Loading...
Publication
Sintonia automática de controladores PID: uma abordagem neuro - genética
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 7.13 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Os controladores PID são extensivamente utilizados em aplicações industriais. A
popularidade deste tipo de controladores advém da sua simplicidade (apenas 3 termos para
sintonizar) e robusto desempenho. Dado que os processos a controlar poderão variar ao
longo do tempo, levando à necessidade de uma nova sintonia do controlador, métodos de
sintonia automática de controladores PID revelam-se de grande interesse tanto prático como
teórico.
Na presente dissertação propõe-se uma técnica de sintonia automática do controlador PID,
com base em metodologias de soft-computing.
A sintonia de controladores pode ser vista como um problema de optimização simultânea
de vários critérios assim, o sistema proposto incorpora algoritmos genéticos para esse
propósito. Destinando-se o método desenvolvido a aplicações em tempo real, houve a
necessidade de se proceder à modelação dos critérios de controlo. Este procedimento tem
vantagem tanto a nível de consumo de tempo, como na prevenção de situações de
instabilidade sempre possíveis durante um processo de optimização, e não sujeita o
processo a perturbações necessárias ao processo de optimização; estes aspectos são de
fundamental importância nas aplicações em tempo real.
O problema de modelação de critérios de controlo bem como de parâmetros optimizados do
controlador foi tratado utilizando-se redes neuronais. Foram testadas as capacidades de
modelação e de adaptação de vários tipos de redes neuronais, nomeadamente perceptrão
multi-camada, redes com funções de base radial e B-Splines.
Os modelos neuronais a que se chegaram, juntamente com os algoritmos genéticos, formam
alguns dos blocos constituintes do modelo de auto-sintonia neuro-genética que se propõe.
Foi construído um simulador que implementa a presente arquitectura e valida a
metodologia proposta.
PID controllers are extensively used in industrial applications. Its popularity comes from the fact that they are very simple -just three terms to tune- as well as from its robust performance. As plants can be time-varying, therefore precluding the need for retuning, methods for PID auto-tuning exhibit great theoretical and practical interest. In this thesis, an automatic tuning technique for PID controllers based on soít-computing methodologies is purposed. The tuning of controllers can be interpreted as a simultaneous optimization problem of several criteria. For this purpose, the proposed system incorporates genetic algorithms. As this method aimed for real time applications, models are used to approximate the control criteria. This has the following advantages: first, there are savings in execution time; secondly, unstabilility situations, which can always occur during an optimization process, are avoided; finally the optimization can be performed avoiding the need of applying perturbations to the actual process, which is always undesirable. Neural networks are used to model the proposed tuning criteria, as well as to model the PID controller optimized parameters. The modeling and adaptation capacities of several neural networks, namely multilayer perceptron, radial basis function networks and B-Spline networks were tested and compared. The chosen neural models and genetic algorithms constitute some of the building blocks of the proposed neuro-genetic auto-tuning method. A Simulator implemented this architecture, in order to validate the proposed methodology.
PID controllers are extensively used in industrial applications. Its popularity comes from the fact that they are very simple -just three terms to tune- as well as from its robust performance. As plants can be time-varying, therefore precluding the need for retuning, methods for PID auto-tuning exhibit great theoretical and practical interest. In this thesis, an automatic tuning technique for PID controllers based on soít-computing methodologies is purposed. The tuning of controllers can be interpreted as a simultaneous optimization problem of several criteria. For this purpose, the proposed system incorporates genetic algorithms. As this method aimed for real time applications, models are used to approximate the control criteria. This has the following advantages: first, there are savings in execution time; secondly, unstabilility situations, which can always occur during an optimization process, are avoided; finally the optimization can be performed avoiding the need of applying perturbations to the actual process, which is always undesirable. Neural networks are used to model the proposed tuning criteria, as well as to model the PID controller optimized parameters. The modeling and adaptation capacities of several neural networks, namely multilayer perceptron, radial basis function networks and B-Spline networks were tested and compared. The chosen neural models and genetic algorithms constitute some of the building blocks of the proposed neuro-genetic auto-tuning method. A Simulator implemented this architecture, in order to validate the proposed methodology.
Description
Tese de dout. em Electrónica e Computação, Faculdade de Ciências e Tecnologia, Univ. do Algarve, 2004
Keywords
Sistemas de controlo Controlador PID