International Journal of Power Electronics and Drive Systems (IJPEDS)

Vol. 14, No. 4, December 2023, pp. 2546~2556

ISSN: 2088-8694, DOI: 10.11591/ijpeds.v14.i4.pp2546-2556  2546

Adaptive control based neural network sliding mode approach

for two links robot

Siham Massou1, Ismail Boumhidi2

1
LABTIC Laboratory, Department of Information Systems and Communication, National School of Applied Science,
University Abdel Malek Essadi, Campus Ziaten, Tangier, Morocco
2
LISAC Laboratory, Department of Physics, Faculty of Sciences Dhar El Mehrez, University Sidi Mohamed Ben Abdellah,
Fes-Atlas, Morocco

Article Info ABSTRACT

Article history: In order to improve the control accuracy of the robot manipulator, the sliding
mode control combined with the adaptive neural network (ANNSMC) is
Received Nov 8, 2022 proposed. Sliding mode control (SMC) is a nonlinear control recognized for
Revised Mar 15, 2023 its efficiency, easy tuning and implementation, accuracy and robustness.
Accepted Mar 30, 2023 However, higher amplitude of chattering is produced due to the higher
switching gain to handle the large uncertainties. For the purpose of reducing
this gain, the uncertain parts of the system are estimated using neural
Keywords: network (NN) with on-line training using back propagation (BP) technique.
The results of the online interconnection weights between the input and the
Adaptive neural network hidden layers and between the hidden and the output layers are injected
Back propagation offline in order to improve the network performance in term of the
Lyapunov stability convergence speed. In order to reduce the response time caused by the
Nonlinear system online training, the obtained output and input weights are updated using the
Sliding mode control adaptive laws derived from the Lyapunov stability approach the proposed
Two links robots control ANNSMC has improved the convergence speed with 41.13% for the
first link and 40.15% for the second link comparing to NNSMC. The
simulation result illustrates the performance of the proposed approach by
using MATLAB and the control action suggested did not manifest any
chattering behavior.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Massou Siham
Department of information systems and communication, National School of Applied Science
University Abdel Malek Essadi
Campus Ziaten, Tangier 90000, Morocco
Email: [email protected]

1. INTRODUCTION
The motion control design for two links robot is an intense area of research that has fascinate
considerable focus, it’s a big challenge because of the nonlinearities of the system and the uncertainties of
their parameters. Furthermore, the right dynamic models are nearly hard to obtain because the system is
described by a nominal model with major uncertainties, to name a few: external disturbance, internal friction,
and payload parameter. Several strategies have been proposed to cope with parameter uncertainties including
neural network (NN) based controls [1]-[22], neural adaptive proportional-integral-derivative (PID) control
[11], the sliding mode control (SMC) [23]–[31], SMC and PID controller tuned by whale optimizer algorithm
(WOA) [29], nonlinear model predictive control tuned by NN [32]–[34] , the adaptive fuzzy control [35]–
[37], the particle swarm optimization (PSO) combined with NN and SMC [38], the combined PID and SMC
[30] were, the self-tuning fuzzy PID-nonsingular fast terminal sliding mode control for robust fault tolerant

Journal homepage: http://ijpeds.iaescore.com

Int J Pow Elec & Dri Syst ISSN: 2088-8694  2547

control of robot manipulators is studied. Sliding mode control (SMC) is an important robust control, it
provides a systematic approach to the problem of uncertainties, bounded external disturbance and
nonlinearities [5], [30], [32]–[37].
Unfortunately, there is an undesirable chattering in the control effort and it involves extremely high
control activity. It is worth mentioning that eliminating this chattering problem several solution are proposed
to perform properly as the boundary layer solution [23]–[25]. However, this method in only efficient for small
uncertainties. In case of large uncertainties, the unknown parts of the two-links robot are estimated by using a
neural network architecture, this approach leads to fewer uncertainties and a lower switching gain as result.
The back propagation technique [4], [38], [39] is used to train the weights of the neural network on-
line. The most important task is to ensure the system’s asymptotical stability and the tracking error’s
convergence to zero. In this study, the online input and output weights are injected and offline changed using
the adaptive rules obtained from the Lyapunov stability theorem [40], therefore the asymptotical stability of
the system is ensured, and also the speed of convergence is widely improved and illustrated in simulation
results. In addition, we used integral squared error ISE [41] to validate numerically the theoretical statements,
and that gives outstanding results.
This paper is organized as follows. The dynamic model of two links robot systems is introduced in
section 2. Section 3 covers the neural network design technique; outlines sliding mode control and illustrates
the adaptive laws for modifying the input and output weights obtained from the Lyapunov synthesis approach
and the stability proof is presented. Section 4 displays the simulation results obtained by using a two-link
robot manipulator. Finally, the conclusion is addressed in section 5.

2. SYSTEM MODEL OF THE TWO LINKS ROBOT

Consider the dynamic model of the two links robot written as (1) [7].

𝑀(𝑞)𝑞̈ + 𝑉𝑚 (𝑞, 𝑞̇ )𝑞̇ + 𝐹𝑞̇ + 𝑓𝑐 (𝑞̇ ) + 𝐺(𝑞) + 𝜏𝑑 = 𝜏 (1)

