Дисертації з теми "Fixed neural network"

This thesis presents the development of small, inexpensive unmanned aerial vehicles (UAVs) to achieve autonomous fight. Fixed wing hobby model planes are modified and instrumented to form experimental platforms. Different sensors employed to collect the flight data are discussed along with their calibrations. The time constant and delay for the servo-actuators for the platform are estimated. Two different data collection and processing units based on micro-controller and PC104 architectures are developed and discussed. These units are also used to program the identification and control algorithms. Flight control of fixed wing UAVs is a challenging task due to the coupled, time-varying, nonlinear dynamic behaviour. One of the possible alternatives for the flight control system is to use the intelligent adaptive control techniques that provide online learning capability to cope with varying dynamics and disturbances. Neural network based indirect adaptive control strategy is applied for the current work. The two main components of the adaptive control technique are the identification block and the control block. Identification provides a mathematical model for the controller to adapt to varying dynamics. Neural network based identification provides a black-box identification technique wherein a suitable network provides prediction capability based upon the past inputs and outputs. Auto-regressive neural networks are employed for this to ensure good retention capabilities for the model that uses the past outputs and inputs along with the present inputs. Online and offline identification of UAV platforms are discussed based upon the flight data. Suitable modifications to the Levenberg-Marquardt training algorithm for online training are proposed. The effect of varying the different network parameters on the performance of the network are numerically tested out. A new performance index is proposed that is shown to improve the accuracy of prediction and also reduces the training time for these networks. The identification algorithms are validated both numerically and flight tested. A hardware-in-loop simulation system has been developed to test the identification and control algorithms before flight testing to identify the problems in real time implementation on the UAVs. This is developed to keep the validation process simple and a graphical user interface is provided to visualise the UAV flight during simulations. A dual neural network controller is proposed as the adaptive controller based upon the identification models. This has two neural networks collated together. One of the neural networks is trained online to adapt to changes in the dynamics. Two feedback loops are provided as part of the overall structure that is seen to improve the accuracy. Proofs for stability analysis in the form of convergence of the identifier and controller networks based on Lyapunov's technique are presented. In this analysis suitable bounds on the rate of learning for the networks are imposed. Numerical results are presented to validate the adaptive controller for single-input single-output as well as multi-input multi-output subsystems of the UAV. Real time validation results and various flight test results confirm the feasibility of the proposed adaptive technique as a reliable tool to achieve autonomous flight. The comparison of the proposed technique with a baseline gain scheduled controller both in numerical simulations as well as test flights bring out the salient adaptive feature of the proposed technique to the time-varying, nonlinear dynamics of the UAV platforms under different flying conditions.

Fixed-object crashes have long time been considered as major roadway safety concerns. While previous relevant studies tended to address such crashes in the context of roadway departures, and heavily relied on police-reported accidents data, this study integrated the SHRP2 NDS and RID data for analyses, which fully depicted the prior to, during, and after crash scenarios. A total of 1,639 crash, near-crash events, and 1,050 baseline events were acquired. Three analysis methods: logistic regression, support vector machine (SVM) and artificial neural network (ANN) were employed for two responses: crash occurrence and severity level. Logistic regression analyses identified 16 and 10 significant variables with significance levels of 0.1, relevant to driver, roadway, environment, etc. for two responses respectively. The logistic regression analyses led to a series of findings regarding the effects of explanatory variables on fixed-object event occurrence and associated severity level. SVM classifiers and ANN models were also constructed to predict these two responses. Sensitivity analyses were performed for SVM classifiers to infer the contributing effects of input variables. All three methods obtained satisfactory prediction performance, that was around 88% for fixed-object event occurrence and 75% for event severity level, which indicated the effectiveness of NDS event data on depicting crash scenarios and roadway safety analyses.
Master of Science

