| 초록 |
Reliable measurement of product quality is one of the most important issues in distillationcolumns. So far, however, most on-line composition analyzers such as gas chromatographyhave many problems : low reliability, high maintenance cost, measurement delay. Although theuse of single temperature measurement has been widely applied to most industries as analternatives, it still has several critical problems especially in high purity columns andmulticomponent columns. For this reason, many researchers (Joseph and Brosilow, 1978;Mejdell and Skogestad, 1991; Kresta et al., 1994) have investigated the estimator based onmultiple measurements in order to overcome these drawbacks. It has been known that theestimator using the multivariate analysis methods such as Principal Component Regression(PCR) and Partial Least Square (PLS) give good predictive power and robustness to noiseand sensor failures in the somewhat small operation range. However, since the distillationcolumns are inherently nonlinear processes (see Fig. 1), the estimation performance of thelinear estimator can be significantly degraded depending on the situation. In this paper we compare the two representative approaches to infer the productcompositions for distillation columns : PLS approach and Artificial Neural Network (ANN).The main objective of this paper is to analyze the pros and cons of the two approaches inorder to provide useful guidelines for selecting the estimator type. We also proposed andevaluated a hybrid type estimator in which PLS and ANN are combined. Simulation studywas performed for a high purity column and several multicomponent columns. In the binary column case, both the PLS estimator and the ANN estimator are relativelymore sensitive to noise compared with the BTX case. Also, the performance of the linearestimator is almost same as that of the nonlinear estimator. It seems that the ANN does notcapture the nonlinearity of the system satisfactorily while sacrificing the robustness. In spiteof the noise, both of the estimators are robust in the case of the BTX column.
|