With 𝑞, 𝑞̇ , 𝑞̈ ∈ ℜ2 represent respectively the joints position, velocity and acceleration. 𝑀(𝑞) ∈ ℜ2×2 is a
symmetric, positive definite inertia matrix, 𝑉𝑚 (𝑞, 𝑞̇ ) ∈ ℜ2×2 is the centripetal and coriolis matrix, 𝐹 ∈ ℜ2
denotes the viscous friction coefficients, 𝑓𝑐 (𝑞̇ ) ∈ ℜ2 is the coefficients of coulomb friction, G (q)   2 is the
vector of gravitation, 𝜏𝑑 ∈ ℜ2 represents unmodeled dynamics and bounded unknown disturbances, 𝜏 ∈ ℜ2
is the input torque vector applied to the servo motor.
𝑉𝑚 𝑉𝑚12 𝑀 𝑀
𝑉𝑚 (𝑞, 𝑞̇ ) = [ 11 ]; 𝑀(𝑞) = [ 11 12 ]
𝑉𝑚21 𝑉𝑚22 𝑀21 𝑀22
𝑚2 𝑙2 4𝑚2 𝑙2 𝑚2 𝑙2 𝑚2 𝑙2 𝑚2 𝑙2
𝑀11 = + + 𝑚2 𝑙 2 𝑐𝑜𝑠( 𝑞2 ) ; 𝑀12 = 𝑀21 = + 𝑐𝑜𝑠( 𝑞2 ); 𝑀22 =
3 3 3 2 3

With 𝑞 = [𝑞1 𝑞2 ]𝑇 the positions; 𝑙 = 𝑙1 = 𝑙2 : the lengths; 𝑚1 , 𝑚2: the masses.

𝑚2 𝑙2 𝑚2 𝑙2
𝑉𝑚11 = − 𝑞̇ 2 𝑠𝑖𝑛( 𝑞2 ), 𝑉𝑚12 = − (𝑞̇ 2 + 𝑞̇ 1 ) 𝑠𝑖𝑛( 𝑞2 )
2 2

𝑚 𝑔𝑙 𝑚 𝑔𝑙
𝑚2 𝑙 2
𝑚2 𝑔𝑙 𝑐𝑜𝑠( 𝑞1 ) + 1 𝑐𝑜𝑠( 𝑞1 ) + 2 𝑐𝑜𝑠( 𝑞1 + 𝑞2 )
2 2
𝑉𝑚21 = 𝑞̇1 𝑠𝑖𝑛( 𝑞2 ) ,𝑉𝑚22 = 0 ;G (𝑞) = [ 𝑚2 𝑔𝑙
]
2
𝑐𝑜𝑠( 𝑞1 + 𝑞2 )
2

With g is the gravitation term.

𝑓𝑑1 𝑞̇ 1 + 𝑘1 𝑠𝑔𝑛( 𝑞̇ 1 )
𝐻 = Fq̇ + 𝑓𝑐 (𝑞̇ ) = [ ]
𝑓𝑑2 𝑞̇ 2 + 𝑘2 𝑠𝑔𝑛( 𝑞̇ 2 )

System model described by its state space is as (2).

𝑥̇1 = 𝑥2
𝑥̇ 2 = 𝑓1 (𝑥) + 𝑔11 (𝑥1 , 𝑥3 )𝑢1 + 𝑔12 (𝑥1 , 𝑥3 )𝑢2 + 𝜁1 (𝑥, 𝑡)
(2)
𝑥̇ 3 = 𝑥4
{𝑥̇ 4 = 𝑓2 (𝑥) + 𝑔21 (𝑥1 , 𝑥3 )𝑢1 + 𝑔22 (𝑥1 , 𝑥3 )𝑢2 + 𝜁2 (𝑥, 𝑡)

Adaptive control based neural network sliding mode approach for two links robot (Siham Massou)
2548  ISSN: 2088-8694

Where: 𝜁1 (𝑥, 𝑡) and 𝜁2 (𝑥, 𝑡) the unknown parts, 𝜏 = 𝑈 = [𝑢1 𝑢2 ]𝑇 = [𝜏1 𝜏2 ]𝑇 𝑞 = [𝑥1 𝑥3 ]𝑇 , 𝑞̇ =
𝑇 𝑔 (𝑥 , 𝑥 ) 𝑔12 (𝑥1 , 𝑥3 )
[𝑥2 𝑥 4 ]𝑇 ,𝑞̈ = [𝑥̇ 2 𝑥̇ 4 ] , 𝑥 = [𝑥1 𝑥2 𝑥3 𝑥4 ]𝑇 and 𝑀−1 = 𝑔(𝑥1 , 𝑥3 ) = ( 11 1 3 ) with
𝑔21 (𝑥1 , 𝑥3 ) 𝑔22 (𝑥1 , 𝑥3 )
𝑔11 (𝑥1 , 𝑥3 ) > 0𝑎𝑛𝑑𝑔22 (𝑥1 , 𝑥3 ) > 0 and 𝑓(𝑥) = [𝑓1 (𝑥) 𝑓2 (𝑥)] = −𝑀 {−𝑉𝑚 [𝑥2 𝑥4 ] − H-G}.
𝑇 −1 𝑇

3. ADAPTIVE NEURAL NETWORK SLIDING MODE CONTROL DESIGN

3.1. Controller design
This study examines the trajectory tracking problem of the robot manipulator mentioned previously.
The challenge of trajectory tracking control is to construct a robot controller based on any of 𝑥(0)𝜀𝑅𝑛 ,
𝑥(𝑡) − 𝑥𝑑 (𝑡) → 0as 𝑡 → ∞. Where 𝑥(𝑡) and 𝑥𝑑 (𝑡) are respectively the trajectory and the desired trajectory of
a robot. In the following, some variables are defined as (3).

𝜉(𝑥, 𝑡) ⥂⥂= [𝜉1 (𝑥, 𝑡) ⥂⥂⥂ 𝜉2 (𝑥, 𝑡)] (3)

The system's output tracking error can be described as (4).

𝑒 = 𝑞 − 𝑞𝑑 = [𝑥1 − 𝑥1𝑑 𝑥3 − 𝑥3𝑑 ]𝑇

𝑒̇ = 𝑞̇ − 𝑞̇ 𝑑 = [𝑥̇1 − 𝑥̇1𝑑 𝑥̇ 3 − 𝑥̇ 3𝑑 ]𝑇
𝑒̈ = 𝑞̈ − 𝑞̈ 𝑑 = [𝑥̈1 − 𝑥̈1𝑑 𝑥̈ 3 − 𝑥̈ 3𝑑 ]𝑇 (4)

With 𝑥1𝑑 and 𝑥4𝑑 are the desired output. 𝑟 = 2 is the relative degree of the system (2), and the sliding surface
is characterized as (5).

