1,721,084 research outputs found
Antifungal and antibacterial activities of indigenous Streptomyces isolates from saline farmlands, prescreening, ribotyping and metabolic diversity
No full textTowards Accurate Loss Prediction In Kahramaa Water Stations
Full text linkQatar General Electricity and Water Corporation (KAHRAMAA, KM) deployed
and started operation of water Supervisory Control and Data Acquisition (SCADA) system
in 2006. The aims of this SCADA system are to increase the control and pumping of water
to customers and reduce the water loss of the network. SCADA collects sensory data such
as reservoirs’ inlet and outlet flows, reservoirs’ levels, reservoirs’ inlet water stock as well
as status of valves and used pumps. These time-stamped data are periodically transmitted
to several back-end servers for logging, storing, and processing. KM water network is
composed of 35 connected stations; each includes from 3 to 12 reservoirs. Currently, KM
lacks the ability to accurately forecast any water loss in the network, except by assuming
that historical loses apply the same in future; causing inaccurate predictions.
Throughout the years, there has been an increasing interest in water loss prediction.
Different techniques are used to analyze and forecast the water loss. These techniques are
classified into three categories, which are: statistical, machine learning and hybrid
modeling approaches. Statistical approach depends on fitting mathematical models to the
observed data. However, these have a disadvantage of high noise error that prevents water
leaks to be accurately detected and forecasted. In machine learning, water loss is predicted through training of various models such as Support Vector Machines, Artificial Neural
Networks and Random Forest. The hybrid approach combines two or more techniques from
the previously mentioned approaches.
This thesis studies methods to accurately predict water loss in KM water stations.
We adopt a knowledge discovery and data mining process and activities that include data
collection, data preprocessing, feature engineering, model training, and validation. This
is the first automated attempt for KM to predict future volumes of water to be lost.
Moreover, several contributions are made to advance prediction accuracy including those
related to data preprocessing (data aggregation, cleaning, and transformation), feature
engineering (feature generation, data windowing), and model training where several
models are optimized for high accuracy using statistically reliable evaluation (crossvalidation).
Experimental results show that the highest water loss prediction accuracy of
the next hour, 12th hour, and 24th hour are 84.78%, 73.01%, and 71.66%, respectively.
These results come with different settings and parameters tuning that are optimized for
each case. Moreover, all of the above results surpass baseline models by 14.78%, 45.32%,
and 11.50%, respectively, in accuracy
Enhancing Knowledge Distillation for Text Summarization
No full textIn the realm of natural language processing, recent advancements have been significantly
shaped by the development of large pretrained Seq2Seq Transformer models,
including BART, PEGASUS, and T5. These models have revolutionized various text
generation applications, such as machine translation, text summarization, and chatbot
development, by offering remarkable improvements in accuracy and fluency.
However, their deployment in text summarization often encounters significant challenges
in environments with limited computational resources. This research proposes
an innovative solution: the development of compact student models. These models are
designed to emulate the capabilities of their larger pretrained counterparts (teacher models)
while ensuring reduced computational demands and increased processing speed,
thus maintaining high performance with greater efficiency.
Knowledge distillation, a popular technique in model optimization, typically employs
two primary techniques: direct knowledge distillation and the use of pseudo-labels. Our
research enhances direct knowledge distillation by introducing an effective behavior
function. This function selectively emphasizes the more certain predictions from the
teacher model, thereby addressing the exposure bias issue that arises from differences
between training and testing environments. In addition to this, we propose a novel
approach to select the most reliable predictions from the teacher model. These highconfidence
predictions are then utilized as pseudo-summaries, optimizing the student
model’s training through the pseudo-label technique. This dual approach mainly focuses
on the confidence of teacher predictions and offers a comprehensive solution to enhance
the model’s performance while maintaining computational efficiency.
