The objective of this thesis is to investigate the possibility of using machine learn- ing at Statistics Sweden within the Foreign Trade in Services (FTS) statistic, to predict the likelihood of a unit to conduct foreign trade in services. The FTS survey is a sample survey, for which there is no natural frame to sample from. Therefore, prior to sampling a frame is manually constructed each year, starting with a register of all Swedish companies and agencies and in a rule- based manner narrowing it down to contain only what is classified as units likely to trade in services during the year to come. An automatic procedure that would enable reliable predictions is requested. To this end, three different machine learning methods have been analyzed, two rule- based methods (random forest and extreme gradient boosting) and one distance- based method (k nearest neighbors). The models arising from these methods are trained and tested on historically sampled units, for which it is known whether they did trade or not. The results indicate that the two rule-based methods perform well in classifying likely traders. The random forest model is better at finding traders, while the extreme gradient boosting model is better at finding non-traders. The results also indicate interesting patterns when studying different metrics for the models. The results also indicate that when training the rule-based models, the year in which the training data was sampled needs to be taken into account. This entails that cross-validation with random folds should not be used, but rather grouped cross-validation based on year. By including a feature that mirror the state of the economy, the model can adapt its rules to this, meaning that the rules learned on training data can be extended to years beyond training data. Based on the observed results, the final recommendation is to further develop and investigate the performance of the random forest model.

The thesis deals with a design and an implementation of software for a detection of logopaedia defects of speech. Due to the need of early logopaedia defects detecting, this software is aimed at a child’s age speaker. The introductory part describes the theory of speech realization, simulation of speech realization for numerical processing, phonetics, logopaedia and basic logopaedia defects of speech. There are also described used methods for feature extraction, for segmentation of words to speech sounds and for features classification into either correct or incorrect pronunciation class. In the next part of the thesis there are results of testing of selected methods presented. For logopaedia speech defects recognition algorithms are used in order to extract the features MFCC and PLP. The segmentation of words to speech sounds is performed on the base of Differential Function method. The extracted features of a sound are classified into either a correct or an incorrect pronunciation class with one of tested methods of pattern recognition. To classify the features, the k-NN, SVN, ANN, and GMM methods are tested.

This research study focuses on the application models of Data Mining and Machine Learning covering cellular network traffic, in the objective to arm Mobile Network Operators with full view of performance branches (Services, Device, Subscribers). The purpose is to optimize and minimize the time to detect service and subscriber patterns behaviour. Different data mining techniques and predictive algorithms have been applied on real cellular network datasets to uncover different data usage patterns using specific Key Performance Indicators (KPIs) and Key Quality Indicators (KQI). The following tools will be used to develop the concept: RStudio for Machine Learning and process visualization, Apache Spark, SparkSQL for data and big data processing and clicData for service Visualization. Two use cases have been studied during this research. In the first study, the process of Data and predictive Analytics are fully applied in the field of Telecommunications to efficiently address users’ experience, in the goal of increasing customer loyalty and decreasing churn or customer attrition. Using real cellular network transactions, prediction analytics are used to predict customers who are likely to churn, which can result in revenue loss. Prediction algorithms and models including Classification Tree, Random Forest, Neural Networks and Gradient boosting have been used with an exploratory Data Analysis, determining relationship between predicting variables. The data is segmented in to two, a training set to train the model and a testing set to test the model. The evaluation of the best performing model is based on the prediction accuracy, sensitivity, specificity and the Confusion Matrix on the test set. The second use case analyses Service Quality Management using modern data mining techniques and the advantages of in-memory big data processing with Apache Spark and SparkSQL to save cost on tool investment; thus, a low-cost Service Quality Management model is proposed and analyzed. With increase in Smart phone adoption, access to mobile internet services, applications such as streaming, interactive chats require a certain service level to ensure customer satisfaction. As a result, an SQM framework is developed with Service Quality Index (SQI) and Key Performance Index (KPI). The research concludes with recommendations and future studies around modern technology applications in Telecommunications including Internet of Things (IoT), Cloud and recommender systems. / Cellular networks have evolved and are still evolving, from traditional GSM (Global System for Mobile Communication) Circuit switched which only supported voice services and extremely low data rate, to LTE all Packet networks accommodating high speed data used for various service applications such as video streaming, video conferencing, heavy torrent download; and for say in a near future the roll-out of the Fifth generation (5G) cellular networks, intended to support complex technologies such as IoT (Internet of Things), High Definition video streaming and projected to cater massive amount of data. With high demand on network services and easy access to mobile phones, billions of transactions are performed by subscribers. The transactions appear in the form of SMSs, Handovers, voice calls, web browsing activities, video and audio streaming, heavy downloads and uploads. Nevertheless, the stormy growth in data traffic and the high requirements of new services introduce bigger challenges to Mobile Network Operators (NMOs) in analysing the big data traffic flowing in the network. Therefore, Quality of Service (QoS) and Quality of Experience (QoE) turn in to a challenge. Inefficiency in mining, analysing data and applying predictive intelligence on network traffic can produce high rate of unhappy customers or subscribers, loss on revenue and negative services’ perspective. Researchers and Service Providers are investing in Data mining, Machine Learning and AI (Artificial Intelligence) methods to manage services and experience. This research study focuses on the application models of Data Mining and Machine Learning covering network traffic, in the objective to arm Mobile Network Operators with full view of performance branches (Services, Device, Subscribers). The purpose is to optimize and minimize the time to detect service and subscriber patterns behaviour. Different data mining techniques and predictive algorithms will be applied on cellular network datasets to uncover different data usage patterns using specific Key Performance Indicators (KPIs) and Key Quality Indicators (KQI). The following tools will be used to develop the concept: R-Studio for Machine Learning, Apache Spark, SparkSQL for data processing and clicData for Visualization. / Electrical and Mining Engineering / M. Tech (Electrical Engineering)