Coal gasification stripped gas liquor (CGSGL) wastewater contains large quantities of complex organic and inorganic pollutants which include phenols, ammonia, hydantoins, furans, indoles, pyridines, phthalates and other monocyclic and polycyclic nitrogen-containing aromatics, as well as oxygen- and sulphur-containing heterocyclic compounds. The performance of most conventional aerobic systems for CGSGL wastewater is inadequate in reducing pollutants contributing to chemical oxygen demand (COD), phenols and ammonia due to the presence of toxic and inhibitory organic compounds. There is an ever-increasing scarcity of freshwater in South Africa, thus reclamation of wastewater for recycling is growing rapidly and the demand for higher effluent quality before being discharged or reused is also increasing. The selection of hybrid fixed-film bioreactor (HFFBR) systems in the detoxification of a complex mixture of compounds such as those found in CGSGL has not been investigated. Thus, the objective of this study was to investigate the detoxification of the CGSGL in a H-FFBR bioaugmented with a mixed-culture inoculum containing Pseudomonas putida, Pseudomonas plecoglossicida, Rhodococcus erythropolis, Rhodococcus qingshengii, Enterobacter cloacae, Enterobacter asburiae strains of bacteria, as well as the seaweed (Silvetia siliquosa) and diatoms. The results indicated a 45% and 79% reduction in COD and phenols, respectively, without bioaugmentation. The reduction in COD increased by 8% with inoculum PA1, 13% with inoculum PA2 and 7% with inoculum PA3. Inoculum PA1 was a blend of Pseudomonas, Enterobacter and Rhodococcus strains, inoculum PA2 was a blend of Pseudomonas putida iistrains and inoculum PA3 was a blend of Pseudomonas putida and Pseudomonas plecoglossicida strains. The results also indicated that a 70% carrier fill formed a dense biofilm, a 50% carrier fill formed a rippling biofilm and a 30% carrier fill formed a porous biofilm. The autotrophic nitrifying bacteria were out-competed by the heterotrophic bacteria of the genera Thauera, Pseudaminobacter, Pseudomonas and Diaphorobacter. Metagenomic sequencing data also indicated significant dissimilarities between the biofilm, suspended biomass, effluent and feed microbial populations. A large population (20% to 30%) of unclassified bacteria were also present, indicating the presence of novel bacteria that may play an important role in the treatment of the CGSGL wastewater. The artificial neural network (ANN) model developed in this study is a novel virtual tool for the prediction of COD and phenol removal from CGSGL wastewater treated in a bioaugmented H-FFBR. Knowledge extraction from the trained ANN model showed that significant nonlinearities exist between the H-FFBR operational parameters and the removal of COD and phenol. The predictive model thus increases knowledge of the process inputs and outputs and thus facilitates process control and optimisation to meet more stringent effluent discharge requirements.
Thesis (PhD)--University of Pretoria, 2017.
Chemical Engineering
PhD
Unrestricted

Дипломна робота магістра присвячена розробленню інформаційної системи для класифікації твердих побутових відходів. У роботі спроектовано та розроблено класифікатор зображень твердих побутових відходів з використанням згорткових нейронних мереж. Вперше використано інтерфейс програмного забезпечення Телеграм для реалізації системи класифікації таких зображень.

Made available in DSpace on 2014-12-17T14:55:47Z (GMT). No. of bitstreams: 1 CarlosAAS_DISSERT_1-60.pdf: 4186909 bytes, checksum: cebf9d80edc07d16ef618a3095ead927 (MD5) Previous issue date: 2010-06-30
This study shows the implementation and the embedding of an Artificial Neural Network (ANN) in hardware, or in a programmable device, as a field programmable gate array (FPGA). This work allowed the exploration of different implementations, described in VHDL, of multilayer perceptrons ANN. Due to the parallelism inherent to ANNs, there are disadvantages in software implementations due to the sequential nature of the Von Neumann architectures. As an alternative to this problem, there is a hardware implementation that allows to exploit all the parallelism implicit in this model. Currently, there is an increase in use of FPGAs as a platform to implement neural networks in hardware, exploiting the high processing power, low cost, ease of programming and ability to reconfigure the circuit, allowing the network to adapt to different applications. Given this context, the aim is to develop arrays of neural networks in hardware, a flexible architecture, in which it is possible to add or remove neurons, and mainly, modify the network topology, in order to enable a modular network of fixed-point arithmetic in a FPGA. Five synthesis of VHDL descriptions were produced: two for the neuron with one or two entrances, and three different architectures of ANN. The descriptions of the used architectures became very modular, easily allowing the increase or decrease of the number of neurons. As a result, some complete neural networks were implemented in FPGA, in fixed-point arithmetic, with a high-capacity parallel processing
Este estudo consiste na implementa??o e no embarque de uma Rede Neural Artificial (RNA) em hardware, ou seja, em um dispositivo program?vel do tipo field programmable gate array (FPGA). O presente trabalho permitiu a explora??o de diferentes implementa??es, descritas em VHDL, de RNA do tipo perceptrons de m?ltiplas camadas. Por causa do paralelismo inerente ?s RNAs, ocorrem desvantagens nas implementa??es em software, devido ? natureza sequencial das arquiteturas de Von Neumann. Como alternativa a este problema, surge uma implementa??o em hardware que permite explorar todo o paralelismo impl?cito neste modelo. Atualmente, verifica-se um aumento no uso do FPGA como plataforma para implementar as Redes Neurais Artificiais em hardware, explorando o alto poder de processamento, o baixo custo, a facilidade de programa??o e capacidade de reconfigura??o do circuito, permitindo que a rede se adapte a diferentes aplica??es. Diante desse contexto, objetivou-se desenvolver arranjos de redes neurais em hardware, em uma arquitetura flex?vel, nas quais fosse poss?vel acrescentar ou retirar neur?nios e, principalmente, modificar a topologia da rede, de forma a viabilizar uma rede modular em aritm?tica de ponto fixo, em um FPGA. Produziram-se cinco s?nteses de descri??es em VHDL: duas para o neur?nio com uma e duas entradas, e tr?s para diferentes arquiteturas de RNA. As descri??es das arquiteturas utilizadas tornaram-se bastante modulares, possibilitando facilmente aumentar ou diminuir o n?mero de neur?nios. Em decorr?ncia disso, algumas redes neurais completas foram implementadas em FPGA, em aritm?tica de ponto fixo e com alta capacidade de processamento paralelo

