Blind Signal Recognition Based On Machine Learning

Signal detection and identification are widely used in various fields.However,traditional single-signal identification methods are incompetent to complex communication environments with severe electromagnetic interferences.It remains a difficult problem to identify the unknown signal in the mixed signals.The traditional blind signal recognition algorithms depend on the effect of blind signal separation.However,it is difficult to separate insufficiently sparse mixed signals in a complex or dynamic environment.This paper focuses on the topic of blind signal recognition based on machine learning,and designs algorithms for the influence of channel fading,frequency offset,and interference signals on the recognition effect,and establishes an end-to-end recognition model to directly extract the features of source signal in the mixed signal.The method has higher recognition accuracy and can better adapt to the dynamic changes of the environment than traditional methods.To solve the problem that the manually extracting features cannot fully characterize,this paper proposes to replace the artificial feature extraction algorithm with data-driven feature learning,and utilizes neural networks to model the singlesignal modulation recognition problem.The features extracted by the CNN convolutional layer can characterize the signal features more completely and the diverse training set can adapt to the dynamic changes of the environment to improve the robustness of the model.Aiming at the problem of the decrease of the recognition accuracy rate caused by the channel change in the practical application of single-signal modulation recognition,this paper proposes to use transfer learning to fine-tune the old channel model to obtain a modulation recognition model that adapts to the new channel environment.Two kinds of transfer learning algorithms are proposed,and experimental simulations are carried out for the limited training samples and sufficient conditions.The research results show that the first type of transfer learning algorithms can fine-tune the pre-training network parameters when the training samples are limited.To achieve better recognition results,the second type of transfer learning algorithm can achieve an effect close to the first type of transfer learning algorithm with sufficient training samples,and the network converges faster and requires less training time.In order to solve the problem of limited number of source signals,this paper proposes to extend the identifiable modulation signals to 11 types,and increase the mixed signal types to five.This paper redefines the problem to a multi-label classification problem,and choose LSTM,GRU,CLDNN to reconstruct mixed signals from the perspective of time and space.This paper carries out experiments from the two perspectives of simplifying network parameters and improving network recognition accuracy,and selects the network model with the best performance.