A Blind Source Separation(BSS) algorithm based on Empirical Mode Decomposition-Non- Linear Principal Component Analysis(EMD-NLPCA) is proposed after studying the BSS algorithm for underdetermined non-linear mixed signals. Firstly, EMD is applied to the observed signal, then high- order statistics are introduced after reconstructing the signal. The principal component analysis is carried out to complete the signal separation. This algorithm can not only deal with the undetermined environment but also solve the problem of non-linear mixing. In the simulation, the results of the algorithm are compared with those of the sparse component analysis, which proves that the proposed algorithm is correct and more universal than the sparse component analysis. Finally, the algorithm is applied to the separation of driving audio signals of unmanned aerial vehicle engines, and it works well.