Τα Τεχνητά Νευρωνικά Δίκτυα είναι μια επιστημονική περιοχή η οποία έχει αναπτυχθεί κατά τις τελευταίες δεκαετίες και επικαλύπτει όλες σχεδόν τις θετικές επιστήμες και την μηχανολογία. Τα Νευρωνικά Δίκτυα αποτελούνται από ένα σύνολο απλών, διασυνδεδεμένων και προσαρμοστικών μονάδων οι οποίες δημιουργούν ένα παράλληλο και πολύπλοκο υπολογιστικό μοντέλο. Στην ουσία είναι προγράμματα που υλοποιούνται στους ηλεκτρονικούς υπολογιστές. Μέχρι σήμερα έχουν χρησιμοποιηθεί σε πολλές εφαρμογές και σε προβλήματα που οι γνωστοί τρόποι αντιμετώπισής τους παρουσιάζουν δυσκολίες, με αποτέλεσμα την αναγκαιότητα των Τεχνητών Νευρωνικών Δικτύων. Η εργασία αυτή αποτελείται από έξι κεφάλαια. Στο πρώτο κεφάλαιο κάνουμε μια εισαγωγή στα Τεχνητά Νευρωνικά Δίκτυα. Αναφέρουμε τις βασικές αρχές τους και την αντιστοιχία τους με τα βιολογικά δίκτυα. Το δεύτερο κεφάλαιο ασχολείται με το δίκτυο Perceptron. Ξεκινάμε με το πιο απλό μοντέλο, τον αισθητήρα και συνεχίζουμε με τα πολυεπίπεδα Νευρωνικά Δίκτυα Perceptron. Αναφέρουμε δύο μεθόδους εκπαίδευσης, τη μέθοδο οπισθοδιάδοσης του λάθους και τον κανόνα Δέλτα. Στο τρίτο κεφάλαιο μελετάμε άλλα είδη δικτύων, όπως τα αναδρομικά δίκτυα, το δίκτυο Hopfield, το δίκτυο SOM και το δίκτυο RBF. Το τέταρτο κεφάλαιο αναφέρεται στον νευρωνικό έλεγχο και στις αρχιτεκτονικές των νευρωνικών ελεγκτών. Στο πέμπτο κεφάλαιο εξετάζουμε κάποιες συγκεκριμένες εφαρμογές των Τεχνητών Νευρωνικών Δικτύων σε διάφορα συστήματα ελέγχου. Στο έκτο κεφάλαιο αναφέρουμε τα συμπεράσματα καθώς και μελλοντικές επεκτάσεις των ΤΝΔ. / Artificial Neural Networks are a research area which has developed over the past decades. Neural Networks consist of a set of simple, interconnected and adaptive plants that create a parallel and complex computational model. They are essentially programs implemented in computers. They have been used in many applications and problems that are very difficult to be solved otherwise. This work consists of six chapters. In the first chapter we make an introduction to Artificial Neural Networks. We mention the basic principles and their correlation with biological networks. The second chapter deals with the network Perceptron. We start with the simplest model, the sensor and continue with the multilayer Neural Network Perceptron. We mention two training methods, the method of error back-propagation and delta rule. In the third chapter we consider other types of networks such as the recurrent networks, Hopfield network, the network SOM and the RBF network. The fourth chapter deals with the neural control and the architectures of neural controllers. In the fifth chapter we examine some specific applications of Artificial Neural Networks in several control systems. The sixth chapter refers to the conclusions of this work and future evolution of ANN.

Time series forecasting and modeling are challenging problems during the past decades, because of its plenty of properties and underlying correlated relationships. As a result, researchers proposed a lot of models to deal with the time series. However, the proposed models such as Autoregressive integrated moving average (ARIMA) and artificial neural networks (ANNs) only describe part of the properties of time series. In this thesis, we introduce a new hybrid model integrated filter structure to improve the prediction accuracy. Case studies with real data from University of Kentucky HealthCare are carried out to examine the superiority of our model. Also, we applied our model to operating room (OR) to reduce the inefficiency cost. The experiment results indicate that our model always outperforms compared with other models in different conditions.

