Abstract
The maintenance of multivariate models used with on-line near-infrared analyzers often consumes significant resources after installation. This paper presents a proposed calibration strategy based on a locally weighted regression (LWR) model with an extended library and its test results on two industrial data sets. The LWR method uses combined spectral and chemical space for distance, which was modified from Wang et al.'s proposed distance measurements, and linear weightings as well as principal component regression (PCR) analysis. The addition of new sample spectra to adapt process changes or to correct for prediction variations is also proposed. Although the maintenance of predictions within performance specifications was the original goal for the proposed strategy, the LWR approach with extended libraries also provides significant accuracy improvement. The root mean squared error of prediction (RMSEP) for the first concentration property in data set 1 from the LWR model on an extended library is 47% better than the one from a global PCR model. The RMSEP for the second concentration in set 1 is 14% better. On the other hand, global models based on the corresponding extended libraries, which were used by the LWR model, do not show any significant performance improvement. The improvement for the physical property of set 2 is 61%, which may be due to the more clustered sample distribution and the nonlinear relationship between the property and spectral response. Current implementations also show that the LWR model provides a convenient platform to include new samples that were collected after installation to adapt process or sample changes.
PDF Article
More Like This
Cited By
You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
Contact your librarian or system administrator
or
Login to access Optica Member Subscription