In this diploma thesis, the reader is introduced to artificial neural networks and convolutional neural networks. Based on that, the design and implementation of a new library for convolutional neural networks is described. The library is then evaluated on widely used datasets and compared to other publicly available libraries. The added benefit of the library, that makes it unique, is its independence on data types. Each layer may contain up to three independent data types - for weights, for inference and for training. For the purpose of evaluating this feature, a data type with fixed point representation is also part of the library. The effects of this representation on trained net accuracy are put to a test.

This master's thesis describes the design of effective working artificial neural network in FPGA Virtex-5 series with the maximum use of the possibility of parallelization. The theoretical part contains basic information on artificial neural networks, FPGA and VHDL. The practical part describes the used format of the variables, creating non-linear function, the principle of calculation the single layers, or the possibility of parameter settings generated artificial neural networks.

In recent years there has been a great interest in neural networks, since neural networks are capable of performing pattern recognition, classification, decision, search, and optimization. A key element of most neural network systems is the massive number of weighted interconnections (synapses) used to tie relatively simple processing elements (neurons) together in a useful architecture. The inherent parallelism and interconnection capability of optics make it a likely candidate for the implementation of the neural network interconnection process. While there are several optical technologies worth exploring, this dissertation examines the capabilities and limitations of using fixed planar holographic interconnects in a neural network system. While optics is well suited to the interconnection task, nonlinear processing operations are difficult to implement in optics and better suited to electronic implementations. Therefore, a hybrid neural network architecture of planar interconnection holograms and opto-electronic neurons is a sensible approach to implementing a neural network. This architecture is analyzed. The interconnection hologram must accurately encode synaptic weights, have a high diffraction efficiency, and maximize the number of interconnections. Various computer generated hologram techniques are tested for their ability to produce the interconnection hologram. A new technique using the Gerchberg-Saxton process followed by a random-search error minimization produces the highest interconnect accuracy and highest diffraction efficiency of the techniques tested. The analysis shows that a reasonable size planar hologram has a capacity to connect 5000 neuron outputs to 5000 neuron inputs and that the bipolar synaptic weights can have an accuracy of approximately 5 bits. To demonstrate the concept of an opto-electronic neural network and planar holographic interconnects, a Hopfield style associative memory is constructed and shown to perform almost as well as an ideal system.

This thesis presents the development of a Neural Network Based Controller (NNBC) for chain grate stoker fired boilers. The objective of the controller was to increase combustion efficiency and maintain pollutant emissions below future medium term stringent legislation. Artificial Neural Networks (ANNs) were used to estimate future emissions from and control the combustion process. Initial tests at Casella CRE Ltd demonstrated the ability of ANNs to characterise the complex functional relationships which subsisted in the data set, and utilised previously gained knowledge to deliver predictions up to three minutes into the future. This technique was then built into a carefully designed control strategy that fundamentally mimicked the actions of an expert boiler operator, to control an industrial chain grate stoker at HM Prison Garth, Lancashire. Test results demonstrated that the developed novel NNBC was able to control the industrial stoker boiler plant to deliver the load demand whilst keeping the excess air level to a minimum. As a result the NNBC also managed to maintain the pollutant emissions within probable future limits for this size of boiler. This prototype controller would thus offer the industrial coal user with a means to improve the combustion efficiency on chain grate stokers as well as meeting medium term legislation limits on pollutant emissions that could be imposed by the European Commission.