Simulations are a good method for the verification of the correct operation of solar-powered sensor nodes over the desired lifetime. They do, however, require accurate models to capture the influences of the loads and solar energy harvesting system. Artificial neural networks promise a simplification and acceleration of the modeling process in comparison to state-of-the-art modeling methods. This work focuses on the influence of the modeling process's different configurations on the accuracy of the model. It was found that certain parameters, such as the network's number of neurons and layers, heavily influence the outcome, and that these factors need to be determined individually for each modeled harvesting system. But having found a good configuration for the neural network, the model can predict the supercapacitor's charge depending on the solar current fairly accurately. This is also true in comparison to the reference models in this work. Nonetheless, the results also show a crucial need for improvements regarding the acquisition and composition of the neural network's training set.

Soil compaction is a widely spread problem in agricultural soils that has negative agronomic and environmental impacts. The former may lead to poor crop growth and yield, whereas the latter may lead to poor hydraulic properties of soils, and high risk to flooding, soil erosion and degradation. Therefore, the elimination of soil compaction must be done on regular bases. One of the main parameters to quantify soil compaction is soil bulk density (BD). Mapping of within field variation in soil BD will be a main requirement for within field management of soil compaction. The aim of this research was to develop a new approach for the measurement of soil BD as an indicator of soil compaction. The research relies on the fusion of data from visible and near infrared spectroscopy (vis-NIRS), to measure soil gravimetric moisture content (ω), with frequency domain reflectometry (FDR) data to measure soil volumetric moisture content (θv). The values of the estimated ω and θv, for the same undisturbed soil samples were collected from selected locations, textures, soil moisture contents and land use systems to derive soil BD. A total of 1013 samples were collected from 32 sites in the England and Wales. Two calibration techniques for vis-NIRS were evaluated, namely, partial least squares regression (PLSR) and artificial neural networks (ANN). ThetaProbe calibration was performed using the general formula (GF), soil specific calibration (SSC), the output voltage (OV) and artificial neural networks (ANN). ANN analyses for both ω and θv properties were based either on a single input variable or multiple input variables (data fusion). Effects of texture, moisture content, and land use on the prediction accuracy on ω, θv and BD were evaluated to arrive at the best experimental conditions for the measurement of BD with the proposed new system. A prototype was developed and tested under laboratory conditions and implemented in-situ for mapping of ω, θv and BD. When using the entire dataset (general data set), results proved that high measurement accuracy can be obtained for ω and θv with PLSR and the best performing traditional calibration method of the ThetaProbe with R2 values of 0.91 and 0.97, and root mean square error of prediction (RMSEp) of 0.027 g g-1 and 0.019 cm3 cm-3, respectively. However, the ANN – data fusion method resulted in improved accuracy (R2 = 0.98 and RMSEp = 0.014 g g-1 and 0.015 cm3 cm-3, respectively). This data fusion approach gave the best accuracy for BD assessment when only vis-NIRS spectra and ThetaProbe V were used as an input data (R2 = 0.81 and RMSEp = 0.095 g cm-3). The moisture level (L) impact on BD prediction revealed that the accuracy improved with soil moisture increasing, with RMSEp values of 0.081, 0.068 and 0.061 g cm-3, for average ω of 0.11, 0.20 and 0.28 g g-1, respectively. The influence of soil texture was discussed in relation with the clay content in %. It was found that clay positively affected vis-NIRS accuracy for ω measurement and no obvious impact on the dielectric sensor readings was observed, hence, no clear influence of the soil textures on the accuracy of BD prediction. But, RMSEp values of BD assessment ranged from 0.046 to 0.115 g cm-3. The land use effect of BD prediction showed measurement of grassland soils are more accurate compared to arable land soils, with RMSEp values of 0.083 and 0.097 g cm-3, respectively. The prototype measuring system showed moderate accuracy during the laboratory test and encouraging precision of measuring soil BD in the field test, with RMSEp of 0.077 and 0.104 g cm-3 of measurement for arable land and grassland soils, respectively. Further development of the prototype measuring system expected to improve prediction accuracy of soil BD. It can be concluded that BD can be measured accurately by combining the vis-NIRS and FDR techniques based on an ANN-data fusion approach.

