A CALIBRATION METHOD FREE OF OPTIMISM FACTOR NUMBER SELECTION FOR AUTOMATED MULTIVARIATE-ANALYSIS - EXPERIMENTAL AND THEORETICAL-STUDY

Several analytical applications of multivariate calibration methods re quire human decisions, the most difficult being the number of factors involved, Thus, eliminating the optimum factor number may contribute t o the improvement of automatic calibration processes. We propose a fac tor analysis method that does not need the factor number, It is partic ularly suitable for indited calibration of a system under indirect obs ervation. The algorithm is based on composing a subspace excluding the contribution from the component of interest and calculating its net a nalyte signal through an orthogonal projection to an orthogonal space, This method is applicable as long as the spectral vector dimension (i .e., the number of data points) is larger than the calibration set siz e, This condition readily satisfied in spectroscopic analysis, The rel evant effects, including the effect of the spectral vector dimension a nd of the calibration set size upon prediction errors, have been inves tigated using extensive computer simulation, The algorithm has been ex emplified by a successful application to the predictions of ethanol co ncentration and of octane number of gasoline samples using near-IR spe ctra. In this example of an indirect calibration, the proposed method, which requires no information on optimum factor number, is of particu lar importance, In most cases, the results obtained by this method are similar to those of the traditional PCR; however, this method does no t fail when the optimal model cannot be correctly determined by automa tic procedures.