Abstract: ‍ ‍Modulation recognition technology is a crucial component of communication signal reconnaissance， serving as an essential prerequisite for classifying and identifying unknown communication signals and extracting information. Existing deep learning-based modulation recognition methods have poor feature extraction capabilities under low signal-to-noise ratio （SNR） conditions. To address this issue， this study proposes a signal modulation recognition algorithm based on multi-scale feature fusion extraction. The algorithm uses a multi-scale convolution module composed of multiple convolution kernels of different sizes to extract multi-scale features of the signal and fuses these features through convolutional layers to extract key features of the signal’s modulation information. Then， a global feature extraction module composed of a multi-head attention mechanism is used to extract global features of the signal， and modulation recognition is achieved through average pooling layers and fully connected layers. To optimize the network parameters and computational complexity， this study proposes replacing the standard convolution with group convolution to simplify the model. Results from experiments on the RadioML2016.10a dataset show that the proposed method can accurately identify various modulation types， with recognition accuracy exceeding 95% for most modulation types under high SNR conditions. Compared to existing neural network recognition methods， the proposed method improves the recognition rates by 1.47% to 7.26%. Under lower SNR conditions （-6 to 0 dB）， it achieves an improvement of 4.73% to 9.09%， demonstrating better noise resistance. Additionally， using group convolution instead of standard convolution reduces the network parameters and computational load by 38.9% and 54.9%， respectively， with minimal performance difference. An ablation study was designed to verify the performance improvement of each module in the proposed algorithm. Experimental results validate the effectiveness of the proposed algorithm in terms of recognition accuracy and noise resistance.