An automatic modulation recognition methodology based on multi-scale features cross fusion is proposed to address the problem of limited recognition accuracy caused by insufficient feature extraction and utilization in existing deep learning-based modulation recognition methods. The automatic modulation recognition methodology proposed in this paper is based on deep learning model. Firstly, the IQ modulation signal is split into I signal and Q signal to form two inputs for the model. Subsequently, a multi-scale features extraction module is designed to effectively suppress signal noise, and on this basis, multi-scale features of the signal were extracted for modulation recognition. Then, a feature fusion module based on cross attention was designed to deeply fuse the multi-scale features extracted from the I signal and Q signal using the cross attention mechanism. Finally, the modulation recognition results were output. The experiments were conducted based on the benchmark dataset RadioML2016.10A and RadioML2016.10B in the field of modulation recognition. Compared to existing mainstream deep learning-based recognition models, the average recognition accuracy of the proposed methodology in the full signal-to-noise ratio (SNR) range is 57.17% and 62.18% respectively on RadioML2016.10A and RadioML2016.10B dataset, representing an improvement of more than 4.92% and 4.16%. The performance under low SNR conditions is also better than existing mainstream recognition models, and the recognition accuracy under high SNR conditions has been improved by 2.74% and 1.43% respectively. The experimental results show that the proposed methodology effectively improves the accuracy of signal automatic modulation recognition and overcomes the shortcomings of existing methods, such as sensitivity to noise and insufficient feature utilization.