Devices like computers that require manual control are difficult to use by disabled people such as having multiple sclerosis (MS), amyotrophic lateral sclerosis (ALS), partial stroke, etc. These people have a very limited movement capability such that they can do most of the interaction with a limited movement of head and the eye. Assistive technologies are very important for people with disabilities not to be dependent on anyone in daily life. In this study, we present a human-computer interaction application that works based on analysis of the head and eye/eyebrow movements from real-time images captured by a visual camera. We propose Difference Between Eye and Eyebrow (DEEB) features to detect the action intention of the user and properly realize the computer keyboard and mouse actions based on eyebrow, eye and head movements. In addition, there are shortcut keys in the designed interface that facilitate access to Instagram, WhatsApp, YouTube etc. which are considered to make it easier for individuals with disabilities to communicate with their social environment. We obtained satisfactory results on the designed interface in the experimental study.

Today, many people die in traffic accidents. Sleeplessness and fatigue of drivers are shown as the most important cause of traffic accidents. For this reason, research on driver performance analysis is of great importance. In this study, a system is designed to analyze driver fatigue using EEG data. As the data set, the EEG signals from sustained-attention driving task prepared by National Chiao Tung University have been used. The data set is divided into four classes to determine the driver's fatigue times and level. In order to determine the frequency ranges that occur during driver fatigue phases, EEG signals are filtered. Principal Component Analysis method has been used to reduce the size of the features matrix. With the Divide and Conquer algorithm, all combinations in which the four classes will be separated best are determined and classification has been done at each step using sub-classifiers. As sub-classifiers, k-Nearest Neighborhood, Support Vector Machines and Linear Discrimination Analysis algorithms are used. As a result of the study, the average classification successes are 87.9% for the k-Nearest Neighborhood algorithm, 88.5% for the Support Vector Machines algorithm and 81.6% for Linear Discrimination Analysis. The highest classification success has been achieved as 93.2% with the Support Vector Machines classifier, between 67.5 - 90 minutes of driving at the 4th grade fatigue level.

In this study, the dual-channel photonic crystal fibre (PCF) based surface plasmon resonance (SPR) sensor has been proposed to sense multi-analyte. Gold and silver used as plasmonic layers. Full vectorial Finite Element Method (FV-FEM) have been used to perform numerical analysis. Also, the spectral interrogation method has been used to calculate the sensitivity of the proposed sensor structure. According to numerical results, the highest sensitivity levels are obtained as 6800 nm/RIU with 1.47x10−5 RIU, where the average sensitivities are calculated as 3500 nm/RIU and 3100 nm/RIU for fixed and variable refractive indices, respectively.

The importance and need for respirators has once again shown itself due to the coronavirus (Covid-19) epidemic, which has recently spread around the world and has been declared as a Pandemic (epidemic worldwide) by the world health organization. Monitoring respiratory activity plays a vital role in detecting respiratory diseases such as Chronic Obstructive Pulmonary Disease (COPD). In this study, a working group of 6 participants was formed to measure respiratory parameters. Each individual was provided to perform FVC, VC, MVV, RR and TV performance. The measurements were carried out simultaneously with the medical spirometer device. A total of 1860 data (1500 data for VC, 360 data for MVV) were sampled and all data were analyzed in Matlab program. It was observed that the results obtained were quite similar to each other (RMVV = 0.998 for MVV; RVC = 0.997 for VC). One of the most important contributions of this study is that the measured data can be sent to the computer and saved to the SD card. Thus, with the thermal printers in standard spirometers, paper wastage was prevented and the data was stored in digital environment. The developed system offers a practical and low cost solution. The developed device is expected to take an important place in biomedical device technology with its ability to measure COPD parameters.

In this study, constant voltage control of a secondary side controlled wireless power transfer (WPT) converter is evaluated based on efficiency. The LC/S compensation network which has perfect constant current characteristic is used in the design of the WPT converter. In order to improve the constant voltage performance of the LC/S compensation network, controlled rectifier is adapted to the converter. The phase shift modulation (PSM) control is applied to the switches of the rectifier as independent of the primary side. Thus, constant voltage regulation is achieved at constant operation frequency in a wide load range. The performance of the proposed converter is verified by a simulation work at 2.5 kW output power and 450 V output voltage. The efficiency values of the converter, as function of the load condition, is extracted by simulation and compared to frequency modulation control. In addition, power loss distribution and its comparison with frequency modulation is also discussed. The maximum efficiency of the converter is obtained 96.4% at full load condition.

