In machine olfaction or electronic nose, sensor optimization is important to enhance pattern recognition efficiency and reduce redundant information. Highly correlated response of one sensor to two different odors implies less contribution of this sensor to the classification of these two odors. Variance difference is a significant index to measure the similarity of sensor responses. A sensor optimization method based on variance difference is proposed in this paper; both the average value of variance difference and cluster analysis of variance difference matrix were considered to identify several possible sensor subsets. Six Chinese herbal medicines and linear discrimination analysis (LDA) were applied to test the classification results in order to determine the best subset. LDA results indicated that the optimized sensor subset performed well in classification of the six Chinese medicines. The proposed sensor array optimization method could be applied to other kinds of odors classification as a novel method.