The growing complexity of neural networks makes their deployment on resource-constrained embedded or mobile devices challenging. With millions of weights and biases, modern deep neural networks can be computationally intensive, with large memory, power and computational requirements. In this thesis, we devise and explore three quantization methods (post-training, in-training and combined quantization) that quantize 32-bit floating-point weights and biases to lower bit width fixed-point parameters while also achieving significant pruning, leading to model compression. We use the total accumulated absolute gradient over the training process as the indicator of importance of a parameter to the network. The most important parameters are quantized by the smallest amount. The post-training quantization method sorts and clusters the accumulated gradients of the full parameter set and subsequently assigns a bit width to each cluster. The in-training quantization method sorts and divides the accumulated gradients into two groups after each training epoch. The larger group consisting of the lowest accumulated gradients is quantized. The combined quantization method performs in-training quantization followed by post-training quantization. We assume storage of the quantized parameters using compressed sparse row format for sparse matrix storage. On LeNet-300-100 (MNIST dataset), LeNet-5 (MNIST dataset), AlexNet (CIFAR-10 dataset) and VGG-16 (CIFAR-10 dataset), post-training quantization achieves 7.62x, 10.87x, 6.39x and 12.43x compression, in-training quantization achieves 22.08x, 21.05x, 7.95x and 12.71x compression and combined quantization achieves 57.22x, 50.19x, 13.15x and 13.53x compression, respectively. Our methods quantize at the cost of accuracy, and we present our work in the light of the accuracy-compression trade-off.
Master of Science
Neural networks are being employed in many different real-world applications. By learning the complex relationship between the input data and ground-truth output data during the training process, neural networks can predict outputs on new input data obtained in real time. To do so, a typical deep neural network often needs millions of numerical parameters, stored in memory. In this research, we explore techniques for reducing the storage requirements for neural network parameters. We propose software methods that convert 32-bit neural network parameters to values that can be stored using fewer bits. Our methods also convert a majority of numerical parameters to zero. Using special storage methods that only require storage of non-zero parameters, we gain significant compression benefits. On typical benchmarks like LeNet-300-100 (MNIST dataset), LeNet-5 (MNIST dataset), AlexNet (CIFAR-10 dataset) and VGG-16 (CIFAR-10 dataset), our methods can achieve up to 57.22x, 50.19x, 13.15x and 13.53x compression respectively. Storage benefits are achieved at the cost of classification accuracy, and we present our work in the light of the accuracy-compression trade-off.

This thesis is concerned with the implementation of Artificial Neural Networks (ANNs) to monitor and control chain grate stoker-fired coal boilers with a view to improving the combustion efficiency whilst minimising pollutant emissions. A novel Neural Network Based Controller (NNBC) was developed following a comprehensive set of experiments carried out on a stoker test facility at the Coal Research Establishment (CRE) Ltd., before being evaluated on an industrial chain grate stoker at Her Majesty's Prison Garth, Leyland. The NNBC mimicked the actions of an expert boiler operator, by providing 'near optimum' settings of coal feed and air flow, as well as taking into account the correct 'staging' sequence of these parameters during load following conditions, before subsequently fine tuning the combustion air under quasi steady- state conditions. Test results from the on-line implementation of the NNBC on both chain grate stoker plants have demonstrated that improved transient and steady state combustion conditions were attained without having any adverse effect on the pollutant emissions nor the integrity of the appliances. A novel combustion monitoring system was also developed during the course of the work that can be used to infer the stability of combustion on the fire bed, following a pilot study of the 'flame front' movement during boiler load changes on the stoker test facility at CRE. This novel low-cost flame front monitor was rigorously tested on the industrial stoker plant, and long hours of successful on-line operation were achieved. It was also demonstrated with the use of ANNs, that the data gathered from the novel flame front monitor can be processed to yield evidence concerning movement of the ignition plane over a short period of time (several minutes). The prototype controller and flame front monitor would thus provide both stoker manufacturers and users with a means of meeting future legislative limits on pollutant emissions as indicated by the European Commission, as well as improving the combustion efficiency of this type of coal firing equipment Finally, ANNs were also used as a simplistic means to represent the complex coal combustion process on the bed of the stoker test facility whilst burning a particular type of coal. The resultant 'black-box' models of the combustion derivatives were able to represent the dynamics of the process and delivered accurate one-step ahead predictions over a wide range of unseen data. The work demonstrated the complex functional mapping capability of ANNs and also addressed the deficiencies in mathematical modelling of the coal combustion process on fixed grate, as indicated in the literature.

