Improved Set-point Tracking Control of an Unmanned Aerodynamic

MIMO System Using Hybrid Neural Networks

David Mohammed Ezekiel 1, Ravi Samikannu 2, MatsebeOduetse3, Karthikeyan

Rajagopal4
1,2
Department of Electrical, Computer and Telecoms Engineering,
Botswana International University of Science & Technology (BIUST), Private bag 16,
Palapye, Botswana.
3
Department of Mechanical, Energy and Industrial Engineering,
Botswana International University of Science & Technology (BIUST), Private bag 16,
Palapye, Botswana.
4
Centre for Nonlinear systems, Chennai Institute of Technology, Chennai, India
Department of Electronics and Communications Engineering, University Centre for
Research andDevelopment, Chandigarh University, Mohali, 140 413, Punjab, India.
*karthikeyan.rajagopal@citchennai.net
Author Correspondence e-mail: karthikeyan.rajagopal@citchennai.net

Abstract: -Artificial neural networks (ANN), an Artificial Intelligence (AI) technique,are both bio-
inspired and nature-inspired models that mimic the operations of the human brain and the
central nervous system that is capable of learning. This paper is based on a system that
optimizes the performance of an uncertainunmanned nonlinear Multi-Input Multi-Output (MIMO)
aerodynamic plant called Twin Rotor MIMO System (TRMS). The pitch and yaw angles which
are challenging to control and optimize in practice, are being used as the input to the Nonlinear
Auto-Regressive with eXogenous (NARX) model, and eventually trained. The training features
use the Matlab Deep Learning Toolbox. The NARX structure has its core in the neural networks
architecture. Data is collected from the TRMS which trains the network. ANN asa
Hybridintelligent control strategy of ANN in combination with Pattern Search and Genetic
Algorithm, is then utilized to optimize the parameters of the neural networks. At the end it was
validated, tested andthe optimized system run in simulation and compared with other intelligent
and conventional controllers.The proposed controller outperformed them, providing a fast-
tracking control, stable, and optimal performance that satisfactorily met all our design
requirements. It was validated, tested, and the optimized system ran in simulation

Key Words:ANN,NN, NARX, TRMS,MIMO, Aerodynamic, Unmanned, Helicopter Model

1. Introduction
The modelling, optimization and control of rigid bodies and flexible structures/systems [1, 2, 3]
(such as plates, shells, beams, frames, etc.) are increasingly gaining a considerable attention
from researchers globally [4, 5, 6]. These bodies and structures are highly essential
manufacturing elements in electro-mechanical, civil, marine and aerospace engineering. In this

1
paper, the application of Feedforward Neural Networks (NN) is applied to the beam of a
nonlinear uncertain system called, the TRMS. It is a highly nonlinear, high-order, complex
system[3, 7, 8, 9, 10], the nonlinearities and complexities emanate from the cross-couplings
between the twin-rotors. These pose as a serious challenge to effectively model, control and
optimize. The modelling, control and optimization of the TRMS can be carried out in either the
model-based/model-driven or data-driven approaches. The data-driven (i.e. black-box
modelling) approaches for which this paper is based, requires some input/output dataset[11]
obtained through system simulation, offline and/or online. This dataset identifies the system
identified through System Identification (SI) techniques. The drawback of SI is that it has
demonstrated a computational inadequacy with nonlinear systems, but much less
uncomplicated with linear systems [1]. Despite of this, it is still an indispensable design strategy,
especially if the system can be linearized about some equilibrium points.
To use SI methodologies,require training of the network used in the design process, which can
be parametric or non-parametric. The non-parametric SI (which is of interest here) involves the
use of Artificial Intelligence, such as ANN[12, 13] or an Adaptive Neuro-Fuzzy Inference System
(ANFIS) [14]. For brevity, ANN is simply referred to as NN. However, they suffer from being
caught in a local minima and a very slow convergence resulting from system complexities of
nonlinear systems. To solve these problems metaheuristic methods are employed for faster
convergence optimization. With this, the solution being trapped in a local minimum/minima is
prevented, thus guaranteeing an accurate solution [6]. Several of these metaheuristics
approaches have been successfully used to train ANNs in engineering and scientific
applications. Some of these methods include Symbiotic Organisms Search (SOS) scheme
employed to train a feed forward NN to solve a classification problem [15], Harmony Search,
Simulated Annealing and Differential Evolution (DE)[16], a hybrid algorithm composed of
Particle Swarm Optimization (PSO) used for optimization of a Convolutional NN to also solve a
classification problem[17], ANN models trained for stock market price predictions/forecasts [18],
the newly developed Stochastic Fractal Search Algorithm SFS by Salimi[19] and used to train
ANNs [20, 21].Also, the successful applications of ANNs in estimating the nonlinear dynamics of
dynamical systems have been reported for kinematics in[22, 23], for dynamics in [24, 25]and
forcontrol in [22, 26].
Due to the extreme and profoundly serious (i.e. massive) nonlinearities the control of Unmanned
Aerial Vehicles (UAVs), of which class the TRMS falls, is a challenging one [27].Rahideh et al.
proposed a Model Inversion Control law to control a 1-DOF pitch model of the TRMS using
ANN[28]. The ANN was used adaptively to tune the system model. The obtained control lawwas
consequently used to achieve control and tracking. The scheme used an adaptive learning for

