Research On Communication Signal Type Recognition Based On Deep Learning

The communication signal recognition technology is a technology based on signal processing and pattern recognition theory.Its primary objective is to analyze received signals and determine their types and features.This technology finds extensive applications in electronic countermeasures,radio reconnaissance,cognitive radio,and network security.With the continuous development of modern communication technology,various new communication devices and protocols have emerged,which has led to a rapid increase in the number of communication signal types.Due to the high similarity between individual communication signals,it has become more challenging to accurately distinguish signal types using traditional signal recognition methods.Therefore,in order to adapt to the high-speed,intelligent,and real-time communication needs of modern communication systems,research on communication signal recognition is of great significance in both military and civilian fields.Based on deep neural networks,this dissertation conducts in-depth research on automatic modulation recognition,channel coding recognition,and specific protocol signal recognition.The main work and innovations are as follows:1.A novel automatic modulation recognition method based on multi-domain feature fusion is proposed.Most existing deep learning-based modulation recognition methods only use a single piece of information such as signal time-domain waveform or timefrequency domain as the network input,ignoring the complementary relationship between different domain features of the signal.This dissertation designs a densely connected multi-domain feature fusion network to simultaneously extract time-domain waveform and time-frequency spectrogram data of signal.Three fusion methods are designed to fuse the extracted multi-domain features to achieve high-performance modulation recognition by utilizing the complementary relationship between multiple-domain features.In addition,in response to the problem of sudden performance degradation caused by changes in signal transmission channel environments,a modulation recognition method based on transfer learning is proposed.The research results show that the multi-domain feature fusion method can effectively improve the accuracy of modulation recognition,and its recognition performance is better than that of the modulation recognition method based on single-domain features;meanwhile,the migration learning-based method can effectively achieve channel migration,resulting in a 3.2% improvement in recognition accuracy compared to the methods without transfer learning.2.A deep learning-based channel coding recognition method is proposed.To address the challenge of dealing with a large candidate set for the recognition of channel coding types,a hierarchical scheme for channel coding type and parameter recognition is proposed.First,a fine-tuned recognition method for channel coding types based on a multi-scale feature fusion network is proposed for the bit stream sequence output by traditional demodulators.This method achieves effective recognition of seven channel coding types.Secondly,an end-to-end multi-task signal processing framework based on convolutional neural networks is proposed for channel coding type recognition of undemodulated signal waveforms,which can recognize the signal’s channel coding type while demodulating the received signal.This avoids the complexity brought by the multilevel signal processing and analysis process and achieves efficient signal demodulation and channel coding type recognition.Furthermore,based on the completion of channel coding type recognition,a convolutional code parameter recognition method based on soft decision and deep residual network is proposed to improve the performance and robustness of convolutional code parameter recognition.3.A modulation and channel coding joint recognition method based on multi-task learning and attention mechanism is proposed.As communication technology continues to advance,traditional modulation and coding hierarchical recognition algorithms can no longer meet current needs.To address this,our paper leverages multi-task learning to construct an end-to-end deep neural network capable of achieving one-step recognition of modulation and channel coding types in received signals.The deep neural network is used to obtain multiscale features of the signals through parallel concatenated dilated convolution layers,and an attention mechanism is also used to optimize the shared features for the purpose of enhancing the useful features and suppressing the invalid ones.By comparing with the one-task method,it is demonstrated that the multi-task joint training strategy can effectively improve the performance of a single task and improve the performance of modulation coding joint recognition by 2.7%;meanwhile,the proposed method can effectively improve the performance of modulation coding joint recognition compared with other methods.4.A method for recognizing specific protocol signal types based on convolutional Transformers is proposed.Existing deep learning-based recognition methods often rely on visual characteristics of time-frequency spectrograms.To overcome this limitation,a network for recognizing specific protocol signal types based on signal waveforms is proposed,which integrates convolutional blocks and Transformer encoders.The network first uses convolutional layers to capture local information of signal waveforms for lowlevel feature extraction,and then introduces attention mechanisms to capture the longterm dependencies of signal waveforms for global feature extraction.Meanwhile,to adapt to the input of Transformer encoders,normalization and embedding methods are designed to preprocess the original one-dimensional signal waveforms.The research results show that the proposed method outperforms other methods in terms of recognition performance and computational complexity.