CoordenaÃÃo de AperfeiÃoamento de Pessoal de NÃvel Superior / This dissertation reports applications of artificial neural networks to detect stator winding interturn fault of three phase induction motor drived by frequency inverter. The artificial neural networks, like Simple and Multilayer Perceptron, served as off-line classifiers to short-circuit fault condition or healthy condition. In the training of Multilayer Perceptron, two different algorithms are used: the error back-propagation, which is a classic algorithm, and the extreme learning machine, as a relative new alternative for the classic back-propagation. The new one is more worthwhile because of its implementation easiness and higher speed of computation. The database used on the training and validation of the networks is created from an experimental setting, therefore it is composed by true data. The attributes used as failuresâ indicators are selected from certain frequencies of the spectrum, based on some theories of current signature analysis. In the second instance, the technique of principal components analysis is employed. The results obtained for the designed classifiers are shown, and some considerations are made on their use in real time embedded applications, which is the most important projection for future researches. / Este trabalho deriva da aplicaÃÃo de redes neurais artificiais para a detecÃÃo de curto-circuito entre espiras em motor de induÃÃo trifÃsico, acionado por inversor de frequÃncia. As redes neurais artificiais, do tipo Perceptron Simples e Multicamadas, sÃo usadas para detectar falhas de curto-circuito no bobinamento estatÃrico de motores de induÃÃo trifÃsicos de forma off-line. Para treinamento do Perceptron Multicamadas sÃo usados dois algoritmos distintos: o error back-propagation, que figura como o algoritmo clÃssico na literatura especializada, e o extreme learning machine, que à uma alternativa, relativamente recente, ao algoritmo clÃssico. Este algoritmo à uma opÃÃo atraente para o desenvolvimento rÃpido de classificadores. O banco de dados usado para treinamento e validaÃÃo das redes à obtido a partir de experimentaÃÃo laboratorial, portanto composto de dados reais. Os atributos utilizados para a detecÃÃo da falha sÃo componentes de frequÃncia do espectro harmÃnico da corrente estatÃrica do motor. O critÃrio de escolha destas componentes, a priori, à fundamentado em resultados de investigaÃÃes prÃvias da assinatura de corrente e, em segunda instÃncia, à aplicada a tÃcnica de anÃlise de componentes principais. SÃo apresentados os resultados obtidospelos classificadores projetados, e feitas algumas consideraÃÃes quanto à utilizaÃÃo destes em aplicaÃÃo embarcada e em tempo real, que à a principal projeÃÃo de futuros trabalhos a partir do atual.

O objetivo deste trabalho foi desenvolver um algoritmo baseado em redes neurais para o reconhecimento facial. O algoritmo contém dois módulos principais, um módulo para a extração de características e um módulo para o reconhecimento facial, sendo aplicado sobre imagens digitais nas quais a face foi previamente detectada. O método utilizado para a extração de características baseia-se na aplicação de assinaturas horizontais e verticais para localizar os componentes faciais (olhos e nariz) e definir a posição desses componentes. Como entrada foram utilizadas imagens faciais de três bancos distintos: PICS, ESSEX e AT&T. Para esse módulo, a média de acerto foi de 86.6%, para os três bancos de dados. No módulo de reconhecimento foi utilizada a arquitetura perceptron multicamadas (MLP), e para o treinamento dessa rede foi utilizado o algoritmo de aprendizagem backpropagation. As características faciais extraídas foram aplicadas nas entradas dessa rede neural, que realizou o reconhecimento da face. A rede conseguiu reconhecer 97% das imagens que foram identificadas como pertencendo ao banco de dados utilizado. Apesar dos resultados satisfatórios obtidos, constatou-se que essa rede não consegue separar adequadamente características faciais com valores muito próximos, e portanto, não é a rede mais eficiente para o reconhecimento facial / This research presents a facial recognition algorithm based in neural networks. The algorithm contains two main modules: one for feature extraction and another for face recognition. It was applied in digital images from three database, PICS, ESSEX and AT&T, where the face was previously detected. The method for feature extraction was based on previously knowledge of the facial components location (eyes and nose) and on the application of the horizontal and vertical signature for the identification of these components. The mean result obtained for this module was 86.6% for the three database. For the recognition module it was used the multilayer perceptron architecture (MLP), and for training this network it was used the backpropagation algorithm. The extracted facial features were applied to the input of the neural network, that identified the face as belonging or not to the database with 97% of hit ratio. Despite the good results obtained it was verified that the MLP could not distinguish facial features with very close values. Therefore the MLP is not the most efficient network for this task

The goal of this master’s thesis is to identify and evaluate data mining algorithms which are commonly implemented in modern Medical Decision Support Systems (MDSS). They are used in various healthcare units all over the world. These institutions store large amounts of medical data. This data may contain relevant medical information hidden in various patterns buried among the records. Within the research several popular MDSS’s are analyzed in order to determine the most common data mining algorithms utilized by them. Three algorithms have been identified: Naïve Bayes, Multilayer Perceptron and C4.5. Prior to the very analyses the algorithms are calibrated. Several testing configurations are tested in order to determine the best setting for the algorithms. Afterwards, an ultimate comparison of the algorithms orders them with respect to their performance. The evaluation is based on a set of performance metrics. The analyses are conducted in WEKA on five UCI medical datasets: breast cancer, hepatitis, heart disease, dermatology disease, diabetes. The analyses have shown that it is very difficult to name a single data mining algorithm to be the most suitable for the medical data. The results gained for the algorithms were very similar. However, the final evaluation of the outcomes allowed singling out the Naïve Bayes to be the best classifier for the given domain. It was followed by the Multilayer Perceptron and the C4.5.

