Research On Modulation Recognition In Wireless Communication Based On Deep Learning Under Low SNR

The contemporary wireless communication technology is rapidly advancing,and to meet the demands of different scenarios,the variety of modulation modes is increasing.Modulation recognition plays a crucial role in signal detection,serving as a key step in communication and signal monitoring.It finds significant applications in military domains such as electronic warfare and intelligence interception,as well as civilian domains like spectrum monitoring and interference identification.The wireless radio spectrum resources are limited and have become increasingly strained in recent years due to exponential growth in the number of frequency-usage devices.This has resulted in a complex and dynamic electromagnetic environment,escalating the uncertainty and complexity of wireless channels.Enhancing the accuracy of modulation recognition under low signal-to-noise ratio(SNR)conditions has become a focal point and a challenge in research.Deep learning methods combined with massive data have significantly improved their feature representation capabilities.They effectively process information such as signal amplitude,phase,and frequency through multi-level feature extraction and abstraction.In recent years,with the advancement of computer hardware and artificial intelligence technology,deep learning has achieved remarkable results in various fields.Based on this,this paper focuses on the research of deep learning-based wireless communication modulation recognition under low SNR conditions.The main contributions and innovations are as follows:(1)A wavelet fusion denoising algorithm based on Optuna is proposed.Traditional wavelet denoising methods rely on manually selecting wavelet basis functions and threshold values,which have certain limitations.In this study,the Optuna hyperparameter optimization framework is used to dynamically search for a set of wavelet basis functions and threshold values that yield good denoising effects during neural network training.The original signals are then denoised using different wavelet basis functions and threshold values,and the denoised signals are fused to generate new input data for the neural network.This approach effectively preserves more signal details while reducing noise.Experimental results demonstrate that this algorithm can effectively improve the accuracy of modulation recognition under low SNR conditions.(2)A modulation mode recognition algorithm based on SET-1DRes Net6 is proposed.To better extract features from the wavelet fusion denoised signals,the proposed algorithm enhances the traditional Res Net18 model.It introduces channel attention mechanisms after the first convolutional layer to weight the signals denoised by different wavelet basis functions,allowing the network to focus more on the denoised signals with higher recognition accuracy.It also incorporates spatial attention mechanisms to weight the spatial information of individual denoised signals,enabling the network to pay more attention to the sampling areas less affected by noise.Furthermore,a transposed compression excitation block is added to the Basic Block to further mitigate the impact of noise on recognition results.To reduce the number of network parameters,the 2D convolutions in Res Net are replaced with 1D convolutions,and the second and fifth down sampling convolutions in Res Net are removed.As a result,the parameter count is reduced to 3.76% of Res Net18.Simulation experiments are conducted using the publicly available RML2016.10 a dataset.Comparative experimental results with other algorithms using the same dataset demonstrate that the improved network not only enhances the recognition accuracy under low SNR conditions but also reduces the parameter count.The transposed compression excitation block only increases the parameter count by 1%,while improving the overall and low SNR recognition accuracy by 4.73% and8.18%,respectively,achieving 65.42% and 45.72% accuracy.