of the memory neuron.

The output of the memory neuron is given by this equation:

 T 

s (k ) = f wi (k )xi (k ) + β(k )v (k − 1 ) (1)
i=1
5146 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 2. Structure of an adaptive PID controller.

The output of the memory unit is given by:

v (k ) = α(k )s (k − 1 ) + (1 − α(k ) )v (k − 1 ) (2)

3. Single memory neuron PID control

The classic PID feedback control structure is shown in Fig. 2.

In Fig. 2, r(k) represents the desired output while y(k) defines the system output.
For finding an optimal control signal, we minimize the following objective function:

1 1
T T
E= e(k )2 = ( r ( k ) − y ( k ) )2 (3)
2 k=1 2 k=1
T is the total number of epochs.
The continuous analog PID controller using the Laplace transform is expressed by:
 
KI
U (s ) = K p + + Kd s E (s ) (4)
s
Using the approximation: s ≈ 1 − z−1 , we obtain the form of the PID controller using the
z transform as:
 
KI  −1

U (z ) = K p + + Kd 1 − z E (z ) (5)
1 − z−1
After a short modification:
     2
1 − z−1 U (z ) = Kp 1 − z−1 + KI + Kd 1 − z−1 E (z ) (6)

By applying the inverse z transform to this equation, the discrete time PID algorithm can
be written as follows:
u (k ) = Kp e(k ) + KI e(k ) + Kd 2 e(k ) (7)
Where:
u (k ) = u (k ) − u (k − 1 ) (8)
We can write:
 e(k ) = (e(k ) ) = (e(k ) − e(k − 1 ) )
2

= (e (k ) − e (k − 1 ) ) − (e (k − 1 ) − e (k − 2 ) ) (9)
= e(k ) − 2e(k − 1 ) + e(k − 2 )
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5147

Fig. 3. Design of the memory neuron controller.

Fig. 4. Block diagram of the memory neuron PID controller.

Finally, Eq. (7) becomes:

u (k ) = Kp [e(k ) − e(k − 1 )] + KI e(k ) + Kd [e(k ) − 2e(k − 1 ) + e(k − 2 )] (10)
In our work, we used the memory neuron as an incremental controller to compute the
increment term u(k).
The memory neuron has three inputs:
⎧
⎪
⎨x 1 ( k ) = e ( k )
x2 (k ) = e(k ) = e(k ) − e(k − 1 ) (11)
⎪
⎩x (k ) = e(k ) − 2e(k − 1 ) + e(k − 2 )
3

And the output of the neuron will represent the increment of the control signal. It is stated
as:
s (k ) = u (k ) = f (inp(k ) ) (12)
The variable inp is the input of the memory neuron given by the following summation:
inp(k ) = w1 (k )x1 (k ) + w2 (k )x2 (k ) + w3 (k )x3 (k ) + β(k )v (k − 1 ) (13)
The control signal applied to the system is stated as:
u (k ) = u (k − 1 ) + u (k ) (14)
The structure of the memory neuron controller is shown in Fig. 3
The different weights wi (k) ( i = 1, 2, 3) represent the parameters KI , Kp and Kd respectively.
The frame form of the PID controller using the memory neuron is illustrated in Fig. 4.
The different weights of the memory neuron were adjusted using the following equations:
∂E (k )
wi (k + 1 ) = wi (k ) − μi (1 ≤ i ≤ 3 ) (15)
∂ wi ( k )
5148 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

∂E (k )
α(k + 1 ) = α(k ) − μα (16)
∂α(k )
∂E (k )
β(k + 1 ) = β(k ) − μβ (17)
∂β(k )
Each weight is adapted with its own learning rate, and μ1 , μ2 and μ3 correspond respec-
tively to the learning rates for the gains Ki , Kp and Kd .
These different learning rates are set by the user, and they are taken in the interval [0, 1].
The partial derivatives of the objective function with respect to the weights are computed
as:
∂E (k ) ∂E (k ) ∂y (k ) ∂s (k ) ∂inp(k )
= (18)
∂ wi ( k ) ∂y (k ) ∂s (k ) ∂inp(k ) ∂ wi (k )
∂E (k ) ∂E (k ) ∂y (k ) ∂s (k ) ∂inp(k ) ∂v (k )
= (19)
∂α(k ) ∂y (k ) ∂s (k ) ∂inp(k ) ∂v (k ) ∂α(k )
∂E (k ) ∂E (k ) ∂y (k ) ∂s (k ) ∂inp(k )
= (20)
∂β(k ) ∂y (k ) ∂s (k ) ∂inp(k ) ∂β(k )
In order to calculate the different error gradients for each weight, we will compute each
partial derivative of Eqs. (18)–(20).
Thus for the weights wi , we will have:

➢ We derive the Eq. (3) to obtain:

∂E (k )
= −e(k ) (21)
∂y (k )
➢ We derive the Eq. (12) to obtain:
∂s (k )
= f  (inp(k ) ) (22)
∂inp(k )
➢ We derive the Eq. (13) to obtain:
∂inp(k )
= xi ( k ) (23)
∂ wi ( k )
With i = 1, 2, 3

For the weight α, we will have:

➢ We derive the Eq. (13) to obtain:

∂inp(k )
= β(k ) (24)
∂v (k )
➢ We derive the Eq. (2) to obtain:
∂v (k )
= s (k − 1 ) − v (k − 1 ) (25)
∂α(k )
For the weight β, we will have:
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5149

Table 1
Ziegler–Nichols tuning rule.

Controller type Kp Ki Kd
P 0.5 Kcr 0 0
1.2 K p
PI 0.45 Kcr Pcr 0
2 Kp
PID 0.6 Kcr Pcr 0.125 Kp Pcr

➢ We derive the Eq. (13) to obtain:

∂inp(k )
= v (k ) (26)
∂β(k )
For simplified calculation, we assume that:
➢ The weight α(k)  1, thus we can do the following approximation v(k) ≈ v(k − 1 ).
➢ The derivative term ∂y(k)
∂s(k)
cannot be computed, it can be replaced approximately by
sgn[ ∂y(k)
∂s(k)
].

Taking into account these simplifications and the different Eqs. (21)–(26), the adaptation
of the weights can be done with these formulas:
 
∂y (k ) 
wi (k + 1 ) = wi (k ) − μi e(k )sgn f (inp(k ) ) xi (k ) (27)
∂s (k )
 
∂y (k ) 
α(k + 1 ) = α(k ) − μα e(k )sgn f (inp(k ) )β(k ) ∗ (s (k − 1 ) − v (k − 1 ) ) (28)
∂s (k )
 
∂y (k ) 
β(k + 1 ) = β(k ) − μβ e(k )sgn f (inp(k ) ) v (k ) (29)
∂s (k )

4. Overview on the Ziegler–Nichols method

The Ziegler–Nichols (ZN) tuning method is one of the most common methods of tuning
the parameters of the PID controller. This technique was developed by John G. Ziegler and
Nathaniel B. Nichols in 1942, its main idea is to determine:

• A critical gain (Kcr ) which causes the oscillation of the system (critically stable).
• A critical period (Pcr ) that defines the period of the obtained oscillation.

The tuning procedure of the ZN method is described below [22]:

(1) The values of Ki and Kd are set to zero.

(2) We use the proportional gain control only and we increase his value gradually until the
system becomes oscillatory. We note the critical value of this oscillation by Kcr .
(3) The period corresponding to the oscillation is calculated and noted by Pcr
(4) We use the Table 1 to find the values of Kp , Ki and Kd .
5150 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 5. Single tank system.

5. Simulation results

In this section, two industrial processes are studied to demonstrate the effectiveness of the
proposed approach. The PID controller is the most used method for this kind of applications.
We have applied our control strategy based on the single memory neuron on two typical
industrial problems: a single tank and a heat exchanger.
Example 1
The first example treated is a system taken from [23], which represents a liquid vessel
described as a first-order model. The flow rate represents the system input and the fluid level
is the output (see Fig. 5).
The dynamic of the system is a discrete-time first-order expressed by the following equa-
tion:
y (k + 1 ) = 0.6y (k ) + 0.01u (k ) (30)
Where u(k) and y(k) represent, respectively, the input and the output of the system at time
(k)
By applying the z transform to this equation, we obtain the following transfer function:

Y (z ) 0.01
H (z ) = = (31)
U (z ) z − 0.6
The activation function used for the SMN is a unipolar sigmoid function stated as:

1
f (x ) = K (32)
1 − e−x
After doing a few number of tests, we chose the following values:
K = 0.1, μ1 = 0.7, μ2 = 0.6, μ3 = 0.4, μα = 0.01, μβ = 0.2
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5151

Table 2
PID parameters for the first system.

