The traditional Feature-Based (FB) signal modulation recognition algorithms suffer from the problems of low recognition accuracy,difficult feature extraction and poor generalization performance.To solve the problems,a signal Automatic Modulation Recognition (AMR) method based on MSPP-CNN is proposed,which combines Convolutional Neural Network (CNN) with Multi-Scale Pyramid Pooling (MSPP).In the proposed method,multi-scale pyramid pooling is used to improve the nonlinear feature extraction ability of the model for different modulation signals,so that the model has better feature representation and generalization performance.In the construction process of the CNN model,different convolution,pooling and activation methods are used to optimize and verify the model,so as to ensure the rationality of model structure and parameters.The experimental results show that the recognition accuracy of the proposed method under signal-to-noise ratio of -18 dB,0 dB and 18 dB reaches 56%,62.98% and 92.04% respectively.In addition,a comparison with other traditional feature extraction algorithms and CNN algorithm is conducted,which verifies the effectiveness and high recognition accuracy of MSPP-CNN.