In this study, the metal tube is heated up to a certain temperature by induction heating method so that the hydrogen stored by metal hydride method in it can be discharged. The voltage fed series resonant inverter (SRI) is used in the power stage of the system and the power switches are turned on under soft-switching conditions. The closed-loop controlled 300 W SRI is designed to set the temperature of the tube to the reference temperature of 250 °C. The power control of SRI is realized by frequency control. The temperature of the tube is controlled by hysteresis on-off control due to its simple structure and easy applicability. 16-bit DSPIC33FJ16GS502 is used in the control of the system.

Random Fourier features (RFF) provide one of the most prominent means for nonlinear classification in especially large scale data settings. However, considering the original proposal of RFF, Fourier features are randomly drawn from a certain distribution and used unoptimized. In this paper, we investigate Fourier features via a single hidden layer feedforward neural network (SLFN) and optimize, i.e., learn, those features (instead of drawing randomly). The learned Fourier features are deduced from the radial basis function (rbf kernel), and implemented in the hidden layer of the SLFN which is followed by the output layer. We present extensive experiments with 10 different classification datasets from various fields, e.g., bioinformatics. The learning of Fourier features is observed to be highly superior over the competing techniques such as perceptron in the rbf kernel space or a greedy forward feature selection strategy. On the other hand, the Fourier feature learning performs comparably with SVM (support vector machines with rbf kernel) while providing substantial computational benefits, and this is even without using the max margin regularization. Moreover, when tested in wireless mesh networks, the SLFN delivers promising smart steering capabilities.

In this study, it is aimed to design discrete time PI, PID and PIR controllers with the dominant pole region assignment method in order to have time domain characteristics of the closed loop system in the desired interval. First of all, determination of the boundary functions for the regions where the dominant and non-dominant poles of the closed-loop system are desired to be located are explained. Here, for the boundary functions, the region where the dominant poles are desired to be located are two circles of constant radius and a constant damping ratio curve in the discrete time domain, and the region where the remaining poles are desired to be located is a circle of constant radius. Then, solution method of dominant pole region assignment problem is given for discrete PI controller. The proposed method is extended for discrete PID and PIR controllers by fixing a parameter of the controller (Kp=kp*). In addition, the design starts with selecting the delay parameter h as a positive integer in the PIR controller. The proposed method is demonstrated for discrete PI, PID and PIR controllers via two systems.

In this study, the new analysis of the introduced hyperchaotic equation set was handled. The equation set was firstly analyzed mathematically and then the results were proven by designing a more efficient circuit with active elements. The aim of the study is offering an effective secure communication application and random number generator application. Hence, based on the new analysis of equation set, secure communication system and random number generation application were proposed. Accordingly, creating a Pseudorandom Number Generator is the halfway house in this study. The signals received from the chaotic oscillator were sampled at low frequency and with a simple post-processing, a bit stream was created. The resulting bit stream passed the NIST test successfully. The other halfway of the study is creating a secure communication system by the synchronization of two chaotic oscillators that are in transmitter and receiver. An identical noise-like signals are generated in both transmitter and receiver parts. At the transmitter part adding a noise-like chaotic signal to the information was done. At the receiver, this same noise-like signal is subtracted from the perceived signal. Thus, the information can be transmitted securely. Spice simulations of both proposed applications have been made and it has been shown that they are compatible with mathematical analysis. The proposed circuits are suitable for realization with commercially available circuit elements.

Turkey's renewable energy sources (RES) can be utilized to increase the amount of electricity generation and significantly reduce emissions from the energy sector. Voluntary carbon markets encourage electricity generation using RES and make greater use of these sources. In this study, the amount of emission reduction and carbon credits to be obtained from the renewable power plants to be connected to the Turkish electricity grid and the revenue to be obtained from carbon credit trading have been calculated. Emission reduction amount has been calculated by using combined margin CO2 emission factors. The amount of CO2 emission reduction that can be achieved through the renewable electricity generation between 2021-2024 is estimated as 454.94 MtCO2. Voluntary carbon credit revenue that can be obtained for the period between 2016 and 2024 is $1.116 billion. A very small part of the carbon reduction potential generated by renewable projects has been the subject of trade through voluntary carbon certificate issuing organizations. Turkey's rich RES potential has not been adequately utilized. The volume of carbon offsets that Turkey can generate by renewable projects is considerably high.

B-Tree based text indexes used in MongoDB are slow compared to different structures such as inverted indexes. In this study, it has been shown that the full-text search speed can be increased significantly by indexing a structure in which each different word in the text is included only once. The Multi-Stream Word-Based Compression Algorithm (MWCA), developed in our previous work, stores word dictionaries and data in different streams. While adding the documents to a MongoDB collection, they were encoded with MWCA and separated into six different streams. Each stream was stored in a different field, and three of them containing unique words were used when creating a text index. In this way, the index could be created in a shorter time and took up less space. It was also seen that Snappy and Zlib block compression methods used by MongoDB reached higher compression ratios on data encoded with MWCA. Search tests on text indexes created on collections using different compression options shows that our method provides 19 to 146 times speed increase and 34% to 40% less memory usage. Tests on regex searches that do not use the text index also shows that the MWCA model provides 7 to 13 times speed increase and 29% to 34% less memory usage.