Activities in the motor cortex are found to be dynamical in nature. Modeling these activities and comparing them with neural recordings helps in understanding the underlying mechanism for the generation of these activities. For this purpose, Recurrent Neural networks or RNNs, have emerged as an appropriate tool. A clear understanding of how the design choices associated with these networks affect the learned dynamics and internal representation still remains elusive. A previous work exploring the dynamical properties of discrete time RNN architectures (LSTM, UGRNN, GRU, and Vanilla) such as the fixed point topology and the linearised dynamics remains invariant when trained on 3 bit Flip- Flop task. In contrast, they show that these networks have unique representational geometry. The goal for this work is to understand if these observations also hold for networks that are more biologically realistic in terms of neural activity. Therefore, we chose to analyze rate networks that have continuous dynamics and biologically realistic connectivity constraints and on spiking neural networks, where the neurons communicate via discrete spikes as observed in the brain. We reproduce the aforementioned study for discrete architectures and then show that the fixed point topology and linearized dynamics remain invariant for the rate networks but the methods are insufficient for finding the fixed points of spiking networks. The representational geometry for the rate networks and spiking networks are found to be different from the discrete architectures but very similar to each other. Although, a small subset of discrete architectures (LSTM) are observed to be close in representation to the rate networks. We show that although these different network architectures with varying degrees of biological realism have individual internal representations, the underlying dynamics while performing the task are universal. We also observe that some discrete networks have close representational similarities with rate networks along with the dynamics. Hence, these discrete networks can be good candidates for reproducing and examining the dynamics of rate networks.
Aktiviteter i motorisk cortex visar sig vara dynamiska till sin natur. Att modellera dessa aktiviteter och jämföra dem med neurala inspelningar hjälper till att förstå den underliggande mekanismen för generering av dessa aktiviteter. För detta ändamål har återkommande neurala nätverk eller RNN uppstått som ett lämpligt verktyg. En tydlig förståelse för hur designvalen associerade med dessa nätverk påverkar den inlärda dynamiken och den interna representationen är fortfarande svårfångad. Ett tidigare arbete som utforskar de dynamiska egenskaperna hos diskreta RNN- arkitekturer (LSTM, UGRNN, GRU och Vanilla), såsom fastpunkts topologi och linjäriserad dynamik, förblir oförändrad när de tränas på 3-bitars Flip- Flop-uppgift. Däremot visar de att dessa nätverk har unik representationsgeometri. Målet för detta arbete är att förstå om dessa observationer också gäller för nätverk som är mer biologiskt realistiska när det gäller neural aktivitet. Därför valde vi att analysera hastighetsnätverk som har kontinuerlig dynamik och biologiskt realistiska anslutningsbegränsningar och på spikande neurala nätverk, där neuronerna kommunicerar via diskreta spikar som observerats i hjärnan. Vi reproducerar den ovannämnda studien för diskreta arkitekturer och visar sedan att fastpunkts topologi och linjäriserad dynamik förblir oförändrad för hastighetsnätverken men metoderna är otillräckliga för att hitta de fasta punkterna för spiknätverk. Representationsgeometrin för hastighetsnätverk och spiknätverk har visat sig skilja sig från de diskreta arkitekturerna men liknar varandra. Även om en liten delmängd av diskreta arkitekturer (LSTM) observeras vara nära i förhållande till hastighetsnäten. Vi visar att även om dessa olika nätverksarkitekturer med varierande grad av biologisk realism har individuella interna representationer, är den underliggande dynamiken under uppgiften universell. Vi observerar också att vissa diskreta nätverk har nära representationslikheter med hastighetsnätverk tillsammans med dynamiken. Följaktligen kan dessa diskreta nätverk vara bra kandidater för att reproducera och undersöka dynamiken i hastighetsnät.

