A new scheme for training Rectified Linear Unit (ReLU) based feedforward neural networks is examined in this thesis. The project starts with the row-by-row updating strategy designed for Single-hidden Layer Feedforward neural Networks (SLFNs). This strategy exploits the properties held by ReLUs and optimizes each row in the input weight matrix individually, under the common optimization scheme. Then the Direct Updating Strategy (DUS), which has two different versions: Vector-Based Method (VBM) and Matrix-Based Method (MBM), is proposed to optimize the input weight matrix as a whole. Finally DUS is extended to Multi-hidden Layer Feedforward neural Networks (MLFNs). Since the extension, for general ReLU-based MLFNs, faces an initialization dilemma, a special structure MLFN is presented. Verification experiments are conducted on six benchmark multi-class classification datasets. The results confirm that MBM algorithm for SLFNs improves the performance of neural networks, compared to its competitor, regularized extreme learning machine. For most datasets involved, MLFNs with the proposed special structure perform better when adding extra hidden layers. / Ett nytt schema för träning av rektifierad linjär enhet (ReLU)-baserade och framkopplade neurala nätverk undersöks i denna avhandling. Projektet börjar med en rad-för-rad-uppdateringsstrategi designad för framkopplade neurala nätverk med ett dolt lager (SLFNs). Denna strategi utnyttjar egenskaper i ReLUs och optimerar varje rad i inmatningsviktmatrisen individuellt, enligt en gemensam optimeringsmetod. Därefter föreslås den direkta uppdateringsstrategin (DUS), som har två olika versioner: vektorbaserad metod (VBM) respektive matrisbaserad metod (MBM), för att optimera ingångsviktmatrisen som helhet. Slutli- gen utvidgas DUS till framkopplade neurala nätverk med flera lager (MLFN). Eftersom utvidgningen för generella ReLU-baserade MLFN står inför ett initieringsdilemma presenteras därför en MLFN med en speciell struktur. Verifieringsexperiment utförs på sex datamängder för klassificering av flera klasser. Resultaten bekräftar att MBM-algoritmen för SLFN förbättrar prestanda hos neurala nätverk, jämfört med konkurrenten, den regulariserade extrema inlärningsmaskinen. För de flesta använda dataset, fungerar MLFNs med den föreslagna speciella strukturen bättre när man lägger till extra dolda lager.

The generalization ability of artificial neural networks (ANN) is highly dependent on their architectures and can be critical to solving a given problem. The current best practice uses fixed architectures determined via a trial-and-error approach. This process can be both computationally and temporally cumbersome and does not guarantee that an optimal topology will even be found. Replacing the user’s role in designing topologies with methods that enable a system to manage its own growth can endow systems with adaptable learning. Constructive algorithms offer the possibility of compact architectures as an alternative to the trial-and-error approach. This class of algorithms grows a network’s topology by incrementally adding units during learning to match task complexity. However, the decision of when to add new units in constructive algorithms heavily depends on user-defined a priori hyperparameters, which can be task-specific. Contrary to having a user fine-tune hyperparameters that govern growth, the intrinsic population dynamics of an ANN could be used to self-govern the growing process. Theoretically, an ANN or each layer comprising the network can be viewed as a set of populations. From this perspective, a hidden layer can be considered the environment in which hidden units exist. In this work, we propose a novel, more self-governed growing algorithm inspired by population dynamics for determining near-optimal topologies of feedforward ANNs. This allows the inclusion of a carrying capacity, the maximum population of hidden units that can be sustained in a hidden layer. Including this constraint in combination with population dynamics provides a built-in mechanism for a dynamic growth rate. The proposed approach is used in parallel with direct performance feedback from the network to modulate the growth rate of the hidden layer, allowing the network to converge to smaller topologies based on the task's demands. More self-governed approaches reduce the number of finely-tuned hyperparameters required to decide when to grow and put more control of the network’s structure and representational capacities in the algorithms themselves, facilitating the emergence of inherent intelligent behaviour. Chapter one introduces a dynamic, more self-governed growing algorithm inspired by population dynamics. Results show that compared to using fixed rules for determining hidden layer sizes; dynamic growth leads to smaller topologies than predicted while still being capable of solving the task. In chapter two, we investigate the algorithm's inherent properties to validate the more self-governed aspect. The results depict that the model’s hyperparameters require less fine-tuning by the user and adhere more toward self-governance. Finally, in chapter three, we investigate the effects of growing hidden layers individually in a sequential fashion or simultaneously in a parallel fashion multilayer context. A modified version of the growing algorithm capable of growing parallel is proposed. Growing hidden layers in parallel resulted in comparable or higher performances than sequential approaches. The growing algorithm presented here offers more self-governed growth, which provides an effective general solution automatically tailored to the task.