2
compensation of the errors due to modelling, thereby identifying the system. The use of the
NARX neural networks based on a Back Propagation (BP) algorithm for network training was
proposed by Tijani et al in [29] to solve a multi-objective optimization problem. The algorithm
used a multi (or many)-objective DE algorithms to identify and control the nonlinear TRMS using
real-time data from experiments.
This study uses Deep Neural Networks architecture using NARX Shallow NN for the ANN
trainingto identify, optimize and control the nonlinear TRMS lab-scaled helicopter.The NARX
model is used here identifies/ capture the nonlinear dynamics of the nonlinear TRMS testbed.
The NARX network is a feedforward neural network composed of 2 layers,with a sigmoid
transfer or activation function in the hidden layer and a linear transfer function in the outer layer.
The network also uses tapped delay lines to store previous values of the input and output
sequences. Here, the outputs are fed back into the inputs through the delay lines, since if is

the output, then

The learning rulesalgorithms employed mostly are theBayesian Regularization (trainbr),

Levenberg-Marquardt (trainlm), and the Scaled Conjugate Gradient (trainscg). The first two
algorithms are based on the Jacobian calculations while the last training method is based on the
gradient calculations. In this paper, the 2 inputs (elevation and azimuth) and the outputs/target
vectors (pitch and yaw) are composed of 181 datasets each, at random, roughly divided into
70% for the training phase, 15% for validation phase and 15% for the test phase to generalize
the network. 2 different set of numbers of hidden neurons of 10 and 1000 were used, with a
tapped delay line of 4.
The paper is organized as follows: Section 2 presents the experimental arrangements as well as
the governing equations of motion; section 3 presents the NN architecture and theoretical
background; section 4gives the training results and final simulations.

2. Setup of Experiment
The figure 1 presented below is the experimental setup:

3
 (a) (b)
Fig. 1. (a) The Real-world experimental setup at theBotswana International University of Science and Technology
(BIUST), with the beam inclined at60 to the horizontal at rest and showing I/O communication cables. (b) Schematic
[30][31] [32]

Fig. 2. The electrical circuit connection of the DC motor of the TRMS

The model of the DC motors is given in [33, 34] as:

where,

4
 is the control voltage input to either the vertical or the horizontal channel,

are respectively the e.m.f, current, resistance and inductance in the armature of the main/tail

motor; are constants; is the flux linkages;

are the magnetic torque and load

torque respectively in the main/tail motor; are the moments of inertias and viscous friction

damping coefficients of the rotors in the main/tail motors.

Fig. 3. Planar (vertical plane) representation of the TRMS, showing the gravity and propulsive forces[28] [35]

2.1 Governing Equations of Motion of the TRMS

As a dynamical system with rotational motion, Newtonian mechanics for rotational dynamics or
Lagrangian mechanics may be used to develop the dynamic equations of motion. Basing on
Newtonian mechanics for rotational dynamics, the dynamic equations of motion (using Newton’s
laws of motion for rotational dynamics) [36, 37, 38]of the TRMS, representing the flight in the
pitch (or vertical) plane and the yaw (or horizontal) plane are respectively given by:

5
where, are the angular/rotational velocities of the rotors for the pitch and yaw

orientations respectively, given by:

where A, B, C, are constants, and are given by:

[39, 40, 41, 42, 43, 44, 45].