Le travail présenté dans cette thèse a pour objectif de proposer et d'évaluer les performances de différentes techniques de suppression de bruit impulsif de type asynchrone adaptées aux transmissions sur courants porteurs en lignes (CPL) indoor. En effet, outre les caractéristiques physiques spécifiques à ce type de canal de transmission, le bruit impulsif asynchrone reste la contrainte sévère qui pénalise les systèmes CPL en terme de QoS. Pour remédier aux dégradations dues aux bruits impulsifs asynchrones, les techniques dites de retransmission sont souvent très utilisées. Bien qu'elles soient efficaces, ces techniques de retransmission conduisent à une réduction de débit et à l’introduction de délais de traitement supplémentaires pouvant être critiques pour des applications temps réel. Par ailleurs, plusieurs solutions alternatives sont proposées dans la littérature pour minimiser l'impact du bruit impulsif sur les transmissions CPL. Cependant, le nombre de techniques, qui permettent d'obtenir un bon compromis entre capacité de correction et complexité d'implantation reste faible pour les systèmes CPL. Dans ce contexte, nous proposons dans un premier temps d'utiliser un filtre linéaire adaptatif : le filtre de Widrow, nommé aussi ADALINE (ADAptive LInear NEuron), que nous utilisons comme méthode de débruitage pour les systèmes CPL. Pour améliorer les performances du débruitage effectué à l'aide d'ADALINE, nous proposons d'utiliser un réseau de neurones (RN) non linéaire comme méthode de débruitage. Le réseau de neurones est un bon outil qui est une généralisation de la structure du filtre ADALINE. Dans un deuxième temps, pour améliorer les performances du débruitage par un réseau de neurones, nous proposons un procédé d'annulation du bruit impulsif constitué de deux algorithmes : EMD (Empirical Mode Decomposition) associé à un réseau de neurones de type perceptron multicouches. L'EMD effectue le prétraitement en décomposant le signal bruité en signaux moins complexes et donc plus facilement analysables. Après quoi le réseau de neurones effectue le débruitage. Enfin, nous proposons une méthode d'estimation du bruit impulsif utilisant la méthode GPOF (Generalized Pencil Of Function). L'efficacité des deux méthodes, EMD-RN et la technique utilisant l'algorithme GPOF, est évaluée en utilisant une chaîne de simulation de transmission numérique compatible avec le standard HPAV. / The aim of our thesis is to propose and to evaluate the performances of some asynchronous impulsive noise mitigation techniques for transmission over indoor power lines. Indeed, besides the particular physical properties that characterize this transmission channel type, asynchronous impulsive noise remains the difficult constraint to overcome on power lines communications (PLC). Usually, the impact of asynchronous impulsive disturbances over power lines is partly compensated by means of retransmission mechanisms. However, the main drawbacks of the use of retransmission solutions for impulsive noise mitigation are the bitrate loss and the induced time delays that may be prohibitive for real-time services. Although several other countering strategies are proposed in the literature, only very few of them have a good compromise between correction capability and implementing complexity for PLC systems. In this context, we proposed an adaptive linear filter, the Widrow filter, also known as ADALINE (Adaptive LInear neurons), as a denoising method for PLC systems. To improve the performance of the denoising method using ADALINE, we proposed to use a neural network (NN) as a nonlinear denoising method. The neural network is a good generalization of the ADALINE filter. In a second step, to improve the performances of denoising by NN, we proposed a combined denoising method based on EMD (Empirical Mode Decomposition) and MLPNN (Multi Layer Perceptron Neural Network). The noised signal is pre-processed by EMD which decomposes it into signals less complex and therefore more easily analyzed. Then the MLPNN denoises it. Finally, we proposed an asynchronous impulsive noise estimation method using the GPOF method (Generalized Pencil Of Function). The performances of the two methods, EMD-MLPNN and GPOF technique, are evaluated using a PLC transmission chain compatible with the HPAV standard.

