Improving specific interval accuracy in multivariate calibration using a net analyte signal-based sample selection method

The normal specification for multivariate calibration is the root mean square error (RMSE), which is computed from the error of all the samples in one set. As a result, condensed samples will inherently have less error than sparse samples. However, this phenomenon is undesirable in monitoring processes where marginal samples should be measured more accurately. Improving the accuracy of a calibration model over a particular interval would have a practical impact. By selecting uniformly distributed samples and including all the focused interval samples to decrease the cluster effect and include more matrix information, the accuracy of the target interval can be improved over that of an unselected calibration set. The selection method is based on a net analyte signal norm value computation and selection. Simulated spectral data and real sample sets are used to test the capability of the presented sample selection method. The experimental results show the method can improve interval accuracy for minor analyte and get almost equal interval accuracy for major analyte.