Parameter PID-ZN PID-SMN

Kp 30 0.6160
Ki 5 47.3232
Kd 0.00001 0.3517

Fig. 6. Output of the single tank system with the PID-ZN controller (unit step reference).

Fig. 7. Input (signal control) of the single tank system with the PID-ZN controller (unit step reference).

The first part of the simulation is done for a unit step reference r(k) = 1 and in the second
part we have used a variable trajectory reference.
In order to evaluate our technique, we compared it with the Ziegler–Nichols method de-
scribed in Section 4, and the obtained critic values Kcr and Pcr are:
Kcr = 50,Pcr = 12.
Parameter values of the PID for the two methods are summarized in Table 2.
Obtained results for the first system with a unit step reference are shown in Figs. 6–9.
5152 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 8. Output of the single tank system with the PID-SMN controller (unit step reference).

Fig. 9. Input (signal control) of the single tank system with the PID-SMN controller (unit step reference).

The adapted PID parameters obtained with the classical Ziegler–Nichols method and the
Single Memory Neuron are used in two controllers (PID-ZN and PID-SMN) for the single
tank system and the results obtained show the performance of each of them.
Figs. 6 and 8 illustrate the difference between the responses obtained with the two con-
trollers. The system output with the PID-SMN controller reaches rapidly its set point for the
time unit 3, while the PID-ZN which does not reach its set point only after the time unit 70.
Figs. 7 and 9 show the control signal graphs for the two controllers. As can be seen in the
case of the PID-SMN controller, the transition time of the control signal to reach the final
value (3 time units) is shorter compared to the PID-ZN controller as the transition time takes
80 time units.
To evaluate the diverse performances of the two controllers, we use the following main
parameters:
Rise time: time required for a signal to reach from 10% to 90% of the final value.
Overshoot: normalized difference between the time response peak and steady output.
Settling time: time required for the response to reach and stay within the specified range
of its final value.
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5153

Table 3
Step response characteristics for the first system.

Parameter PID- ZN PID-SMN

Settling time 46.5840 2.5763
Rise time 24.2036 1.4974
Overshoot (%) 0 0

Fig. 10. Graph of the variable reference.

Fig. 11. Output of the single tank system with the PID-ZN controller (variable reference).

From the Table 3 we can see that the settling and rise times are significantly reduced by
the PID-SMN controller.
These results confirm the effectiveness of the proposed controller.
Figs. 10–12 represent, respectively, the variable reference, the output of the PID-ZN and
the PID-SMN controllers in the variable reference case.
In the case of a variable reference, the PID-SMN controller reaches its reference with
quick rise time. It has also no steady state error in the entire trajectory which not the same
with PID_ZN.
5154 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 12. Output of the single tank system with the PID-SMN controller (variable reference).

Table 4
MSE values for the first.

PID-ZN PID-SMN
Unit step reference 8.75 × 10–3 2.69 × 10–3
Variable reference 0.1932 0.0524

We summarize the different graphical results by using the Mean Squared Error (MSE)
defined by:

1
T
MSE = ( r ( k ) − y ( k ) )2 (33)
T k=1

Where T is the total number of epochs used in simulations

Table 4 shows the different values of the MSE for the first system using the two controllers
From this table, we can see clearly that the overall error computed with the PID-SMN
controller is considerably lower than that obtained with the PID-ZN controller.
To exhibit the performances of both controllers, the disturbance rejection test was done.
We add a disturbance in the fluid level measurement of 10% of its nominal value at the
time unit 80. The obtained results for both controllers are shown in Fig. 13.
As shown in Fig. 13, after injecting the disturbance, the PID-SMN controller was able to
track quickly the desired reference in 2 time units. By contrast, the PID-ZN controller tooks
50 time units to return to the desired trajectory. This proves the superiority of our controller
for the disturbance rejection.
Example 2
The second application represents a boiler and heat exchanger system studied in [24].
The boiler and heat exchanger set-up system covers two parts: an electrical boiler and a
heat exchanger. The electrical boiler is operated at 4 bar pressure at 144 °C temperature giving
a steam used in the heat exchanger and operated in counter current flow of fluids (Fig. 14).
Counter-current heat exchanger is used more often because it allows more efficient transfer
of energy.
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5155

Fig. 13. Output of the single tank system using the two controllers with noisy measurement.

A displacement pump is used for the boiler tank to pump the water to be heated. The
generated steam flows to the heat exchanger through an equal percentage valve that controls
the flow of steam.
Pumping water is done using a cold water tank with a centrifugal pump. The condensate
formed in the heat exchanger during the heat transfer is collected in a separate condensate
tank. Hot water is collected in a hot water tank that can also be mixed in the cold water
tank or can be added partially / fully to the boiler feed to change the temperature of the feed
water.
Measurements were conducted at several locations in the system. Each location in paren-
theses corresponds to its tag in the diagram.

(1) Level of water in the boiler drum (LT-1)

(2) Drum pressure (PT-1)
(3) Steam temperature (TT-1)
(4) Boiler feed water temperature (TT-5)
(5) Boiler feed water flow ((FT-2)
(6) Temperature from water from hot water tank for mixing with boiler feed water (TT-4)
(7) Steam flow (FT-3)
(8) Heat exchanger steam inlet temperature (TT-2)
(9) Heat exchanger water inlet temperature (TT-4 )
(10) Heat exchanger water outlet temperature (TT-3)
(11) Heat exchanger water flow (FT-1)

To highlight the performance of our controller, we chose in our study to control the
temperature of the heat exchanger. Thus, the process control can be seen as a SISO system
with the input is the heat exchanger water inlet temperature (TT-4), and the output is the heat
exchanger water outlet temperature (TT-3).
The purpose of the command is to use the PID controller so that the output of the system
follows a desired reference.
The control scheme of the heat exchanger is given in Fig. 15.
5156 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 14. The boiler and heat exchanger system.

 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5157

Fig. 15. Control scheme of the heat exchanger temperature using the PID controller.

Fig. 16. Output of the heat exchanger system with the PID-ZN controller (unit step reference).

The dynamic of the system is described by the following differential equation:

d 2 y (t ) dy (t ) du (t )
+ 0.06689 + 0.005244y (t ) = 0.2304 + 0.01265u (t ) (34)
dt 2 dt dt
Where u(t) and y(t) represent, respectively, the water inlet temperature and the water outlet
temperature of the heat exchanger.
Using the Laplace transform to this equation, we obtain the following second order transfer
function:
Y (s ) 0.2304s + 0.01265
H (s ) = = 2 (35)
U (s ) s + 0.06689s + 0.005244
As for the first example we use the unipolar sigmoid function as activation function.
After many tests, we selected the following values:
K = 0.8, μ1 = 0.15, μ2 = 0.2, μ3 = 0.1, μα = 0.01,
μβ = 0.1
The first part of simulation was done for a unit step reference r(k) = 1 and in the second
part we used a variable trajectory reference.
A comparison is done with the Ziegler–Nichols method like for the first system, and the
obtained critic values Kcr and Pcr are given by:
Kcr = 4.712,Pcr = 5.6544.
5158 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Table 5
PID parameters for the second system.

Parameter PID-ZN PID-SMN

Kp 2.8272 29.2079
Ki 1 1.7009
Kd 0.25 0.0327

Table 6
Step response characteristics for the second system.

Parameter PID- ZN PID-SMN

Settling time (s) 15.1079 0.5698
Rise time (s) 1.9667 0.3242
Overshoot (%) 18.6722 0.5421

Fig. 17. Input (signal control) of the heat exchanger system with the PID-ZN controller (unit step reference).

Parameter values for the two methods are summarized in Table 5.

