Selection of variables for interpreting multivariate gas sensor data Original Research Article

In this work, methods to select relevant variables from a large set of available gas sensor parameters were examined. Two data sets, containing a large number of variables, were studied. The objective was to find the best descriptors, which could predict interesting properties of the measurements. Using a forward selection procedure, applying the root mean square error from a multilinear regression model as the selection criterion, it was possible to get good prediction accuracy from a back-propagation neural network (ANN). This procedure was fast, compared to the usual trial and error variable selection, objective, and can be made fully automated. In addition, the use of principal component analysis (PCA) and partial least squares (PLS) score vectors used as descriptors were examined. The ANN models constructed with either the PCA or the PLS score vectors as input gave for the first, rather smooth, data set, errors of the same size or smaller than for the forward selected parameters. For the second data set, containing many noisy variables, the forward selected parameters outperformed the other two data reduction methods.