Research work presented within this thesis propose a neoteric variant of deep convolutional neural network architecture, CondenseNeXt, designed specifically for ARM-based embedded computing platforms with constrained computational resources. CondenseNeXt is an improved version of CondenseNet, the baseline architecture whose roots can be traced back to ResNet. CondeseNeXt replaces group convolutions in CondenseNet with depthwise separable convolutions and introduces group-wise pruning, a model compression technique, to prune (remove) redundant and insignificant elements that either are irrelevant or do not affect performance of the network upon disposition. Cardinality, a new dimension to the existing spatial dimensions, and class-balanced focal loss function, a weighting factor inversely proportional to the number of samples, has been incorporated in order to relieve the harsh effects of pruning, into the design of CondenseNeXt’s algorithm. Furthermore, extensive analyses of this novel CNN architecture was performed on three benchmarking image datasets: CIFAR-10, CIFAR-100 and ImageNet by deploying the trained weight on to an ARM-based embedded computing platform: NXP BlueBox 2.0, for real-time image classification. The outputs are observed in real-time in RTMaps Remote Studio’s console to verify the correctness of classes being predicted. CondenseNeXt achieves state-of-the-art image classification performance on three benchmark datasets including CIFAR-10 (4.79% top-1 error), CIFAR-100 (21.98% top-1 error) and ImageNet (7.91% single model, single crop top-5 error), and up to 59.98% reduction in forward FLOPs compared to CondenseNet. CondenseNeXt can also achieve a final trained model size of 2.9 MB, however at the cost of 2.26% in accuracy loss. Thus, performing image classification on ARM-Based computing platforms without requiring a CUDA enabled GPU support, with outstanding efficiency.<br>

The thesis shows the historical development of the South Moravian electricity system and the development of Brno´s network is discussed in more detail, especially the current form of the network from the outskirts of the city to its historical centre. The lattice network, as a subject of the analysis, is described in terms of the operation and reliability of the power supply. It also informs about the gradual development of the technology applied in the lattice network Brno-střed. The practical part performs the analysis of theoretical data and recorded data with respect to steady-state and faulty-state operation of the lattice network Brno-střed. The theoretical analysis is based on the simulation and calculation of the lattice network model in the software PAS DAISY Bizon and the monitored parameter is the transformers power load. In addition, the analysis of the real data includes the assessment of the difference average phase current values, as well as voltage, between the transformer phases. There is also the evaluation of the energy flow from the low voltage side to the high voltage side of the analysed network.

This thesis deals with analysis of T-wave Alternans (TWA), periodical changes of T wave in ECG signal. Presence of these alternans may predict higher risk of sudden cardiac death. From the several possible ways of TWA classification, the training algorithms of self organizing maps are used in this thesis. Result of the thesis is a program, which in the first step detects QRS complexes in the signal. Then, in the next step, gained reference points are used for T-waves detection. Detected waves are represented by a vector of significant points, which is used as an input for artificial neural network. Final output of the program is a decision about presence of TWA in the signal and its rate.

As it is already apparent from a title of my work, I will be engaged in proposal of container VN/NN distributor using components from company ABB. This work will be after final making also used by company ABB to simplify procedure of proposal of any kind of container distributor. In my labor I will be generally deal with individual products, which will after that be used in distributor. Next I will be engaged in particular making of distributor, which should be made. The distributor will be designed for a separate part of the transformer and for a transformer as a unit, which is not separated from substation.

