The microseismic monitoring signals which need to be determined in mines include those caused by both rock bursts and by blasting. The blasting signals must be separated from the microseismic signals in order to extract the information needed for the correct location of the source and for determining the blast mechanism. The use of a convolutional neural network (CNN) is a viable approach to extract these blast characteristic parameters automatically and to achieve the accuracy needed in the signal recognition. The Stockwell Transform (or S-Transform) has excellent two-dimensional time-frequency characteristics and thus to obtain the microseismic signal and blasting vibration signal separately, the microseismic signal has been converted in this work into a two-dimensional image format by use of the S-Transform, following which it is recognized by using the CNN. The sample data given in this paper are used for model training, where the training sample is an image containing three RGB color channels. The training time can be decreased by means of reducing the picture size and thus reducing the number of training steps used. The optimal combination of parameters can then be obtained after continuously updating the training parameters. When the image size is 180 × 140 pixels, it has been shown that the test accuracy can reach 96.15% and that it is feasible to classify separately the blasting signal and the microseismic signal based on using the S-Transform and the CNN model architecture, where the training parameters were designed by synthesizing LeNet-5 and AlexNet.