Dissertations / Theses on the topic 'Neural network inversion'

Inversion of the artificial neural network mapping is a relatively unexplored field of science. By inversion we mean that a search is conducted to find what input patterns that corresponds to a specific output pattern according to the analysed network. In this report, an evolutionary algorithm is proposed to conduct the search for input patterns. The hypothesis is that the inversion with the evolutionary search-method will result in multiple, separate and equivalent input patterns and not get stuck in local optima which possibly would cause the inversion to result in erroneous answer. Beside proving the hypothesis, the tests are also aimed at explaining the nature of inversion and how the result of inversion should be interpreted. At the end of the document a long list of proposed future work is suggested. Work, which might result in a deeper understanding of what the inversion means and maybe an automated analysis tool, based on inversion.

In this thesis, adaptive neural network controllers are designed for a high subsonic cruise missile. Two autopilot designs are included in the study using adaptive neural networks, namely an altitude hold autopilot designed for the longitudinal channel and a directional autopilot designed for heading control. Aerodynamic coefficients are obtained using missile geometry
a 5-Degree of Freedom (5-DOF) simulation model is obtained, and linearized at a single trim condition. An inverted model is used in the controller. Adaptive Neural Network (ANN) controllers namely, model inversion controllers with Sigma-Pi Neural Network, Single Hidden Layer Neural Network and Background Learning implemented Single Hidden Layer Neural Network, are deployed to cancel the modeling error and are applied for the longitudinal and directional channels of the missile. This approach simplifies the autopilot designing process by combining a controller with model inversion designed for a single flight condition with an on-line learning neural network to account for errors that are caused due to the approximate inversion. Simulations are performed both in the longitudinal and directional channels in order to demonstrate the effectiveness of the implemented control algorithms. The advantages and drawbacks of the implemented neural network based controllers are indicated.

Prediction of rock and fluid properties such as porosity, clay content, and water saturation is essential for exploration and development of hydrocarbon reservoirs. Rock and fluid property maps obtained from such predictions can be used for optimal selection of well locations for reservoir development and production enhancement. Seismic data are usually the only source of information available throughout a field that can be used to predict the 3D distribution of properties with appropriate spatial resolution. The main challenge in inferring properties from seismic data is the ambiguous nature of geophysical information. Therefore, any estimate of rock and fluid property maps derived from seismic data must also represent its associated uncertainty. In this study we develop a computationally efficient mathematical technique based on neural networks to integrate measured data and a priori information in order to reduce the uncertainty in rock and fluid properties in a reservoir. The post inversion (a posteriori) information about rock and fluid properties are represented by the joint probability density function (PDF) of porosity, clay content, and water saturation. In this technique the a posteriori PDF is modeled by a weighted sum of Gaussian PDF’s. A so-called mixture density network (MDN) estimates the weights, mean vector, and covariance matrix of the Gaussians given any measured data set. We solve several inverse problems with the MDN and compare results with Monte Carlo (MC) sampling solution and show that the MDN inversion technique provides good estimate of the MC sampling solution. However, the computational cost of training and using the neural network is much lower than solution found by MC sampling (more than a factor of 104 in some cases). We also discuss the design, implementation, and training procedure of the MDN, and its limitations in estimating the solution of an inverse problem. In this thesis we focus on data from a deep offshore field in Africa. Our goal is to apply the MDN inversion technique to obtain maps of petrophysical properties (i.e., porosity, clay content, water saturation), and petrophysical facies from 3D seismic data. Petrophysical facies (i.e., non-reservoir, oil- and brine-saturated reservoir facies) are defined probabilistically based on geological information and values of the petrophysical parameters. First, we investigate the relationship (i.e., petrophysical forward function) between compressional- and shear-wave velocity and petrophysical parameters. The petrophysical forward function depends on different properties of rocks and varies from one rock type to another. Therefore, after acquisition of well logs or seismic data from a geological setting the petrophysical forward function must be calibrated with data and observations. The uncertainty of the petrophysical forward function comes from uncertainty in measurements and uncertainty about the type of facies. We present a method to construct the petrophysical forward function with its associated uncertainty from the both sources above. The results show that introducing uncertainty in facies improves the accuracy of the petrophysical forward function predictions. Then, we apply the MDN inversion method to solve four different petrophysical inverse problems. In particular, we invert P- and S-wave impedance logs for the joint PDF of porosity, clay content, and water saturation using a calibrated petrophysical forward function. Results show that posterior PDF of the model parameters provides reasonable estimates of measured well logs. Errors in the posterior PDF are mainly due to errors in the petrophysical forward function. Finally, we apply the MDN inversion method to predict 3D petrophysical properties from attributes of seismic data. In this application, the inversion objective is to estimate the joint PDF of porosity, clay content, and water saturation at each point in the reservoir, from the compressional- and shear-wave-impedance obtained from the inversion of AVO seismic data. Uncertainty in the a posteriori PDF of the model parameters are due to different sources such as variations in effective pressure, bulk modulus and density of hydrocarbon, uncertainty of the petrophysical forward function, and random noise in recorded data. Results show that the standard deviations of all model parameters are reduced after inversion, which shows that the inversion process provides information about all parameters. We also applied the result of the petrophysical inversion to estimate the 3D probability maps of non-reservoir facies, brine- and oil-saturated reservoir facies. The accuracy of the predicted oil-saturated facies at the well location is good, but due to errors in the petrophysical inversion the predicted non-reservoir and brine-saturated facies are ambiguous. Although the accuracy of results may vary due to different sources of error in different applications, the fast, probabilistic method of solving non-linear inverse problems developed in this study can be applied to invert well logs and large seismic data sets for petrophysical parameters in different applications.

Adaptive control designs using neural networks (NNs) based on dynamic inversion are investigated for aerospace vehicles which are operated at highly nonlinear dynamic regimes. NNs play a key role as the principal element of adaptation to approximately cancel the effect of inversion error, which subsequently improves robustness to parametric uncertainty and unmodeled dynamics in nonlinear regimes. An adaptive control scheme previously named composite model reference adaptive control is further developed so that it can be applied to multi-input multi-output output feedback dynamic inversion. It can have adaptive elements in both the dynamic compensator (linear controller) part and/or in the conventional adaptive controller part, also utilizing state estimation information for NN adaptation. This methodology has more flexibility and thus hopefully greater potential than conventional adaptive designs for adaptive flight control in highly nonlinear flight regimes. The stability of the control system is proved through Lyapunov theorems, and validated with simulations. The control designs in this thesis also include the use of pseudo-control hedging techniques which are introduced to prevent the NNs from attempting to adapt to various actuation nonlinearities such as actuator position and rate saturations. Control allocation is introduced for the case of redundant control effectors including thrust vectoring nozzles. A thorough comparison study of conventional and NN-based adaptive designs for a system under a limit cycle, wing-rock, is included in this research, and the NN-based adaptive control designs demonstrate their performances for two highly maneuverable aerial vehicles, NASA F-15 ACTIVE and FQM-117B unmanned aerial vehicle (UAV), operated under various nonlinearities and uncertainties.

