Method for detecting information in signals: application to two-dimensional time domain NMR data†

Abstract

Time domain (TD) NMR is used in industry for quality control. Like near-infrared (NIR) spectrometry, it has many advantages over wet chemistry including speed, ease of use and versatility. Unlike NIR, TD-NMR can generate a wide range of responses depending on the particular pulse sequences used. The resulting relaxation curves may vary as a function of the physico-chemical properties or even the biological and geographical origin of the product. The curves are usually decomposed into sums of exponentials and the relaxation parameters are then used in regression models to predict water content, iodine number, etc. The diversity of possible signals is both an advantage and disadvantage for TD-NMR as it broadens the range of potential applications of the technique but also complicates the development and optimisation of new analytical procedures. It is shown that univariate statistical techniques, such as analysis of variance or chi-squared, may be used to determine whether a signal contains any information relevant to a particular application. These techniques are applied to 2D TD-NMR signals acquired for a series of traditional and ‘light’ spreads. Once it has been demonstrated that the signals contain relevant information, partial least-squares (PLS) regression is applied directly to the signals to create a predictive model. The Durbin–Watson function is shown to be a means characterising the signal-to-noise ratio of the vectors calculated by PLS to select the components to be used in PLS regression.