A design optimization based on coupled solid–fluid analysis is investigated in this work to achieve specific flow rate through a peristaltic micropump. A micropump consisting of four pneumatically actuated nozzle/diffuser shaped moving actuators on the sidewalls is considered for numerical study. These actuators are used to create pressure difference in the four pump chambers, which in turn drives the fluid through the pump in one direction. Genetic algorithms along with artificial neural networks are used for optimizing the pump geometry and the actuation frequency. A simple example with moving walls is considered for validation by developing an exact analytical solution of Navier–Stokes equation and comparing it with numerical simulations. Possible applications of these pumps are in microelectronics cooling and drug delivery. Based on the results obtained from the fluid–structure interaction analysis, three optimized geometries result in flow rates which match the predicted flow rates with 95% accuracy. These geometries need further investigation for fabrication and manufacturing issues.

Land cover classification using artificial neural networks Abstract This Diploma thesis deals with automatic classification of the satellite high spatial resolution image in the field of land cover. The first half of the work contains the theoretical information about remote sensing and classification methods. The biggest attention is given to the artificial neural networks. In practical part of Diploma thesis are these methods used for the classification of SPOT satellite image. Keywords: remote sensing, image classification, artificial neural networks, SPOT

The thesis introduces Artificial Neural Networks (ANN) to the field of financial durations. We begin by reviewing the findings about financial durations and models applied to analyze them. ANNs are then surveyed and one of the possible network architectures is selected for the forecasting. The selected ANN is a feed-forward network, with one hidden layer, a sigmoid activation function and a genetic algorithm for optimization. We use original and diurnally adjusted data for estimation and in contrast to other duration models, ANNs do not require data pre-processing. Therefore forecasts are estimated in one step without removing seasonalities for raw data. The estimates of the ANN are compared to estimates of the Autoregressive Conditional Duration (ACD) model, which serves as a benchmark for forecasting capabilities of the ANNs. The findings confirm that ANNs can be used to model durations with a similar accuracy as the ACD model. In the case of raw data the model slightly outperforms the ACD model, while the opposite is true for adjusted data, however the forecasting ability difference is not significant.

Cette thèse a été réalisée dans le cadre d'un projet industriel (BMCI), dont l'objectif est d'augmenter la disponibilité des équipements sur les navires. Dans cette thèse, nous proposons une approche qui met à contribution deux approches différentes, à savoir une approche à base de données pour la détection des défauts et une approche à base de connaissances d'experts pour l'isolation des défauts. Cette approche se veut générique et applicable à différents sous-systèmes du moteur ainsi qu'à divers moteurs et offre une ouverture pour une éventuelle application sur d'autres équipements. De plus, elle est tolérante vis-à-vis des éventuels changements au niveau de l'instrumentation disponible. Cette approche a été testée sur la détection et l'isolation des défauts les plus fréquents et aux conséquences graves auxquels les moteurs Diesel sont sujets. Tous les sous-systèmes du moteurs Diesel sont inclus et l'approche de diagnostic prend en considération les interactions existantes entre les sous-systèmes. L'approche de diagnostic a été testée sur un banc d'essai et sur le navire militaire Adroit de DCNS. Les défauts réalisés sur divers circuits du banc moteur et les défauts apparus en fonctionnement sur certains moteurs de l'Adroit, ont été majoritairement détectés et isolés avec succès. De plus, pour pallier à l'incertitude et au caractère flou des relations expertes utilisées dans la procédure d'isolation, une validation des relations de cause à effet a été réalisée, dans le cadre de cette thèse, par la réalisation d'un modèle analytique de simulation de défauts. / This thesis is carried out within an industrial framework (BMCI) which aims to enhance the availability of equipments on board ships. In this work, a data-based method for fault detection is combined with a knowledge-based method for fault isolation. The presented approach is generic and characterized by the ability to be applied to all the Diesel engine subsystems, to different kind of Diesel engines and can also be extended to other equipments. Moreover, this approach is tolerant regarding differences in instrumentation. This approach is tested upon the detection and isolation of the most hazardous and frequent faults which subject Diesel engines. This approach intends to make diagnosis upon the entire Diesel engine including all the subsystems and the existing interactions between the subsystems. The proposed approach is tested upon a test bench and upon the Diesel engines of the DCNS military vessel textquotedblleft Adroit". Most of the introduced faults on the test bench and the appeared faults on the Adroit engines have been successfully detected and isolated. In addition, to deal with uncertainties and fuzziness of the causal relationships given by maintenance experts, a model is developed. This model aims to validate these causal relationships used in the isolation part of the diagnosis approach.