In this thesis the predictive models the multilayer perceptron and random forest are evaluated to predict cloud top pressure. The dataset used in this thesis contains brightness temperatures, reflectances and other useful variables to determine the cloud top pressure from the Advanced Very High Resolution Radiometer (AVHRR) instrument on the two satellites NOAA-17 and NOAA-18 during the time period 2006-2009. The dataset also contains numerical weather prediction (NWP) variables calculated using mathematical models. In the dataset there are also observed cloud top pressure and cloud top height estimates from the more accurate instrument on the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite. The predicted cloud top pressure is converted into an interpolated cloud top height. The predicted pressure and interpolated height are then evaluated against the more accurate and observed cloud top pressure and cloud top height from the instrument on the satellite CALIPSO. The predictive models have been performed on the data using different sampling strategies to take into account the performance of individual cloud classes prevalent in the data. The multilayer perceptron is performed using both the original response cloud top pressure and a log transformed repsonse to avoid negative values as output which is prevalent when using the original response. Results show that overall the random forest model performs better than the multilayer perceptron in terms of root mean squared error and mean absolute error.

Konfliktprediktering handlar om att bedöma risken för våld i ett geografiskt område vid en given tid. Uppgiften lämpar sig bra för datorer som med hjälp av matematiska modeller kan hitta mönster i stora mängder data. Att prediktera konflikthändelser går att göra med olika metoder. Syftet med studien var att utvärdera multilayer perceptron (MLP), en typ av artificiella neuronnät, som metod för konfliktprediktering i relation till två andra metoder. I studien beskrivs hur MLP-neuronnätet konstruerades och hur prestationsmått togs fram för dess prediktioner. De värdena jämfördes senare med prestationsmått från andra studier för de två andra metoderna. Prediktionerna grundade sig på data om konflikthändelser, samt ekonomiska och demografiska faktorer för länder i världen. Jämförelsen visade att MLP är användbar som metod för konfliktprediktering och hade, under de förutsättningar som rådde, i viktiga avseenden högre prediktiv förmåga än de andra metoderna. Studien presenterar även fyra faktorer som kan påverka vilken modelleringsmetod som en modellerare borde använda för konfliktprediktering.

No description available.

L’objectif de ces travaux de doctorat est de développer un modèle prédictif capable de prévoir correctement les concentrations en polluants du jour pour le lendemain en Corse. Nous nous sommes intéressés aux PM10 et à l’ozone, les deux polluants les plus problématiques sur l’île. Le modèle devait correspondre aux contraintes d’un usage opérationnel au sein d’une petite structure, comme Qualitair Corse, l’association locale de surveillance de la qualité de l’air.La prévision a été réalisée à l’aide de réseaux de neurones artificiels. Ces modèles statistiques offrent une grande précision tout en nécessitant peu de ressources informatiques. Nous avons choisi le Perceptron MultiCouche (PMC), avec en entrée à la fois des mesures de polluants, des mesures météorologiques, et des sorties de modèles de chimie-transport (CHIMERE via la plate-forme AIRES) et de modèles météorologiques (AROME).La configuration des PMC a été optimisée avant leur apprentissage automatique, en conformité avec le principe de parcimonie. Pour en améliorer les performances, une étude de sélection de variables a été au préalable menée. Nous avons comparé l’usage d’algorithmes génétiques, de recuits simulés et d’analyse en composantes principales afin d’optimiser le choix des variables d’entrées. L’élagage du PMC a été également mis en œuvre.Nous avons ensuite proposé un nouveau type de modèle hybride, combinaison d’un classifieur et de plusieurs PMC, chacun spécialisé sur un régime météorologique particulier. Ces modèles, qui demandent un large historique de données d’apprentissage, permettent d’améliorer la prévision des valeurs extrêmes et rares, correspondant aux pics de pollution. La classification non-supervisée a été menée avec des cartes auto-organisatrices couplées à l’algorithme des k-means, ainsi que par classification hiérarchique ascendante. L’analyse de sensibilité à été menée grâce à l’usage de courbes ROC.Afin de gérer les jeux de données utilisés, de mener les expérimentations de manière rigoureuse et de créer les modèles destinés à l’usage opérationnel, nous avons développé l’application « Aria Base », fonctionnant sous Matlab à l’aide de la Neural Network Toolbox.Nous avons également développé l’application « Aria Web » destinée à l’usage quotidien à Qualitair Corse. Elle est capable de mener automatiquement les prévisions par PMC et de synthétiser les différentes informations qui aident la prévision rendues disponibles sur internet par d’autres organismes. / The objective of this doctoral work is to develop a forecasting model able to correctly predict next day pollutant concentrations in Corsica. We focused on PM10 and ozone, the two most problematic pollutants in the island. The model had to correspond to the constraints of an operational use in a small structure like Qualitair Corse, the local air quality monitoring association.The prediction was performed using artificial neural networks. These statistical models offer a great precision while requiring few computing resources. We chose the MultiLayer Perceptron (MLP), with input data coming from pollutants measurements, meteorological measurements, chemical transport model (CHIMERE via AIRES platform) and numerical weather prediction model (AROME).The configuration of the MLP was optimized prior to machine learning, in accordance with the principle of parsimony. To improve forecasting performances, we led a feature selection study. We compared the use of genetic algorithms, simulated annealing and principal component analysis to optimize the choice of input variables. The pruning of the MLP was also implemented.Then we proposed a new type of hybrid model, combination of a classification model and various MLPs, each specialized on a specific weather pattern. These models, which need large learning datasets, allow an improvement of the forecasting for extreme and rare values, corresponding to pollution peaks. We led unsupervised classification with self organizing maps coupled with k-means algorithm, and with hierarchical ascendant classification. Sensitivity analysis was led with ROC curves.We developed the application “Aria Base” running with Matlab and its Neural Network Toolbox, able to manage our datasets, to lead rigorously the experiments and to create operational models.We also developed the application “Aria Web” to be used daily by Qualitair Corse. It is able to lead automatically the prevision with MLP, and to synthesize forecasting information provided by other organizations and available on the Internet.

