Development of a robust calibration model for nonlinear in-line process data

A comparative study involving a global linear method (partial least squares ), a local linear method (locally weighted regression), and a nonlinear met hod (neural networks) has been performed in order to implement a calibratio n model on an industrial process. The models were designed to predict the w ater content in a reactor during a distillation process, using in-line meas urements from a near-infrared analyzer. Curved effects due to changes in te mperature and variations between the different batches make the problem par ticularly challenging, The influence of spectral range selection and data p reprocessing has been studied. With each calibration method, specific proce dures have been applied to promote model robustness. In particular, the use of a monitoring set with neural networks does not always prevent overfitti ng. Therefore, we developed a model selection criterion based on the determ ination of the median of monitoring error over replicate trials. The back-p ropagation neural network models selected were found to outperform the othe r methods on independent test data.