We evaluated our methods using BART on the CNN/DM dataset and Pegasus on
the XSUM dataset. The findings of these assessments revealed that our approaches not
only successfully achieved the knowledge distillation objectives, but also significantly
surpassed the performance of the teacher models
Accurate Classification of Partial Discharge Phenomena in Power Transformers in the Presence of Noise
Full text linkThe objective of this research is to accurately classify different types of Partial
Discharge (PD) phenomenon that occurs in transformers in the presence of noise. A PD is
an electrical discharge or spark that bridges a small portion of the insulation in electrical
equipment, which causes progressive deterioration of high voltage equipment and could
potentially lead to flashover. The data for the study is generated from a laboratory setup
and it is 300 time series signals each with 2016 attributes corresponding to 3 types of PDs;
namely: Porcelain, Cable and Corona. The data is collected from two sensors with different
bandwidths, in which Channel A signals refer to the data collected from the higher
frequency sensor and signals from Channel B refer to data of the lower frequency sensor.
Different feature engineering approaches are investigated in order to find the set of the
most discriminant features which help to achieve high levels of classification accuracy for
Channel A and Channel B signals. First, features that describe the shape and pulse of
signals in the time domain are extracted. Then frequency domain based statistical features
are generated. In comparison with classification accuracies using frequency domain
features, time domain based features gave higher accuracy of more than 90% on average
for both channels in the absence of noise while frequency domain features allowed
classification accuracy up to 80% on average. However, in the presence of noise, both
methods degraded. To overcome this, Regularization techniques were applied on the
features from the frequency domain which helped to maintain classification accuracy even
in the presence of high levels of noise
EVALUATION OF 2D AND 3D TECHNIQUES FOR SENTIMENT VISUALIZATION
Full text linkWith the rise of user generated content on the Internet, sentiment visualization is being highly researched and practiced. Advances in information visualization, such as the use of three-dimensional visualizations need to be applied to sentiment visualization. However, minimal efforts were taken in the literature, to address when two-dimensional (2D) and three-dimensional (3D) visualization techniques can be used for sentiment visualization. In this thesis, we investigate the 2D and 3D visualization techniques based on the visual variables which represent sentiment in sentiment visualization and perform a comparative empirical study. We conduct a task-based evaluation to measure the performance and cognitive load of visualizations where sentiment is represented by different visual variables in both 2D and 3D visualizations. The objective of this work is to find when 2D and 3D visualization techniques can be used for sentiment visualization and which visual variable is comparatively well-suited for visual representation of sentiment in 2D and 3D. We use scatterplot and bar chart in 2D and 3D for case-study. While the results reflect the known fact that 2D has better performance and lower cognitive load, we investigate different scenarios involving the visual representation of sentiment in 2D and 3D visualizations. Additionally, we discuss the trade-offs of using 2D and 3D visualizations for sentiment visualization. We expect this study to help data analysts, sentiment analysis and visualization researchers and developers make an informed decision of when 3D visualization can be used for sentiment visualization
Two-Timescale Multi-Objective Volt/Var Optimization Considering Distributed Energy Resources In Active Distribution Networks
Full text linkThe high penetration of distributed energy resources (DERs) introduces several challenges to the power network. They may cause a high level of voltage variation, sudden over/under-power generation, high power losses, and negatively impacteddistribution assets. Thus, there is a vital need for volt/var optimization (VVO) schemes that integrate utility-owned assets with inverter-interfaced resources to overcome these challenges. This thesis addresses the above-mentioned challenges by proposing a comprehensive two-timescale multi-objective VVO algorithm. The slow timescaleutilizes utility-owned assets to minimize system losses and maximize asset lifetime ina three-step methodology. This stage incorporates the utility operator's direct input for setting the utility-owned assets. At the faster scale, the algorithm optimizes the reactive power of DERs to minimize voltage variations and system losses. The proposed VVO is solved using conventional optimization and reinforcement learning algorithms. The IEEE 33-bus system is modified and used to demonstrate the effectiveness of the proposed algorithm
Time-Aware Workload Charactrization And Prediction For Proactive Auto-Scaling Of Web Applications
Full text linkProactive auto-scaling techniques aim to predict the future workload of web
applications to provision the required resources, such as virtual machines (VMs), ahead
of time. Nevertheless, deciding the optimal number of resources to allocate is a
challenging task due to the dynamic nature of workload characteristics and the
difficulty of predicting them. Most of the existing workload approaches only consider
one workload feature which is typically the volume of requests to characterize and
predict the workload. In this thesis, we report the design and development of a time
aware workload prediction model that considers the request time features in order to
achieve better workload characterization and prediction. We explore two different
approaches, namely Time-Aware Single-Modeling and Time-Aware Multi-Modeling.
