1,721,011 research outputs found
Direct torque control of induction motor drives using space vector modulation (DTC-SVM)
No full textDirect Torque Control is a control technique used in AC drive systems to obtain high performance torque control. The conventional DTC drive contains a pair of hysteresis comparators, a flux and torque estimator and a voltage vector selection table. The torque and flux are controlled simultaneously by applying suitable voltage vectors, and by limiting these quantities within their hysteresis bands, de-coupled control of torque and flux can be achieved. However, as with other hysteresis-bases systems, DTC drives utilizing hysteresis comparators suffer from high torque ripple and variable switching frequency. The most common solution to this problem is to use the space vector depends on the reference torque and flux. The reference voltage vector is then realized using a voltage vector modulator. Several variations of DTC-SVM have been proposed and discussed in the literature. The work of this project is to study, evaluate and compare the various techniques of the DTC-SVM applied to the induction machines through simulations. The simulations were carried out using MATLAB/SIMULINK simulation package. Evaluation was made based on the drive performance, which includes dynamic torque and flux responses, feasibility and the complexity of the systems
Task-space dynamic control of underwater robots
No full textThis thesis is concerned with the control aspects for underwater tasks performed by
marine robots. The mathematical models of an underwater vehicle and an underwater
vehicle with an onboard manipulator are discussed together with their associated
properties.
The task-space regulation problem for an underwater vehicle is addressed where the
desired target is commonly specified as a point. A new control technique is proposed
where the multiple targets are defined as sub-regions. A fuzzy technique is used to
handle these multiple sub-region criteria effectively. Due to the unknown gravitational
and buoyancy forces, an adaptive term is adopted in the proposed controller.
An extension to a region boundary-based control law is then proposed for an underwater
vehicle to illustrate the flexibility of the region reaching concept. In this novel
controller, a desired target is defined as a boundary instead of a point or region. For a
mapping of the uncertain restoring forces, a least-squares estimation algorithm and the
inverse Jacobian matrix are utilised in the adaptive control law.
To realise a new tracking control concept for a kinematically redundant robot, subregion
tracking control schemes with a sub-tasks objective are developed for a UVMS.
In this concept, the desired objective is specified as a moving sub-region instead of a
trajectory. In addition, due to the system being kinematically redundant, the controller
also enables the use of self-motion of the system to perform sub-tasks (drag
minimisation, obstacle avoidance, manipulability and avoidance of mechanical joint
limits)
Region boudary-based control scheme for an underwater vehicle with an edge-based segmentation approach
No full textReinforcement Learning Approach for Commodity Market Trading Strategy
No full textThe apparent phenomena of commodity price fluctuations significantly affect the cost of living. Most current studies utilize datasets collected before the Russo-Ukrainian War and Covid-19. Moreover, many people are focusing on fund investment, exploring avenues such as commodity trading in addition to stocks and forex investments. However, most research for price prediction in commodities does not cover the periods of Covid-19 and the Russo-Ukrainian war. The aim of this project is to develop trading strategy models to predict whether to buy or sell a commodity, and to evaluate the potential rewards and profits. The dataset used contains daily historical prices of various types of commodities from the year 2000 until March 2022. Furthermore, a real-world dataset, specifically the gold trading dataset from Nasdaq, will be used to validate the performance of the best-performing trading models. The algorithms employed are reinforcement learning-based: Advantage Actor Critic (A2C) and Proximal Policy Optimization (PPO). Evaluation performance across six rounds of experiments has shown that the A2C model in a forex environment, using 80% of the dataset for training and 20% for testing, achieved the best results, with a Sharpe ratio of 0.63, a Sortino ratio of 1.0, an Omega ratio of 1.24, and a Calmar ratio of 0.55. The best-performing trading models in Objective 2 and Objective 3 are similar but employ different window sizes. Window size specifies the timesteps that will serve as reference points for the trading model to determine the next trade. Different datasets may require different window sizes, an issue that necessitates further refinement. This refinement is crucial as it involves tailoring the window size to align with the unique characteristics and volatility patterns of each dataset, thereby ensuring that the model's predictive accuracy is optimized for varied market conditions and historical trends. In conclusion, the best-performing trading model is the Advantage Actor Critic (A2C) model in a forex environment.</p
Alzheimer’s Disease and Frontotemporal Dementia:Differential Diagnosis Using Electroencephalogram Signal
No full textDementia is a neurological disorder that affects a person’s cognitive and social skills, leading to a decline in their overall mental functioning. Frontotemporal dementia (FTD) and Alzheimer’s disease (AD) are two common types of dementia. Frontotemporal dementia mostly affects the frontal and temporal lobes of the brain. These areas are responsible for executive functions, decision-making, language, behavior regulation, and personality traits. Alzheimer’s disease primarily damages the cerebral cortex, which is responsible for higher cognitive functions such as memory, language, and perception. The electroencephalogram (EEG) signal has many advantages in diagnosing these disorders, including low cost and high temporal resolution. This study compares AD and FTD patients with healthy subjects by extracting specific features from EEG signals. Two machine learning algorithms were used for the separation, Support Vector Machines (SVMs) and k-Nearest Neighbors (KNNs), and 10-Fold Cross-Validation was applied to validate the performance of this method and an accuracy of 91.2% (sd = 7.8) was achieved using the SVM classifier for diagnosing the disease and 71.1% (sd = 8.6) for classifying AD and FTD.</p
Fractal Dimension Analysis Demonstrates Overestimation and Underestimation of Time in EEG Signal