Obtained results for the second system with a unit step reference are shown in Figs. 16–19.
In the second part of simulation, we have applied our tuning PID parameters method to
the boiler and heat exchanger.
Figs. 16 and 18 illustrate, respectively, the responses obtained with the two controllers. The
system output with the PID-SMN controller reaches rapidly its set point at the time 0.5 s,
while for the PID-ZN the set point is reached after 20 s.
Figs. 17 and 19 illustrate the control signal graphs for the two controllers. As can be seen
in the case of the PID-SMN controller, the transition time of the control signal to reach the
final value (0.5 s) is less important compared to the PID-ZN controller when the transition
time takes 15 s.
As for the first system, some main parameters are shown in Table 6 to estimate the diverse
performance of the two controllers.
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5159

Fig. 18. Output of the heat exchanger system with the PID-SMN controller (unit step reference).

Fig. 19. Input (signal control) of the heat exchanger system with the PID-SMN controller (unit step reference).

As shown on Table 6, the temperature in the PID-ZN case presents an overshoot ≈

18.6722% and a rise time ≈ 1.9667 s. But for the PID-SMN there is a reduction of the
overshoot to 0.5421% and the rise time ≈ 0.3242 s.
These results prove the efficiency of the PID-SMN controller.
Figs. 20–22 represent, respectively, the variable reference, the output of the PID-ZN and
the PID-SMN controllers in the variable reference case.
For the variable reference output, Figs. 21 and 22 show clearly the superiority of the
PID-SMN where the output follows accurately its variable reference.
Table 7 shows the different values of the MSE for the first system using the two controllers
5160 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 20. Graph of the variable reference.

Fig. 21. Output of the heat exchanger system with the PID-ZN controller (variable reference).

Table 7
MSE values for the second.

PID-ZN PID-SMN
Unit step reference 2.47 × 10–2 8.2 × 10–3
Variable reference 0.51 0.137

The Mean Squared Error values exposed in Table 7 show that the PID-SMN controller
provides a small overall error than that obtained with the PID-ZN controller for the two cases
of references constant and variable.
After introducing a disturbance in the temperature of 10% of its nominal value at the time
25 s, we obtained results for both controllers are shown in Fig. 23.
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5161

Fig. 22. Output of the heat exchanger system with the PID-SMN controller (variable reference).

Fig. 23. Output of the heat exchanger system using the two controllers with disturbance.

After adding the disturbance, the PID-SMN controller was able to track rapidly the desired
reference within 0.4 s. While for the PID-ZN controller, the time necessary to reach the
set point is approximately 09 s. These results demonstrate the advantage of the PID-SMN
controller for the disturbance rejection.
If we consider that there is no noise in output measurement or in the input control, we
will test the response of the two controllers in presence of noise at the input system as well
as at its output.
Let consider Fig. 24, a closed loop control with noise at input and output of the system
With:

d(k): the plant noise signal.

g(k): the measurement noise signal.
5162 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 24. Closed loop system with noise additive.

Fig. 25. Noisy input (signal control) of the single tank system with the PID-ZN controller.

Fig. 26. Noisy output of the single tank system with the PID-ZN controller.

un (k): the noised input control signal.

yn (k): the noised (measured) output signal.

We repeated our past simulations by adding the different noises d(k) and g(k) for the two
systems with the two controllers PID-ZN and PID-SMN.
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5163

Fig. 27. Noisy input (signal control) of the single tank system with the PID-SMN controller.

Fig. 28. Noisy output of the single tank system with the PID-SMN controller.

We used d(k) and g(k) as white Gaussian noises with variance = 0.1 and mean = 0. The
reference r(k) was taken as a unit step reference.
Figs. 25 and 26 show, respectively, the signal control and the plant output both with noise
using the PID-ZN controller for the first system .
Figs. 27 and 28 show, respectively, the signal control and the plant output both with noise
using the PID-SMN controller always for the first system.
Figs. 29 and 30 show, respectively, the signal control and the plant output both with noise
using the PID-ZN controller for the second system .
Figs. 31 and 32 show, respectively, the signal control and the plant output both with noise
using the PID-SMN controller for the second system.
As we can see from Figs 25 to 32, the two controllers cannot track the step reference due
to the presence of noise.
Thus, we need a filter block to estimate the true output from the noisy input and output.
5164 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 29. Noisy input (signal control) of the heat exchanger system with the PID-ZN controller.