The aerodynamic propulsive forces, are produced by the main/tail rotors in the

vertical/horizontal planes respectively and are given by:

are the sum of moments in the pitch and yaw planes and are the sum of the

moments of inertias in the vertical and horizontal planes respectively, is the frictional

torque developed, for the pitch plane; are the Composite mass of Main/Tail rotor

plusMain/Tail DC motor, are the Beam’s mass of Main/Tail portion, is the mass for

the counter weight, are the Main/Tailshield mass, is the counterbalance beam

length, is the Distance from the pivot joint to the counter-balance-or counter weight, are

the Beam’s Main/Tail portion length, are the angles for the pitch and yaw respectively.

3. The Neural Network Structure

6
The NN structure presents the layers’arrangements and the no. of neurons in each layer. our
plant's two planes/axes (i.e. the pitch and yaw) are presented with a Feedforward NN structure.

3.1 The Feedforward Neural Network

(a) (b)
Fig. 4. The NARX shallow NN for the (a) pitch (b) yaw angles[28] [46]

weights w weights w
z z
p p

weights w weights w
z z
p p

weights w weights w
z z
p p

weights w weights w
z z
p p

weights w weights w
z z
p p

Mux
Mux weights w
weights w z
z p
p

weights w
weights w z
z p
p

weights w
weights w z
z p
p

weights w
weights w z
z p
p

weights w
weights w z
z p
p

(a) (b)

7
 Process Input 1 Layer 1 a{1}

Layer 2 a{2}

a{1} Layer 3

a{2} Process Output 1

(c)

0 W

p{1} Delays 1 IW{1,1} +

b a{1}

b{1}

(d)
Fig. 5 The actual NARX shallow (1-layered) NN controllers for the (a) pitch (b) yaw, each having 10 neurons with
network weights attached as seen above (c) the configuration showing the single-input single-output network (d) the
feedforward configuration for both pitch and yaw angles.

For the First NARX shallow Neural Networks(fig. 4a)having one hidden layer with 10 hidden
neurons, since 4 input tapped delays are employed for this research work, then the pitch angle
of the beam is the inputs to the NNat the current time, and delayed 1, 2, 3 and 4 samples, i.e.

In a similar vein, the second NARX NN model is

for the yaw angle with inputs at the current time instant, and those delay by 1,2 and 3 samples,
i.e. . For both models, the output is

expressed as:

8
where, the input vector, the network weights, the biases matrices the activation

function[47].

3.2 Training the Network

The neural network must be trained for biases and weights adjustments to obtain the optimum
system parameters. This training could be carried out in offline or online scenarios, but here the
offline training is adopted. The network weights and biases are updated/adjusted with the main
aim of minimizing the tracking error response of the plant (TRMS). This adjustment is done
according to the following formulations [47]:

where represent the network learning rates, and with ensures tracking the error of

the system and the neural networks weights are bounded uniformly. is given by:

9
where, is the Lyapunov candidate solution for the nonlinear equation:

where must be a positive-definite matrix (i.e. are matrices for the tracking error (

), given by:

3.3 Validation of the Model

The tools used for validating the nonlinear model of the TRMS include, One Step-Ahead (OSA)
Prediction, Mean Squared Error (MSE), Correlations Tests (Autocorrelation and cross-
correlation functions) and Normalization.

3.4 OSA Prediction

The OSA prediction for the NARX network or model occurs when the feedback loop of the
network is open. The NARX network/model thus, predicts the next value of the output

from the previous ones of and the input . For a Multi-Step-Ahead prediction, the

feedback loop must necessarily be closed. The OSA is a measure of the accuracy in modelling.
It is given by:

where, is a nonlinear function approximator, represent the input/output respectively,

is the prediction value and represents pitch or yaw. The OSA is an extension of the

NARX model, where the NARX model is given by:

10
The residual prediction error is given by:

3.5 Mean Squared Error (MSE)

The MSE is a validation test and it gives an average of thesum of mean squares of the

difference between the actual and predicted outputs of the TRMS system. The

inputs and the optimized parameters of the network generate the outputs. MSE is given by:

where, no. of input/output samples.

The MSE algorithm helps in adjusting the network weights and biases minimizing the MSE.
Fortunately,the MSE performance indicator is a quadratic function that will either have a global
minimum, or a weak minimum or no minimum determined by the nature of the input vectors.
Hence a unique solution may or may not exist. The MSE algorithm or Widrow-Hoff learning
algorithm [48]approximates MSE based on the steepest descent algorithm at each iteration.
Taking the partial derivatives of the MSE w.r.t weights & biases at the th iteration, we get:

and

Taking the partial derivative w.r.t error (

or

where, th element of the input vector at the th iteration.