[pt] O presente trabalho apresenta o desenvolvimento de um modelo neural para a inferência da qualidade do polietileno de baixa densidade (PEBD) a partir dos valores das variáveis de processo do sistema reacional. Para tal, fez- se uso de dados operacionais de uma empresa petroquímica, cujo pré-processamento incluiu a seleção de variáveis, limpeza e normalização dos dados selecionados e preparação dos padrões. A capacidade de inferência do modelo neural desenvolvido neste estudo foi comparada com a de dois modelos fenomenológicos existentes. Para tal, utilizou-se como medida de desempenho o valor do erro médio absoluto percentual dos modelos, tendo como referência valores experimentais do índice de fluidez. Neste contexto, o modelo neural apresentou-se como uma eficiente ferramenta de modelagem da qualidade do sistema reacional de produção do PEBD. / [en] This work comprises the development of a neural network- based model for quality inference of low density polyethylene (LDPE). Plant data corresponding to the process variables of a petrochemical company`s LDPE reactor were used for model development. The data were preprocessed in the following manner: first, the most relevant process variables were selected, then data were conditioned and normalized. The neural network- based model was able to accurately predict the value of the polymer melt index as a function of the process variables. This model`s performance was compared with that of two mechanistic models developed from first principles. The comparison was made through the models` mean absolute percentage error, which was calculated with respect to experimental values of the melt index. The results obtained confirm the neural network model`s ability to infer values of quality-related measurements of the LDPE reactor.

Les réseaux de neurones multicouches ont été proposés pour la première fois il y a plus de trois décennies, et diverses architectures et paramétrages ont été explorés depuis. Récemment, les unités de traitement graphique ont permis une formation très efficace sur les réseaux neuronaux et ont permis de former des réseaux beaucoup plus grands sur des ensembles de données plus importants, ce qui a considérablement amélioré le rendement dans diverses tâches d'apprentissage supervisé. Cependant, la généralisation est encore loin du niveau humain, et il est difficile de comprendre sur quoi sont basées les décisions prises. Pour améliorer la généralisation et la compréhension, nous réexaminons les problèmes de paramétrage du poids dans les réseaux neuronaux profonds. Nous identifions les problèmes les plus importants, à notre avis, dans les architectures modernes : la profondeur du réseau, l'efficacité des paramètres et l'apprentissage de tâches multiples en même temps, et nous essayons de les aborder dans cette thèse. Nous commençons par l'un des problèmes fondamentaux de la vision par ordinateur, le patch matching, et proposons d'utiliser des réseaux neuronaux convolutifs de différentes architectures pour le résoudre, au lieu de descripteurs manuels. Ensuite, nous abordons la tâche de détection d'objets, où un réseau devrait apprendre simultanément à prédire à la fois la classe de l'objet et l'emplacement. Dans les deux tâches, nous constatons que le nombre de paramètres dans le réseau est le principal facteur déterminant sa performance, et nous explorons ce phénomène dans les réseaux résiduels. Nos résultats montrent que leur motivation initiale, la formation de réseaux plus profonds pour de meilleures représentations, ne tient pas entièrement, et des réseaux plus larges avec moins de couches peuvent être aussi efficaces que des réseaux plus profonds avec le même nombre de paramètres. Dans l'ensemble, nous présentons une étude approfondie sur les architectures et les paramétrages de poids, ainsi que sur les moyens de transférer les connaissances entre elles / Multilayer neural networks were first proposed more than three decades ago, and various architectures and parameterizations were explored since. Recently, graphics processing units enabled very efficient neural network training, and allowed training much larger networks on larger datasets, dramatically improving performance on various supervised learning tasks. However, the generalization is still far from human level, and it is difficult to understand on what the decisions made are based. To improve on generalization and understanding we revisit the problems of weight parameterizations in deep neural networks. We identify the most important, to our mind, problems in modern architectures: network depth, parameter efficiency, and learning multiple tasks at the same time, and try to address them in this thesis. We start with one of the core problems of computer vision, patch matching, and propose to use convolutional neural networks of various architectures to solve it, instead of manual hand-crafting descriptors. Then, we address the task of object detection, where a network should simultaneously learn to both predict class of the object and the location. In both tasks we find that the number of parameters in the network is the major factor determining it's performance, and explore this phenomena in residual networks. Our findings show that their original motivation, training deeper networks for better representations, does not fully hold, and wider networks with less layers can be as effective as deeper with the same number of parameters. Overall, we present an extensive study on architectures and weight parameterizations, and ways of transferring knowledge between them