In this paper, we propose a new technique for feature extraction/selection based on the projection of sensor features in class space while taking into account the sensor variance. The proposed technique is inspired by the organization of the early stages in the biological olfactory system. The algorithm proves to be highly suitable for high-dimensional feature vectors. The performance shows robustness with problems where only a small number of samples are available as a training dataset. We demonstrate the method on experimental data from two metal oxide sensors driven by a sinusoidal temperature profile.