Further simplification yields:

11
3.6 The Correlations Tests
These tests are statistical tests for bivariate dataset composed of the autocorrelation and cross-
correlation functions [33], given by:

where, are the autocorrelation and cross-correlation functions between

, and is the error of prediction sequence.

3.7 Normalization
In practice, the correlations computed are normalized to ensure all the values fall within a given
bandwidth and/or range. The normalized correlation function between 2 sequences

is given by:

where the normalized correlation function must lie within a bandwidth/range of

regardless of or notwithstanding the signal strengths.

4. Results and Discussion

4.1 The Nonlinear TRMS Modelling

The method employed in this research involves using a time-domain closed-loop control
approach. Here, the NARX shallow Neural Networks modelling is used as a compensator to
train the network and provide the closed-loop control signal.
The input signals to the TRMS and the NN are a uniform random signal, exciting both the pitch
and yaw subsections of the plant/system. The Simulink model of the TRMS and the signals are
given below.

12
 (a) (b) (c)

(d) (e)
Fig. 6. Simulation of the TRMS (a) Pitch Random signals for the input (b) Its Log/Magnitude (c) Yaw Random signals
for the input (d) Its Log/Magnitude (e) Simulink model of the nonlinear TRMS

4.2 Shallow NN Modelling

4.2.1Training with Levenberg-Marquardt (LM)
Response of Output Element 1 for Time-Series 1
Response of Output Element 1 for Time-Series 1 0.4
0.4
Training Targets
Training Targets 0.3
0.3 Training Outputs
Training Outputs
Validation Targets
Validation Targets 0.2
0.2 Validation Outputs
Validation Outputs
Output and Target
Output and Target

Test Targets Autocorrelation of Error 1

Test Targets 0.1 10-7
0.1 Test Outputs
Test Outputs
Errors Errors Correlations
0 0
Response 8 Zero Correlation
Response
Confidence Limit
-0.1 -0.1

-0.2 6
-0.2
Correlation

-0.3 -0.3
4
-0.4 -0.4

-0.5 -0.5
2
0.01 0.02
Targets - Outputs Targets - Outputs

0 0.01
Error
Error

0
-0.01 0

-0.02 -0.01 -2
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 -20 -15 -10 -5 0 5 10 15 20
Time Time Lag

(a) (b) (c)

10-7 Autocorrelation of Error 1 Correlation between Input 1 and Error 1 = Target 1 - Output 1
10-5
4
Correlations
Correlations
20 Zero Correlation Zero Correlation
3
Confidence Limit Confidence Limit
2
15
1
Correlation

Correlation

0
10
-1

5 -2

-3

0 -4

-5
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Lag Lag
(d) (e)

13
 Best Validation Performance is 2.7204e-05 at epoch 176 Best Validation Performance is 4.2084e-05 at epoch 113 Best Validation Performance is 3.1104e-05 at epoch 141
Train Train Train
Validation Validation Validation
Test
100 100
Test
100
Test
Best Best
Mean Squared Error (mse)

Mean Squared Error (mse)

Best

Mean Squared Error (mse)

10-2 10-2 10-2

10-4 10-4 10-4

0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 0 20 40 60 80 100 120 140

182 Epochs 119 Epochs 147 Epochs

(g) (h) (i)

Best Validation Performance is 0.00012509 at epoch 122
101 Best Validation Performance is 0.00011389 at epoch 91 Best Validation Performance is 6.876e-05 at epoch 100
Train 100
Validation Train Train
Test Validation Validation
100 Best Test Test
Mean Squared Error (mse)

10-1 Best 100

Mean Squared Error (mse)

Best

Mean Squared Error (mse)

10-1
10-2

10-2
10-2
10-3

10-3 10-4
10-4

-4
10 10-5
0 20 40 60 80 100 120 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 100
128 Epochs 97 Epochs 106 Epochs

(j) (k) (l)

(m)
Fig. 7 Levenberg-Marquardt training with (a) 10 (b) 1000 hidden neurons; Autocorrelation Error for (c) 10 (d) 1000
hidden neurons; Input-Error Corelation for (e) 10 (f) 1000 hidden neurons; Best validation performance for (g)-(i) 10 (j)
– ( l) 1000 hidden neurons (m) Matlab nntool NARX NN GUI programming and execution.