𝑆 = 𝑒̇ + 𝛽𝑒 (5)

𝛽 0
β is designed diagonal matrix as 𝛽 = ( 11 ). The selection of  must satisfy the following Hurwitz
0 𝛽22
polynomial as (6).

𝑠 (2) + 𝛽11 𝑠 (1) = 0

𝑠 (2) + 𝛽22 𝑠 (1) = 0 (6)

𝑑 𝑖 (𝑠)
With 𝑠 (𝑖) = . The sliding variable derivative is (7).
𝑑𝑡 𝑖

𝑥̇
𝑆̇ = 𝑓𝑛 + 𝜉(𝑥, 𝑡) + 𝑔𝑛 𝑢 − ( 2𝑑 ) + 𝛽𝑒̇ (7)
𝑥̇ 4𝑑

The following is the robust adaptive controller that was used (8).

𝑢 = 𝑢𝑛 + 𝑢𝑠 (8)

with 𝑢𝑠 = −𝑘𝑠𝑖𝑔𝑛(𝑆). The control law that respect (8) is (9),

𝑥̇
𝑢 = 𝑔𝑛−1 (𝑥) (−𝑓𝑛 (𝑥) + ( 2𝑑 ) − 𝛽𝑒̇ ) − 𝑘𝑠𝑖𝑔𝑛(𝑆) (9)
𝑥̇ 4𝑑

where 𝑠𝑖𝑔𝑛( )is the sign function given by:

1if𝑆 > 0
𝑠𝑖𝑔𝑛(𝑆) = {0if𝑆 = 0
−1if𝑆 < 0.

In order to compensate the uncertainties, the positive switching gain k is designed as 𝐵 < 𝑘. With 𝐵 the
upper bound of uncertainties given by ‖𝜉(𝑥, 𝑡)‖ < 𝐵. The boundary layer method may be utilized to remove
the chattering effect induced by the discontinuous control law. The control is now as (10),

Int J Pow Elec & Dri Syst, Vol. 14, No. 4, December 2023: 2546-2556
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2549

𝑥̇
𝑢 = 𝑔𝑛−1 (𝑥) (−𝑓𝑛 (𝑥) + ( 2𝑑 ) − 𝛽𝑒̇ ) − 𝑘𝑠𝑎𝑡(𝑆) (10)
𝑥̇ 4𝑑

with the saturation function is described by (11),

𝑆/𝛿if‖𝑆‖ < 𝛿
𝑠𝑎𝑡(𝑆) = { (11)
𝑠𝑔𝑛( 𝑆)otherwise

with 𝛿 is the boundary layer thickness.

This approach is suitable for systems with minor uncertainty. We suggest using NN to estimate the
uncertain terms of the system presented in (1) for large uncertain systems, in order to keep system
uncertainties manageable. Let denote the prediction of unknown nonlinear functions parts as: 𝜉̂ (𝑥, 𝑡) =
𝑇
[𝜉̂1 (𝑥, 𝑡) 𝜉̂2 (𝑥, 𝑡)] , with as (12):

𝜀(𝑥, 𝑡) = 𝜉(𝑥, 𝑡) − 𝜉̂ (𝑥, 𝑡) (12)

and: ‖𝜉(𝑥, 𝑡)‖ < 𝜀 ∗ where 𝜀 ∗ is the network prediction errors' upper bound.
- Theorem 1: Consider the robot manipulator stated in (1) in the context of significant uncertainty. If the
system control is programmed as (13):

𝑥̇
𝑢 = 𝑔𝑛−1 (𝑥) (− (𝑓𝑛 (𝑥) + 𝜉̂ (𝑥, 𝑡)) + ( 2𝑑 ) − 𝛽𝑒̇ − 𝑘𝑠𝑎𝑡(𝑆)) (13)
𝑥̇ 4𝑑

with 𝜀 ∗ < 𝑘 and 𝜉̂(𝑥, 𝑡) is predicted by the proposed off-line NN structure, then the trajectory tracking
errors converge to zero in limited time.
1
- Proof. Consider the potential Lyapunov function: 𝑉 = 𝑆 𝑇 𝑆 then 𝑉̇ = 𝑆 𝑇 𝑆̇ replacing the expression of 𝑆̇
2
given in (7) we have:

𝑥̇
𝑉̇ = 𝑆 𝑇 (𝑓𝑛 + 𝜉(𝑥, 𝑡) + 𝑔𝑛 𝑢 + ( 2𝑑 ) + 𝛾𝑒̈ + 𝛽𝑒̇ )
𝑥̇ 4𝑑

by substituting the expression of u provided in the theorem we get:

𝑉̇ = 𝑆 𝑇 (𝜉(𝑥, 𝑡) − 𝜉̂ (𝑥, 𝑡) − 𝑘𝑠𝑎𝑡(𝑆)) = 𝑆 𝑇 𝜀(𝑥, 𝑡) − 𝑘𝑆 𝑇 𝑠𝑎𝑡(𝑆) ≤ ‖𝑆 𝑇 ‖‖𝜀(𝑥, 𝑡)‖ − 𝑘𝑆 𝑇 𝑠𝑎𝑡(𝑆)

≤ ‖𝑆 𝑇 ‖𝜀 ∗ − 𝑘𝑆 𝑇 𝑠𝑎𝑡(𝑆)

by selectingε* < k, with k s a minor gain that is just responsible for compensating network faults forecast,
we have: if ‖𝑆‖ ≥ 𝛿 then we take 𝑠𝑎𝑡(𝑆) = 𝑠𝑖𝑔𝑛(𝑆) for any𝛿 > 0 and the function 𝑉̇ = (𝜀 ∗ − 𝑘)‖𝑆‖ <
𝑆
0. However, we take 𝑠𝑎𝑡(𝑆) = as continuous function in a boundary layer (a small 𝛿 -vicinity of the
𝛿
origin), as a result the system trajectories are restricted to a sliding mode manifold boundary layer. 𝑆 = 0.
Basing on (10), (8) and (4) we can write as (14).