In this thesis the major objective is the implementation of the inverse neural network concept in the design of printed lens (transmitarray) antenna. As it is computationally extensive to perform full-wave simulations for entire transmitarray structure and thereafter perform optimization, the idea is to generate a design database assuming that a unit cell of the transmitarray is situated inside a 2D infinite periodic structure. This way we generate a design database of transmission coefficient by varying the unit cell parameters. Since, for the actual design, we need dimensions for each cell on the transmitarray aperture and to do this we need to invert the design database. The major contribution of this thesis is the proposal and the implementation of database inversion methodology namely inverse neural network modelling. We provide the algorithms for carrying out the inversion process as well as provide check results to demonstrate the reliability of the proposed methodology. Finally, we apply this approach to design a transmitarray antenna, and measure its performance.

PONTIFÍCIA UNIVERSIDADE CATÓLICA DO RIO DE JANEIRO
COORDENAÇÃO DE APERFEIÇOAMENTO DO PESSOAL DE ENSINO SUPERIOR
PROGRAMA DE EXCELENCIA ACADEMICA
A inversão sísmica é o processo de transformar dados de Sísmica de Reflexão em valores quantitativos de propriedades petroelásticas das rochas. Esses valores, por sua vez, podem ser correlacionados com outras propriedades ajudando os geocientistas a fazer uma melhor interpretação que resulta numa boa caracterização de um reservatório de petróleo. Existem vários algoritmos tradicionais para Inversão Sísmica. Neste trabalho revisitamos a Inversão Colorida (Impedância Relativa), a Inversão Recursiva, a Inversão Limitada em Banda e a Inversão Baseada em Modelos. Todos esses quatro algoritmos são baseados em processamento digital de sinais e otimização. O presente trabalho busca reproduzir os resultados desses algoritmos através de uma metodologia simples e eficiente baseada em Redes Neurais e na pseudo-impedância. Este trabalho apresenta uma implementação dos algoritmos propostos na metodologia e testa sua validade num dado sísmico público que tem uma inversão feita pelos métodos tradicionais.
Seismic inversion is the process of transforming Reflection Seismic data into quantitative values of petroleum rock properties. These values, in turn, can be correlated with other properties helping geoscientists to make a better interpretation that results in a good characterization of an oil reservoir.There are several traditional algorithms for Seismic Inversion. In this work we revise Color Inversion (Relative Impedance), Recursive Inversion, Bandwidth Inversion and Model-Based Inversion. All four of these algorithms are based on digital signal processing and optimization. The present work seeks to reproduce the results of these algorithms through a simple and efficient methodology based on Neural Networks and pseudo-impedance. This work presents an implementation of the algorithms proposed in the methodology and tests its validity in a public seismic data that has an inversion made by the traditional methods.

Thesis (M.S.)--West Virginia University, 2005.
Title from document title page. Document formatted into pages; contains xii, 82 p. : ill. (some col.), maps (some col.). Includes abstract. Includes bibliographical references (p. 80-82).

The aim of inverting artificial neural networks (ANNs) is to find input patterns that are strongly classified as a predefined class. In this project an ANN is inverted by an evolutionary algorithm. The network is retrained by using the patterns extracted by the inversion as counter-examples, i.e. to classify the patterns as belonging to no class, which is the opposite of what the network previously did. The hypothesis is that the counter-examples extracted by the inversion will cause larger updates of the weights of the ANN and create a better mapping than what is caused by retraining using randomly generated counter-examples. This hypothesis is tested on recognition of pictures of handwritten digits. The tests indicate that this hypothesis is correct. However, the test- and training errors are higher when retraining using counter-examples, than for training only on examples of clean digits. It can be concluded that the counter-examples generated by the inversion have a great impact on the network. It is still unclear whether the quality of the network can be improved using this method.

[. . . ]Certaines caractéristiques concernant la qualité et l’hydrogéologie des eaux souterraines, dans de nombreux cas, ne sont pas directement mesurables et doivent être évaluées en fonction d'autres paramètres directement mesurables. Le problème de la détermination des paramètres inconnus du modèle est généralement dénommé "problème inverse". La résolution du problème inverse pour la modélisation de l'écoulement et le transport des contaminants dans les eaux souterraines est l'objectif principal de ce travail de recherche. Quant à la résolution du problème inverse, dans le présent document, nous avons pour objectif la définition d’une méthodologie qui permette l’identification des caractéristiques dans l’espace et le temps des sources inconnues de contaminations. Dans ce travail de recherche, le problème inverse est résolu sur la base de mesures de concentrations du contaminant dans les puits de surveillance situés dans un domaine d'intérêt. Une fois connu l’effet d’un certain phénomène, nous cherchons à reconstruire la cause qui l’a généré. Ainsi, la recherche a-t-elle été élaborée selon les points suivants : - Modélisation de la contamination des eaux souterraines par l'utilisation d'un logiciel non-commercial pour la modélisation des flux et le transport des contaminants dans les milieux poreux. -Modélisation des relations cause-effet de la contamination des eaux souterraines par les Réseaux de Neurones Artificiels (RNA). - Application des RNA pour la résolution du problème inverse dans deux cas de contamination des eaux souterraines étudiés. [. . . ]
[. . . ]In many cases, some hydrogeological and groundwater quality characteristics, are not directly measurable and must be physically assessed in function of directly measurable parameters. The problem of determining the unknown model parameters is usually identified as "inverse problem". Solving the inverse problem is the main goal of modeling groundwater flow and contaminant transport. The validity of an aquifer forecasting model is closely related to the reliability and accuracy of the parameters assessment. With respect to the resolution of the inverse problem, this work aims at defining a methodology that allows to identify the features in space and time of unknown contamination sources. In our case, the inverse problem is solved on the basis of measurements of contaminant concentrations in monitoring wells located in the studied areas. In the framework of this thesis, the research is developed under the following themes: - groundwater contamination modeling using a non-commercial software for the flux and transport model in porous media. - modeling of the cause and effect relationships in groundwater contamination with Artificial Neural Networks (ANN) technology. - application of ANN to solve the inverse problem in two cases of groundwater contamination. Over the past decades, Artificial Neural Networks (ANN) have become increasingly popular as a problem solving tool and have been extensively used as a forecasting tool in many disciplines. […]

