1 |
Projection of a Markov Process with Neural NetworksFolkesson, John January 2001 (has links)
In this work we have examined an application from the insurance industry. We first reformulate it into a problem of projecting a markov process. We then develop a method of carrying out the projection many steps into the future by using a combination of neural networks trained using a maximum entropy principle. This methodology improves on current industry standard solution in four key areas: variance, bias, confidence level estimation, and the use of inhomogeneous data. The neural network aspects of the methodology include the use of a generalization error estimate that does not rely on a validation set. We also develop our own approximation to the hessian matrix, which seems to be significantly better than assuming it to be diagonal and much faster than calculating it exactly. This hessian is used in the network pruning algorithm. The parameters of a conditional probability distribution were generated by a neural network, which was trained to maximize the log-likelihood plus a regularization term. In preparing the data for training the neural networks we have devised a scheme to decorrelate input dimensions completely, even non-linear correlations, which should be of general interest in its own right. The results we found indicate that the bias inherent in the current industry-standard projection technique is very significant. This work may be the only accurate measurement made of this important source of error.
|
2 |
Hybrid neural net and physics based model of a lithium ion batteryRefai, Rehan 12 July 2011 (has links)
Lithium ion batteries have become one of the most popular types of battery in consumer electronics as well as aerospace and automotive applications. The efficient use of Li-ion batteries in automotive applications requires well designed battery management systems. Low order Li-ion battery models that are fast and accurate are key to well- designed BMS. The control oriented low order physics based model developed previously cannot predict the temperature and predicts inaccurate voltage dynamics. This thesis focuses on two things: (1) the development of a thermal component to the isothermal model and (2) the development of a hybrid neural net and physics based battery model that corrects the output of the physics based model.
A simple first law based thermal component to predict the temperature model is implemented. The thermal model offers a reasonable approximation of the temperature dynamics of the battery discharge over a wide operating range, for both a well-ventilated battery as well as an insulated battery. The model gives an accurate prediction of temperature at higher SOC, but the accuracy drops sharply at lower SOCs. This possibly is due to a local heat generation term that dominates heat generation at lower SOCs.
A neural net based modeling approach is used to compensate for the lack of knowledge of material parameters of the battery cell in the existing physics based model. This model implements a neural net that corrects the voltage output of the model and adds a temperature prediction sub-network. Given the knowledge of the physics of the battery, sparse neural nets are used. Multiple types of standalone neural nets as well as hybrid neural net and physics based battery models are developed and tested to determine the appropriate configuration for optimal performance. The prediction of the neural nets in ventilated, insulated and stressed conditions was compared to the actual outputs of the batteries. The modeling approach presented here is able to accurately predict voltage output of the battery for multiple current profiles. The temperature prediction of the neural nets in the case of the ventilated batteries was harder to predict since the environment of the battery was not controlled. The temperature predictions in the insulated cases were quite accurate. The neural nets are trained, tested and validated using test data from a 4.4Ah Boston Power lithium ion battery cell. / text
|
3 |
Estudo da predição da circularidade e rugosidade de peças retificadas utilizando as redes neurais artificiaisFrança, Thiago Valle [UNESP] 19 January 2005 (has links) (PDF)
Made available in DSpace on 2014-06-11T19:24:47Z (GMT). No. of bitstreams: 0
Previous issue date: 2005-01-19Bitstream added on 2014-06-13T18:21:15Z : No. of bitstreams: 1
franca_tv_me_bauru.pdf: 6693171 bytes, checksum: a1e62f81ea86d1eeed97e598cb2bb8f5 (MD5) / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP) / Atualmente, a fabricação é caracterizada pela sua complexidade, pluralidade de disciplinas e crescente demanda de novas ferramentas e técnicas para a solução de difíceis problemas. As redes neurais artificiais oferecem uma nova e diferente alternativa para investigar e analisar os desafiadores tópicos relacionados à manufatura. Desta forma, estudou-se neste trabalho os assuntos relacionados à aplicação das redes neurais na predição da circularidade e rugosidade da peça retificada pela análise de algumas variáveis de saída do processo. Foram empregados nos ensaios de usinagem: um fluido de corte (óleo emulsionável), um rebolo superabrasivo de