The invention provides a single-discrimination-based extraction method for features of an intelligent olfaction spectrum used for characterization of variation of honey. The method is characterized in that pattern recognition is carried out on every feature point one by one and difference between discrimination accuracies obtained when every feature point is used as a single feature is compared. The method embeds pattern recognition into feature selection and can realize prediction of a sample at each signal location through combined usage of a discrimination method. The method is dependent on a selected pattern recognition method, and selection results accordingly changes with the changes of the discrimination method. Difference between accuracies of different sensors in single vector selection is small, difference between different signal locations in a same sensor at different detection times is great, and the variation tendency of Bayes discrimination accuracy of different sensors at points of time is consistent with the variation tendency of variance ratios; that is to say, signal discrimination accuracy is high at an initial stage of detection, which mainly happens at initial 30 s of the detection time of each sensor, detection effects at a later stage are poor. If signal locations with discrimination accuracy greater than 60% are selected as feature points, there are 598 feature points altogether; and prediction accuracies of honey of rape flowers, basswood honey and acacia honey are 20/23, 13/17 and 34/39, respectively.