Weighted partial least squares method to improve calibration precision for spectroscopic noise-limited data
Multivariate calibration methods have been applied extensively to the quantitative analysis of Fourier transform infrared (FT-IR) spectral data. Partial least squares (PLS) methods have become the most widely used multivariate method for quantitative spectroscopic analyses. Most often these methods are limited by model error or the accuracy or precision of the reference methods. However, in some cases, the precision of the quantitative analysis is limited by the noise in the spectroscopic signal. In these situations, the precision of the PLS calibrations and predictions can be improved by the incorporation of weighting in the PLS algorithm. If the spectral noise of the system is known (e.g., in the case of detector-noise-limited cases), then appropriate weighting can be incorporated into the multivariate spectral calibrations and predictions. A weighted PLS (WPLS) algorithm was developed to improve the precision of the analysis in the case of spectral-noise-limited data. This new PLS algorithm was then tested with real and simulated data, and the results compared with the unweighted PLS algorithm. Using near-infrared (NIR) calibration precision when the WPLS algorithm was applied. The best WPLS method improved prediction precision for the analysis of one of the minor components by a factor of nearly 9 relative to the unweighted PLS algorithm.
- Research Organization:
- Sandia National Labs., Albuquerque, NM (United States)
- Sponsoring Organization:
- USDOE, Washington, DC (United States)
- DOE Contract Number:
- AC04-94AL85000
- OSTI ID:
- 531089
- Report Number(s):
- SAND--97-2226C; CONF-970812--1; ON: DE97009353
- Country of Publication:
- United States
- Language:
- English
Similar Records
Classical least squares multivariate spectral analysis
Comparisons of prediction abilities of augmented classical least squares and partial least squares with realistic simulated data : effects of uncorrelated and correlated errors with nonlinearities.