Fig. 30. Noisy output of the heat exchanger system with the PID-ZN controller.

The schematic closed loop with the filter block is shown in Fig. 33.
yf (k): the filtered (estimated) output signal
In this work we used a Multi Layer Memory Neuron Filter (MLMNF) as a filter to estimate
the true plant output. The MLMN network consists of three layers:

➢ an input layer with two external inputs: un (k) and yn (k).

➢ a single hidden layer of N memory neurons.
➢ an output layer with a single output memory neuron yf (k).

The structure of the MLMNF is shown in Fig. 34.

Where:
w1ji (k) : the weight of the connection from the ith input in the input layer to the jth input
of the hidden layer at time k
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5165

Fig. 31. Noisy input (signal control) of the heat exchanger system with the PID-SMN controller.

Fig. 32. Noisy output of the heat exchanger system with the PID-SMN controller.

w12 j (k) : the weight of the connection from the jth input in the hidden layer to the single
input of the output layer at time k. (i ∈ {1, 2} ), (1 ≤ j ≤ N )

The design of the filtrage procedure was done in two steps.

In the first step, we trained the MLMNF offline (disconnected from the closed loop) so its
output yf (k) corresponded to the desired output. For our case we had to train the MLMNF
for each system and each controller, so we had four training procedure:

➢ For the first system:

• The desired output is the graph of the fluid level shown in Fig. 6 (with the PID-ZN
controller).
• The desired output is the graph of the fluid level shown in Fig. 8 (with the PID-SMN
controller).
5166 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 33. Closed loop noisy system with filter.

Fig. 34. Structure of the multi layer memory neuron filter.

Fig. 35. The measured output of the single tank system with the PID-ZN controller.

➢ For the second system:

• The desired output is the graph of the temperature shown in Fig. 16 (with the PID-ZN
controller).
• The desired output is the graph of the temperature shown in Fig. 18 (with the PID-SMN
controller).
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5167

Fig. 36. The filtered output of the single tank system with the PID-ZN controller.

Fig. 37. The control input of the single tank system with the PID-ZN controller.

Fig. 38. The measured output of the single tank system with the PID-SMN controller.
5168 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 39. The filtered output of the single tank system with the PID-SMN controller.

Fig. 40. The control input of the single tank system with the PID-SMN controller.

Fig. 41. The measured output of the heat exchanger system with the PID-ZN controller.
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5169

Fig. 42. The filtered output of the heat exchanger system with the PID-ZN controller.

Fig. 43. The control input of the heat exchanger system with the PID-ZN controller.

We used d(k) and g(k) as white Gaussian noises with variance = 0.1 and mean = 0
In the second step, we placed the trained MLMNF in the closed loop (Fig. 33) then we
visualize the filtered output. In order to evaluate the effectiveness of our approach we use
d(k) and g(k) as white Gaussian noises with variance = 0.5 and mean = 0.
For the simulation, we chose the number of neurons in the hidden layer as N = 6.
The activation function used for the memory neurons in the hidden layer is a unipolar
sigmoid function given by Eq. (32) with K = 1.
The activation function used for the single memory neurons in the output layer is a linear
function stated as:
f (x ) = x (36)
The training procedure was done using Eqs. (27) to (29).
The measured output, the filtered output and the control input for the first system using
the PID-ZN controller are shown, respectively, in Figs 35–37.
5170 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

Fig. 44. The measured output of the heat exchanger system with the PID-SMN controller.

Fig. 45. The filtered output of the heat exchanger system with the PID-SMN controller.

