Radio Signal Recognition Based On Deep Learning

The recognition of radio signal channel coding and modulation has a wide range of applications in spectrum detection,channel estimation,interference identification and other fields,and it is a pre-requisite step for signal decoding and demodulation.However,traditional recognition methods are mostly based on shallow machine learning methods,there are three problems in the process of implementation: First,traditional methods have high dependence on artificial features and need to rely on complex artificial feature extraction to meet the needs of different signal recognition;Second,traditional methods are less robust,the complex electromagnetic environment has a great impact on the recognition result.Third,traditional methods have high model complexity and cannot meet the lightweight deployment requirements of signal recognition.In view of the above issues,this paper innovatively applies deep learning technology to radio signal recognition field,and designs three radio signal recognition methods based on deep learning to solve the radio signal recognition method's high dependence on artificial features,difficulty on model deploying and low model robustness.The proposed methods achieve accurate identification of radio modulation and channel coding signals.The main research contents and achievements of this article are as follows:1.Aiming at the multi-scale time-frequency characteristics of signal time-frequency images,a radio signal modulation recognition method based on multi-scale 2D convolutional network is proposed.First,a Wigner-Ville Distribution Map(WVDM)is constructed as the network input.Second,for the small difference characteristics between WVDM classes and fine granularity of features,multi-scale kernel and multi-channel convolution sub-networks are designed to separately extract the fine-grained multi-scale,multi-level features in WVDM,and used for identification.Experiments are performed on the EMC dataset,the results show that the recognition accuracy of the proposed method is improved by 10% compared with the existing methods in the signal-to-noise ratio range of [-6,20] d B,and the proposed method can effectively identify the untrained-10 d B WVDM,the experiments results verify the effectiveness and robustness of the method.2.Aiming at the difficulty of high-complexity training and the large amount of parameters of the method mentioned in the previous chapter,combined with the time-series characteristics of radio signals,a sequence-sequence encode-decode recurrent neural network modulation recognition method is proposed.First,the dual-channel IQ data of the original signal is used as the input,and the signal is characterized through the encode network to the hidden layer feature vector,then the hidden layer feature vector is decoded by the decode network to be uniformly characterized,finally the classification network is used for classification.Experiments are performed on the Radio ML2016.10 b dataset.The results show that the method improves the recognition accuracy by 5% on the [-20,+18]d B compared with the existing method,and the parameter amount decreases by nearly 90% compared with the previous chapter,which proved the lightweight,effectiveness and robustness of proposed method.3.Aiming at the problems such as the large amount of parameters of the existing deep learning model and the difficulty on deployment in practical applications,a joint recognition method of channel coding and modulation based on lightweight one-dimensional deep convolutional network is proposed.First,the original one-dimensional signal is directly used as the network input;Second,a lightweight deep convolutional neural network is constructed,and multi-scale convolution kernels and pyramidal network structure are designed to extract the deep hierarchical features.Experiments on the TMC dataset show that the method can achieve an overall accuracy of 96%,and the amount of parameters used is lower than stateof-art deep learning methods by 80%,which proves that the method has good robustness,effectiveness and lightweight.