Abstract:

To solve the problem of low classification accuracy of electronic nose ( E-nose) caused by sensor drift in various applications，a sample distribution weighting cross domain extreme learning machine model was proposed． Considering the different contributions of a single sample to the global distribution discrepancy measure，it uses the maximum mean discrepancy based on the sample distribution weighting as a measure of the sample distribution discrepancy between domains． The data of source domain and target domain were projected onto a high-dimensional extreme learning machine feature space． Then a suitable projection direction was found，by which the data was projected onto a common subspace． Thus the source domain and target domain data in the subspace had a similar distribution． Matlab was used to simulate this algorithm，and the effects of different number of hidden layer nodes on the recognition rate of the algorithm were compared so as to verify the feasibility of the algorithm． The results show that the model proposed in this paper can significantly reduce the distribution discrepancies between the two domains，and can meet the distribution requirements of traditional classification algorithms for training and test data，thus to improve the classification accuracy of E-nose．

Key words: extreme learning machine, subspace learning, sample distribution weighting, drift compensation, electronic nose