14
 (a) (b)
Fig. 8 Control system designed to implement step input signal tracking control using (a) conventional PID controller
(b) neural networks controllers generated using the ‘gensim’ command

(a) (b)
Fig. 9. The tracking control for the elevation (pitch)-red and azimuth (yaw)-blue trackings of the TRMS using PID controllers for a
simulation time of (a) 50 (b) 100 seconds

(a) (b)
Fig. 10. The final acutal output shown for the elevation (pitch)-red and azimuth (yaw)-blue trackings of the TRMS
using ANN controllers realized using the “getsim”command for a simulation time of (a) 50 (b) 100 seconds

15
 (a) (b)
Fig. 11 Improved step input tracking control for the pitch and yaw angles using neural networks controllers optimized
using Pattern Search + Latin Hyperbole for a simulation time of (a) 10 (b) 100 seconds

(a) (b)
Fig. 12 Final Improved step input tracking control for the pitch and yaw angles using neural networks controllers
optimized using Pattern Search + GA for a simulation time of (a) 10 (b) 100 seconds

(a1) (a2)

16
 (b1) (b2)
Fig. 13 Comparisons of the different controllers employed above for a step input tracking response for the (a) pitch
and (b) yaw angles, for a simulation time of (a) 50 (b) 100 seconds
Table 1 Quantitative comparison of time-domain specifications between the proposed ANN + Pattern Search + GA
controller, and the 3 other controllers design strategies for pitch & yaw angles

Controller Method Horizontal plane Vertical plane

PID 0.24 100 0.04 0.48 25.00 0.00

ANN 1.18 14.5 0.02 1.32 56 0.00
ANN + PS + LH 1.55 9.9 0.99 4 37 0.11
ANN + PS + GA 3.85 14.42 0.03 2.79 5.14 0.002
Note: PS Pattern Search; LH Latin Hyperbole;

Table 2Quantitative comparison of performance indices between the proposed ANN + Pattern Search + GA
controller, and 3 other controllers design strategiesfor pitch & yaw angles

Controller Method Horizontal plane Vertical plane

ISE RMSE ISE RMSE

PID 1.402 0.087 1.135. 0.001
ANN 1.862 0.016 1.556 0.223
ANN+PS+LH 1.835 0.002 2.243 0.123
ANN+PS+GA 2.034 0.032 72.36 0.758
Note: ISE Integral Squared Error; RMSE Root Mean Squared error

The TRMS plant is a highly unceertain and highly nonlinear plant with high frequency
oscillations, particularly with the pitch angle. This can pose as a serious control challenge in
removing the rippling oscilations. This is shown in the scope of the designed control system
using the conventional PID controllers (fig. 9). For the real system, this evidently would affect
the plant operation in the plant's inability to settle within acceptable limits specified for effective

17
control. Theneed therfore for improved tracking becomes indespensable for such safety-critical
system. This informed theuse of the neural networks (NN) controllers (fig. 10). The NN
controllers were able to eliminate the undesirable oscillations or ripples in the final outputsforthe
yaw angle at the first glance (fig. 10), though with a large overshoot. When the network weights
of the ANN controllers were optimized using intelligent schemes of Pattern Search with Latin
Hyperbole (ANN + PS + LH) and Pattern Search with Genetic Algorithm (ANN + PS + GA), the
GA-based ANN controllers completely eliminated the ripples for both angles and brought
thesystem within acceptable bandwidths of control (fig. 12) and fast tracking simulation time of
10 seconds. The ANN + PS + LH also performed well with fast tracking response but with a very
high overshoot for the yaw angle and a large steady state error (fig.11). The best

performing algorithm as evidenced in table 2 above is the proposed ANN + PS + GA, with best
settling time for the pitch angle and 2nd best settling time for the yaw angle, as well

as good rise times, i.e. very fast responses and good steady-state errors for both the pitch and
yaw angles.

Conclusion
Results for the NARX Feedforward NN methodology to the modelling of the nonlinear TRMS
has been presented in this report. It has also been shown that different solutions are obtained
for every NN training undergone. This is due to the differing initial weights conditions and biases
and the arbitrary (i.e. random) division of the dataset into training, validation and testing in the
giving ratios of 0.7, 0.15. 0.15.To ensure accuracy of the modelling results retraining should be
performed several times.The final output results for the ANN and Pattern Search with Genetic
Algorithm (ANN + PS + GA) shows a satisfactory control for the optimized performance of the
nonlinear plant, outperforming three other controllers employed as proven by the statistical and
graphical results presented. The proposed controller met all our design requirements of within
5% of settling time, below 1% of overshoot, and excellent rise times, i.e. very fast

(aggressive) responses for both the pitch and yaw angles. and no steady-state error for the
pitch angle a negligible steady-state error for the yaw angle. Also of note was the minimum
control energyused by the controller in achieving these objectives.