The Time-Aware Single-Modeling approach builds one model for the entire time-space
and has three variations: multivariate regression, univariate Long Short-Term Memory
Neural Networks (LSTM), and multivariate LSTM neural network model. While, Time
Aware Multi-Modeling approach develops a prediction model for each time partition
discovered using a periodicity detection component.
The proposed solutions are evaluated using two real workload datasets: Library
portal at Qatar University and NewsLink portal in Pakistan. The results demonstrate
that the time-aware approaches achieve more accurate predictions of the workload patterns compared to other existing approaches. Also, it has been shown that the
achieved improvements are statistically different than existing approaches
Home energy management system embedded with a multi objective demand response optimization model to benefit customers and operators
Full text linkThis thesis aims to develop a Home Energy Management System (HEMS) that optimizes the load demand and distributed energy resources considering utility price signal, customer satisfaction, and distribution transformer condition. The electricity home demand considers Electric Vehicles (EVs), PV-based renewable energy resources, Energy Storage Systems (ESSs), and all types of fixed, shiftable, and controllable appliances. A multi-objective demand/generation response is presented to optimize the scheduling of various loads/supplies based on the pricing schemes. The
customers' behavior comfort-level and a degradation cost that reflects the distribution transformer Loss-of-Life (LoL) are integrated into the multi-objective optimization problem. First, conventional optimization approaches are utilized to solve the multi objective optimization problem. To overcome the conventional optimization limitations, a data-driven analysis, which utilizes deep reinforcement learning (DRL),is used. The results show that the DRL-based HEMS is more efficient in minimizing the energy cost while adapting to the user comfort within the desired level
ADVANCING FORENSIC AGE ESTIMATION: INTEGRATING CBCT IMAGING WITH MACHINE LEARNING AND DEEP LEARNING APPROACHES
No full textDental age estimation plays a vital role in forensic odontology, contributing to legal, clinical, and identification processes. While previous studies predominantly focused on maxillary teeth, mandibular teeth, known for their greater durability in adults, have been underexplored for this purpose. This research addresses this gap by utilizing cone beam computed tomography (CBCT) images of mandibular teeth alongside advanced machine learning and deep learning techniques to enhance predictive accuracy. The study utilized a dataset of 118 CBCT images collected for diagnostic purposes at the Jordan University of Science and Technology Dental Teaching Center. The dataset, comprising 50 males and 68 females with ages ranging from 18 to 72 years (mean age: 34.58 ± 14.20 years), was preprocessed to ensure consistency and reliability. A forward feature selection approach identified key age predictive features, including periodontal recession, axial crown area, and pulp-tooth volume ratio. Notably, axial plane measurements emerged as particularly informative, with three of the top ten age-correlated features derived from axial imaging. The machine learning framework was designed to accommodate any of four mandibular teeth (canines and second premolars), enhancing flexibility in forensic scenarios where specific teeth may be missing, damaged, or impacted. An ensemble model combining Decision Tree, K-Nearest Neighbors, XGBoost, and CatBoost regressors achieved a mean absolute error (MAE) of 4.19 years, outperforming individual regression models. To further explore automated dental age estimation, deep learning models were applied to 2,384 orthopantomogram (OPG) images, including both original OPGs and reconstructed OPGs derived from CBCT scans. Among the 37 evaluated deep learning models, DenseNet169 and ResNeXt101-64x4d demonstrated the highest accuracy, achieving an MAE of 5.35 years when tested on the 118 CBCT dataset, a result comparable to traditional machine learning methods. However, deep learning models exhibited reduced interpretability compared to machine learning approaches, which explicitly utilized handcrafted features for prediction. Guided class activation mapping (Guided CAM) visualizations confirmed that deep learning models relied primarily on the lower molar and premolar regions for age estimation, aligning with known dental aging indicators. Despite their lower interpretability, deep learning methods offered the advantage of automatic feature extraction, reducing the need for manual preprocessing. Overall, this study underscores the adaptability of machine learning and deep learning techniques to diverse dental structures. While ensemble machine learning models demonstrated superior accuracy, deep learning approaches provided a more automated yet less interpretable solution. The results highlight the potential for integrating both methodologies to enhance the reliability and precision of dental age estimation in forensic applications
- …