Estimates of the physical parameters of the soil surface, namely moisture content and surface roughness, are important for hydrological and agricultural studies, as they appear to be the two major parameters for runoff forecasting in an agricultural watershed. Radar has high potentiality for the remote measurement of soil surface parameters. In particular, the investigation of the radar backscattering response of bare soil surfaces is an important issue in remote sensing because of its capacity for retrieving the desired physical parameters of the surface. The objective of this study is to formulate and to constrain a methodology for solving the inverse problem for the operational retrieval of soil surface roughness and moisture. To separate the effects of the different parameters on the measured signal over complex areas, multi-technique concepts (multi-polarization, multi-angular, multi-sensor, multi-frequency, and multi-temporal) are the main solution. In this work, based on a simulation study, three different configurations, multi-polarization, multi-frequency and multi-angular, are compared to obtain the best configuration for estimating surface parameters and the multi-angular configuration gives the best results. Based on these results, this study was continued according to five different phases: (1) A new index, the NBRI (Normalized radar Backscatter soil Roughness Index), using the multi-angular approach was presented. This index can estimate and classify surface roughness in agricultural fields using two radar images with different incidence angles. (2) A new linear empirical model to estimate soil surface moisture using RADARSAT-1 data was proposed. This model can provide soil moisture with reduced errors of estimation compared to other linear models. (3) Inversion of the surface parameters using nonlinear classical methods. In this case, the Newton-Raphson method, an iterative numerical method, was used in the retrieval algorithm to solve the inverse problem. (4) In this phase, the neural network technique, with a dynamic learning method, was applied to invert the soil surface parameters from the radar data. The results were obtained through performance testing on two different input schemes (one and two data series) and two different databases (theoretical and empirical). The advantage of the multi-angular set with measured data is apparent. These results are the best in this study. (5) Finally, a novel genetic algorithm (GA) was developed to retrieve soil surface parameters. In this study, it is shown that the genetic algorithms, as an optimization technique, can estimate simultaneously soil moisture and surface roughness from only one radar image.

Master of Science
Department of Geology
Abdelmoneam Raef
Matthew W. Totten
This work focuses on the use of advanced seismically driven technologies to estimate the distribution of key reservoir properties which mainly includes porosity and hydrocarbon reservoir pay. These reservoir properties were estimated by using a multitude of seismic attributes derived from post-stack high resolution inversions, spectral imaging and volumetric curvature. A pay model of the reservoir in the Wierman field in Ness County, Kansas is proposed. The proposed geological model is validated based on comparison with findings of one blind well. The model will be useful in determining future drilling prospects, which should improve the drilling success over previous efforts, which resulted in only few of the 14 wells in the area being productive. The rock properties that were modeled were porosity and Gamma ray. Water saturation and permeability were considered, but the data needed were not available. Sequential geological modeling approach uses multiple seismic attributes as a building block to estimate in a sequential manner dependent petrophysical properties such as gamma ray, and porosity. The sequential modelling first determines the reservoir property that has the ability to be the primary property controlling most of the other subsequent reservoir properties. In this study, the gamma ray was chosen as the primary reservoir property. Hence, the first geologic model built using neural networks was a volume of gamma ray constrained by all the available seismic attributes. The geological modeling included post-stack seismic data and the five wells with available well logs. The post-stack seismic data was enhanced by spectral whitening to gain as much resolution as possible. Volumetric curvature was then calculated to determine where major faults were located. Several inversions for acoustic impedance were then applied to the post-stack seismic data to gain as much information as possible about the acoustic impedance. Spectral attributes were also extracted from the post-stack seismic data. After the most appropriate gamma ray and porosity models were chosen, pay zone maps were constructed, which were based on the overlap of a certain range of gamma ray values with a certain range of porosity values. These pay zone maps coupled with the porosity and gamma ray models explain the performance of previously drilled wells.

Les échanges d'eau au sein du cycle global hydrologique sont déterminés par des contraintes mécaniques et thermodynamiques complexes qui forment les bases du système de dynamique du climat. La salinité océanique (S) est l’une des variables de ce cycle les plus délicates à observer. Les mesures in-situ ne permettent pas une bonne connaissance de la structure interne de l'océan car ne couvrent que des résolutions spatiales et temporelles souvent limitées. Alors que les mesures de surface offrent une bonne couverture spatio-temporelle. L’objectif principal de cette thèse est de reconstruire le profil de S océanique à partir des données de surface (SSH: Sea Surface Height, l’ADT : Absolute Dynamic Topography, la SST: Sea Surface Temperature, la SSS: Sea Surface Salinity) en Atlantique tropical. Le modèle d'inversion (INV2) que nous avons développé est décliné en 2 parties. La partie exploratoire a permis de savoir, non seulement, que la latitude est liée aux S de subsurface et de profondeur alors que la SSS est liée au S de proche surface et de surface mais aussi que la SSS n’est linéairement liée à aucun des paramètres de surface étudiés alors que la latitude est fortement liée à la SSH (ADT) et à la SST. Partant de cette étude préalable, nous avons mis en place INV2 qui est basé sur un algorithme séquentiel à 3 étapes (2 projections sur une SOM et une recherche optimale). INV2 a pu reconstruire les profils de S simulés par le modèle DRAKKAR avec des erreurs absolues moyennes inférieures à 0.08 psu sur presque tout le profil. Avec les mesures in situ Coriolis, les erreurs moyennes de INV2 tournent autour de 0.172 psu sur les profils individuels et 0.037 psu sur les profils moyens. Les plus fortes erreurs ont été notées au niveau de l’halocline et dans les zones de MSW. Après une évaluation, un algorithme de complétion de profils de S a été également proposé. Les performances obtenues en utilisant le modèle prouve que cette approche est adéquate pour la reconstruction 3D de S océanique quand la connaissance se limite à la surface
Water exchanges within the global hydrological cycle are determined by complex constraints that form the basis of climate dynamics system.Ocean salinity (S) is one of the most delicate variables to monitor in this cycle. The in-situ measurements cover only limited resolutions. While surface measurements have good spatio- temporal coverage. In this thesis, we reconstruct ocean S profile in the tropical Atlantic from surface parameters (SSH-ADT, SST, SSS). The developed inversion model (INV2) comes in 2 parts. The exploratory part allowed to know that the latitude is related to the S of the subsurface and the depth while the SSS is linked to S of surface and near surface but not to the studied surface parameters. The latitude is strongly linked to the SSH (ADT) and the SST. Based on this preliminary study, we have implemented INV2 which is based on a sequential 3 steps algorithm (2 projections on a SOM and an optimal research). INV2 could reconstructed the S profiles simulated by the DRAKKAR model with mean absolute errors less than 0.08 psu on almost all the profile. With the in situ Coriolis data , the rmse are around 0.172 psu on individual profiles 0.037 psu on average profiles. The largest errors were noted at the halocline and in MSW areas. After the evaluation, a S profiles completion algorithm was also proposed. The INV2 performances obtained showed that this neural approach is adequate for 3D reconstruction of oceanic S when knowledge is limited to the surface