DECLARATIONS
Ethics approval and consent to participate
The study involves no animal or human being, ethics approval and consent to participate is not
applicable.
Consent for publication

18
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Availability of data and materials
The data used to support the findings of this paper are included within the manuscript.
Funding
This work is partially funded by Centre for Nonlinear Systems, Chennai Institute of Technology,
Chennai, Tamilnadu, India vide funding number CIT/CNS/2023/RD/016
Authors' contributions
The contributions of the team members are as follows: DME played a role in the writing process,
which encompassed reviewing and editing. RS provided valuable supervision and contributed to
conceptualization and methodology. MO's involvement included visualization and investigation.
KR took charge of writing the initial draft, validation, and played a role in original draft
preparation.
Acknowledgements

References
[1] S.M. Ahmad, A.J. Chipperfield and M.O. Tokhi, "Dynamic Modelling and Control of a 2 DOF Twin Rotor Multi-
Input Multi-Output System," in 26th Ann. Conf. IEEE Indu. Electronics Society ((2000 IEEE Int. Conf. on Indu.
Electronics, Control & Instr.), Nagoya, Japan, 2000.
[2] S.M. Ahmad, A.J. Chipperfield and M.O. Tokhi, "Dynamic Modelling and Optimal Control of a Twin Rotor MIMO
System," in Proc. IEEE 2000 Nat'l Aerospace & Electronics Conf. NAECON 2000., Dayton, OH, USA, 2000.
[3] M. Mones and T. Diaa-Eldeen, "Experimental Nonlinear Identification of a Lab-Scale Helicopter System using
MLP Neural Network," in 13th Int. Comp. Eng'g Conf. (ICENCO), Cairo, Egypt, 2017.
[4] Tavakolpour AR, Mailah M, Mat Darus IZ, Tokhi O, "Self-learning active vibration control of a flexible plate
structure with piezoelectric actuator," Simulation Modelling Practice and Theory, vol. 18, pp. 516-532, 2010.
[5] Nasir ANK, Tokhi MO, "A novel hybrid bacteria-chemotaxis spiral-dynamic algorithm with application to
modelling of flexible systems," Engineering Applications of Artificial Intelligence, vol. 33, pp. 31-46, 2014.
[6] T. A. Z. Rahman TAZ, As’arry A, Abdul Jalil NA, Kamil R, "Dynamic Modelling of a Flexible Beam Structure
Using Feedforward Neural Networks for Active Vibration Control," International Journal of Automotive and
Mechanical Engineering, vol. 16, no. 1, pp. 6263-6280, March 2019.
[7] S.F. Toha and M.O. Tokhi, "Dynamic Nonlinear Inverse-Model-Based Control of a Twin Rotor System using
Adaptive Neuro-Fuzzy Inference System," in 3rd UKSim Euro. Symp. Comp. Modeling & Sim., Athens, Greece,
2009.
[8] M.S. Alam, F.M. Aldebrez and M.O. Tokhi, "Adaptive Command Shaping using Genetic Algorithms for Vibration
Control," in Proc. IEEE SMC UK-RI 3rd W/shop on Intel. Cybernatic Sys., Ulster, UK, 2004.
[9] S.F. Toha M.O. Tokhi, "Augmented Feedforward and Feedback Control of a Twin Rotor System using Real-
Coded MOGA," in IEEE Congress on Evolu. Computation, Barcelona, Spain, 2010.
[10] M. Ahmad, A. Ali and M.A. Choudhry, "Fixed-Structure H∞ Controller Design for Two-Rotor Fixed-Structure H∞
Controller Design for Two-Rotor," Arabian J. Sci. & Eng'g, vol. 41, no. 9, p. 3619–3630, 2016.
[11] Ljung L, System Identification: Theory for the User, 1st ed., New Jersey,: Prentice Hall, 1987.
[12] Sjöberg J, Hjalmarsson H, Ljung L, "Neural Networks in System Identification," IFAC Proceedings Volumes, vol.
27, pp. 359-382, 1994.
[13] Chu SR, Shoureshi R, Tenorio M, "Neural networks for system identification," IEEE Control Systems Magazine,
vol. 10, pp. 31-35, 1990.
[14] Castillo E, Guijarro-Berdiñas B, Fontenla-Romero O, Alonso-Betanzos A, Bengio Y. A, "Very Fast Learning
Method for Neural Networks Based on Sensitivity Analysis," Journal of Machine Learning Research, vol. 7, pp.
1159-1182, 2006.

