Calibration model design based on weighted nearest correlation spectral clustering

Calibration models have been widely used for estimating product quality or other key variables with near-infrared spectroscopy (NIRS), and it is important to select appropriate input variables (wavelengths) for building a highly accurate calibration model. A novel input variable selection method based on nearest correlation spectral clustering (NCSC), which is a correlation-based clustering method, was proposed, and it is referred to as NCSC-based variable selection (NCSC-VS). In NCSC-VS, some variable groups are clustered by NCSC, and a few variable groups are selected by their contribution to estimates. Although variable selection performance of NCSC-VS depends on variable group clustering by NCSC, its clustering results easily fluctuate according to measurement noise. The present work proposes an improved version of NCSC that can cope with measurement noise by introducing a weighting function into affinity matrix construction. In addition, the proposed clustering method, referred to as weighted NCSC (WNCSC), is applied to variable selection in calibration model design. WNCSC-VS can achieve a higher estimation performance than NCSC-VS. The usefulness of the proposed WNCSC-VS is demonstrated through an application to calibration model design for a pharmaceutical process.