Quadcopter Control using Intelligent Control Methods

Abstract

Best Poster Presentation Award (Deep Learning Indaba): Unmanned Aerial Vehicles (UAV) are currently being explored for a wide range of applications in both military and civilian spheres. That being said a quadcopter UAV is a highly unstable, multiple-input, multiple output (MIMO) system with coupled translational and rotational dynamics . The system is under actuated with four motor thrusts (inputs) controlling six degrees of freedom (DoF).

Consequently, optimized controller design is crucial to achieve hover stability.
PID, Fuzzy Logic and both Offline and Online Trained Neural Network controllers are proposed as solutions.

Full text

Quadcopter Control using Intelligent Control Methods
Nabeel Seedat & Prof Anton Van Wyk
1. Introducon
Unmanned Aerial Vehicles (UAV) are currently being explored for a
wide range of applications in both military and civilian spheres [1].
That being said a quadcopter UAV is a highly unstable, multiple-input,
multiple output (MIMO) system with coupled translational and
rotational dynamics [6]. The system is under actuated with four motor
thrusts (inputs) controlling six degrees of freedom (DoF) [1].
Consequently, optimized
controller design is crucial to
achieve hover stability.
PID, Fuzzy Logic and both
Oine and Online Trained
Neural Network controllers are
proposed as solutions.
2. Quadcopter UAVs
A quadcopter is a UAV with 4 motor inputs and
6 degrees of freedom (DoF). The primary control of a
quadcopter involves control of aitude (Euler Angles—
Roll, Pitch and Yaw) and altitude. The structure of the
quadcopter is illustrated in Figure 2.
The mathematical model [3] of the quadcopter is pre-
sented below, where Φ (Roll angle), ϴ (Pitch Angle), ψ
(Yaw Angle). x,y,z represent the position of the quad-
copter in the reference frame.
In order, to stabilize the quadcopter a controller must be developed. The primary aim of aitude con-
trol (3 Euler Angles) and altitude control necessitates the development of 4 separate controllers.
However, the generalized closed loop control structure is illustrated in Figure 3.
3. Methods & Control Algorithms
A controller performance evaluation was conducted by simulating the system in a MATLAB-
SIMULINK environment. The quadcopter is subjected to a 45° initial angular displacement in the
roll, pitch and yaw axes. The initial altitude is 100m. The simulation was conducted under “ideal”
conditions, as well as, real-world conditions which subjected the UAV to a constant 6.2m/s simulated
wind along the x,y,z axis, as well as, quantization and sampling eects of sensors and an MCU being
incorporated. The following control architectures were evaluated to optimize performance:
(1) PID tuned by Ziegler-Nichols methods
(2) Fuzzy Logic Controller (25 Rule Mamdani)
(3) Oine Trained 3-Layer Articial Neural Network (ANN)
(4) Online Trained Multi-Layer Perceptron (MLP) ANN
(5) Online Trained Radial-Basis Function (RBF) ANN
(6) Online Trained ADALINE ANN
The 4 neural networks are 3-layer networks, with one hidden layer consisting of 10
neurons for comparison across neural network architectures. Figure 4 shows the structure
of the neural network utilized, whilst Figure 5 indicates ANN controller implementation.
4. Results
The results for Roll, Pitch, Yaw and Altitude tests is shown in Figure 6 for ideal conditions and in Figure 7 for non-ideal conditions.
The nal results displayed in the above Figures 6-7 reveal the degraded performance of all controllers under non-ideal simulated conditions. Firstly, the Fuzzy Logic
controller matched ANN performance under ideal conditions, but reduced in performance under non-ideal conditions. The oine trained ANN generally had poor
performance across both ideal and non-ideal conditions due to inability to update neuron weights to the changing error signal. Moreover, whilst the RBF ANN
architecture had the “best” performance with regard to seling time and SSE, under non-ideal conditions failed due to motor saturation. Thus, the online trained MLP
architecture can be considered the most viable controller due to sustained impressive performance across both ideal and non-ideal simulations
5. Conclusions
Quadcopters despite the potential use in many areas of society,
present an interesting and challenging control design challenge.
The non-linearity and coupled dynamics add to the complexity.
Intelligent controllers were proposed to optimize controller per-
formance in terms of reduced seling time and SSE.
The initial performance results presented here suggest that
online trained neural networks have a signicant advantage
under both ideal and non-ideal test scenarios by adjusting
neuron weights online. The disturbance rejection capability is a
key advantage.
Moreover, the online trained MLP ANN had the most superior
performance under both conditions and it is postulated that
the performance could be optimized by increasing the number
of hidden layers to account beer for non-linearity's and
un-foreseen dynamics.
References
[1] E. H. Khadija, E. K. Abdeljalil, M. Mostafa and A. Hassan, "Adapting parameters for ight control of a quadcopter using reference model
and fuzzy logic," 2015 Third World Conference on Complex Systems (WCCS), Marrakech, 2015, pp. 1-6.
[2] "Flying personal drone quadcopter", Flickr, 2017. [Online]. Available: hps://www.ickr.com/photos/ackab/15899607032. [Accessed: 09- Sep- 2017].
[3] L. Ascorti. “An application of the extended Kalman lter to the aitude control of a quadrotor.” Masters Thesis, Poletinico di Milano, 2013
Contact Details
Nabeel Seedat (seedatnabeel@gmail.com)
Prof. A. Van Wyk (anton.vanwyk@wits.ac.za)
School of Electrical and Information Engineering
University of the Witwatersrand, South Africa
Figure 3: Closed-Loop Control Structure of the entire quadcopter system
Figure 2: Quadcopter Structure
Figure 4: ANN Structure
Figure 5: ANN Controller Implementation
Figure 7: Controller Comparison (NON-IDEAL CONDITIONS) Figure 6: Controller Comparison (IDEAL CONDITIONS)
Figure 1: Quadcopter UAV [2]