Made available in DSpace on 2017-07-10T18:07:58Z (GMT). No. of bitstreams: 1 Lucas Maycon Hoff Zeni.pdf: 1421377 bytes, checksum: 75c6a9407a955e26c7fd4db2939b1b79 (MD5) Previous issue date: 2012-02-24
Coordenação de Aperfeiçoamento de Pessoal de Nível Superior
Configuration of fixed bed that operates with biphasic flow is used in industrial operations such as the Fischer-Tropsch, hydrogenation, and residual water treatments. Vital information for the project and operation of this type of bed is in its characteristics fluid-dynamic and among these characteristics the flow regime because these have a direct influence transferring heat and mass present in the bed. In the two-phase flow with ascendant flow through fixed bed, three distinct regimes can be identified: the bubble regime, for low gas flow; pulsating regime, for moderate liquid and gas flow; and spray regime; for low flow of liquid and high flow rates of gas. Although there are different techniques to determine flow regimes, the most used is the visual identification. Thus, this research aims to develop, by using artificial neural networks (ANNs) a way to determine, for a given set of liquid-gas flow what out-flow regime the bed presents. To do so, firstly, the out-flow regime were identified by using water and air, respectively flux mass flowing varying from 2 to 16.5 kg.m-2.s-1 and from 0 to 0.6 kg.m-2.s-1, flowing up-words through a fixed bed packed with glass spheres measuring from 2.7 to 3.5 mm of diameter. The network proposed to identify the regimes contains Multiple Layers Perceptron architecture (PML) trained by the back propagation algorithm put together by applying the Multiple Back-Propagation (MBP) software, version 2.2.3 consistently with two input neurons, two intermediate layers, and four output neurons. The number of neurons of the intermediate layers was assorted to find out the best configuration. As activation of function, logistic, tangent, hyperbolic, and Gaussian were tested. Observed results showed that it is possible the identification of regimes through neural networks and among those tested the one that showed the best performance was the one that used the hyperbolic-tangent activation function; 10 neurons in the first hidden layer, and 12 neurons in the second hidden layer.
A configuração de leito fixo que opera com escoamento bifásico é muito utilizada em operações industriais, tais como síntese de Fischer-Tropsch, hidrogenação e tratamento de águas residuais. Uma informação vital para projeto e operação deste tipo de leito está nas características fluidodinâmicas, e dentre estas características podem ser citados os regimes de escoamento, pois estes influenciam diretamente nas transferências de calor e massa presentes no leito. No escoamento bifásico com fluxo ascendente através de leito fixo podem ser identificados três regimes distintos: regime bolha, para baixas vazões de gás; regime pulsante, para vazões moderadas de líquido e gás; e regime spray, para baixas vazões de líquidos e altas vazões de gás. Apesar de haver diferentes técnicas para a determinação dos regimes de escoamento, a mais empregada é a identificação visual. Sendo assim, esta pesquisa tem por objetivo desenvolver, por meio da utilização de redes neurais artificiais (RNA s), uma maneira de determinar, para um dado conjunto de vazões gás-líquido, qual regime de escoamento o leito apresenta. Para isto, os regimes de escoamento primeiramente foram identificados utilizando água e ar, respectivamente com fluxo mássico variando de 2 a 16,5 kg.m-2.s-1 e de 0 a 0,6 kg.m-2.s-1, escoando em fluxo ascendente por meio de um leito fixo recheado com esferas de vidro de diâmetro entre 2,7 e 3,5 mm. A rede proposta para a identificação dos regimes possui arquitetura perceptron de múltiplas camadas (MLP) treinada pelo algoritmo backpropagation e foi montada utilizando o programa freeware Multiple Back-Propagation (MBP) versão 2.2.3 sempre com dois neurônios de entrada, duas camadas intermediárias e quatro neurônios de saída. O número de neurônios das camadas intermediárias foi variado a fim de descobrir a melhor configuração. Como função de ativação, foram testadas as funções logística, tangente hiperbólica e gaussiana. Os resultados observados mostram que é possível a identificação dos regimes por meio de redes neurais e dentre as configurações testadas, a que apresentou melhor desempenho foi a rede que utilizou a função de ativação tangente hiperbólica, 10 neurônios na primeira camada oculta e 12 neurônios na segunda camada oculta.

