Abstract: To solve the problem of effective recognition of atmospheric low-frequency acoustic signal in the low-frequency acoustic monitoring of the total nuclear test, a method of using convolution neural network is proposed. It converts low-frequency acoustic signal into images, then puts images into convolution neural network. The method is compared with SVM method based on artificial design features. The experimental results show that, when the training data set is not large, the convolution neural network with improved learning process can mine the potential features of signals, it has the same recognition ability as SVM.