No description available.

Active structural acoustic control (ASAC) in conjunction with the adaptive feedforward control has been proved to be an efficient practical approach to reduce structure-borne sound. ASAC works on the principles of reducing the vibration amplitude of the structure (modal reduction), as well as changing the vibration distributions of the structure so that the vibration distributions of each structural modes destructively interfere with one another in their associated radiating acoustic field (modal restructuring). Based on these observations, two different but related design strategies, namely the non-volumetric design and the minimum supersonic wavenumber design, were developed for designing efficient ASAC system. The eigenassignment method for feedforward control system serves as the fundamental design tool for both formulations. In this study, the dynamic characteristics of a multiple-input, multiple-output (MIMO) feedforward controlled system was investigated both analytically and experimentally on a simply supported plate under harmonic excitation. It was demonstrated that, when the control system has equal number of control inputs and error sensor outputs, the feedforward controller can effectively modify the system dynamics (i.e., resonance frequencies and mode shapes). This provides the theoretical basis for the eigenassignment method. For the non-volumetric design, the single-input, single-output (SISO) eigenassignment technique is used to modify the eigenproperties of a planar structure using structure actuators and sensors so that all the controlled modes are non-volumetric (inefficient sound radiators at low frequencies, i.e., k_0a << 1), leading large global sound attenuation in the far field. The effectiveness of this formulation was demonstrated through numerical simulations for the control of radiation from simply supported and clamped-free beams. The experimental validation of the non-volumetric design was also carried out on a simply supported beam using PZT actuators and shaped PVDF film as error sensor. The filtered-x LMS algorithm was used in the experiment. Excellent global sound attenuation was achieved in the low frequencies. The minimum supersonic wavenumber design stems from the fact that only supersonic wavenumber components of the structural velocity spectra radiate to the far field. A SISO eigenassignment technique is used to modify the eigenproperties of a planar structure so that the eigenfunctions of the controlled system have minimum supersonic wavenumber in the frequency range of study. The sound pressure or sound power radiated by the structure is therefore reduced. The design was demonstrated on a simply supported beam to minimize the supersonic wavenumber components contributed by the odd-order modes only. Significant global sound attenuation was achieved in the frequency range of study. The main advantage of the proposed design methods is that they do not depend on the characteristics of the external disturbance, such as the form, location and frequency contents. Also, the error sensor and control input are optimized simultaneously, resulting in better acoustic control performance. The practical implementations of the proposed designs require accurate system modeling, this is the major limitation of the proposed designs. / Ph. D.

The key conclusions of this research are: 1. The EFXLMS algorithm demonstrated superior performance than the FXLMS algorithm during fast adaptive processes, in particular for non-stationary inputs. 2. Good attenuation of the peak. and root-mean-square (rms) values of the structural responses using the hybrid control system were observed for most of the real accelerograms. It was also observed that the hybrid control system always improved the performance of the passive control system. 3. Peak. structural responses during sensor failures were observed to increase, in particular for the failure of the first error sensor. However, the peak. responses during these failures never reached the peak. values obtained by the system with passive control. This demonstrated the ability of the control system to contain the structural responses in light of the error sensor failures. 4. Peak. structural responses during control actuator failures were also observed to increase, in particular for the failure of the fIrst control actuator ADVAI. However, the peak. responses during these failures again never reached the peak. values obtained by the system using passive control. This demonstrated the ability of the control system to contain the structural responses in light of the control actuator failures. 5. The use of the adaptive control showed greater attenuation of the structural responses than that observed using the non-adaptive system. This demonstrated that the adaptive control system corrected for uncertainties and errors in the modeling process. / Master of Science

Adaptive FeedforWard algorithms have been successfully applied in the active control of sound and vibrations~ Current approaches, such as the Filtered-X Least Mean Square algorithm, are limited in adaptability and applicability by the computationally intensive system identification process. The Time Average approach is a technique that promises to be a major advance in the active control of sound and vibration. This approach is based on statistical estimates of the error signals, and does not require having previous knowledge of the system. A new minimization algorithm has been developed, which enhances the performance of the Time Average controllers. A new technique to estimate the cost function is also presented. This technique is called the Real Time Cost Function Evaluation method. Based upon these new controller design concepts, single channel and multi-channel controllers were developed and experimentally tested. The experiments included control of a harmonically excited, simply supported beam, and the control of inlet tonal noise radiated from a JT1SD-l turbofan engine. / Master of Science