CBN com ligante vitrificado e peças temperadas e revenidas de aço VC-131. Este trabalho também utilizou outras tecnologias de otimização do processo de retificação, tais como: a utilização de defletores aerodinâmicos para a quebra da camada de ar e a refrigeração otimizada por meio de um jato de fluido direcionado. Os ensaios de usinagem foram realizados para gerar a base de dados utilizada nos testes das redes neurais (ensaios computacionais). Fez-se portanto, diversos experimentos variando-se a velocidade de avanço, ou mergulho do rebolo na peça. As variáveis de saída analisadas que serviram de dados de entrada para a RNA foram: a força tangencial de corte (Ft), a energia específica de retificação (u), o desgaste diametral do rebolo, o parâmetro DPO e a emissão acústica (EA). A rugosidade e circularidade foram utilizadas para o treinamento das RNA s. Nos testes computacionais, foram analisadas duas bases de dados: a primeira referente às médias de todos os 40 ciclos de retificação, já a segunda utilizou todos os valores destes 40 ciclos. Ainda foram examinadas diferentes combinações de dados de entrada para verificar a influência do parâmetro DPO na predição. Os resultados... / Nowadays, the manufacturing is characterized by its complexity, plurality of subjects and increasing demand of new tools and techniques for the solution of difficult problems. Artificial neural nets propose a new and different alternative to investigate and analyze the challenging topics related to the manufacturing. The objective of this work is to study the use of artificial neural nets in the prediction of roundness and roughness of a ground workpiece. It was used a CBN wheel, emulsion oil and workpieces made of VC-131 steel. This work also used other technologies of grinding optimization, such as: the use of a coolant shoe to break the air curtain layer in addition and the high pressure fluid jet. Grinding tests had been carried through to generate the database used in the artificial neural nets (computational tests). Different feed rates were used in these experiments to generate outputs such as: tangential cutting force (Ft), specific energy of grinding (u), diametrical wear of the wheel, DPO parameter and acoustic emission (EA). The roughness and roundness were used to train the RNA's. In the computational tests, it was verify the influence of the DPO parameter in the prediction as well as two different databases. The results suggest that this parameter (DPO) was not able to substitute the tangential cutting force (Ft) and the acoustic emission (EA) in the prediction. Moreover, it was verify the need of an input that represents the dynamic stiffness of the machine-tool-workpiece system to improve the roundness prediction.
|
4 |
Využití umělých neuronových sítí pro řešení úloh kombinatorické optimalizace / Using artificial neural networks to solve problems in combinatorial optimizationDvořák, Marek January 2014 (has links)
This thesis discusses combinatorial optimization problems, its characteristics and solving methods. Different types of such problems are presented here and I hint at solution using classical heuristical algorithms. In the next part, I focus on artificial neural networks, their description and classification. In the last part, I'm comparing two neural network approaches for solving a travelling salesman problem on several examples.
|
5 |
Techniky klasifikace proteinů / Protein Classification TechniquesDekrét, Lukáš January 2020 (has links)
Main goal of classifying proteins into families is to understand structural, functional and evolutionary relationships between individual proteins, which are not easily deducible from available data. Since the structure and function of proteins are closely related, determination of function is mainly based on structural properties, that can be obtained relatively easily with current resources. Protein classification is also used in development of special medicines, in the diagnosis of clinical diseases or in personalized healthcare, which means a lot of investment in it. I created a new hierarchical tool for protein classification that achieves better results than some existing solutions. The implementation of the tool was preceded by acquaintance with the properties of proteins, examination of existing classification approaches, creation of an extensive data set, realizing experiments and selection of the final classifiers of the hierarchical tool.
|
6 |
Optické zpracování dotazníkových dat / Optical processing of questionnairesNožka, Tomáš January 2011 (has links)
This master thesis deals with the principles of form design, form printing and form processing. Three different types of forms and applications for their detection are created with the reference of these principles. This application provides to create a new type of form and to print out a form. The application itself is implemented in C++ with the use of OpenCV library. This work describes the classification methods of direction finding marks, identification numbers and submission numbers, bar codes EAN-13, page numbers, answer fields and single answers. The classification of all the handwritten numbers is implemented by neural nets.