The measured output, the filtered output and the control input for the first system using
the PID-SMN controller are shown, respectively, in Figs 38–40.
The measured output, the filtered output and the control input for the second system using
the PID-ZN controller are shown, respectively, in Figs 41–43.
The measured output, the filtered output and the control input for the second system using
the PID-SMN controller are shown, respectively, in Figs 44–46.
As shown in Figs. 35 to 46, we can see that after using the MLMN filter, the control effect
was improved distinctly, both on the two systems, especially in the case of the PID-SMN.
An interesting remark is that with the MLMN filter the system output reaches the set point
faster than it does in the case without noise (Figs. 36, 39, 42 and 45). The reason is that
during the training of the MLMNF its output was very closed with the desired output.
The noise affects also the control input when we compare, for the first system, the right
sentence is "for the first system, Fig. 7 to Fig. 37 in the case of the PID-ZN and Fig. 9 to
 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172 5171

Fig. 46. The control input of the heat exchanger system with the PID-SMN controller.

Fig. 40 for the PID-SMN. For the second system the comparison was done between Figs.
17 and 43 for the PID-ZN controller, and Figs. 19 and 46 for the PID-SMN controller. This
effect is caused by adding noise d(k) to the control signal.

6. Conclusion

PID controller is widely used in several industrial applications to achieve a desired refer-
ence trajectory. The main objective in a PID controller is how to find optimal values for its
parameters (Kp , Ki and Kd ).
In this paper, we propose a method for the PID control based on a Single Memory Neuron
(SMN). The PID gains tuning procedure was depicted using the SMN and tested on two
classic industrial processes: a single tank and a boiler & heat exchanger.
According to the simulation results, the comparison made between the two PID controllers
illustrate the effectiveness of the PID-SMN in terms of changing of set point, good tracking
precision and overshoot cancellation.
In the case of noise rejection we used a filter block to reduce the effect of noise in the
closed loop. This block was formed by a Multi Layer Memory Neuron Filter (MLMNF) so
as to estimate the true plant output. The obtained results for the two examples are satisfactory
especially for the PID-SMN controller.
As future work, we plan to extend our work to various industrial systems and use other
neural architectures.

References

[1] M. Salazar, F. Méndez, PID control for a single-stage transcritical CO2 refrigeration cycle, Appl. Therm. Eng.
67 (2014) 429–438, doi:10.1016/j.applthermaleng.2014.03.052.
[2] P.V.M. Maalini, G. Prabhakar, S. Selvaperumal, Modelling and control of ball and beam system using PID
controller, in: Proceedings of the International Conference on Advanced Communication and Computing Tech-
nologies, 2016, pp. 322–326, doi:10.1109/ICACCCT.2016.7831655.
5172 S. Ladjouzi and S. Grouni / Journal of the Franklin Institute 357 (2020) 5143–5172

