DL Massart | PubMed Health

PubMed2000

Supersaturated designs for robustness testing

Supersaturated designs are factorial designs in which the number of factors examined exceeds the number of experiments performed. They do not allow an estimation of individual effects, since even the main effects are confounded. However, the total variance of a response estimated from such a design can in principle be used as a measure for the robustness of the method. A number of case studies were examined to determine whether the variance estimated from a supersaturated design is similar to the one from a Plackett-Burman design. This was found to be the case, which means that the estimated variance describes well the variation in the response caused by the variation in the factors.

Journal Article↗

PubMed2000

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 method (neural networks) has been performed in order to implement a calibration model on an industrial process. The models were designed to predict the water content in a reactor during a distillation process, using in-line measurements from a near-infrared analyzer. Curved effects due to changes in temperature and variations between the different batches make the problem particularly challenging. The influence of spectral range selection and data preprocessing has been studied. With each calibration method, specific procedures have been applied to promote model robustness. In particular, the use of a monitoring set with neural networks does not always prevent overfitting. Therefore, we developed a model selection criterion based on the determination of the median of monitoring error over replicate trials. The back-propagation neural network models selected were found to outperform the other methods on independent test data.

Journal Article↗