Vehicle acoustic signals have long been considered as unwanted traffic noise. In this research acoustic signals generated by each vehicle will be used to detect its presence and classify its type. Circular arrays of microphones were designed and built to detect desired signals and suppress unwanted ones. Circular arrays with multiple rings have an interesting and important property that is constant sidelobe levels. A modified genetic algorithm that can work directly with real numbers is used in the circular array design. It offers more effective ways to solve numerical problems than a standard genetic algorithm. In classifier design two main paradigms are considered: multilayer perceptrons and adaptive fuzzy logic systems. A multilayer perceptron is a network inspired by biological neural systems. Even though it is far from a biological system, it possesses the capability to solve many interesting problems in variety fields. Fuzzy logic systems, on the other hand, were inspired by human capabilities to deal with fuzzy terms. Its structures and operations are based on fuzzy set theory and its operations. Adaptive fuzzy logic systems are fuzzy logic systems equipped with training algorithms so that its rules can be extracted or modified from available numerical data similar to neural networks. Both fuzzy logic systems and multilayer perceptrons have been proved to be universal function approximators. Since there are approximations in almost every stage, both of these system types are good candidates for classification systems. In classification problems unequal learning of each class is normally encountered. This unequal learning may come from different learning difficulties and/or unequal numbers of training data from each class. The classifier tends to classify better for a well-learned class while doing poorly for other classes. Classification costs that may be different from class to class can be used to train and test a classifier. An error backpropagation algorithm can be modified so that the classification costs along with unequal learning factors can be used to control classifier learning during its training phase. / Ph. D.

At present day, distributed computing is a widely used technique, where volunteers support different computing power needs organizations might have. This thesis sought to benchmark distributed computing performance limited to mobile device support since this type of support is seldom done with mobile devices. This thesis proposes two approaches to harnessing computational power and infrastructure of a group of mobile devices. The problems used for benchmarking are small instances of deep learning training. One requirement posed by the mobile devices’ non-static nature was that this should be possible without any significant prior configuration. The protocol used for communication was HTTP. The reason deep-learning was chosen as the benchmarking problem is due to its versatility and variability. The results showed that this technique can be applied successfully to some types of problem instances, and that the two proposed approaches also favour different problem instances. The highest request rate found for the prototype with a 99% response rate was a 2100% increase in efficiency compared to a regular server. This was under the premise that it was provided just below 2000 mobile devices for only particular problem instances. / För närvarande är distribuerad databehandling en utbredd teknik, där frivilliga individer stödjer olika organisationers behov av datorkraft. Denna rapport försökte jämföra prestandan för distribuerad databehandling begränsad till enbart stöd av mobila enheter då denna typ av stöd sällan görs med mobila enheter. Rapporten föreslår två sätt att utnyttja beräkningskraft och infrastruktur för en grupp mobila enheter. De problem som används för benchmarking är små exempel på deep-learning. Ett krav som ställdes av mobilenheternas icke-statiska natur var att detta skulle vara möjligt utan några betydande konfigureringar. Protokollet som användes för kommunikation var HTTP. Anledningen till att deeplearning valdes som referensproblem beror på dess mångsidighet och variation. Resultaten visade att denna teknik kan tillämpas framgångsrikt på vissa typer av probleminstanser, och att de två föreslagna tillvägagångssätten också gynnar olika probleminstanser. Den högsta requesthastigheten hittad för prototypen med 99% svarsfrekvens var en 2100% ökning av effektiviteten jämfört med en vanlig server. Detta givet strax under 2000 mobila enheter för vissa speciella probleminstanser.

No description available.