The Lower Cretaceous Mannville Group in East-Central Alberta is one of the main targets for hydrocarbon exploration in Western Canada, typically associated with stratigraphic controls. The depositional environment is diverse, ranging from fluvio-continental to shallow marine. Mannville Group strata overlie a major unconformity that separates them from predominantly carbonate rocks of the Paleozoic (Devonian). This study sought to integrate well and 3-D seismic data to create a stratigraphic architecture for the zone of interest. Seismic inversion was applied to the data and proved to be an excellent tool for mapping the unconformity, clearly distinguishing between the clastic rocks above, and the units below. The inversion result was included as part of a seismic attribute study applied to the dataset. Stepwise regression and validation testing indicated that a combination of six attributes was found to provide the best set able to predict log-derived physical properties related to lithology (gamma-ray) using the seismic data. This result was used to train a neural network, and finally, a pseudo-lithology volume able to estimate the distribution of lithology within the Mannville Group over the whole 3-D survey was generated. New stratigraphic features that were non-apparent in the original amplitude seismic version were discovered using the inversion and the pseudo-lithology volumes. The results of this study show how comparing the three different 3-D seismic versions can be useful for understanding the stratigraphic complexity of the Mannville. The extensive methodology approach presented herein can be used for analog purposes in Western Canada as well as in any other geological setting.
La Formation Mannville (Crétacé Inférieur) du Centre Est de l'Alberta est l'un des principaux objectifs concernant l'exploration des hydrocarbures, typiquement associés à un contrôle stratigraphique. Les environnements de dépôts sont variés, allant du domaine fluvio-continental au domaine marin. Les dépôts de la Formation Mannville reposent sur une discordance stratigraphique majeure qui les sépare des dépôts carbonatés du Paléozoïque (Devonien). Nous avons cherché à intégrer des données de puits pour reconstituer l'architecture stratigraphique dans la zone d'étude. L'inversion sismique à été appliquée aux données existantes, et s'est trouvée être un excellent outil pour cartographier les discordances, faisant clairement la distinction entre les dépôts silico-clastiques des dépôts carbonatés en dessous. Les résultats de l'inversion ont ensuite été utilisés comme faisant partie des attributs sismiques appliqués au jeu de données dans le but de rechercher si la distribution des sédiments silici-clastiques dans la zone d'étude peut être distinguée d'un point de vu sismique. A travers une analyse de régression point par point, une combinaison de six attributs à fourni le meilleur jeu capable de prédire l'information lithologique dans les données sismiques dérivées des données de puits (Gamma-Ray). Ce résultat a été transféré dans un réseau de neurones, qui utilise cette relation pour s'entraîner à prédire la réponse du Gamma-Ray au delà des points de contrôle dans le volume sismique. Cela conduit à la génération d'un volume informé en pseudo-lithologie capable d'estimer la distribution de la lithologie à l'intérieur de le Formation Mannville sur l'ensemble de l'étude 3-D. Ce nouveau volume est capable de mettre en évidence les caractéristiques stratigraphiques invisibles avec l'amplitude sismique et dans les volumes d'inversion. Ce

Esta pesquisa consiste na solução do problema inverso de transferência radiativa para um meio participante (emissor, absorvedor e/ou espalhador) homogêneo unidimensional em uma camada, usando-se a combinação de rede neural artificial (RNA) com técnicas de otimização. A saída da RNA, devidamente treinada, apresenta os valores das propriedades radiativas [ω, τ0, ρ1 e ρ2] que são otimizadas através das seguintes técnicas: Particle Collision Algorithm (PCA), Algoritmos Genéticos (AG), Greedy Randomized Adaptive Search Procedure (GRASP) e Busca Tabu (BT). Os dados usados no treinamento da RNA são sintéticos, gerados através do problema direto sem a introdução de ruído. Os resultados obtidos unicamente pela RNA, apresentam um erro médio percentual menor que 1,64%, seria satisfatório, todavia para o tratamento usando-se as quatro técnicas de otimização citadas anteriormente, os resultados tornaram-se ainda melhores com erros percentuais menores que 0,04%, especialmente quando a otimização é feita por AG.
This research consists in the solution of the inverse problem of radiative transfer for a participating media (emmiting, absorbing and/or scattering) homogeneous one-dimensional in one layer, using the combination of artificial neural network (ANN), with optimization techniques. The output of the ANN, properly trained presents the values of the radiative properties [w, to, p1 e p2] that are optimized through the following techniques: Particle Collision Algorithm (PCA), Genetic Algorithm (GA), Greedy Randomized Adaptive Search Procedure (GRASP) and Tabu Search (TS). The data used in the training are synthetics, generated through the direct problem without the introduction of noise. The results obtained by the (ANN) alone, presents an average percentage error minor than 1,64%, what it would be satisfying, however, for the treatment using the four techniques of optimization aforementioned, the results have become even better with percentage errors minor than 0,03%, especially when the optimization is made by the GA.

Les méthodes géophysiques aéroportées sont très utilisées pour la prospection du sous-sol à l'échelle régionale, permettant ainsi de couvrir de grandes surfaces en particulier les zones difficiles d'accès. Le sujet de thèse concerne le développement de techniques d'interprétation des données géophysiques pour le problème des ressources naturelles et de l'environnement. La première partie de cette thèse concerne le développement d'une méthode de calcul direct des composantes des champs gravimétrique et magnétique à partir d'une structure (corps) géologique modélisé(e) par une grille à base de tétraèdres, ce qui permet de représenter des modèles géologiques très complexes particulièrement en présence de zones faillées et hétérogènes avec un nombre d'éléments optimal. Plusieurs techniques d'inversions utilisent des contraintes mathématiques pour la résolution du problème inverse en modélisation. Ces contraintes permettent de réduire le nombre de modèles possibles. Cependant les solutions proposées appelées aussi « le modèle le plus probable » présentent des solutions lisses, ce qui est loin de représenter la réalité géologique. Pour éluder ce problème, les deuxièmes et troisièmes parties de la thèse proposent des améliorations majeures du processus d'inversion par l'utilisation des méthodes géostatistiques telles que la Simulation Gaussienne Séquentielle ou la Co-Simulation dans le cas d'une inversion conjointe afin d'estimer les probabilités a posteriori des modèles simulés. La quatrième partie de ce mémoire présente une alternative à la simulation de plusieurs variables. L'apprentissage du réseau de neurones supervisé par un certain nombre de points permet d'établir une relation entre les différentes variables
In recent years with the technology developments, airborne geophysical methods (gravity, magnetic, and electromagnetic) are widely used in the natural resource exploration at the regional scale. It covers large areas particularly in the areas with difficult access. The first part of this thesis consist on the development of new forward modeling algorithm for the calculation of the components of the gravity and magnetic fields based on a tetrahedron grid. The tetrahedral mesh allows the representation of very complex geological models holding many heterogeneous and faulted zones with an optimal number of elements, this reduces significantly the time calculation. Several inversion techniques use mathematical constraints for the resolution of the inverse problem in order to reduce the number of possible models. However the proposed solutions called also "the most probable model" provide a smooth solutions that cannot represent the geological reality. To circumvent this problem in the second and the third parts of this thesis, we made two major improvements. The first, we integrate Sequential Gaussian Simulation into the inversion procedure to determine a possible distributions of a single property. The second is that we used the Co-Simulation in the case of joint inversion to estimate a posteriori probabilities of the simulated models. The last part of this thesis presents an alternative to the several variables simulation, supervised learning of neural networks allows to establish a relationship between the different variables