Orientadores: Ana Maria Frattini Fileti, Fernando Palu / Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Química / Made available in DSpace on 2018-08-05T22:21:40Z (GMT). No. of bitstreams: 1 Eyng_Eduardo_M.pdf: 616784 bytes, checksum: 5124ded82b67d3d54bfa60a9d0b1a28e (MD5) Previous issue date: 2006 / Resumo: O etanol perdido por evaporação durante o processo de produção por fermentação pode ser recuperado por uma coluna de absorção, a qual requer um sistema de controle robusto de acordo com sua complexidade. Sendo assim, no presente trabalho foi proposto e testado, o emprego de um controlador feedforward, baseado em um modelo inverso de redes neurais, para realizar a manipulação da vazão de solvente que é alimentado a esta coluna. O desempenho do controlador feedforward proposto foi superior ao proporcionado pelo Controle por Matriz Dinâmica (DMC) (Palú, 2001), para situações nas quais uma perturbação degrau foi inserida na composição do gás à entrada da coluna. Outro teste realizado foi o de implementar um distúrbio não previsto na etapa de treinamento da rede neural, e analisar o comportamento do sistema de controle RNA. Houve uma queda de eficiência, sendo gerado off set. Para solucionar este problema, agregou-se ao sistema de controle, um dispositivo de segurança. Também se procurou testar a tolerância do sistema de controle a distorções nos dados de entrada da rede. Neste sentido, acrescentou-se às variáveis de entrada do tipo concentração, um ruído. O desempenho do controlador RNA foi prejudicado, no entanto, a faixa de oscilação da variável controlada foi reduzida à aproximadamente 50% da faixa de ruído empregada. Aplicando-se um filtro aos dados distorcidos, esta redução foi ainda maior, cerca de 75% / Abstract: Some of ethanol lost by evaporation during its fermentation production process may be recovered using an absorption column, which requires a robust control system. In this work the use of a feedforward controller (NN controller), based on a neural network inverse model, was proposed and tested to manipulate the solvent flow, which was fed in the analyzed column. The feedforward controller performance was better than the Dynamic Matrix Control (Palú, 2001) one, when step disturbances were inserted on the inlet gas composition. Another test done was the insert of a not trained disturbance. The NN controller performance decrease, and a off set could be observed. To solve this problem, a security device was added to the control system. The control system tolerance to modified input neural network data also was tested. In this way, a noise was added to input data. The NN control performance decreased, but the amplitude of the control variable oscillation band was reduced to the half of the noise amplitude employed. When the modified data was filtered the NN control performance decrease was smaller, about 75% of the noise amplitude / Mestrado / Sistemas de Processos Quimicos e Informatica / Mestre em Engenharia Química

Le travail présenté dans cette thèse s'inscrit dans le cadre de l'étude de l'efficacité de la commande CRONE avec effet anticipatif robuste dans les problèmes de poursuite et de régulation pour les systèmes monovariables et multivariables. L'objectif de l'anticipation est de concevoir des algorithmes de commande capables de minimiser un critère quadratique basé sur l'erreur entre la sortie du système et le signal de référence avec modération du niveau du signal de commande. La solution proposée repose sur l'association d'une commande robuste de type feedback à une commande anticipative de type feedforward. L'action feedforward utilise un filtre anticipatif synthétisé dans le domaine fréquentiel en utilisant des critères d'optimalité et de contraintes de type $mathcal{H}_2$ et $mathcal{H}_infty$. La réduction du nombre de paramètres du filtre anticipatif en utilisant une période d'échantillonnage plus lente a été traitée. Des exemples académiques illustrent les développements théoriques. Une application pratique sur un système de régulation hydraulique a montré la pertinence de cette nouvelle approche. / The work presented in this thesis is part of the study of the effectiveness of CRONE control with anticipative effect in the problems of tracking and control for monovariable and multivariable systems. The anticipation objective is to design a control algorithm that minimizes a quadratic error between the reference and the output of the system and at the same time achieve a good level of the control signal. The proposed solution combines a robust feedback control with a feedforward control. The feedforward action uses optimality criteria and an anticipative filter which is designed in the frequency domain using a mix of $mathcal{H}_2$ et $mathcal{H}_infty$ constraints. The reduction of the number of parameters of the anticipative filter by using a slower sampling period for the anticipative filter is treated. Academic examples highlight the theoretical developments. A practical application on a hydraulic control system has shown the efficiency of the developed approach.

