ABSTRACT

This thesis aims to find the optimal paths for multi-mobile robots working in the same static complex environment, as well as control their motion. To achieve this goal, three stages are proposed. In the first stage, a path planning method provides the shortest and smoothest path with collision avoidance between the starting and the target points in a static robot environment. This work proposes an intelligent hybrid optimization method that combines two algorithms; the first algorithm is called the Quarter Orbits (QO) algorithm, which tries to enhance the behavior of the cell decomposition path planning algorithm in terms of reducing the path length required to reach the target. The second algorithm uses the Particle Swarm Optimization (PSO) algorithm to get a more efficient, smoother, and shorter path for the mobile robot to reach the target in a complex environment. The generated hybrid method is called the Quarter Orbits Particle Swarm Optimization (QOPSO). In the second stage, a proposed Inverse Differential Kinematic Neural Network Trajectory Tracking (IDKNNTT) controller based on a Modified Elman Recurrent Neural Network (MERNN) is used. This proposed controller is used to control the nonlinear kinematics mobile robots’ system to smoothly and quickly generate the left and right wheels’ velocities of the multimobile robots, which are used to control the orientation and position of each mobile robot. In particular, the controller guarantees that all the mobile robots will follow their desired paths quickly and correctly. Moreover, at the third stage, an adaptive PID-PSO low-level controller is proposed to control the left and right wheels velocities of the DC motor with minimum velocity error to guarantee that each mobile robot follows its desired path and to design this controller a programmable integrated circuit (IC) will be used. This IC was implemented using a field programmable gate array (FPGA) board Spartan-3E kit. All simulation results clearly show that the proposed hybrid path-planning algorithm provided the shortest path with collision avoidance. The proposed hybrid QOPSO algorithm achieves enhancement on the path length equal to 29.42% compared to the VCD algorithm and 24.25% compared to the RCD algorithm for different types of simple and complex maps. To validate the numerical simulation results of the proposed control strategy they were compared to those of other types of controllers in terms of the maximum error enhancement in the X-position and the Y-position. Particularly, when the proposed controller was compared to the Convolutional Neural Network Trajectory Tracking (CNNTT) controller, the comparison results show that the proposed controller improves the tracking error rate on the X-axis by 75.5% and enhances the tracking error rate on the Y-axis by 21.2%. In addition, the proposed controller was compared to the MIMO-PID-MENN controller, and the comparison results show that the proposed controller improves the tracking error rate on the X-axis by 33.3% and on the Y-axis by 40.6%. Moreover, the simulation and emulation result of using the low-level FPGA-PID-PSO controller provides a minimum velocity error of approximately zero value.

ABSTRACT

Network intrusion detection is essential for protecting computer systems and data in the age of increasingly sophisticated cyber-attacks. Classic intrusion detection systems have many issues adapting to evolving and complex attack patterns; hence, there is a need for new approaches. The problem is with traditional intrusion detection systems (that) cannot cope with complex and fluid cyberattacks. The objective of this study is to investigate the use of Machine Learning (ML) and Deep Learning (DL) algorithms to reduce the limitations of current approaches and improve network intrusion detection.

The main goal of this study was to build accurate and scalable models for anomaly detection and classification. Then, we test the effectiveness of machine learning and deep learning approaches and compare them with respect to their results and the impact of preprocessing measures on model performance and the resulting accuracies. Two popular datasets, NSL-KDD and UNSW-NB15, were used to perform exhaustive experiments for this task purpose. These datasets provide an overview of various network setups and attack cases. The proposed system uses a pre-process in which feature selection and dimensionality reduction are used to optimize the classifiers.

Both the machine learning and deep learning strategies yielded promising outcomes, as determined by the analysis. On both the NSL-KDD and UNSW-NB15 datasets, machine learning classifiers (such as Gaussian Naive Bayes) performed very well in accuracy, precision, recall, and F1-measure. These classifiers had a good balance of accuracy while being computationally efficient enough to serve as real-time anomaly detectors. Machine learning, on the other hand, failed compared to deep learning models, specifically Convolutional Neural Networks (CNNs), as they capture more complex patterns within network traffic data. The results show that the CNN models proved to be very effective in detecting network intrusion even in large and complex datasets with high accuracy (99 %), precision, recall and f1-score.

Overall, this study brings substantial improvement in the area of network intrusion by evaluating all the different machine-learning algorithms and deep-learning approaches. It also benchmarks their performance and examines preprocessing techniques to improve their accuracy and efficiency. These results highlight the potential of these methods for fortifying network security and managing emerging risks.