Nesta tese é proposta uma implementação para a solução do problema inverso com as estimativas das propriedades radiativas (o albedo de espalhamento simples, a espessura ótica do meio e as reflectividades difusas) a partir dos valores das intensidades de radiação que deixam o meio participante utilizando uma abordagem híbrida de sistemas neuro-fuzzy (SNF), o qual combina a utilização de sistemas de inferência fuzzy com as redes neurais artificiais. Busca-se com a utilização desse sistema híbrido integrar a habilidade dos sistemas fuzzy no tratamento de informações inexatas, imprecisas, e vagas, e a capacidade das redes neurais artificiais de tratar o aprendizado por experiência e a generalização do conhecimento. É proposta também uma metodologia de máquinas de comitês neuro-fuzzy na solução deste problema inverso em transferência radiativa. Foi observado paralelamente que a solução dos sistemas neuro-fuzzy e dos sistemas híbridos de máquinas de comitê neuro-fuzzy, apresentam baixa qualidade nos resultados quando são utilizados os dados experimentais com os menores coeficientes de sensibilidade para os parâmetros que serão estimados. Por outro lado, quando são utilizados dados com maior sensibilidade, são obtidos melhores resultados. Esta abordagem procura evitar a possibilidade da não convergência desses métodos.
In this thesis is proposed an implementation for solving the inverse problem with the estimates of radiative properties (the single scattering albedo, the optical thickness of the media and the diffuse reflectivities) by the values of the intensities of radiation that leaves the participant medium using a hybrid approach of neuro-fuzzy systems, which combines the use of fuzzy inference systems with artificial neural networks. The use of this hybrid system try to include the ability of fuzzy systems in the treatment of inaccurate, imprecise, and vague data, and the ability of artificial neural networks to deal with learning from experience and widespread knowledge. Also is proposed a methodology for machines committees in neuro-fuzzy solution of this inverse problem in radiative transfer. It was observed in parallel that the solution of neuro-fuzzy systems and hybrid systems neuro-fuzzy committee machines, have a poor quality results when using the experimental data with the lowest sensitivity coefficients for the parameters that will be estimated. Moreover, when data are used with greater sensitivity, better results are obtained. This approach seeks to avoid the possibility of non-convergence in such methods.

Esta tese tem por objetivo propor uma metodologia para recuperação de perfis verticais de temperatura na atmosfera com nuvens a partir de medidas de radiância feitas por satélite, usando redes neurais artificiais. Perfis verticais de temperatura são importantes condições iniciais para modelos de previsão de tempo, e são usualmente obtidos a partir de medidas de radiâncias feitas por satélites na faixa do infravermelho. No entanto, quando estas medidas são feitas na presença de nuvens, não é possível, com as técnicas atuais, efetuar a recuperação deste perfil. É uma perda significativa de informação, pois, em média, 20% dos pixels das imagens acusam presença de nuvens. Nesta tese, este problema é resolvido como um problema inverso em dois passos: o primeiro passo consiste na determinação da radiância que atinge a base da nuvem a partir da radiância medida pelos satélites; o segundo passo consiste na determinação do perfil vertical de temperaturas a partir da informação de radiância fornecida pelo primeiro passo. São apresentadas reconstruções do perfil de temperatura para quatro casos testes. Os resultados obtidos mostram que a metodologia adotada produz resultados satisfatórios e tem grande potencial de uso, permitindo incorporar informações sobre uma região mais ampla do globo e, consequentemente, melhorar os modelos de previsão do tempo.
This thesis presents a methodology for retrieving vertical temperature profiles in the atmosphere with clouds from radiance measurements made by satellite, using artificial neural networks. Vertical temperature profiles are important initial conditions for numerical weather prediction models, and are usually obtained from measurements of radiance using infrared channels. Though, when these measurements are performed in the atmosphere with clouds, it is not possible to retrieve the temperature profile with current techniques. It is a significant loss of information, since on average 20% of the pixels of the images have clouds. In this thesis, this problem is solved as a two-step inverse problem: the first step is an inverse problem of boundary condition estimation, where the radiance reaching the cloud basis is determined from radiance measured by satellite; the second step consists in determining the vertical temperature profile from the boundary condition estimated in the first step. Reconstructions of temperature profile are presented for four test cases. The results show that the proposed methodology produces satisfactory results and has great potential for use, allowing to incorporate information from a wider area of the planet and thus to improve numerical weather prediction models.