Resumo: Este trabalho apresenta o desenvolvimento de uma análise teórica do desempenho de um sistema de controle ativo utilizando mancais magnéticos como atuadores de não-contato para a redução de vibrações em rotores. São analisados três modelos de rotores, sendo que em um deles aplica-se apenas o controle feedback e os outros são suportados apenas por mancais magnéticos, os quais também são os atuadores do controlador. Assim, Luna arquitetura de controle tipo feedforward é empregada sobreposta ao sistema de controle feedback dos atuadores nestes dois modelos, sendo urna análise realizada em relação ao desempenho do sistema de controle quanto às diferentes geometrias de distribuição de massa acopladas ao eixo do rotor. O enfoque principal deste trabalho é voltado para a análise do desempenho do sistema de controle em função da posição e quantidade dos sensores de erro (onde se deseja minimizar as vibrações) em relação à posição dos atuadores e das forças de excitação. As excitações são do tipo síncronas e sub-síncronas que normalmente aparecem em rotores com elevadas velocidades de rotação, como as turbomáquinas. Também é realizada urna análise das forças de controle necessárias a serem aplicadas pelos atuadores para se obter urna redução dos níveis de vibração do rotor na posição dos sensores de erro do sistema feedforward. A análise é executada empregando modelos de rotores desenvolvidos pelo método da matriz de impedância. Esta pesquisa também apresenta Luna aplicação da técnica de controle Feedforward em acústica, que realiza a depuração da voz para comunicação em ambientes ruidosos. / Abstract: This research work brings a theoretical analysis of a control system performance that uses magnetic bearings as non-contact actuators to reduce rotor vibrations. It is analyzed three rotor models, in which one of them operates under the feedback control only. The other models are supported by magnetic bearings only, which also are the controller system actuators. Thus, a feedforward control scheme is applied over the feedback control inherent to the AMB control circuit. The analysis is carried out over these two last models regarding to the control performance for different geometry of mass distribution along the rotor. The focus of this work is to analyze the controller performance according to the sensor quantity and placement (where the vibrations are desired to be minimized) regarding to the actuator position and to the exciting forces. The subsynchronous and synchronous excitations are considered here since they frequently occur in high rotating speed rotors, as in the turbomachinery scenario. Also, the control force required by the actuators is monitored according to the sensors placement to reduce the local vibrations level and the analysis was carried out using the impedance matrix rotor modeling. Further, this work brings a modeling and an application of the feedforward active control scheme in the acoustics field used for voice extraction for communication in noisy environments. / Orientador: Luiz de Paula do Nascimento / Coorientador: Vicente Lopes Junior / Banca: Gilberto Pechoto de Melo / Banca: Kátia Luchese Cavalca Dedini / Mestre

Em um sistema came-seguidor a dinâmica do sistema seguidor faz com que o movimento final se desvie daquele especificado. Este efeito pode ser compensado considerando-se o modelo dinâmico inverso do sistema seguidor no projeto da lei de movimento do came. Considerando-se constante a velocidade do came, o sistema seguidor tem dinâmica linear. Entretanto, devido à razão de transmissão variável, e devido a outros efeitos não lineares, o sistema de acionamento como um todo é não linear, e procedimentos não lineares devem ser utilizados para se ajustar a lei de movimento do came. Uma análise teórica, suportada por simulações, mostra o potencial deste procedimento, ao menos no caso de uma dinâmica simples do sistema seguidor: uma considerável redução do erro de movimento, e uma boa robustez relativa a erros na freqüência de ressonância e razão de amortecimento estimadas. Experimentos com o acionamento por um servomotor sub-dimensionado, como é de se esperar, mostram resultados diferentes, devido à velocidade angular não constante. A flutuação observada na velocidade angular em torno do valor constante é então levada em consideração para o projeto da lei de movimento com nonlinear feedforward. / In a cam-follower system the dynamics of the follower-train mechanism cause an actual motion which deviates from the desired one. This effect can be compensated by taking into account the inverse dynamic model of the follower-train in the design of the cam motion law. By considering a constant cam velocity, the follower-train has linear dynamics. However, due to the varying transmission ratio, and due to other nonlinear effects, the whole drive train is a nonlinear system, and nonlinear procedures should be used to fit a motion law. A theoretical analysis with only the linear feed-forward compensation, supported by simulation results, has shown the potential of this approach, at least in the case of simple follower-train dynamics: a considerable reduction of the motion error, and a good robustness with respect to errors in the estimated resonance frequency and damping ratio of the follower-train. Experiments with a small driving servomotor, as expected, show differerit results, due to the non-constant angular velocity. The observed cam angular velocity ripple is then taken into account to design a complete nonlinear feedforward motion law.