As seismic signals and artificial blasting signals exhibit high similarity in time–frequency domain features, resulting in insufficient recognition accuracy, we propose a self-organizing map (SOM) neural network classification model based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) multiscale distribution entropy (MDE) feature extraction and Ant Lion Optimization (ALO) algorithm improvement. The multiscale decomposition of the original seismic and blasting signals was carried out using CEEMDAN, and the distribution entropy values of the obtained multiple intrinsic mode functions were calculated to construct multidimensional feature inputs containing complexity information in the time–frequency domain. The ALO algorithm optimized the key parameters of the SOM neural network (competing layer dimensions and number of training iterations), with the root mean squared error serving as the fitness function. The optimal solution obtained by ALO optimization replaced the hyperparameter values in the original model, and multiple prediction rounds were performed on the seismic data test set to address unstable classification performance caused by random initialization in the traditional SOM network. The results revealed that the recognition performance of the CEEMDAN–MDE combined with the ALO–SOM model was significantly improved compared with machine learning models, such as linear discriminant analysis (LDA), decision tree, support vector machine, probabilistic LDA, and AdaBoost. Its recognition accuracy, recall, and F1-score were 99.3373%, 99.1479%, and 99.4557%, respectively, suggesting that this method can serve as a reliable approach for accurately differentiating between natural earthquakes and artificial blasting events, with important application value for seismic monitoring and blasting event exclusion.