[EN] «Study of neural networks and quadrupole models for the solution of the inverse problem in non-destructive testing by eddy currents. Application control of thickness on metal surfaces with multilayer protection» Abstract Currently the use of test methods using eddy current testing has become widespread in the industry, not only in the field of non-destructive testing, to study the behavior of cracks in components subject to fatigue and to detect failures metal internal structure, but also as measuring instruments to obtain information from the characterization of materials, composition of the surface layers, surface oxidation extent of the thickness of the surface layers treated for hardening. However, the application of currents induced to meet sizing surface layers can be solved by direct problem but not the reverse, obtained from measurements of electromagnetic parameters, the dimensions of the laminated layers on the metal surface. This unique case is trying to solve this thesis so that replaces modeling using different physical layers of the material, quadrupole or quadrupole elements, whose behavior can set the parallel appearing in induced surface currents metal under real conditions. For this, they have taken several sets of plates with different known thicknesses are measured and their characteristic impedances by suitable sensors. The results obtained are used to train an artificial neural network, and therefore should allow its use to solve the inverse problem, i.e., from readings measured as input to the neural network to obtain the dimensions wanted to check out. The work begins with a detailed study of the solutions to equations and harmonic electromagnetic fields for the model proposed. The method of the magnetic vector potential vector with Coulomb contrast conditions used. Signal propagation near the surface and in the border region are studied. The basic formulation of the induced currents is established and studied in detail two specific problems and their models, the finite thickness plate and a cylinder. However, this method of solving the equations of the problem is proposed by using numerical series in replacing the integrals with infinite upper limit. Results to calculate the resulting model parameters based on the obtained test frequency and for a model settled down. We proposes a simulation model ANSYS two solutions, the first a plane model is used PLANE53 and second dimensional model SOLID236. From the latter assay sensitivity of currents induced by benchmarking the environmental conditions change, permeability and conductivity are studied. A model similar to the propagation of a transmission line in telecommunications models for that quadrupoles are set and are daisy-chained signal is also proposed theoretical results are obtained even different frequencies and are used to show a method of measuring or control of different thicknesses.
[ES] Estudio de redes neuronales y modelos de cuadripolos para la solución del problema inverso en ensayos no destructivos, por medio de corrientes inducidas. Aplicación para el control de espesores en superficies metálicas con protección multicapa Resumen En la actualidad el uso de métodos de prueba empleando ensayos de corrientes inducidas se ha generalizado en la industria, no sólo en el campo de los ensayos no destructivos, para estudiar el comportamiento de las grietas en elementos sujetos a fatiga así como para detectar fallas en la estructura interna metálica, sino también como instrumentos de medida para obtener información de la caracterización de los materiales, composición de las capas superficiales, magnitud de la oxidación superficial, espesor de las capas superficiales tratadas para su endurecimiento. Sin embargo, la aplicación de corrientes inducidas para conocer el dimensionado de capas superficiales puede resolverse mediante el problema directo pero no así el inverso, obtener a partir de las mediciones de parámetros electromagnéticos, las dimensiones de las capas estratificadas en la superficie del metal. Este caso singular es el que trata de resolver esta tesis de manera que empleando un modelado que sustituye las distintas capas físicas del material, por elementos cuadripolares o cuadripolos, de cuyo comportamiento se puede establecer el paralelismo con que aparecen las corrientes inducidas en la superficie del metal en condiciones reales. Para esto, se han tomado varios juegos de placas con distintos espesores conocidos y se han medido sus impedancias características mediante sensores adecuados. Los resultados obtenidos sirven para entrenar una red neuronal artificial, y por tanto deben de permitir su empleo para resolver el problema inverso, es decir, a partir de las lecturas de la medición como entrada de la red neuronal obtener las dimensiones buscadas a la salida. Los trabajos se inician con un estudio detallado de las soluciones a las ecuaciones de campos electromagnéticos armónicos y para el modelo que se propone. Se usa el método vectorial del potencial vector magnético con las condiciones de contraste de Coulomb. Se estudian la propagación de señales en las proximidades de la superficie y en la región frontera. Se establece la formulación básica de las corrientes inducidas y se estudian con detalle dos problemas concretos y sus modelos, el de una placa de espesor finito y el de un cilindro. Con todo ello se propone un método de resolución de las ecuaciones del problema mediante el empleo de series numéricas en la sustitución de las integrales con límite superior infinito. Se obtienen resultados para el cálculo de los parámetros resultantes en función de la frecuencia de ensayo y para un modelo establecido. Se propone un modelo de simulación ANSYS con dos soluciones, en la primera se emplea un modelo plano PLANE53 y en la segunda un modelo tridimensional SOLID236. A partir de este último se estudia la sensibilidad del ensayo de corrientes inducidas por comparación de resultados al modificar las condiciones del medio, permeabilidad y conductividad. También se propone un modelo de propagación de señal similar al de una línea de transmisión en telecomunicaciones, para ello se establecen modelos de cuadripolos y se conectan en cadena, los resultados teóricos se obtienen par las distintas frecuencias y sirven para mostrar un método de medida u control de los distintos espesores
[CAT] «Estudi de xarxes neuronals i models de quadripols per a la solució del problema invers en assajos no destructius, per mitjà de corrents induïts. Aplicació per al control de gruixos en superfícies metàl·liques amb protecció multicapa» Resum En l'actualitat l'ús de mètodes de prova emprant assaigs de corrents induïts s'ha generalitzat en la indústria, no només en el camp dels assajos no destructius, per estudiar el comportament de les esquerdes en elements subjectes a fatiga així com per detectar falles en l'estructura interna metàl·lica, sinó també com a instruments de mesura per obtenir informació de la caracterització dels materials, composició de les capes superficials, magnitud de l'oxidació superficial, espessor de les capes superficials tractades per al seu enduriment. No obstant això, l'aplicació de corrents induïts per conèixer el dimensionament de capes superficials pot resoldre mitjançant el problema directe però no així l'invers, obtenir a partir dels mesuraments de paràmetres electromagnètics, les dimensions de les capes estratificades en la superfície del metall. Aquest cas singular és el que tracta de resoldre aquesta tesi de manera que emprant un modelatge que substitueix les diferents capes físiques del material, per elements quadrupolars o quadripols, del comportament es pot establir el paral·lelisme amb què apareixen els corrents induïdes en la superfície del metall en condicions reals. Per això, s'han pres diversos jocs de plaques amb diferents gruixos coneguts i s'han mesurat les seves impedàncies característiques mitjançant sensors adequats. Els resultats obtinguts serveixen per entrenar una xarxa neuronal artificial, i per tant han de permetre la seva ocupació per resoldre el problema invers, és a dir, a partir de les lectures del mesurament com a entrada de la xarxa neuronal obtenir les dimensions buscades a la sortida. Els treballs s'inicien amb un estudi detallat de les solucions a les equacions de camps electromagnètics harmònics i per al model que es proposa. Es fa servir el mètode vectorial del potencial vector magnètic amb les condicions de contrast de Coulomb. S'estudien la propagació de senyals en les proximitats de la superfície hi ha la regió frontera. S'estableix la formulació bàsica dels corrents induïts i s'estudien amb detall dos problemes concrets i els seus models, el d'una placa de gruix finit i el d'un cilindre. Amb tot això es proposa un mètode de resolució de les equacions del problema mitjançant l'ús de sèries numèriques en la substitució de les integrals amb límit superior infinit. S'obtenen resultats per al càlcul dels paràmetres resultants en funció de la freqüència d'assaig i per a un model establert. Se proposa un model de simulació ANSYS amb dues solucions, en la primera es fa servir un model pla PLANE53 i en la segona un model tridimensional SOLID236. A partir d'aquest últim s'estudia la sensibilitat de l'assaig de corrents induïts per comparació de resultats en modificar les condicions del medi, permeabilitat i conductivitat. També es proposa un model de propagació de senyal similar al d'una línia de transmissió en telecomunicacions, per això s'estableixen models de quadripols i es connecten en cadena, els resultats teòrics s'obtenen bat les diferents freqüències i serveixen per mostrar un mètode de mesura o control dels diferents gruixos.
Corbí Cabrera, A. (2016). Estudio de redes neuronales y modelos de cuadripolos para la solución del problema inverso en ensayos no destructivos por medio de corrientes inducidas. Aplicación para el control de espesores en superficies metálicas con protección multicapa [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/63448
TESIS

Este trabalho fundamenta-se no conceito de máquina de comitê de redes neurais artificiais e tem por objetivo resolver o problema inverso de transferência radiativa em um meio unidimensional, homogêneo, absorvedor e espalhador isotrópico. A máquina de comitê de redes neurais artificiais agrega e combina o conhecimento adquirido por um certo número de especialistas aqui representados, individualmente, por cada uma das redes neurais artificiais (RNA) que compõem a máquina de comitê de redes neurais artificiais. O objetivo é atingir um resultado final melhor do que o obtido por qualquer rede neural artificial separadamente, selecionando-se apenas àquelas redes neurais artificiais que apresentam os melhores resultados na fase de generalização descartando-se as demais, o que foi feito neste trabalho. Aqui são utilizados dois modelos estáticos de máquinas de comitê, usando a média aritmética de conjunto, que se diferenciam entre si apenas na composição do combinador de saída de cada máquina de comitê. São obtidas, usando-se máquinas de comitê de redes neurais artificiais, estimativas para os parâmetros de transferência radiativa, isto é, a espessura óptica do meio, o albedo de espalhamento simples e as refletividades difusas. Finalmente, os resultados obtidos com ambos os modelos de máquina de comitê são comparados entre si e com aqueles encontrados usando-se apenas redes neurais artificiais do tipo perceptrons de múltiplas camadas (MLP), isoladamente. Aqui essas redes neurais artificiais são denominadas redes neurais especialistas, mostrando que a técnica empregada traz melhorias de desempenho e resultados a um custo computacional relativamente baixo.
This work is based on the concept of neural networks committee machine and has the objective to solve the inverse radiative transfer problem in one-dimensional, homogeneous, absorbing and isotropic scattering media. The artificial neural networks committee machine adds and combines the knowledge acquired by an exact number of specialists which are represented, individually, by each one of the artificial neural networks (ANN) that composes the artificial neural network committee machine. The aim is to reach a final result better than the one obtained by any of the artificial neural network separately, selecting only those artificial neural networks that presents the best results during the generalization phase and discarding the others, what was done in this present work. Here are used two static models of committee machines, using the ensemble arithmetic average, that differ between themselves only by the composition of the output combinator by each one of the committee machine. Are obtained, using artificial neural networks committee machines, estimates for the radiative transfer parameters, that is, medium optical thickness, single scattering albedo and diffuse reflectivities. Finally, the results obtained with both models of committee machine are compared between themselves and with those found using artificial neural networks type multi-layer perceptrons (MLP), isolated. Here that artificial neural networks are named as specialists neural networks, showing that the technique employed brings performance and results improvements with relatively low computational cost.