Made available in DSpace on 2017-07-10T18:08:16Z (GMT). No. of bitstreams: 1 Thiago DArisbo.pdf: 2108504 bytes, checksum: 7b8aad29ec7d75a6fd370e54a95cd849 (MD5) Previous issue date: 2011-02-24
Coordenação de Aperfeiçoamento de Pessoal de Nível Superior
Ion exchange is a process that is used in the treatment of aqueous industrial effluents containing organic compounds and heavy metals. The fixed bed columns are longer applied by allowing the process to occur continuously (cycles of regeneration). The design and process optimization of the ion exchange column requires the use of mathematical models. Phenomenological models of these systems involve the solution of partial differential and algebraic equations. The equilibrium data for ion exchange processes are usually described by the Mass Action Law (MAL), which can be considered non-ideality of aqueous and solid phases. Artificial Neural Networks (ANN) are being used successfully for the study of equilibrium data because they are empirical models and don t demand a mathematical rigor. This work aimed to evaluate the applicability of the hybrid model to describe the dynamics of ion exchange in fixed beds of binary systems. This system consists of partial differential equations obtained from mass balance in fluid phases in the ion exchanger and ANN to describe the balance. LAM was adjusted to experimental data of ion exchange equilibrium and then were generated 4200 data sets for each binary pair studied, which served as training for RNA. We tested networks with different structures, with one and two input layers. The 3-3-2 structure was used in the simulations of the hybrid model because it was the best represented the systems during the training phase. The differential equations were solved by the lines method. A computer program in FORTRAN language was developed for solving the model equations. DASSL subroutine was used to solve the equations. The performance of the hybrid model was evaluated from the results obtained with the phenomenological model, in which case the equilibrium description was made with the use of MAL. It also was the analysis of results from the comparison of experimental data. To evaluate the model we used data from the literature of ion exchange in Amberlite IR 120 resin on the systems Cu-Na and Zn-Na and in NaY zeolite on Fe-Na and Zn-Na. Both models were efficient to describe the dynamics of ion-exchange fixed bed columns, and the hybrid model had the advantage of the reduced computational time (82% reduction on average) as a result of not needing to solve a nonlinear equation.
A troca iônica é um processo muito utilizado no tratamento de efluentes industriais aquosos contendo compostos orgânicos e metais pesados. As colunas de leito fixo são mais aplicadas por permitir que o processo ocorra de maneira contínua (ciclos de regeneração). O projeto e a otimização de processos de troca iônica em coluna requer o uso de modelos matemáticos. Os modelos fenomenológicos destes sistemas envolvem a resolução de equações diferenciais parciais e algébricas. Os dados de equilíbrio de processos de troca iônica geralmente são descritos pela Lei da Ação das Massas (LAM), na qual podem ser consideradas as não idealidades das fases aquosa e sólida. As Redes Neurais Artificiais (RNA) estão sendo utilizadas com sucesso para o estudo destes dados de equilíbrio por serem modelos empíricos e não demandarem tal rigor matemático. Esta dissertação teve por objetivo avaliar a aplicabilidade do modelo híbrido para descrever a dinâmica do processo de troca iônica em leito fixo de sistemas binários. Este sistema é constituído de equações diferenciais parciais obtidas por meio de balanço de massa nas fases fluida e no trocador iônico e de RNA para descrever o equilíbrio. A LAM foi ajustada a dados experimentais de equilíbrio de troca iônica e, então, foram gerados conjuntos de 4200 dados para cada par binário estudado, os quais serviram como treinamento para a RNA. Foram testadas redes com diferentes estruturas, com uma e com duas camadas de entrada. A estrutura 3-3-2 foi utilizada nas simulações do modelo híbrido, pois foi a que melhor representou os sistemas na etapa de treinamento. As equações diferenciais foram resolvidas pelo método das linhas. Um programa computacional em linguagem FORTRAN foi desenvolvido para a resolução das equações do modelo. Foi utilizada a sub-rotina DASSL para resolver as equações. O desempenho do modelo híbrido foi avaliada a partir dos resultados obtidos com o modelo fenomenológico, sendo que neste caso a descrição do equilíbrio foi feita pelo uso da LAM. Também foi feita a análise dos resultados a partir da comparação dos dados experimentais. Para avaliar o modelo foram utilizados dados da literatura de troca iônica em resina Amberlite IR 120 dos sistemas Cu-Na e Zn-Na e na zeólita NaY dos sistemas Fe-Na e Zn-Na. Ambos os modelos foram eficientes para descrever a dinâmica de troca iônica de colunas de leito fixo, sendo que o modelo híbrido apresentou como vantagem o menor tempo computacional (82% de redução em média) em decorrência de não necessitar resolver a equação não-linear.

碩士
國立臺北科技大學
生產系統工程與管理研究所
90
The characteristics of services include intangibility、inseparability、heterogeneity and perishability. Due to these characteristics, it is hard to make zero-defect services. Because service failure is unavoidable, service recovery can enhance customer satisfaction. Since 1998 the telecommunication market deregulation and liberalization, the market structure is different from before. The same target of every company is how to defeat the competitor and increase market shares. The most important thing is to understand the situation of service failure and then recovery it. This research discusses the kinds of service failure and recovery、the behavior of customer complain and the relation between service failure、service recovery、customer satisfaction and repurchase intention. We choice the research business is fixed network. By questionnaire, the input data is the variable of the kinds of service failure、its criticality and population. The output data is the transaction that customers hope to recovery. The model is established by statistical（Regression Analysis）and neural network methods（Back-Propagation Network、General Regression Neural Network）. And then choice the smallest MSE model builds up the diagnosis system of service recovery. The research results showed that the slow speed of service occurs very often. The company recovery is the smallest cost and simple way to deal with customer complain. But customers that hope to recovery ways are fixed or improved the service failure、communicate rating discount. We choice General Regression Neural Network model and Visual Basic 6.0 to design the diagnosis system of service recovery. By this recovery system, company can do appropriate recovery strategy.