𝑒̈ = 𝐴𝑒̇ + 𝐵𝜁̃(𝑥, 𝑡) − 𝐵𝑔𝑛 (𝑥)𝑢𝑠 (14)

The neural controller aim is to drive the system output to match the defined intended reference trajectory
as closely as feasible. The suggested neural network's design technique will be described in the following
paragraph.

3.2. Neural network representation

We investigate a NN with two layers of tunable weights in this study [8]. The architecture used is
illustrated in Figure 1 with one hidden layer, where x is the state input variables: the joints position, and the
output variables are the unknown parts given by (2): 𝑦1 = 𝜉̂1 (𝑥, 𝑡) 𝑎𝑛𝑑 𝑦2 = 𝜉̂2 (𝑥, 𝑡).
𝑦𝑘 = 𝑊𝑘𝑇 𝜎(𝑊𝑗𝑇 𝑥) 𝑘 = 1,2 Where 𝜎()is the activation function for hidden-layer considered as a
1 𝑇
sigmoid 𝜎(𝑠) = function given by 𝑊𝑘 = [𝑊𝑘1 𝑊𝑘2 𝑊𝑘𝑁 ]𝑇 and 𝑊𝑗 = [𝑊𝑗1 𝑊𝑗2 𝑊𝑗𝑁 ] are the connectivity
1+𝑒 −𝑠

Adaptive control based neural network sliding mode approach for two links robot (Siham Massou)
2550  ISSN: 2088-8694

weights between the hidden and output layers, as well as between the input and hidden layers. The actual
output 𝑦𝑑𝑘 (𝑥) which is the difference between the actual and nominal functions is (15).

𝑦𝑑𝑘 (𝑥) = 𝑦𝑘 (𝑥) + 𝜀(𝑥) (15)

Where the approximation error of NN is 𝜀(𝑥). The weights of the network are modified during the offline
implementation. The technique used is based on the BP algorithm, it’s a gradient-descent method and widely
used. The back propagation technique is used to determine the necessary modifications after randomly
selecting the network weights. The algorithm may be broken down into five steps:
i) Forward propagation of operating signal
ii) Calculate loss function
iii) Back propagation of error signal
iv) Calculate gradients
v) Weight updates
When the value of the error function becomes sufficiently minimal, the algorithm is terminated. It worth
mentioning that BP algorithm converges slowly. In order to improve the speed of convergence, the adaptive
principles obtained from the Lyapunov stability theorem is used to update online the obtained inputs and
outputs weights. The detailed steps of the proposed algorithm are illustrated in [40].

Figure 1. A multilayer neural network's design for predicting unknown components

3.3. Implementation of adaptation laws

The network weights are adjusted using the hybrid BP algorithm which takes important time to have
a result, to deal with this time response weights are adjusted offline. In this case the output of ANN with 5
hidden nodes can be presented by (16).

𝜁̃(𝑥, 𝑡) = 𝑊𝑘 ∗ 𝜎𝑘 (𝑥, 𝑊𝑗 ) (16)

The parameters 𝑊𝑘 and 𝑊𝑗 need to be adjusted further for the purpose to minimize approximation errors. The
adaptive rules for them were developed as (17) [23],

𝑊̇𝑘 = −𝜂1 𝜎𝑘𝑇 𝐵𝑇 𝑃𝑒̇

{ (17)
𝑊̇𝑗 = −𝜂2 𝑥 𝑇 𝐵𝑇 𝑃𝑒̇

where 𝜂1 and 𝜂2 are constants that are always positive. 𝑃 is the positive and symmetric definite matrix that
corresponds to:

𝐽 = −(𝐴𝑇 𝑃 + 𝑃𝐴) (18)

where the designer selected J as a asymmetric definite matrix.

Int J Pow Elec & Dri Syst, Vol. 14, No. 4, December 2023: 2546-2556
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2551

- Theorem 2: Suppose the nonlinear system described by (1). If the adaptive neural control rule mentioned
in (13) is used with the parameter adaptation laws (17), as a result, the tracking errors converge to zero
as 𝑡 → ∞ and all signals in the closed-loop system are limited.
- Proof: Take into consideration the possible Lyapunov function, which is:
1 1
𝑉̇ = 𝑒̇ 𝑇 𝑃𝑒̇ + 𝑊𝑘𝑇 𝑊𝑘
2 2𝜂1

The Lyapunov function's derivative is stated as:

1 1
𝑉̇ = (𝑒̈ 𝑇 𝑃𝑒̇ + 𝑒̇ 𝑇 𝑃𝑒̈ ) + 𝑊̇𝑘𝑇 𝑊𝑘
2 𝜂1

using in (14) we have:

1 1
𝑉̇ = ((𝐴𝑒̇ + 𝐵𝜁̃(𝑥, 𝑡) − 𝐵𝑔𝑛 (𝑥)𝑢𝑠 )𝑇 𝑃𝑒̇ + 𝑒̇ 𝑇 𝑃(𝐴𝑒̇ + 𝐵𝜁̃(𝑥, 𝑡) − 𝐵𝑔𝑛 (𝑥)𝑢𝑠 )) + 𝑊̇𝑘𝑇 𝑊𝑘
2 𝜂1

applying (16) and (17) we get:

1
𝑉̇ = ((𝐴𝑒̇ + 𝐵𝜁̃(𝑥, 𝑡) − 𝐵𝑔𝑛 (𝑥)𝑢𝑠 )𝑇 𝑃𝑒̇ + 𝑒̇ 𝑇 𝑃(𝐴𝑒̇ + 𝐵𝜁̃(𝑥, 𝑡) − 𝐵𝑔𝑛 (𝑥)𝑢𝑠 )) +
2
1
(−𝜂1 𝜎𝑘𝑇 𝐵𝑇 𝑃𝑒̇ )𝑇 𝑊𝑘
𝜂1

1 1 1
𝑉̇ = 𝑒̇ 𝑇 (𝐴𝑇 𝑃 + 𝑃𝐴)𝑒̇ + (𝜁̃𝑇 (𝑥, 𝑡)𝐵𝑇 𝑃𝑒̇ + 𝑒̇ 𝑇 𝑃𝐵𝜁̃(𝑥, 𝑡)) − (𝑢𝑠𝑇 𝐵𝑇 𝑔𝑛𝑇 (𝑥)𝑃𝑒̇ + 𝑒̇ 𝑇 𝑃𝐵𝑔𝑛 (𝑥)𝑢𝑠 ) −
2 2 2
𝑒̇ 𝑇 𝑃𝐵 𝜎𝑘 𝑊𝑘

𝑃 is symmetric, we get:
1 1
𝑉̇ = 𝑒̇ 𝑇 (𝐴𝑇 𝑃 + 𝑃𝐴)𝑒̇ − 𝑢𝑠𝑇 𝐵𝑇 𝑔𝑛𝑇 (𝑥)𝑃𝑒̇
2 2

1 1
𝑉̇ ≤ 𝑒̇ 𝑇 𝐽𝑒̇ − ‖𝑢𝑠𝑇 ‖𝐵𝑇 𝑔𝑛𝑇 (𝑥)𝑃‖𝑒̇ ‖ ≤ 0
2 2

Hence 𝑉̇ is negative semi definite, the signals 𝑒̇ and 𝑊𝑘 are all bounded. The parameters 𝑊𝑘 and 𝑊𝑗
described in (17) are adjusted using the projection algorithm as (19):
−𝜂1 𝜎 𝑇 𝐵 𝑇 𝑃𝑒 𝑖𝑓 ‖𝑊𝑘 ‖< 𝑀𝐵 𝑜𝑟
‖𝑊 ‖= 𝑀 𝑎𝑛𝑑
𝑖𝑓 { 𝑇 𝑘 𝑇 𝐵
̇
𝑊𝑘 = 𝑒 𝑃𝐵𝜎 𝑊𝑘 ≥ 0

𝑒 𝑇 𝑃𝐵𝜎𝑇 𝑊 ‖𝑊 ‖= 𝑀 𝑎𝑛𝑑
−𝜂 𝜎 𝑇 𝐵 𝑇 𝑃𝑒 − ‖𝑊 ‖2 𝑘 𝑊𝑘 𝑖𝑓 { 𝑇 𝑘 𝑇 𝐵
{ 1 𝑘 𝑒 𝑃𝐵𝜎 𝑊𝑘 < 0
−𝜂2 𝑥 𝑇 𝐵 𝑇 𝑃𝑒𝑖𝑓 ‖𝑊𝑗 ‖< 𝑀𝐵2 𝑜𝑟
‖𝑊𝑗 ‖= 𝑀𝐵2 𝑎𝑛𝑑
𝑖𝑓 { 𝑇
𝑒 𝑃𝐵𝑥 𝑇 𝑊𝑗 ≥ 0
𝑊̇𝑗 =
𝑒 𝑇 𝑃𝐵𝑥 𝑇 𝑊𝑗 ‖𝑊𝑗 ‖= 𝑀𝐵2 𝑎𝑛𝑑
−𝜂2 𝑥 𝑇 𝐵 𝑇 𝑃𝑒 − 2 𝑊𝑗 𝑖𝑓 {
{ ‖𝑊𝑗 ‖ 𝑒 𝑇 𝑃𝐵𝑥 𝑇 𝑊𝑗 < 0 (19)

The utilization of projection algorithm has a good performance on the tracking trajectory and also in the
control law illustrated in the next section.

3.4. Integral squared error (ISE)

In advance of talking about how to configure a controller, we must first define what makes a
satisfactory response. It is a great challenge to take up. In reality, there are several metrics that may be used
to compare the quality of regulated replies. The control measure described in this section is integral squared
error (ISE) which is used to illustrate the success of the designed control ANNSMC. According to [41], the
mathematical expression for ISE is (20),

ISE=∫ e2 dt (20)

where t is the simulation time, and e is the system's output tracking error.

Adaptive control based neural network sliding mode approach for two links robot (Siham Massou)
2552  ISSN: 2088-8694

4. SIMULATION RESULT
In this part, we put the suggested control strategy to the test on a two-link robot represented by the
𝜋
model (1). The aim of the control is to keep the system tracking the desired angle trajectory:𝑥1𝑑 = −
2
𝜋
𝑐𝑜𝑠( 0.5𝑡) and 𝑥3𝑑 = (𝜋/3) 𝑠𝑖𝑛( 0.5𝑡). The criteria are thought to be :m1 = m2 = 1; l1 = l2 = 1 the initial
3
𝑇
conditions are: 𝑞(0) = [𝑥1 (0) 𝑥3 (0)]𝑇 = 0 ,𝑞̇ (0) = [𝑥2 (0) 𝑥 4 (0)] = 0. The uncertainties under
consideration are vector random noise with a value of unity |𝜏𝑑 | ≤ 1. The following are the parameters
100 0
connected with the controller design: 𝐽 = ( ) 𝑀 = 1.35, 𝑀𝐵2 = 2,𝜂1 = 6, 𝜂2 = 1.5, A=[1 0;0 1],
0 100 𝐵
B=[1 0;0 1], P=[-50 0;0 -50], the number of hidden nodes is 5, the gain is 𝑘1 = 𝑘2 = 1.5.
The simulation results show position tracking for links 1 and 2, which are depicted respectively in
Figure 2 and Figure 3, where the reference signal is represented by the dashed line (red), the results using
adaptive law for weigh (ANNSMC) is represented by solid line (green), the results without adaptive law for
weigh (NNSMC) is represented by dashed line (blue) and the results using conventional SM is represented by
dot line (black). According to Figure 2, the gap between the ANNSMC outcomes position of link 1 and the
reference value is high at first, but the system remains stable, and the system outputs faster converge to the
intended trajectory than NNSMC, however the SMC result position don’t converge as the ANNSMC result
position. Similarly, the ANNSMC position of link 2 closely matches the reference signals and quickly than
NNSMC. The corresponding control torque signals given in Figure 4 and Figure 5 are smooth even in the
presence of significant uncertainty, there is no oscillatory behavior.