[3] S. Ahmadi, S. Abdi, M. Kakavand, Maximum power point tracking of a proton exchange membrane fuel cell
system using PSO-PID controller, Int. J. Hydrogen Energy 42 (2017) 20430–20443, doi:10.1016/j.ijhydene.
2017.06.208.
[4] L. Osorio, J. Mendes, R. Araujo, T. Matias, A comparison of adaptive PID methodologies controlling a DC
motor with a varying load, in: Proceedings of the 18th International Conference on Emerging Technologies and
Factory Automation, 2013, pp. 1–8, doi:10.1109/ETFA.2013.6648093.
[5] J.G. Ziegler, N.B. Nichols, Optimum settings for automatic controllers, J. Dyn. Syst. Meas. Control 115 (1993)
220–222, doi:10.1115/1.2899060.
[6] V. Bobál, P. Dostál, M. Sysel, Delta modification of self-tuning pole placement PID controllers, in: Proceedings
of the 12th IFAC Symposium on System Identification SYSID, Santa Barbara, 2000 Paper ThMD1-3, doi:10.
1016/s1474- 6670(17)39806- 3.
[7] M.A. Da Silva, F.A.C. Gomide, W.C. Amaral, A rule based procedure for self tuning PID controllers, in:
Proceedings of the 27th Conference on Decision and Control, 1988, pp. 1947–1951, doi:10.1109/CDC.1988.
194671.
[8] M.I. Solihin, L.F. Tack, M.L. Kean, Tuning of pid controller using particle swarm optimization (PSO), in:
Proceedings of the International Conference on Advanced Science, Engineering and Information Technology,
2011, pp. 458–461, doi:10.18517/ijaseit.1.4.93.
[9] Y. Miao, Y. Liu, Y. Chen, H. Jin, PID controller parameter tuning based on improved particle swarm optimization
algorithm, in: Proceedings of the International Conference on Electrical, Electronics and Mechatronics, 2015,
pp. 104–107, doi:10.2991/iceem-15.2015.26.
[10] A. Ayman, PID parameters optimization using genetic algorithm technique for electrohydraulic servo control
system, in: Proceedings of the International Conference on Intelligent Control and Automation, 2, 2011, pp. 69–
76, doi:10.4236/ica.2011.22008.
[11] E. Sameni, E. Ebrahimi, F. Koohneshin, M.R. D.astranj, in: Design of optimal PID control of dc motor using
genetic algorithm, 4, 2012, pp. 433–435, doi:10.7763/ijcte.2012.v4.500.
[12] M. Li, L. Wang, J. Liu, J. Ye, Method study on fuzzy-PID adaptive control of electric-hydraulic hitch system,
in: Proceedings of the International Conference on Advances in Materials, Machinery, Electronics, 2017, doi:10.
1063/1.4977350.
[13] K. Singh, P. Vasant, I. Elamvazuthi, R. Kannan, PID tuning of servo motor using bat algorithm, J. Procedia
Comput. Sci. 60 (2015) 1798–1808, doi:10.1016/j.procs.2015.08.290.
[14] Q. Sheng, Z. Xianyi, W. Changhong, X.Z. Gao, L. Zilong, Design and implementation of an adaptive PID
controller using single neuron learning algorithm, in: Proceedings of the International Conference on 4th World
on Intelligent Control and Automation, 2002, pp. 2279–2283, doi:10.1109/WCICA.2002.1021495.
[15] J. Minzhi, G. Chao, S. Xiaomin, Study on the application of the single neuron adaptive PID controller in
prestressed tension device, in: Proceedings of the 10th International Conference on Electronic Measurement &
Instruments, 4, 2011, pp. 16–19, doi:10.1109/ICEMI.2011.6037941.
[16] S. Zuo, Y. Song, L. Wang, Z. Zhou, Neuron-adaptive PID based speed control of SCSG wind turbine system,
Abst. Appl. Anal. (2014), doi:10.1155/ 2014/ 376259.
[17] L. Hang, L. Yu, S. Quan, L. Huang, Q. Chen, Y. Xiong, J. Quan, Design of voltage loop for three-phase PWM
rectifier based on single neuron adaptive PID control, in: Proceedings of the 32nd Youth Academic Annual
Conference of Chinese Association of Automation, 2017, pp. 171–175, doi:10.1109/YAC.2017.7967399.
[18] P. Poddar, K.P. Unnikrishnan, Memory neuron networks: a prolegomenon. Tech re. GMR-7493, General Motors
Research Laboratories, 1991.
[19] P.S. Sastry, G. Santharam, K.P. Unnikrishnan, Memory neuron networks for identification and control of dy-
namical systems, IEEE Trans. Neural Netw. 5 (2) (1994) 306–319, doi:10.1109/72.279193.
[20] S. Suresh, S.N. Omkar, V. Mani, Parallel implementation of memory neuron network for identification of
dynamical system, Adv. Vib. Eng. 2 (2) (2003) 1–9.
[21] K. Patan, Artificial Neural Networks For the Modeling and Fault Diagnosis of Technical Processes, Springer
Verlag, Berlin, 2008, doi:10.1007/978- 3- 540- 79872- 9.
[22] K. Ogata, Modern Control Engineering, 5th ed., Pearson, Upper Saddle River, 2010.
[23] S.G. Ramdani, Y. Soufi, Application of predictive controller tuning and a comparison study in terms of PID
controllers, Int. J. Hydrogen Energy 41 (29) (2016) 12454–12464, doi:10.1016/j.ijhydene.2016.05.102.
[24] J.S. Lahane, M.A. Khandekar, System identification and controller design for boiler and heat exchanger set-up,
Int. J. Adv. Res. Electr. Electron. Instrum. Eng. 3 (7) (2014) 10708–10714, doi:10.15662/ijareeie.2014.0307060.