Repository@USM

Welcome to closed access digital repository of Universiti Sains Malaysia (USM).

This repository contains multiple types of scholarly materials, especially USM theses and exam papers.

To access the full text, please log in with your USM email account.

Photo by @clarissemeyer

Research outputs

4064

Projects

People

Recent Additions

Brain tumor segmentation with mri images
(2023-08)
Khoo, Rong Quan
This thesis addresses the critical issue of brain tumor segmentation in 3D magnetic resonance imaging (MRI) scans, aiming to enhance early detection and diagnosis for improved patient outcomes. Brain tumors, among the most serious and life-threatening types, necessitate accurate identification and treatment planning. However, manual human-assisted classification from a vast number of MRI scans is time-consuming and error-prone, underscoring the need for automated segmentation methods. The study adopts deep learning techniques, with a focus on the U-Net architecture, which effectively captures both low-level details and high-level semantic information through CNN-based skip connections, facilitating precise segmentation. During the training process, the model benefits from a combination of Dice loss and categorical focal loss functions, enabling a comprehensive optimization approach for segmentation tasks. To evaluate the model's effectiveness, essential metrics such as sensitivity, specificity, accuracy, Intersection over Union (IoU), and F1 score are utilized. These metrics offer a holistic assessment, capturing segmentation accuracy, class imbalance, and overall correctness in identifying positive and negative instances. The optimal development of an efficient and accurate brain tumor segmentation model, as demonstrated by the evaluation metrics Sensitivity: 0.9784, Specificity: 0.9930, Accuracy: 0.9342, IoU, and F1 score has promising implications for medical image processing. Automating the segmentation process through this research contributes to the advancement of early detection and diagnosis, potentially improving patient recovery rates and treatment planning for individuals affected by brain tumors in the future medical field. This research paves the way for transformative advancements in the medical field, harnessing the power of deep learning techniques for accurate brain tumor segmentation. As technology continues to evolve, such automated approaches hold immense potential in revolutionizing medical imaging practices, ultimately contributing to better healthcare and a positive impact on patients' lives
Balancing control of battery banks in electric vehicles using a DC-DC converter
(2023-07)
Wong, Wai Kin
Nowadays, battery cell is widely used and developed for various reasons. This technology had been seeming like a new way for the user to work any devices or machines without power supply. The main problem of the battery cell is balancing control. Without balancing control, the health of the battery cell will decrease and the cost of maintenance and changing it will be very high. Therefore, DC-DC converter is needed to make sure the balancing control work. Zeta converter provides a positive output voltage from an input voltage that varies above and below the output voltage. The balancing control method is active balancing. Active balancing is work for all kind of battery cells. To make sure the balancing control work, controller is needed. This thesis starts with modelling Zeta converter and followed by connection with a controller. The input power of this system is a connection of 10 batteries in series. Therefore, this paper deals with balancing control of battery cell in electric vehicle using DC-DC converter. This simulation-based project is done by using MATLAB R2022a software.
Automatic ground vehicle for pick and place coloured box
(2023-07)
Divesh a/l S.Ramu
Nowadays the creations of Automatic Mobile Robot (AMR) model can be found from all over the countries, as it gives many advantages in our lives. It works just like a robot as it is able to sense and response to the environment. Considering that, AMRs should be well developed to optimize its benefits to our own living. The primary objective of the thesis is to design and construct a custom hardware platform capable of automatic navigation and manipulation of coloured boxes. The vehicle incorporates a range of sensors, including cameras, to capture and process visual information in real-time and ultrasonic sensors, this sensor is essential in detecting colored boxes positioned in front of the vehicle, triggering a halt in its movement and activating the gripper mechanism. The computer vision algorithms are developed to detect and identify different coloured boxes, enabling the vehicle to make informed decisions during pick and place operations. In order to realize this goal, a robotic prototype is constructed using MyRio and LabVIEW. The findings of the research demonstrate the successful realization of a robust and efficient automatic ground vehicle capable of accurately detecting, picking up, and placing coloured boxes. The system's performance is evaluated based on metrics such as task completion time, success rate, and overall efficiency. The tasks of material handling, especially in hazardous environments, are made easier with a well-developed AMR to enhance efficiency in the delivery of items, especially at centres for palliative care, the nuclear industry and other hazardous environments. Therefore, this project involves of designing and fabrication of the hardware and circuitry.
Assessment on the current transformer performance under the effect of power system harmonics
(2023-07)
Muhammad Adli bin Mohd Ribi
This research was implemented to study the use of current transformers as instrumental transformers. Generally, overall finding from this project is known as the current transformer will fail to operate correctly when a certain level of harmonic percentage is present in power system. The implementation of non-linear loads in electrical systems today has increased along with the latest technology. However, the use of nonlinear loads greatly contributes to higher harmonic distortion occurring in the electrical distribution system. Examples of non-linear loads that exist today are rectifiers, alternating current to direct current power converters. Main issue regarding the implementation of higher harmonic distortion in power system will cause the performance of current transformer fail to convert current values stated on the current transformer itself. Therefore, this study will study the performance of CT under linear and non-linear load and do an assessment of the performance and effects on the CT with different levels of THD percentage. The error percentage ratio for current transformers is measured to observe the functionality of current transformer. In this project, an experiment was carried out at the Power Laboratory of the School of Electrical and Electronics Engineering and a software simulation is done using MATLAB Simulink to simulate the operating current transformer with linear and non-linear load. The current transformer with a ratio of 7.5/1 A and a secondary load or CT burden for the current transformer is selected in this study. With this study, more understanding is acquired regarding the effect of the power system harmonics on the current transformer operation.
Arrhythmia electrocardiogram (ECG) signal classification using long short term memory, lstm (deep learning).
(2023-07)
Kang, Xian Jie
Arrhythmias, which are irregularities in the heartbeat's rhythm, can seriously harm a person's health. The type of rhythm disruption and the region of the heart where the disturbance arises are used to classify arrhythmias. However, it is nearly impossible to obtain a 100% accurate test due to the complexity of the cardiac conduction system through human eye. Recurrent Neural Networks (RNNs) is a type of artificial neural network that is commonly used to analysing ECG data due to its ability to analyse sequences of inputs. However, the vanishing gradient problem, which affects RNNs, causes delayed convergence and makes learning long term dependencies difficult as the gradient values are very small during backpropagation. Thus, LSTM which is specifically designed to overcome this limitation of traditional RNN is chosen in this project to classified arrhythmias from a dataset. A dataset of ECG signals is collected from PhysioNet Database which contains the recordings of several types of arrhythmias such as sinus arrhythmia, atrial fibrillation, ventricular tachycardia, and ventricular fibrillation. The raw data of ECG signals is first pre-processed to filter out the noise and normalizing the signals. Then, the data is then segmented into individual heartbeat and extracted relevant features for classification. Next, an LSTM model is developed to classify the ECG signals into different classes. Finally, the performance output of the model is evaluated using confusion matrix to assess the effect of different model parameters. Our findings show that LSTM neural networks show better results in classifying ECG arrhythmias compared to traditional RNN.