Figure 2. Angle response 𝑥1𝐴𝑁𝑁𝑆𝑀𝐶 using adaptive law, 𝑥1𝑁𝑁𝑆𝑀𝐶 angle response without adaptive law
𝑥1𝑆𝑀𝐶 angle response with conventional SMC and desired trajectory 𝑥1𝑑

Figure 3. Angle response 𝑥3𝐴𝑁𝑁𝑆𝑀𝐶 using adaptive law, 𝑥3𝑁𝑁𝑆𝑀𝐶 angle response without adaptive law 𝑥3𝑆𝑀
angle response using conventional SMC and desired trajectory 𝑥3𝑑

Int J Pow Elec & Dri Syst, Vol. 14, No. 4, December 2023: 2546-2556
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2553

Figure 4. Control 𝑢1 (input torque of join actuator 1)

Figure 5. Control 𝑢2 (input torque of join actuator 2)

In addition, the suggested adaptive neural control method's performance evaluation in term of the
speed of convergence is investigated in Figure 6 and Figure 7, that represent the position tracking for link 1
and link 2 where the reference signals is represented by dash red line and the results without adaptive law for
weight is represented by solid line in blue, by comparing the time response made by the results using this
approach, Figure 2 and Figure 3, and the results without it, Figure 6 and Figure 7, it’s obvious that the time
response is widely improved. Furthermore, the error response is also reduced by using the proposed control
and illustrated in Figure 8 and Figure 9.

Figure 6. Angle response 𝑥1𝑁𝑁𝑆𝑀𝐶 without adaptive law, desired trajectory 𝑥1𝑑

Adaptive control based neural network sliding mode approach for two links robot (Siham Massou)
2554  ISSN: 2088-8694

Figure 7. Angle response 𝑥3𝑁𝑁𝑆𝑀𝐶 without adaptive law, desired trajectory 𝑥3𝑑

Figure 8. Error response 𝑒1𝑁𝑁𝑆𝑀𝐶 , 𝑒1𝐴𝑁𝑁𝑆𝑀𝐶 and 𝑒1𝑆𝑀𝐶 of angle 𝑞1 , respectively without and with adaptive
law and SMC

Figure 9. Error response 𝑒2𝑁𝑁𝑆𝑀𝐶 , 𝑒2𝐴𝑁𝑁𝑆𝑀𝐶 and 𝑒2𝑆𝑀𝐶 of angle 𝑞2 , respectively without and with adaptive
law and SMC

The performance comparison of the ANNSMC control in Table 1 shows that it outperforms the
neural network sliding mode control (without the adaptive law derived based on Lyapunov theory). The
control system designed to minimize ISE tend to rapidly reduce huge mistakes, this leads to fast responses
This is seen in Figures 2, 3, 6, and 7.

Int J Pow Elec & Dri Syst, Vol. 14, No. 4, December 2023: 2546-2556
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2555

Table 1. Comparative ISE for ANNSMC and SMC control

NNSMC ANNSMC SMC
𝑒1 𝑒2 𝑒1𝑎 𝑒2𝑎 𝑒1𝑠 𝑒2𝑠
ISE 6.1398 80.2373 2.8940 47.241 9.6930 108.6011

5. CONCLUSION
The robust reference tracking problem for two-link robot manipulators was solved in this study. The
developed method is a hybrid of the sliding mode control SMC technology and the adaptive neural network.
ANN used online to extract weights adjusted firstly with BP algorithm. However, these weights are injected
to be adjusted secondly by the adaptation rule obtained from the Lyapunov stability theorem. The ability of
ANN to approximate fatly the uncertainties and external interference were assessed through comparison with
an traditional SMC methods. Simulations conducted on computers of a two-link robot manipulator confirm
outstanding results and show that the proposed algorithm can quickly approach the desired trajectory and
effectively suppresses the chattering. Furthermore, the measure of ISE is used in addition to verify the
theoretical statements achieved.