Recently, there has been an increasing interest in the use of artificial intelligence techniques in different fields. In this work is aimed to use different machine learning algorithms (MLA) to predict the elongation at break from the mechanical properties of cable sheath materials in halogen-free flame retardant (HFFR) cables. In order to be used in the developed prediction models, tensile test was applied to the samples and the percent elongation values were determined. Obtained experimental results were used in different artificial intelligence prediction models. Absolute percentage errors of support vector machine (SVM) and artificial neural network (ANN) methods were obtained at a quite acceptable level with a limited number of data obtained from HFFR cable samples. The estimations obtained by these methods were compared with the data of the estimations obtained by performing regression analysis with the MS Excel program. According to the statistical results, with the use of SVM and ANN in this area, the successful prediction rate was 87.5%, and the average success rate for the predictions made was 92%. The use of MEA in this area will largely end the uncertainty in the trial and error production and reduce the rate of unsuccessful production.

Weather has a significant impact on human life and activities. As abrupt changes in air temperature negatively affect daily life and various industries, the importance of weather forecast accuracy is increasing day by day. Current weather forecasting methods can be divided into two main groups: numerical-based and machine learning-based approaches. Numerical-based weather forecasting methods use complex mathematical formulas that significantly increase the computational cost. On the other hand, machine learning-based methods have been preferred more in recent years due to their lower computational costs. In this study, the next day's maximum and minimum air temperature are estimated for Seoul, South Korea by using 12 different regression methods together with the boosting-based machine learning algorithms developed in recent years, as well as traditional machine learning methods. Furthermore, since tuning of hyperparameters affects the process time and performance of machine learning algorithms, all 12 methods have been extensively studied in terms of time and hyperparameters. The square correlation coefficient (R^2), which is frequently adopted in the literature, is used to compare the performances of the methods. According to the observed results, the boosting-based XGBoost and LightGBM methods are the most successful machine learning algorithms in predicting the maximum and minimum air temperature for all years with both statistical test analysis and the highest R^2 score

Copy-move forgery, in which copied a region of the image and pasted onto another region on the same image, is the most encountered image forgery technique recently. Many frameworks have been presented to detect such forgeries. The main drawback with these approaches is their performance can be degraded when the duplicated image has undergone some attacks. In this work, it is aimed to propose a hybrid approach, which uses deep features and DCT-based block features in a combined manner, to achieve higher detection performance even if under various attack scenarios. The proposed method uses a global contrast correction technique called LDR during the preprocessing phase and then extracts deep features from the image patches using a deep neural network. The method also obtains block features from the image to robustness against JPEG compression attacks. Hybrid features (deep and block-based features) are matched using Patch Match and then the proposed post-processing operation is realized on the matching results to minimize false matches. According to empirical studies performed on available databases, the proposed scheme gives better results when compared to both keypoint-based and block-based references even under attacks with challenging parameters.

Inspection of the reliability of videos is one of the most important issues in recent years. In order to determine the reliability of the videos, the changes made on them should be investigated. Editing can be made on the videos to hide an event or an object. In this study, a new method for the detection of duplication fraud, between video frames is proposed. The developed method consists of three stages. First of all, feature vectors are obtained with the edge information extracted from the frames. Then, similarity analysis is performed based on the correlation information between the frame groups of feature vectors. Finally, the false positive locations obtained are eliminated. A new method that can automatically decide whether the video is fake or original with high accuracy has been contributed to the literature.

Today, as a result of the inadequacies of the standard sentiment analysis, aspect-based sentiment analysis (ABSA) studies have great attracting interest. ABSA reveals detailed sentiment and opinion about every term/attribute in a text. The most important sub-stage of the ABSA method is the process of extracting the aspect terms from a text. This process becomes more difficult in texts with agglutinative language structures such as Turkish. In this study, we proposed an ensemble approach that uses statistical (TF-IDF), topic modeling (LDA and NMF), and rule-based methods together to extract aspect terms from Turkish user comments. The proposed method strategically combines the candidate aspect term obtained by different methods and determines the final aspect term lists. The proposed method has been tested on the SemEval-2016 ABSA benchmarking dataset, which consists of Turkish restaurant reviews. The experimental results of the proposed method were compared with previous studies on the same dataset.