الفهرس 
Chapter 1: INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 189 |  |  |
| | | from | 189 | | |
Abstract
Robotics replaces humans in more and more real-world applications every day,
taking the place of them in previously manual jobs. Their use is improving
performance, productivity, and safety in hazardous environments. The nature and
magnitude of hazards present special challenges for the accomplishment of desired
tasks in hazardous environments. Among the hazards may be toxic contamination,
potential explosions or radiation. But, the mobile robots often exhibit nonlinearities
due to the complexity of their dynamics and their interaction with their environment.
In this concern, there are some common sources of nonlinearity in mobile robots such
as: nonlinear kinematics, nonlinear dynamics, actuator dynamics, sensor
nonlinearities, terrain irregularities and environmental disturbances. In mobile robots,
uncertainties are a major challenge, and several sources contribute to them, including
localization errors, environmental variability, actuator uncertainties, sensor noise, and
dynamic environment changes. Also, the mobile robot is considered a multi-input
multi-output (MIMO) system.
To solve these problems of the mobile robot, it requires a precision controller.
Learning controllers are effective controllers for solving complex problems that may be
challenging to address through traditional rule-based methods. A learning controller is
suitable for dynamic and time-sensitive environments requiring real-time decisionmaking since it adapts quickly for changing conditions. The types of learning controllers
include supervised learning controllers, unsupervised learning controllers,
reinforcement learning controllers, deep learning controllers, and neural network
controllers. This thesis presents a proposal for deep learning controllers.
In this thesis, three controllers are proposed. The first proposal is a feedforward neural network deep learning controller (FFNNDLC). The proposed FFNNDLC
combines the principles of both the multilayer feed-forward neural network (MLFFNN)
and the restricted Boltzmann machine (RBM). The second proposal is a hybrid deep
learning neural network controller (HDLNNC). The proposed HDLNNC is composed of
two parts, the first of which is the MLFFNN, which is the main controller. A self- organizing map of Kohonen (SOMK) procedure and Hebbian learning are used to
initialize the weights of the FFNN, in which the initial weights are equal to zero for all
controllers. The third proposal is a hybrid deep learning diagonal recurrent neural
network controller (HDL-DRNNC). The HDL-DRNNC structure consists of a diagonal
recurrent neural network (DRNN), whose initial values can be obtained through DL.
The DL algorithm is performed based on SOMK and RBM. The updating weights and
learning rate for the proposed algorithms are developed using the Lyapunov stability
criterion.
Simulation tasks are performed on both mathematical and physical systems,
which are classified as single input single output (SISO) and MIMO, respectively. The 4-
wheels skid steering mobile robot (4-WSSMR) is considered a physical system. The
proposed controllers have been simulated and compared with other existing
controllers in the previous publications. The proposed controllers have been designed
and implemented practically based on an embedded Arduino kit for controlling a real
4-WSSMR. The performance of the proposed controllers is measured using some
performance indices such as root mean square error (RMSE) and mean absolute error
(MAE). The simulation and practical results show good and significant improvement in
the performance of the proposed controllers to respond the system uncertainties and
nonlinearities compared with other existing controllers. On the other hand, the RMSE
and MAE values, which are obtained for the proposed controllers, are lower than those
obtained for other existing controllers.