碩士
高雄工學院
電機電力研究所
85
Due to the very significance of power system load forecasting to electric util ity company, a wide variety of procedures for load forecasting have been propo sed in the last two decades. In recent years, the neural network (NN) technolo gy has been applied widely in this area based on its excellent learning abilit y. In most of these studies, NN structure utilized is fixed that means the NN keeps same size during the training and testing phases. It automatically devel ops an internal non-linear, complex relationship between power load and its in fluencing factors such as weather information through a training process on th e historical data. Then, the trained NN can be used to carry out the forecast no matter the training is appropriate or not. However, the correlations betwee n load and its influencing factors are various, depended very much on geograph ics, seasons, and the behavior of consumption of customers. the improper infor mation will make fixed NN to an ill-learning and cause a poor forecasting.In t his research, the NN structure utilized is non-fixed that means the NN's size keeps changing based on different situations during its learning and testing p rocesses. The correlations of load and its influencing factors on historical d ata are analyzed precisely. The problem of improper influencing is handled bef ore NN's training. Therefore, it makes non-fixed NN develop a more accurate mo del and then carry out a better forecasting. Furthermore, the phenomenon of ov ertraining is always happened in NN's learning with a non-stationary environme nt that makes an ill result of load forecasting . In this research, the modifi cation of learning rate based on fuzzy theory is also investigated. A proper p rocedure how to solve this training problem is proposed. From the point of com mercialization of NN, we hope to make NN technology utilizing in this field ha s a real potential and more promising.

碩士
亞洲大學
光電與通訊學系
107
In this thesis, the research on the learning performances of fixed-point and float-ing-point implementations in single-layer and double-layer recurrent neural net-works is proposed. The recursive neural network is in fact the combination of feed-forward neural networks and infinite impulse response (IIR) filters. We have devel-oped the optimized filter structure and investigated its learning behaviors in finite-precision digital devices. In this paper, we test the robustness of fixed-point numbers and floating-point numbers with different finite precisions, and optimize the effect of finite precision on the state-space structure, so that the sensitivity of system pa-rameters with finite-precision can be effectively reduced in shorter word length. Once the optimal structure is synthesized, the RNN system can be stabilized with shorter word-length. Then, the performance of the finite precision of the single-layer and double-layer recurrent neural networks is compared. The results show that the fixed-point number before the optimization is slightly worse than the floating-point number. After optimization, the double-layer RNN has better learning performance than the single-layer RNN under finite precision. Finally, we verify its effectiveness by numerical examples.

碩士
國立成功大學
電腦與通信工程研究所
107
With the increasing popularity of mobile devices and the effectiveness of deep learning-based algorithms, people try to put deep learning models on mobile devices. However, it is limited by the complexity of computational and software overhead. We propose an efficient framework for inference to fit resource-limited devices with about 1000 times smaller than Tensorflow in code size, and a layer-wised quantization scheme that allows inference computed by fixed-point arithmetic. The fixed-point quantization scheme is more efficient than floating point arithmetic with power consumption reduced to 8% left in cost grained evaluation and reduce model size to 40%~25% left, and keep TOP5 accuracy loss under 1% in Alexnet on ImageNet.

The intermittent nature of renewable energy, variations in energy demand, and fluctuations in oil and gas prices have all contributed to variable demand for power generation from coal-burning power plants. The varying demand leads to load-follow and on/off operations referred to as cycling. Cycling causes transients of properties such as pressure and temperature within various components of the steam generation system. The transients can cause increased damage because of fatigue and creep-fatigue interactions shortening the life of components. The data-driven model based on artificial neural networks (ANN) is developed for the first time to estimate properties of the steam generator components during cycling operations of a power plant. This approach utilizes data from the Coal Creek Station power plant located in North Dakota, USA collected over 10 years with a 1-hour resolution. Cycling characteristics of the plant are identified using a time-series of gross power. The ANN model estimates the component properties, for a given gross power profile and initial conditions, as they vary during cycling operations. As a representative example, the ANN estimates are presented for the superheater outlet pressure, reheater inlet temperature, and flue gas temperature at the air heater inlet. The changes in these variables as a function of the gross power over the time duration are compared with measurements to assess the predictive capability of the model. Mean square errors of 4.49E-04 for superheater outlet pressure, 1.62E-03 for reheater inlet temperature, and 4.14E-04 for flue gas temperature at the air heater inlet were observed.