Automatic Modulation Recognition Method Under Dynamic SNR Conditions Based On Machine Learning

As the core technology of software radio and cognitive radio,AMR(Automatic Modulation Recognition)is widely used in today's intelligent communication systems.The performance of the current AMR methods are sensitive to SNR,and most existing methods are lack of generalization under dynamic SNR conditions.While under the circumstances like deep space communication and high speed aircraft communication,due to cosmic rays and high speed movement of aircraft,the SNRs are in large dynamic range and changes rapidly at the receiver,which makes the performance of AMR highly depend on the speed and accuracy of channel estimation.This causes AMR methods not feasible under dynamic SNR circumstances.This paper studies on automatic modulation recognition methods based on machine learning under dynamic SNR conditions for singlesignal and multi-signal cases,which aims at improving the classification and generalization ability of AMR,reduce the dependence on channel estimation.First,under the condition of non-cooperative communication systems,when the available amount of samples is few,to solve the problem of lack of generalization ability caused by redundant features in extracted feature set,an automatic modulation recognition method based on robust feature selection for few-shot learning is proposed.At first,K-means clustering is applied to select noise-robust features through the statistical information of features under different SNRs,then attribute reduction based on rough set is used to remove redundant features.A robust feature set that suitable for classification under dynamic SNR condition is obtained,and is then used to train the classifier.The method is able to greatly reduce the dimension of the feature set,and the generalization ability is greatly improved while maintaining the original recognition performance with few training samples.Second,under the condition of cooperative communication systems,when enough amount of samples are available,AMR methods based on feature learning using deep learning is studied.Deep learning is appied to solve more complex conditions when samples are enough.In total,two different conditions are considered:(1)Under the condition of white Gaussian noise channels,to solve the insufficient generalization ability caused by the lack of effective features under dynamic SNR conditions,an AMR method for complex modulation sets based on robust feature combination is proposed.Since feature selection can't handle the problems caused by insufficient features,SAE(Stacked Auto-Encoder)is applied to train and learn from all the features extracted in the SNR range,under the premise of being able to be classified under fixed SNR.An set of new features is obtained after feature combination.This method greatly reduces the feature dimension,and at the same time improves the feature set's recognition and generalization ability under dynamic SNR conditions.(2)Under the condition of multi-path fading channels,signals can't be effectively recognized under fixed SNR under multi-path fading channels,and cannot be deployed under dynamic SNR condition.To solve this,an AMR method for complex modulation set using CNN(convolutional neural network)is proposed.When the feature set cannot be identified under a fixed SNR,finding new effective features is a time-consuming,laborious and performance-unguaranteed task.CNN is applied to directly learn features suitable for recognition from a large amount of samples through feature learning under dynamic SNR condition.Through the visualization of the intermediate layers and the output features,it is verified CNN learned features can provide excellent generalization ability.This method does not rely on manual search and design of feature sets,and can effectively handle the multi-path fading channels.Last,Under the condition of overlapped inseparable multisignals,they need to be classified at one time.To solve this,an AMR method for overlapped signals under dynamic SNR condition is proposed.Through the combination of capsule network and CNN,the network is able to achieve multi-signal classification.Threshold decision is proposed to achieve the detection of signal number;a corresponding loss function is proposed to improve the speed of convergence.At the same time,The principals of results and performance evaluation brought by unknown number of signals are also defined.Compared with existing methods,the number of modulation types and performance of recognition is significantly improved,and under the condition that 3 signals exist simultaneously,the recognition is achieved the first time.