MULTIVARIATE REGRESSION APPLIED TO TIME-DOMAIN NUCLEAR-MAGNETIC-RESONANCE SIGNALS - DETERMINATION OF MOISTURE IN MEAT-PRODUCTS

This study demonstrates the feasibility of applying multivariate regre ssion techniques directly to Time Domain-Nuclear Magnetic Resonance (T D-NMR) signals to obtain predictive models for the quantification of s ample constituents. Applications of such multivariate methods are rare in TD-NMR where simple univariate regression models based on calculat ed relaxation parameters (T-1, T-2, initial magnetisations) are usuall y used. The method was applied to the determination of moisture in a m eat product, i.e. frankfurter sausages, dried to a water content range of between 10% and 61% wet basis. The NMR signals are the relaxation curves obtained using Free Induction Decay, Carr-Purcell-Meiboom-Gill and Inversion Recovery sequences. The best model was obtained by Parti al Least Squares regression on the three combined signals, using two F actors. This model was tested vith a second set of samples, anti by a Jack-knifing test to estimate the robustness of its regression coeffic ients. These coefficients glare information on the regions of the NMR signals which contain information. This multivariate signal regression method gave a Standard Error of Prediction of 3.163 and an R-2 of 0.9 86. Compared with the regression results obtained using the calculated NMR parameters, it was shown to be less subject to error, faster and easier.