REFERENCES
[1] H. D. Patiño, R. Carelli, and B. R. Kuchen, “Neural networks for advanced control of robot manipulators,” IEEE Transactions on
Neural Networks, vol. 13, no. 2, pp. 343–354, 2002, doi: 10.1109/72.991420.
[2] M. A. Hussain and P. Y. Ho, “Adaptive sliding mode control with neural network based hybrid models,” Journal of Process
Control, vol. 14, no. 2, pp. 157–176, 2004, doi: 10.1016/S0959-1524(03)00031-3.
[3] P. X. Liu, M. J. Zuo, and M. Q. H. Meng, “Using neural network function approximation for optimal design of continuous-state parallel-
series systems,” Computers and Operations Research, vol. 30, no. 3, pp. 339–352, 2003, doi: 10.1016/S0305-0548(01)00100-9.
[4] S. Sefriti, J. Boumhidi, R. Naoual, and Y. Boumhidi, “Adaptive neural network sliding mode control for electrically-driven robot
manipulators,” Control Engineering and Applied Informatics, vol. 14, no. 4, pp. 27–32, 2012.
[5] Y. J. Liu, J. Li, S. Tong, and C. L. P. Chen, “Neural network control-based adaptive learning design for nonlinear systems with
full-state constraints,” IEEE Transactions on Neural Networks and Learning Systems, vol. 27, no. 7, pp. 1562–1571, 2016, doi:
10.1109/TNNLS.2015.2508926.
[6] Z. Liu, C. Chen, Y. Zhang, and C. L. P. Chen, “Adaptive neural control for dual-arm coordination of humanoid robot with
unknown nonlinearities in output mechanism,” IEEE Transactions on Cybernetics, vol. 45, no. 3, pp. 507–518, 2015, doi:
10.1109/TCYB.2014.2329931.
[7] S.Ge, T. H Lee, and C. J. Harris, Adaptive neural network control of robotic manipulators. Singapore: World Scientific, 1998.
[8] F. W Lewis, S. Jagannathan, and A. Yesildirak, Neural Network Control of Robot Manipulators and Nonlinear Systems. London:
Taylor and Francis, 1999.
[9] L. Fu, Neural Networks in Computer Intelligence. New York: McGraw-Hill, 1995.
[10] H. N. Rahimi, I. Howard, and L. Cui, “Neural impedance adaption for assistive human–robot interaction,” Neurocomputing, vol.
290, pp. 50–59, 2018, doi: 10.1016/j.neucom.2018.02.025.
[11] H. Rahimi Nohooji, “Constrained neural adaptive PID control for robot manipulators,” Journal of the Franklin Institute, vol. 357,
no. 7, pp. 3907–3923, May 2020, doi: 10.1016/j.jfranklin.2019.12.042.
[12] H. N. Rahimi, I. Howard, and L. Cui, “Neural adaptive tracking control for an uncertain robot manipulator with time-varying joint
space constraints,” Mechanical Systems and Signal Processing, vol. 112, pp. 44–60, 2018, doi: 10.1016/j.ymssp.2018.03.042.
[13] Y. Song, B. Zhang, and K. Zhao, “Indirect neuroadaptive control of unknown MIMO systems tracking uncertain target under
sensor failures,” Automatica, vol. 77, pp. 103–111, 2017, doi: 10.1016/j.automatica.2016.11.034.
[14] K. Yong, M. Chen, and Q. Wu, “Constrained adaptive neural control for a class of nonstrict-feedback nonlinear systems with
disturbances,” Neurocomputing, vol. 272, pp. 405–415, 2018, doi: 10.1016/j.neucom.2017.07.015.
[15] Z. Chen, Z. Li, and C. L. P. Chen, “Adaptive Neural Control of Uncertain MIMO Nonlinear Systems With State and Input Constraints,”
IEEE Transactions on Neural Networks and Learning Systems, vol. 28, no. 6, pp. 1318–1330, 2017, doi: 10.1109/TNNLS.2016.2538779.
[16] R. Yang, C. Yang, M. Chen, and A. S. Annamalai, “Discrete-time optimal adaptive RBFNN control for robot manipulators with
uncertain dynamics,” Neurocomputing, vol. 234, pp. 107–115, 2017, doi: 10.1016/j.neucom.2016.12.048.
[17] D. Zhang and B. Wei, “A review on model reference adaptive control of robotic manipulators,” Annual Reviews in Control, vol.
43, pp. 188–198, 2017, doi: 10.1016/j.arcontrol.2017.02.002.
[18] S. S. Ge and C. Wang, “Adaptive Neural Control of Uncertain MIMO Nonlinear Systems,” IEEE Transactions on Neural
Networks, vol. 15, no. 3, pp. 674–692, May 2004, doi: 10.1109/TNN.2004.826130.
[19] Z. L. Tang, S. S. Ge, K. P. Tee, and W. He, “Adaptive neural control for an uncertain robotic manipulator with joint space
constraints,” International Journal of Control, vol. 89, no. 7, pp. 1428–1446, 2016, doi: 10.1080/00207179.2015.1135351.
[20] Z. Yang, J. Peng, and Y. Liu, “Adaptive neural network force tracking impedance control for uncertain robotic manipulator based
on nonlinear velocity observer,” Neurocomputing, vol. 331, pp. 263–280, 2019, doi: 10.1016/j.neucom.2018.11.068.
[21] C. Cao, F. Wang, Q. Cao, H. Sun, W. Xu, and M. Cui, “Neural network–based terminal sliding mode applied to position/force
adaptive control for constrained robotic manipulators,” Advances in Mechanical Engineering, vol. 10, no. 6, 2018, doi:
10.1177/1687814018781288.
[22] S. Zhang, Y. Dong, Y. Ouyang, Z. Yin, and K. Peng, “Adaptive Neural Control for Robotic Manipulators with Output Constraints
and Uncertainties,” IEEE Transactions on Neural Networks and Learning Systems, vol. 29, no. 11, pp. 5554–5564, 2018, doi:
10.1109/TNNLS.2018.2803827.
[23] J.J.E.~Slotine, “Sliding controller design for non-linear systems,” International Journal of control, vol. 40, no. 2, pp. 421–434,
1984.
[24] V. I. Utkin, Sliding Modes in Control and Optimization, 1st ed. Berlin: Springer Berlin Heidelberg, 1992. doi: 10.1007/978-3-
642-84379-2.
[25] J. J. Slotine and S. S. Sastry, “Tracking Control of Non-Linear Systems Using Sliding Surfaces With Application To Robot
Manipulators.,” Proceedings of the American Control Conference, vol. 1, pp. 132–135, 1983, doi: 10.23919/acc.1983.4788090.

