Identificação de sistemas "on-line", otimização e controle avançado com o filtro de Kalman estendido / On line system identification, advanced control and optimization with the (Extended) Kalman filter
AUTOR(ES)
Ramon Scheffer
DATA DE PUBLICAÇÃO
2006
RESUMO
In the continuing competition between it will be more and more necessary to optimize current chemical processes in real time. To be able to optimize a plant in real time, there have to be various aspects to be fulfilled, such as measurement, reliability of the measurement and prediction of the process behaviour. In this work some of the aspects of such an advanced control are studied and are measurement monitoring, on-line non-linear system identification (recurrent neural networks) and constrained non-linear optimisation. It is wanted that this system can work under measurement noise, unmeasured disturbance and process changes such as a catalyst deactivation. All these tools were developed in the FORTRAN programming language and are available at the laboratory LOPCA/UNICAMP. Validated models were used to simulate the processes, but in some cases real industrial and pilot-plant data were used to study the algorithms developed. The fractional Gaussian noise (fGn) and fractional Brownian motion (fBm) were thought to be models suitable as measurement predictors, and applied to pilot plant data of an airlift reactor, whose pressure signal presents a complex non-white behaviour. It was shown that the fGn does describe part of the measured signals and is able to do some prediction of the time series, but the other part could be explained well by a (4,3) Auto-Regressive and Moving Average (ARMA) model. It was noted that the fGn and fBm lack parameters to be adjusted and cannot be used for processes having a sinus type of autocorrelation function (ACF). Therefore an extension of ARMA models known as the fractional ARMA (FARMA) models can be used as a measurement monitoring tool, allowing the possibility to develop a general diagnostic tool. It is shown a various cases (from theoretical to practical industrial data) that the MEKA Kalman filter algorithm is a quite fast training algorithm for recurrent neural network training, but especially results in better generalisation properties of the neural network trained than the other sequential training algorithms (standard backpropagation (with momentum)). It was shown that the Kalman filter can be successfully used in unconstrained and constrained optimisation. The unconstrained optimisation of the Rosenbrock function demonstrates that a very fast optimisation can be obtained by manipulating the process noise covariance matrix. The applicability to constrained optimisation was shown in a large scope of different test problems and one real industrial problem
ASSUNTO(S)
kalman advanced control filtragen de redes neurais (computação) controle em tempo real non-linear constrained optimization otimização matematica neural network controle de processos quimicos kalman filter
ACESSO AO ARTIGO
http://libdigi.unicamp.br/document/?code=vtls000434555Documentos Relacionados
- An Extended Kalman Filter-Based Technique for On-Line Identification of Unmanned Aerial System Parameters
- Extended Kalman filter applied to electrical impedance tomography.
- ESTIMAÇÃO DE VELOCIDADE DO MOTOR COM CONTROLE VETORIAL SEM SENSOR, UTILIZANDO FILTRO ESTENDIDO DE KALMAN COM ESTIMAÇÃO DA COVARIÂNCIA DOS RUÍDOS
- Integração da otimização em tempo real com controle preditivo.
- A HYBRID APPROACH FOR SIMULTANEOUS LOCALIZATION AND MAPPING WITH SONAR BASED ROBOTS AND EXTENDED KALMAN FILTER