No full textKeeping track of time is regarded as an essential human behavior. The question of how the brain deals with temporal information remains a subject of scholarly debate. The current investigation aims to explore the mechanism underlying time perception by extracting fractal dimension from an electroencephalogram (EEG) signal and its frequency bands. To accomplish this, Higuchi’s fractal dimension was calculated for 42 healthy subjects’ electroencephalogram (EEG) signal and its sub-bands during the time perception task. The EEG signal was recorded from 19 channels. Subsequently, a statistical analysis was conducted to compare participants who underestimated versus those who overestimated the elapsed time and significantly different channels were presented. The findings suggest an elevated level of fractal dimensionality in persons who displayed a tendency to overestimate time. The EEG signal and Gamma rhythm emerged as the most distinguishing signals between the two cohorts. The contrast in fractal dimension between the two groups was predominantly apparent in the parietal and central channels. To summarize, an increased level of complexity is discernible in the EEG signal and high-frequency rhythms when there is an overestimation of temporal duration. It can be asserted that the employment of FD yields presents an exceptional approach to comprehending cerebral functionalities, notably temporal perception.</p
Online Conversation-Based Social Engineering Detection Using Machine Learning
No full textSocial engineering through online conversations can occur via phone calls, Skype, or Google Meet, among others. This paper presents a machine learning-based classifier for detecting scam conversations in various online formats, including live call conversations. However, selecting an appropriate dataset and the optimal vectorization technique for the algorithm remains challenging, and many fraudulent scams remain undetectable in online conversations. Consequently, six experiments were conducted to apply a machine learning classifier, resulting in 108 outcomes. All six experiments demonstrate that different classifiers possess unique strengths and weaknesses when applied to different scenarios. When compared to Doc2Vec, the vectorization techniques of Universal Sentence Encoder yield excellent results. Among various clustering methods, K-Means and the EM algorithm perform exceptionally well. The results reveal that Random Forest and CatBoost classifiers outperform others in terms of accuracy, precision, recall, and F1-score across all cases. These findings can contribute to enhancing the detection of scam attempts in live call conversations, thus helping protect individuals from falling victim to scams.</p
Automating Galaxy Image Classification in Galaxy Zoo:A Comparative Study of Deep Learning Models
No full textThis study compares the effectiveness of Artificial Neural Networks (ANNs) and Logistic Regression in classifying galaxy images from Galaxy Zoo 1. We propose Convolutional Neural Network (CNN) and Autoencoder models as potential solutions to mitigate the burden of manual classification. Both models are analyzed, and results reveal that ANN surpasses Transferred Learning Logistic Regression in terms of accuracy and runtime. Further investigation highlights the impact of activation functions, neuron count, hidden layers, and algorithm ensembling on ANN's classification performance. Additionally, we explore training time complexity reduction through learning rate, optimization algorithms, and batch size. The findings provide valuable insights for galaxy image classification tasks.</p
Automated Sleep Staging Classification System Based on Convolutional Neural Network Using Polysomnography Signals
No full textSleep is a crucial bodily process that plays a vital role in maintaining overall health and well-being. When diagnosing and treating sleep disorders, the initial step is sleep staging. However, manual sleep staging by physicians can be complicated, leading to a growing interest in computer-aided sleep stage classification algorithms. In this research, a method was introduced for automatically classifying sleep stages by extracting distinctive representations from single-channel EEG signals. PSG signals are selected exclusively for the project because they directly capture the essential physiological changes needed for sleep staging, ensuring both data relevance and quality. This choice also aligns with the project’s goal of feasibility and computational efficiency while avoiding potential ethical and privacy issues linked to audio and video data. Furthermore, it conforms to established practises in the field, ensuring consistency in benchmarking. A filterbank is applied by dividing the range of the frequency signal into two 15 sub epochs. The activity of the signal within distinct frequency ranges during different sleep stages was fully comprehended by computing the standard deviation as a single characteristic from different frequency subbands of the EEG. These characteristics served as the input for a two-stream convolutional neural network (CNN) that was trained using a two-stage learning methodology for classification. These characteristics served as the input for a two-stream convolutional neural network (CNN) that was trained using a two-stage learning methodology for classification.</p
Utilizing Color Space Information as Features for Deep Learning for a Heterogeneous Food Recognition System
No full textFood recognition serves as one of the most promising applications of visual object recognition, as it can be used to calculate food calories and analyze people's eating habits and dietary practices for health-care purposes. It forms the primary and most vital step in developing an application capable of providing nutritional recommendations. In this work, the CNN, ResNet, and AlexNet architectures have been proposed, given their popularity among researchers. A merged dataset, comprising Food-101 and UEC-Food256, was employed to evaluate the proposed models. To ascertain the optimal color spaces for food recognition, eight selection criteria were proposed. Among the various color space selection criteria, RGB showed superior performance when used with the Custom CNN model. To enhance performance, an upsampling method was implemented to increase the number of samples in the minority classes. This was achieved by duplicating existing samples using an augmentation process, thereby balancing the merged dataset. The final results suggest that EfficientNetB0, a CNN-based pre-trained model, performs better with the RGB color space, increasing accuracy from 40.48% (Custom CNN) to 73.11% (EfficientNetB0).</p
- …