Adaptive control based neural network sliding mode approach for two links robot (Siham Massou)
2556  ISSN: 2088-8694

[26] H. Li, P. Shi, D. Yao, and L. Wu, “Observer-based adaptive sliding mode control for nonlinear Markovian jump systems,”
Automatica, vol. 64, pp. 133–142, 2016, doi: 10.1016/j.automatica.2015.11.007.
[27] X. Yin, L. Pan, and S. Cai, “Robust adaptive fuzzy sliding mode trajectory tracking control for serial robotic manipulators,”
Robotics and Computer-Integrated Manufacturing, vol. 72, 2021, doi: 10.1016/j.rcim.2019.101884.
[28] Z. Zhou, H. Ji, and Z. Zhu, “Online sequential fuzzy dropout extreme learning machine compensate for sliding-mode control
system errors of uncertain robot manipulator,” International Journal of Machine Learning and Cybernetics, vol. 13, no. 8, pp.
2171–2187, 2022, doi: 10.1007/s13042-022-01513-x.
[29] F. Loucif, S. Kechida, and A. Sebbagh, “Whale optimizer algorithm to tune PID controller for the trajectory tracking control of
robot manipulator,” Journal of the Brazilian Society of Mechanical Sciences and Engineering, vol. 42, no. 1, 2020, doi:
10.1007/s40430-019-2074-3.
[30] M. Van, X. P. Do, and M. Mavrovouniotis, “Self-tuning fuzzy PID-nonsingular fast terminal sliding mode control for robust fault
tolerant control of robot manipulators,” ISA Transactions, vol. 96, pp. 60–68, 2020, doi: 10.1016/j.isatra.2019.06.017.
[31] H. Shi, Y. Liang, and Z. Liu, “An approach to the dynamic modeling and sliding mode control of the constrained robot,”
Advances in Mechanical Engineering, vol. 9, no. 2, 2017, doi: 10.1177/1687814017690470.
[32] A. A. Adeniran and S. El Ferik, “Modeling and Identification of Nonlinear Systems: A Review of the Multimodel Approach—
Part 1,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 47, no. 7, pp. 1149–1159, Jul. 2017, doi:
10.1109/TSMC.2016.2560147.
[33] M. A. Ahmad, M. Z. M. Tumari and A. N. Kasruddin Nasir, “Composite Fuzzy Logic Control Approach to a Flexible Joint
Manipulator,” International Journal of Advanced Robotic Systems, vol. 10, 2012, doi: 10.5772/52562.
[34] L. Zhang, H. Liu, D. Tang, Y. Hou, and Y. Wang, “Adaptive Fixed-Time Fault-Tolerant Tracking Control and Its Application for
Robot Manipulators,” IEEE Transactions on Industrial Electronics, vol. 69, no. 3, pp. 2956–2966, 2022, doi:
10.1109/TIE.2021.3070494.
[35] J. Zhou and Q. Zhang, “Adaptive Fuzzy Control of Uncertain Robotic Manipulator,” Mathematical Problems in Engineering, vol.
2018, 2018, doi: 10.1155/2018/4703492.
[36] F. Qu, S. Tong, and Y. Li, “Observer-based adaptive fuzzy output constrained control for uncertain nonlinear multi-agent
systems,” Information Sciences, vol. 467, pp. 446–463, 2018, doi: 10.1016/j.ins.2018.08.025.
[37] A. Karamali Ravandi, E. Khanmirza, and K. Daneshjou, “Hybrid force/position control of robotic arms manipulating in uncertain
environments based on adaptive fuzzy sliding mode control,” Applied Soft Computing Journal, vol. 70, pp. 864–874, 2018, doi:
10.1016/j.asoc.2018.05.048.
[38] S. Massou and I. Boumhidi, “Optimal neural network-based sliding mode adaptive control for two-link robot,” International
Journal of Systems, Control and Communications, vol. 8, no. 3, p. 204, 2017, doi: 10.1504/ijscc.2017.10006534.
[39] D. E. Rumelhart and J. L. McClelland, “Learning Internal Representations by Error Propagation,” in Parallel Distributed
Processing: Explorations in the Microstructure of Cognition: Foundations, 1987, pp. 318–362.
[40] H. J. Rong, J. T. Wei, J. M. Bai, G. S. Zhao, and Y. Q. Liang, “Adaptive neural control for a class of MIMO nonlinear systems
with extreme learning machine,” Neurocomputing, vol. 149, no. Part A, pp. 405–414, 2015, doi: 10.1016/j.neucom.2014.01.066.
[41] B. Dhanasekaran, S. Siddhan, and J. Kaliannan, “Ant colony optimization technique tuned controller for frequency regulation of
single area nuclear power generating system,” Microprocessors and Microsystems, vol. 73, 2020, doi:
10.1016/j.micpro.2019.102953.

BIOGRAPHIES OF AUTHORS

Siham Massou is assistant professor of signals and information technology in

department of information system and communication in National School of Applied
Sciences, University Abdel Malek Essaadi. She received engineering degree from faculty of
sciences and technology, University Abdel Malek Essaadi, Tangier, and Ph.D. degree from
faculty Dhar El Mahraz, University Sidi Mohamed Ben Abdellah, Fes. Her research interests
focus in the areas of fuzzy logic, neural network, optimization, nonlinear control systems and
robust control. She can be contacted at email: [email protected].

Ismail Boumhidi is a full Professor of Electronics and automation in department

of physics in the Faculty of Sciences Dhar El Mahrez, University Sidi Mohamed ben
Abdellah, Fez Morocco. He received advanced graduate studies degree and his Ph.D. from
Faculty of Sciences Dhar El Mahrez. His research areas include control system synthesis,
fuzzy control, fuzzy systems, nonlinear control systems, wind turbines, continuous time
systems, reduced order systems. He can be contacted at email: [email protected].

Int J Pow Elec & Dri Syst, Vol. 14, No. 4, December 2023: 2546-2556