Signal Modulation Recognition Algorithm Using Deep Learning in Non-cooperative Communication

This paper proposes a signal modulation recognition algorithm based on deep learning that integrates multi-scale features and hybrid attention mechanism. The algorithm uses multi-scale convolution kernels to construct feature extraction modules. The 3×3 convolution kernel focuses on capturing local detail features of the signal, while the 5×5 and 7×7 convolution kernels are used to obtain broader context information. Through feature concatenation and weighted fusion, the effective extraction of features at different levels of the signal is achieved. In addition, hybrid attention mechanism can be introduced into weight features of channel and space dimension in order to improve the ability of focusing on key features. The RadioML 2016.10A Standard Data Set was used in experiment, and comprehensive testing was carried out in the SNR range between -20dB and 20dB. The results show that the recognition precision reaches 83.6% at low SNR -2dB, and it increases by 12.3% compared with traditional convolutional neural networks. When the signal-to-noise ratio increases to 10 dB, its precision reaches 97.8% and increases by 5.1% compared with the advanced ResNet-based algorithm. In the simulated multipath fading environment test, the average F1 value of the algorithm is 0.92, which is significantly better than other comparison algorithms. Through ablation experiments, it is verified that the multi-scale feature fusion module improves the accuracy of the algorithm by 7.8%, and the hybrid attention mechanism contributes to a 5.2% increase in accuracy.