19
[15] Wu H, Zhou Y, Luo Q, Basset MA, "Training Feedforward Neural Networks Using Symbiotic Organisms Search
Algorithm," Computational Intelligence and Neuroscience, vol. 2016, pp. 1-14, 2016.
[16] Rere LMR, Fanany MI, Arymurthy AM, "Metaheuristic Algorithms for Convolution Neural Network,"
Computational Intelligence and Neuroscience, vol. 2016, pp. 1-13, 2016.
[17] Yaghini M, Khoshraftar MM, Fallahi M, "A hybrid algorithm for artificial neural network training," Engineering
Applications of Artificial Intelligence, vol. 26, pp. 293-301, 2013.
[18] Ghasemiyeh R, Moghdani R, Sana SS, "A Hybrid Artificial Neural Network with Metaheuristic Algorithms for
Predicting Stock Price," Cybernetics and Systems, vol. 48, pp. 365-92, 2017.
[19] Salimi H, Stochastic Fractal Search: A powerful metaheuristic algorithm. Knowledge-Based Systems, vol. 75,
2015, pp. 1-18.
[20] Mosbah H, El-Hawary ME, "Optimization of neural network parameters by Stochastic Fractal Search for dynamic
state estimation under communication failure," Electric Power Systems Research, vol. 147, pp. 288-301, 2017.
[21] Khishe M, Mosavi MR, Moridi A, "Chaotic fractal walk trainer for sonar data set classification using multi-layer
perceptron neural network and its hardware implementation," Applied Acoustics, vol. 137, pp. 121-139, 2018.
[22] Xia Y, Wang J, "A dual neural network for kinematic control of redundant robot manipulators," IEEE
Transactions on Systems, Man and Cybernetics, Part B (Cybernetics), vol. 31, no. 1, p. 147–154, 2001.
[23] Yoo SJ, Choi YH, Park JB, "Generalized predictive control based on selfrecurrent wavelet neural network for
stable path tracking of mobile robots: Adaptive learning rates approach," IEEE Transactions on Circuits and
Systems I. Regular Papers, vol. 53, no. 6, p. 1381–1394, 2006.
[24] Lin S, Goldenberg AA, "Neural-network control of mobile manipulators," Neural Networks, IEEE Transactions,
vol. 12, no. 5, p. 1121–1133, 2001.
[25] El-Fakdi A, Carreras M, "Two-step gradient-based reinforcement learning for underwater robotics behavior
learning," Robotics and Autonomous Systems, vol. 61, no. 3, p. 271–282, 2013.
[26] Wai R-J, "Tracking control based on neural network strategy for robot manipulator," Neurocomputing, vol. 51, p.
425–445, 2003.
[27] Agand P, Shoorehdeli MA, Khaki-Sedigh A, "Adaptive recurrent neural network with Lyapunov stability learning
rules for robot dynamic terms identification," Engineering Applications of Artificial Intelligence, vol. 65, pp. 1-11,
2017.
[28] Rahideh A, Bajodah AH, Shaheed MH, "Real time adaptive nonlinear model inversion control of a twin rotor
MIMO system using neural networks," Engineering Applications of Artificial Intelligence, vol. 25, p. 1289–1297,
2012.
[29] Tijani IB, Akmeliawati R, Legowo A, Budiyono A, "Nonlinear identification of a small scale unmanned helicopter
using optimized narx network with multiobjective differential evolution," Engineering Applications of Artificial
Intelligence, vol. 33, p. 99–115, 2014.
[30] O.W. Abdulwahhab and N.H. Abbas, "A New Method to Tune a Fractional-Order PID Controller for a Twin Rotor
Aerodynamic System," Arabian J. Sci. & Eng'g, vol. 42, no. 12, pp. 5179-5189, 19 June 2017.
[31] Ezekiel DM, Ravi S, & Matsebe O, "Modelling of the Twin Rotor MIMO System (TRMS) Using the First
Principles Approach," IEEE: 2020 Int. Conf. Comp. Comm & Informatics (ICCCI), pp. 726-732, 24 1 2020.
[32] Ezekiel DM, Ravi S, & Matsebe O, "Pitch and Yaw Angular Motions (Rotations) Control of the 1-DOF and 2-DOF
TRMS: A Survey," Springer: Archives of Computational Methods in Engineering, 4 2020.
[33] Darus IZM, Lokaman ZA, "DYnamic Modelling of the Twin Rotor MIMO System in Horizontal Motion," Jurnal
Mekanikal, vol. 31, pp. 17-29, 2010.
[34] Rahideh A, Shaheed MH, Huijberts HJC, "Dynamic modelling of a TRMS using analytical and empirical
approaches," Control Eng. Pract, vol. 16, p. 241–259, 2008.
[35] Rahideh A, Shaheed MH, and Bajodah AH, "David Mohammed Ezekiel 1, Ravi Samikannu 2, Matsebe Oduetse
3," in 16th IEEE Int Conf on Control Applications, Kuala Lumpur, Singapore, 2007.
[36] Coelho J, Matos Neto R, Lebres C, Fachada H, Fonseca Ferreira NM, Solteiro Pires EJ, Tenreiro Machado JA,
"Application of Fractional Algorithms in the Control of an Helicopter System," 2007.
[37] Coelho J, Matos Neto R, Lebres C, Fachada H, Fonseca Ferreira NM, Solteiro Pires EJ, Tenreiro Machado JA,
"Helicopter System Modelling and Control With Matlab," Academia, 2008.
[38] Coelho J, Matos Neto R, Lebres C, Santos V, Fonseca Ferreira NM, Solteiro Pires EJ, Tenreiro Machado JA,
Coimbra Codex, "Application of Fractional Algorithms in the Control of A Twin Rotor Multiple Input-Multiple
Output System," in 6th EUROMECH, Nonlinear Dynamics Conf. ENOC 2008, Saint Petersburg, Russia, 2008.
[39] Coelho, J., Matos Neto, R., Lebres, C., Fachada, H., Fonseca Ferreira, N.M., Solteiro Pires, E.J., and Tenreiro
Machado, J.A., "Application of Fractional Algorithms in the Control of an Helicopter System," 2007.
[40] Chalupa P, Pikryl J, Novk J, "Modelling of Twin Rotor MIMO System," 25th DAAAM Int'l Symp. on Intel
Manufacturing & Automation 2014: Procedia Eng'g, vol. 100, pp. 249-258, 2015.
[41] Kumar, G.V., Rao, G.V.S.K, "Control of TRMS and Modelling Using LPV," Int'l Res. J. Eng'g & Tech. (IRJET),
vol. 3, no. 8, pp. 912-918, Aug 2016.
[42] Juang, J-G., and Tu, K-T., "Design and Realization of a Hybrid Intelligent Controller for a Twin Rotor MIMO