|
7 |
Age-Suitability Prediction for Literature Using Deep Neural NetworksBrewer, Eric Robert 30 July 2020 (has links)
Digital media holds a strong presence in society today. Providers of digital media may choose to obtain a content rating for a given media item by submitting that item to a content rating authority. That authority will then issue a content rating that denotes to which age groups that media item is appropriate. Content rating authorities serve publishers in many countries for different forms of media such as television, music, video games, and mobile applications. Content ratings allow consumers to quickly determine whether or not a given media item is suitable to their age or preference. Literature, on the other hand, remains devoid of a comparable content rating authority. If a new, human-driven rating authority for literature were to be implemented, it would be impeded by the fact that literary content is published far more rapidly than are other forms of digital media; humans working for such an authority simply would not be able to issue accurate content ratings for items of literature at their current rate of production. Thus, to provide fast, automated content ratings to items of literature (i.e., books), we propose a computer-driven rating system which predicts a book's content rating within each of seven categories: 1) crude humor/language; 2) drug, alcohol, and tobacco use; 3) kissing; 4) profanity; 5) nudity; 6) sex and intimacy; and 7) violence and horror given the text of that book. Our computer-driven system circumvents the major hindrance to any theoretical human-driven rating system previously mentioned--namely infeasibility in time spent. Our work has demonstrated that mature content of literature can be accurately predicted through the use of natural language processing and machine learning techniques.
|
8 |
ASCAT Wind Estimation at 2.5 km Resolution Supported by Machine Learning Rain DetectionKjar, Joshua Benjamin 01 December 2022 (has links)
The Advanced Scatterometer (ASCAT) is a C-band scatterometer designed to be less sensitive to rain contamination than other higher frequency scatterometers. However, the radar backscatter is still affected by rain which increases error during wind estimation. The error can be reduced in rainy conditions by combining a rain backscatter model with the existing wind only (WO) backscatter model to perform simultaneous wind and rain (SWR) estimation. I derive and test several 2.5 km resolution rain backscatter models for ASCAT data which are used with the WO model to estimate the near surface winds. Various rain models optimal for different purposes are discussed. The best rain model for estimating wind speed lowers the root mean square error (RMSE) in the presence of rain by 13.6% when compared to using the WO model alone. The rain model which best predicts rain rates has a RMSE of 7.9 mm/h. A neural network (NN) is designed to discriminate the presence of rain using ASCAT's backscatter measurements. Such a NN enables the SWR algorithm to be used only on rainy samples and thus improves estimation. By removing all samples identified by the NN as rain, the WO algorithm's speed estimate improved by 2.83%.
|
9 |
A real-time neural-net computing approach to the detection and classification of underwater acoustic transientsHemminger, Thomas Lee January 1992 (has links)
No description available.
|
10 |
Elaboração de rankings por meio do uso de técnicas estruturadas: uma aplicação no setor de seguros privados / Preparation of rankings through the use of structured techniques: an application in the sector of private insuranceAraujo, Pedro Henrique de Sousa Leão 26 November 2008 (has links)
A demanda por metodologias para classificação de empresas que possuam características em comum e que componham um mesmo setor de atividade tem instigado pesquisadores a avaliar alternativas que sejam fidedignas à representação da realidade, e que façam uso reduzido de quesitos voltados à subjetividade de julgamento. Por isso, adotou-se como objetivo desta pesquisa a elaboração de rankings utilizando as técnicas de análise por envoltória de dados e redes neurais artificiais, com aplicação no setor de seguros privados, setor este de forte influência na economia nacional. Como dados para a aplicação das duas técnicas propostas, foram considerados alguns indicadores, via de regra adotados pelo setor, para avaliar o desempenho das empresas no cumprimento de suas atividades. Como resultado obtido, foi verificado que a ponderação direta de acordo com a importância de cada indicador não representa a única forma de apresentar uma ordenação justa das empresas consideradas com base em seus desempenhos. Por meio das técnicas utilizadas, foi observado que empresas que mantiveram um resultado satisfatório na maioria das variáveis consideradas obtiveram os melhores posicionamentos nos rankings. A rede neural, mesmo requerendo um maior tempo de processamento e oferecendo uma complexidade de aplicação maior que a técnica DEA, apresentou resultados mais consistentes. / The demand for methodologies and procedures to classify companies that have some characteristics in common and that are part of the same activity sector has instigated researchers to evaluate alternatives that represent the real situation according to their performance as business units, by making use of reduced amount of subjectivity in the performance judgment. Therefore, this research has as its main goal the objective to set up some rankings using the techniques of analysis and data envelopment by artificial neural networks, by making applying these techniques in the insurance sector, a activity with great influence in national economy. As data for the implementation of both techniques proposed, some indicators well known by specialists were considered to evaluate the performance of companies in their activities. As a result, it was found that the direct weighting used to enforce the importance of each indicator is not the only way to make a fair ranking of the insurance companies. About the techniques used, it was observed that companies that have maintained a satisfactory performance in most of the variables considered occupied best positions in the rankings. The neural network, even though requiring a longer processing time, and offering a greater complexity of application than DEA technique, showed some more consistent results.
|
Page generated in 0.0649 seconds