Research On Deep Learning Methods For Spectrum Recognition

With the large-scale commercial use of 5G and the rapid development of the Internet of Things,human society has entered a new stage of "Internet of Everything".The significant increase in frequency-using equipment has made the insufficient spectrum resources increasingly scarcer.And the perception,understanding and utilization of frequencies have become increasingly complex.Accurate understanding of the electromagnetic environment through the identification of the spectrum characteristics of wireless communication and efficient use of spectrum resources have become an important part of ensuring the quality of wireless communication and improving user satisfaction.Traditionally,in order to identify typical spectrum features such as modulation methods or radiation source fingerprints,the industry has developed methods including likelihood ratio detection and pattern recognition,which have achieved certain results.However,these methods still have inherent defects such as low recognition accuracy,narrow application range,and high labor cost.In view of the great success of deep learning technology in the fields of image recognition,video understanding,natural language processing,etc.,the industry has begun to apply deep learning technology to spectrum feature recognition and carried out preliminary explorations.In this context,this dissertation expects to design a deep learning method for spectrum feature recognition and use it in specific scenarios such as automatic modulation recognition and radio frequency fingerprint recognition.Secondly,a data generation method is proposed based on Generative Adversarial Network,in order to solve the challenges in small sample scenarios;finally,based on the concept of federated learning,a distributed spectrum feature recognition architecture was developed and a typical deep learning model was verified.Specifically,the research content of this dissertation includes 3 aspects:1.Aiming at the disadvantages of the traditional algorithms and deep learning algorithms used for spectrum feature recognition,such as high manual participation,low recognition accuracy,and poor generalization ability,a deep learning algorithm for spectrum feature recognition,namely Multi-level attention CNN Bi-LSTM(MCBL)is proposed.MCBL effectively integrates Convolutional Neural Network,Recurrent Neural Network(RNN)and attention mechanism to more accurately extract the spatial,temporal and saliency information of the spectrum signal.In order to verify the performance of the MCBL network,comparative experiments were carried out on the public data sets of modulation recognition and radio frequency fingerprint recognition.Compared with 9 common deep learning algorithms,the recognition accuracy of the MCBL model was up to 93%,and It can still achieve better recognition results when SNR is low,and its overall performance is better than current methods.2.Aiming at the problem of under-fitting or over-fitting of the model when performing spectral feature recognition based on deep learning in small sample scenarios,and then the recognition accuracy is severely reduced,a data enhancement method based on generative adversarial networks is proposed,Temporal Data Generative Adversarial Network(TDGAN).The network obtains the timing characteristics of the spectrum signal,trains and generates high-quality sample data,and then uses it for deep network training,which significantly enhances the training effect of the deep learning model and improves its robustness.Finally,the TDGAN network is applied to the fields of modulation recognition and radio frequency fingerprint recognition,and the generated data is added to the original data for experimentation.The results show that TDGAN can accurately fit the characteristics of a given small sample,and the generated data samples are similar to the characteristics of the given samples.After the generated data is used for deep network training,the training accuracy can be improved by nearly 50%,and good data enhancement effect are achieved.3.For distributed platforms with limited computing and storage resources,a distributed spectrum feature recognition algorithm based on federated learning is proposed.Communication platforms with limited capabilities are often difficult to independently run deep learning networks with tens of millions of parameters,and the running time is long or even unable to converge.Therefore,in view of the features of the embedded platform with limited capabilities,the idea of federated learning is adopted to realize a lightweight distributed spectrum feature recognition deep neural network architecture,relying on multiple platforms with limited capabilities to jointly complete neural network training and obtain relatively accurate recognition Model.Then,a lightweight deep neural network is implemented in the proposed distributed architecture,and it is applied in the fields of modulation recognition and radio frequency fingerprint recognition.Experimental results show that the training accuracy of a single distributed platform with limited resources is very low.By applying the proposed distributed training architecture,the recognition accuracy can be significantly improved by more than 15%.