O ensaio de resposta em frequência em geradores síncronos vem ganhando espaço nas últimas décadas, porém o alto custo dos equipamentos empregados para realização do ensaio ainda é um empecilho, tanto para fabricantes como para consumidores finais. Este trabalho tem por objetivo complementar trabalhos anteriores, através do uso de redes neurais artificiais para identificação de padrões em ensaios realizados com equipamentos de baixo custo e baixa resolução. Através das redes neurais artificiais utilizadas é possível estimar novos pontos de ensaio sem que seja necessário um novo ensaio ou até mesmo um ensaio com equipamentos mais caros. Através de uma combinação de algoritmos de tratamento de dados, é possível a aproximação de modelos ensaiados com modelos teóricos e através desses resultados obter parâmetros elétricos dos geradores síncronos.
The frequency response test on synchronous generators has been increasing in recent decades, but the high cost of equipment used for conducting the test is still a stumbling block for both manufacturers and end-consumers. This dissertation aims to complement previous work, through the use of artificial neural networks to identify patterns in tests conducted with low-cost and low-resolution equipment. Through the artificial neural networks used it is possible to estimate a new set of test points without retesting is necessary or even a more expensive assay equipments. Through a combination of data processing algorithms, the approach tested models with theoretical models is possible and through these results to obtain electrical parameters of synchronous generators.

Le travail présenté dans ce mémoire de thèse s'inscrit dans le cadre de la commande non-linéaire et robuste des avions de transport. Le but de cette thèse est de coupler les propriétés de la commande non-linéaire (adaptation aux non-linéarités de l'avion, synthèse de correcteurs explicites, facilité de réglage une fois la synthèse réalisée, généricité des lois de commande obtenues) à des propriétés de robustesse indispensables à l'activité aéronautique. En effet, pour garantir la sécurité des vols tant en pilotage manuel qu'en pilotage automatique, les lois de commande doivent présenter des propriétés fortes de stabilité et de performances robustes. Après une introduction au contexte industriel et de recherche du sujet de thèse, une partie "techniques, méthodes et outils" nous permet de mettre en avant les contributions du travail de thèse dans les domaines de la commande non-linéaire robuste et de la modélisation automatique. La technique de commande non-linéaire robuste présentée, appelée commande RMI (Robust Multi-Inversion) s'appuie sur la technique désormais classique d'inversion de la dynamique, notamment étudiée à Airbus depuis quelques années (Fabrice VILLAUME, Jean DUPREZ) et qu'elle robustifie par l'adjonction d'une boucle supplémentaire d'observation. Nous présentons aussi un outil de génération automatique de modèles non-linéaires, multivariables et embarquables, ainsi que les méthodes afférentes basées sur les réseaux de neurones. Cet outil est nécessaire à l'industrialisation des lois de commandes non-linéaires basées modèles. La partie applicative de la thèse souligne ensuite les particularités du système "avion" et propose des architectures de lois de commande, des trajectoires de référence associées, et la validation avancée de l'ensemble par simulations sur simulateur certifié. Enfin, après une conclusion sur le bilan de la thèse et les perspectives envisageables, nous proposons des annexes permettant d'approfondir certains aspects de notre étude
The work presented in this PhD thesis report is situated within the framework of the nonlinear and robust control of transport aircrafts. The purpose of this thesis is to couple the properties of nonlinear controllers (adaptation to the aircraft nonlinearities, explicit controllers synthesis, easy and decoupled setting once the synthesis is achieved, genericity of the obtained control laws) with essential robustness properties. Indeed, to guarantee the flight safety, both in manual handling and in automatic control, the control laws have to present strong robust stability and performances properties. After an introduction to the industrial and research context, a "techniques, methods and tools" part allows us to point out the thesis contributions in the nonlinear robust control and automatic modelling domains. The nonlinear robust control technique presented, called RMI control (for "Robust Multi-Inversion") is based on the now classical Nonlinear Dynamic Inversion (NDI) technique, notably studied at Airbus for some years (Fabrice VILLAUME, Jean DUPREZ), and is robustified by adding a complementary observation loop. We also present an automatic tool creating nonlinear, multivariable and embeddable models, as well as neural networks correlated methods. This tool is mandatory for the industrialization of our model-based flight control laws. Then the applicative part of the thesis underlines the specificities of the "aircraft" system and proposes flight control laws architectures, associated reference trajectories, and the advanced validation of the whole system by simulations performed on Airbus' certified simulator. Finally, after a conclusion on the main results and perspectives linked to the thesis, we propose annexes allowing to go further into the details of certain parts of our study

碩士
國立成功大學
造船及船舶機械工程學系碩博士班
91
A new approach for measuring acoustic impedance is developed by using Artificial Neural Network(ANN) algorithm. Instead of using impedance tube, a rectangular room or a box is simulated with known boundary conditions at some boundaries and a unknown acoustic impedance at one side of the wall. The training data basis fort the ANN algorithm is evaluated by Similar Source Method which was developed earlier by Too[1999] for the estimation of interior and exterior sound field. The training data basis is constructed by evaluating of acoustic pressure at a field point with various acoustic impedance conditions at one side of the wall. The simulation result indicates that the prediction of acoustic impedance is very accurate with error percentage under 1%. Also, one point field measure-ment in the present approach provides a straightforward and easy evaluation than that in the two points measurement of the impedance tube approach.

I present results from quantitative evaluation of the capability of hosting and trapping CO₂ of a carbonate brine reservoir from Dickman Field, Kansas. The analysis includes estimation of some reservoir parameters such as porosity and permeability of this formation using pre-stack seismic inversion followed by simulating flow of injected CO₂ using a simple injection technique. Liner et at (2009) carried out a feasibility study to seismically monitor CO₂ sequestration at Dickman Field. Their approach is based on examining changes of seismic amplitudes at different production time intervals to show the effects of injected gas within the host formation. They employ Gassmann's fluid substitution model to calculate the required parameters for the seismic amplitude estimation. In contrast, I employ pre-stack seismic inversion to successfully estimate some important reservoir parameters (P- impedance, S- impedance and density), which can be related to the changes in subsurface rocks due to injected gas. These are then used to estimate reservoir porosity using multi-attribute analysis. The estimated porosity falls within a reported range of 8-25%, with an average of 19%. The permeability is obtained from porosity assuming a simple mathematical relationship between porosity and permeability and classifying the rocks into different classes by using Winland R35 rock classification method. I finally perform flow simulation for a simple injection technique that involves direct injection of CO₂ gas into the target formation within a small region of Dickman Field. The simulator takes into account three trapping mechanisms: residual trapping, solubility trapping and mineral trapping. The flow simulation predicts unnoticeable changes in porosity and permeability values of the target formation. The injected gas is predicted to migrate upward quickly, while it migrates slowly in lateral directions. A large amount of gas is concentrated around the injection well bore. Thus my flow simulation results suggest low trapping capability of the original target formation unless a more advanced injection technique is employed. My results suggest further that a formation below our original target reservoir, with high and continuously distributed porosity, is perhaps a better candidate for CO₂ storage.
text