There is a difference between the raw sensory input to the brain and our stable perception of entities in the environment. A first approach to investigate perception is to study relationships between properties of currently presented stimuli and biological correlates of perceptual processes. However, it is known that such processes are not only dependent on the current stimulus. Sampling of information and the concurrent neuronal processing of stimulus content rely on contextual relationships in the environment, and between the environment and the body. Perceptual processes dynamically adjust to relevant context, such as the current task of the organism and its immediate history. To understand perception, we have to study how processing of current stimulus content is influenced by such contextual factors. This thesis investigates the influence of such factors on visual processing. In particular, it investigates effects of temporal context in early visual processing and the effect of task context in eye movement control. To investigate effects of contextual factors on early visual processing of current stimulus content, we study neuronal processing of visual information in the primary visual cortex. We use real-time optical imaging with voltage sensitive dyes to capture neuronal population activity in the millisecond range across several millimeters of cortical area. To characterize the cortical layout concerning the mapping of orientation, previous to further investigations, we use smoothly moving grating stimuli. Investigating responses to this stimulus type systematically, we find independent encoding of local contrast and orientation, and a direct mapping of current stimulus content onto cortical activity (Study 1). To investigate the influence of the previous stimulus as context on processing of current stimulus content, we use abrupt visual changes in sequences of modified natural images. In earlier studies, investigating relatively fast timescales, it was found that the primary visual cortex continuously represents current input (ongoing encoding), with little interference from past stimuli. We investigate whether this coding scheme generalizes to cases in which stimuli change more slowly, as frequently encountered in natural visual input. We use sequences of natural scene contours, comprised of vertically and horizontally filtered natural images, their superpositions, and a blank stimulus, presented with 10 or 33 Hz. We show that at the low temporal frequency, cortical activity patterns do not encode the present orientations but instead reflect their relative changes in time. For example, when a stimulus with horizontal orientation is followed by the superposition of both orientations, the pattern of cortical activity represents the newly added vertical orientations instead of the full sum of orientations. Correspondingly, contour removal from the superposition leads to the representation of orientations that have disappeared rather than those that remain. This is in sharp contrast to more rapid sequences for which we find an ongoing representation of present input, consistent with earlier studies. In summary, we find that for slow stimulus sequences, populations of neurons in the primary visual cortex are no longer tuned to orientations within individual stimuli but instead represent the difference between consecutive stimuli. Our results emphasize the influence of the temporal context on early visual processing and consequentially on information transmission to higher cortical areas (Study 2). To study effects of contextual factors on the sampling of visual information, we focus on human eye movement control. The eyes are actively moved to sample visual information from the environment. Some traditional approaches predict eye movements solely on simple stimulus properties, such as local contrasts (stimulus-driven factors). Recent arguments, however, emphasize the influence of tasks (task context) and bodily factors (spatial bias). To investigate how contextual factors affect eye movement control, we quantify the relative influences of the task context, spatial biases and stimulus-driven factors. Participants view and classify natural scenery and faces while their eye movements are recorded. The stimuli are composed of small image patches. For each of these patches we derive a measure that quantifies stimulus-driven factors, based on the image content of a patch, and spatial viewing biases, based on the location of the patch. Utilizing the participants’ classification responses, we additionally derive a measure, which reflects the information content of a patch in the context of a given task. We show that the effect of spatial biases is highest, that task context is a close runner-up, and that stimulus-driven factors have, on average, a smaller influence. Remarkably, all three factors make independent and significant contributions to the selection of viewed locations. Hence, in addition to stimulus-driven factors and spatial biases, the task context contributes to visual sampling behavior and has to be considered in a model of human eye movements. Visual processing of current stimulus content, in particular visual sampling behavior and early processing, is inherently dependent on context. We show that already in the first cortical stage, temporal context strongly affects the processing of new visual information and that visual sampling by eye movements is significantly influenced by the task context, independently of spatial factors and stimulus-driven factors. The empirical results presented provide foundations for an improved theoretical understanding of the role of context in perceptual processes.

In this thesis, the robustness of deep learning techniques in the field of visual odometry is investigated, with a specific focus on the impact of motion blur. A comparative study is conducted, evaluating the performance of state-of-the-art deep convolutional neural network methods, namely DF-VO and DytanVO, against ORB-SLAM3, a well-established non-deep-learning technique for visual simultaneous localization and mapping. The objective is to quantitatively assess the performance of these models as a function of motion blur. The evaluation is carried out on a custom synthetic dataset, which simulates a camera navigating through a forest environment. The dataset includes trajectories with varying degrees of motion blur, caused by camera translation, and optionally, pitch and yaw rotational noise. The results demonstrate that deep learning-based methods maintained robust performance despite the challenging conditions presented in the test data, while excessive blur lead to tracking failures in the geometric model. This suggests that the ability of deep neural network architectures to automatically learn hierarchical feature representations and capture complex, abstract features may enhance the robustness of deep learning-based visual odometry techniques in challenging conditions, compared to their geometric counterparts.