20
 System," J. Marine Sci. & Tech., vol. 21, no. 3, pp. 333-341, Jun 2013.
[43] Coelho, J., Matos Neto, R., Lebres, C., Fachada, H., Fonseca Ferreira, N.M., Solteiro Pires, E.J., and Tenreiro
Machado, J.A., "Helicopter System Modelling and Control With Matlab," Academia, 2008.
[44] Coelho, J., Matos Neto, R., Lebres, C., Santos, V., Fonseca Ferreira, N.M., Solteiro Pires, E.J., Tenreiro
Machado, J.A., and Coimbra Codex, "Application of Fractional Algorithms in the Control of A Twin Rotor Multiple
Input-Multiple Output System," in 6th EUROMECH, Nonlinear Dynamics Conf. ENOC 2008, Saint Petersburg,
Russia, 2008.
[45] Bouguerra, A., Saigaa, D., Kara, K., Seghlache, S., and Loukal, K., "Fault-Tolerant Control of a 2 DOF
Helicopter (TRMS System) Based on H∞," in Int'l Conf. Control, Eng'g & Info. Tech., CEIT 2013, Sousse,
Tunisia, 2013.
[46] Rahideh A, Shaheed MH, and Bajodah AH, "Adaptive Nonlinear Model Inversion Control of a Twin Rotor
System Using Artificial Intelligence," in 16th IEEE Int Conf on Control Applications, Kuala Lumpur, Singapore,
2007.
[47] A. Rahideh a,b, A.H.Bajodah Shaheed MH, "Real time adaptive nonlinear model inversion control of a twin rotor
MIMO system using neural networks," Engineering Applications of Artificial Intelligence, vol. 25, p. 1289–1297,
2012.
[48] W. B. a. H. M.E, "Adaptive switching circuits,," in IRE WESCON Convention Record, New York, 1960.
[49] Demuth H, Beale M, "Neural Network Toolbox: For Use with MATLAB," The MathWorks Inc., Natick, MA, USA,
2000.

21