A run-of-mine ore milling circuit is primarily used to grind incoming ore containing precious metals to a particle size smaller than a specification size. A traditional run-of-mine (ROM) ore single-stage closed milling circuit comprises of the operational units: mill, sump and cyclone. These circuits are difficult to control because of significant nonlinearities, large time delays, large unmeasured disturbances, process variables that are difficult to measure and modelling uncertainties. A nonlinear model predictive controller with state estimation could yield good control of the ROM ore milling circuit despite these difficulties. Additionally, the ROM ore milling circuit is an energy intensive unit and a controller or power optimizer could bring significant cost savings. A nonlinear model predictive controller requires good state estimates and therefore a neural network for state estimation as an alternative to the particle filter has been addressed. The neural network approach requires fewer process variables that need to be measured compared to the particle filter. A neural network is trained with three disturbance parameters and used to estimate the internal states of the mill, and the results are compared with those of the particle filter implementation. The neural network approach performed better than the particle filter approach when estimating the volume of steel balls and rocks within the mill. A novel combined neural network and particle filter state estimator is presented to improve the estimation of the neural network approach for the estimation of volume of fines, solids and water within the mill. The estimation performance of the combined approach is promising when the disturbance magnitude used is smaller than that used to train the neural network. After state estimation was addressed, this work targets the implementation of a nonlinear controller combined with full state estimation for a grinding mill circuit. The nonlinear controller consists of a suboptimal nonlinear model predictive controller coupled with a dynamic inversion controller. This allows for fast control that is asymptotically stable. The nonlinear controller aims to reconcile the opposing objectives of high throughput and high product quality. The state estimator comprises of a particle filter for five mill states as well as an additional estimator for three sump states. Simulation results show that control objectives can be achieved despite the presence of noise and significant disturbances. The cost of energy has increased significantly in recent years. This increase in price greatly affects the mineral processing industry because of the large energy demands. A run-of-mine ore milling circuit provides a suitable case study where the power consumed by a mill is in the order of 2 MW. An attempt has been made to reduce the energy consumed by the mill in the two ways: firstly, within the nonlinear model predictive control in a single-stage circuit configuration and secondly, running multiple mills in parallel and attempting to save energy while still maintaining an overall high quality and good quantity. A formulation for power optimization of multiple ROM ore milling circuits has been developed. A first base case consisted not taking power into account in a single ROM ore milling circuit and a second base case split the load and throughput equally between two parallel milling circuits. In both cases, energy can be saved using the NMPC compared to the base cases presented without significant sacrifice in product quality or quantity. The work presented covers three topics that has yet to be addressed within the literature: a neural network for mill state estimation, a nonlinear controller with state estimation integrated for a ROM ore milling circuit and power optimization of a single and multiple ROM ore milling circuit configuration.
Dissertation (MEng)--University of Pretoria, 2015.
Electrical, Electronic and Computer Engineering
Unrestricted

碩士
國立交通大學
資訊學院碩士在職專班資訊組
95
In well logging inversion problem, a non-linear mapping exists between the synthetic logging measurements and the true formation conductivity. Without complexity of theoretic computation, neural network is able to approximate the input-output mapping through training with the iterative adjustment of connection weights. In our study, we develop the higher-order feature neural nets on the basis of neural network, and then apply on well logging inversion. The usually used training algorithm for neural network is gradient descent, which is easy to get trapped at local minimum, so we adopt a method that combine with genetic algorithm to improve the training efficiency. In addition, the convergence of gradient descent is slow, so we adopt the conjugate gradient to speed up the convergence. In order to make network more non-linear, we proposed higher-order feature neural nets that use functions to expand the input feature to higher degree. In order to use more training patterns and increase the convergence efficiency, we test various network architectures that use different number of input nodes. Besides, the experimental results show that the convergence efficiency of the network with 1 hidden layer is better than that without hidden layer, so we adopt the network with 1 hidden layer. We use 31 synthetic logging datasets. Each has 200 input features and corresponding outputs. The performance of network is evaluated by comparing the mean absolute error between the actual outputs and desired outputs. Leave-one-out validation method is used in experiments. Each time 30 datasets are used in training, the trained network is then tested with the left 1 dataset. After 31 trials, the network performance is computed by averaging these testing results. To validate the effectiveness of higher-order feature neural nets, the network size is 30-36-10 (not include bias), we train the network using conjugate gradient with synthetic logging datasets, and the trained network is then tested with real field logs. Results obtained from our experiments have shown that the proposed higher-order feature neural nets can be used effectively to process the well logging inversion. Our study shows an effective architecture of neural network to apply on well logging data inversion.

碩士
國立交通大學
資訊科學與工程研究所
96
Multilayer perceptron is adopted for well log data inversion. The gradient descent method is used in the back propagation learning rule and a hybrid method, which combines differential evolution (DE) and gradient descent method, are used in training process respectively. The input of the neural network is the apparent conductivity (Ca) of the well log and the desired output is the true formation conductivity (Ct). The higher order of the input features and the original features are the network input for training. From our experimental results, we find the expanding input features with back propagation learning rule can get fast convergence in training and decrease the mean absolute error between the desired output and the actual output. The hybrid method will provide more precise results, despite it takes a longer training time. The multilayer perceptron network, which is trained by the hybrid method, with 10 input features, the expanding input features to the third order, 8 hidden nodes, and 10 output nodes can get the smallest average mean absolute error on simulated well log data. And then the system is applied on the real well log data.

Title: Toward concept visualization through image generation Author: Tien Dat Nguyen Department: Institute of Formal and Applied Linguistics Supervisors: Pavel Pecina (Charles University in Prague), Angeliki Lazaridou, Raffaella Bernardi, Marco Baroni (University of Trento), Abstract: Computational linguistic and computer vision have a common way to embed the semantics of linguistic/visual units through vector representation. In addition, high-quality semantic representations can be effectively constructed thanks to recent advances in neural network methods. Nevertheless, the under- standing of these representations remains limited, so they need to be assessed in an intuitive way. Cross-modal mapping is mapping between vector semantic embedding of words and the visual representations of the corresponding objects from images. Inverting image representation involves learning an image inversion of visual vectors (SIFT, HOG and CNN features) to reconstruct the original one. The goal of this project is to build a complete pipeline, in which word represen- tations are transformed into image vectors using cross modal mapping and these vectors are projected to pixel space using inversion. This suggests that there might be a groundbreaking way to inspect and evaluate the semantics encoded in word representations by...