Research On Modulation Classification Methods In Non-cooperative Communication Based On Deep Learning

Modulation classification in non-cooperation communication that aims at achieving modulation classification under no prior knowledge is an important process in blind signal processing and an effective guarantee of demodulation,decoding,and information restoration,and a basis of the radio monitoring,electronic countermeasure and spectra electronic warfare,which has played a significant role in many military and civil fields.It is a challenging and practical research subject with a highly complex and changeable electromagnetic environment,insufficient knowledge and large amount of data.Modulation classification based on traditional methods including likelihood ratio testing and expert features that requires extensive domain knowledge can not achieve good classification performance with insufficient representation of features and is unable to cope with huge amount of data.Deep learning with end-to-end learning ability has powerful mass data processing capability and deep feature extraction capability,which provides a solution to modulation classification in complex environments,and achieves better classification results than traditional algorithms.However,because deep learning has poor interpretation and need mass labeled data,modulation classification based on deep learning still faces many challenges.In the dissertation,a series of solutions are proposed to solve the problems such as weak representation of input features,insufficient information,unreasonable network structure,and low model robustness by means of multi-feature fusion,network structure combination,and the introduction of distance metrics to establish an adaptive layer.Abundant experiments on simulation and actual data are conducted to verify the effectiveness of the algorithms.The main work and contributions of this dissertation are summarized as follows:1.Modulation classification using mean Gaussian probability distribution of phase and amplitude based on CNN(Convolutional Neural Network,CNN)and LSTM(Long Short-term Memory,LSTM)for symbol synchronous complex signals is proposed.Firstly,according to the symmetry,the signal is preliminarily preprocessed using fold.Secondly,the signal is mapped into the signal feature distribution domain with Gaussian kernel density estimation whose representation is significant and the distributions are processed using average,which increases the signal length implicitly,achieves smoother input and enhances completeness of modulation information.Then,the constellation matrix distribution is used to realize the conversion of IQ signal from the time domain to the space domain.Finally,according to the different input characteristics,a reasonable network structure is designed to complete modulation classification and the experimental results show that the algorithm is superior to others.2.Modulation classification based on multi-dimensional CNN and LSTM using multi-feature fusion is proposed for baseband complex signals.It takes advantage of the orthogonal characteristics and high-order sparsity of IQ signals and effectively improves classification performance.Firstly,according to the orthogonal characteristics of IQ signal that I,Q and IQ signal have modulation information,a multidimensional CNN module(MD-CNN)is proposed to fully exploit features from IQ signals using different dimension convolutional filters.1×8 convolutional filters are used to extract individual features and 2×8 filters are used to extract interactive features.Secondly,LSTM layer is used after the convolution layer to further extract deeper modulation features and exploit intrinsic sequential information.Then based on the cyclical periodicity and high-order sparsity characteristics,a hierarchical multi-feature fusion(HMF)scheme is proposed to enhance feature diversity,where IQ signal and the biquadrate higher order statistics are utilized as two parallel branches of input.Finally,the effectiveness of the proposed method is verified by abundant experiments that are conducted on the dataset RadioML2016.10a and RadioML2016.10b and includes discussions about hyperparameter configuration,enhanced feature diversity,and performance comparison with different classifiers and varieties.Better performance compared with state-of-the-art works is presented.3.In view of the weak representation,incompletion information of the input features and unreasonable network design of the current algorithms,modulation classification based on attention mechanism and hybrid networks using multi-features fusion is further proposed.Firstly,the IQ signal is mapped into the spatial domain with the characteristic that the time-frequency spectrogram has both signal time and frequency information.Then,based on that ConvLSTM(Convolutional Long Short-term Memory,ConvLSTM)simultaneously learns space features and sequence features,a network is proposed that consists of CNN and ConvLSTM to exploit intrinsic information and stronger representation features.Meanwhile,attention mechanism is introduced to improve the classification performance by weighted features.Afterwards,the IQ signal is mapped into the phase Gaussian probability distribution that is related to the modulation type because the time-frequency spectrogram does not contain phase features to enhance completeness of information.And the cascade of CNN and LSTM is used to extract intrinsic features.Finally,a series of experiments including enhanced feature diversity,and performance comparison with different classifiers and varieties are conducted on the dataset RadioML2016.10a.It is proved that the classification performance of the proposed algorithm is better than others’ in the whole signal-to-noise ratio.Especially,the proposed method achieves a breakthrough in the classification accuracy of low signal-to-noise ratio,and has an increment of more than 15%compared with other methods.4.Considering weak robustness and difficulty in labeling,modulation classification based on distance measure and CNN is proposed.It is the first time to introduce unsupervised domain adaptation into modulation classification to alleviate dataset shift caused by signal parameters and improve generalization robustness with unseen signal settings that labor intensive labeling on target domain is not required.Firstly,a CNN model is proposed to learn features and classify modulation for source signal.Afterward,an adaption layer with MK-MMD(Multiple Kernel Maximum Mean Discrepancy,MK-MMD)and CORAL(Correlation Alignment,CORAL)is involved for linking source data and unlabeled target data that mitigates domain discrepancy of learned CNN features.Finally,a series of experiments are conducted,while a real world signal dataset consisting of eight digital modulation schemes is provided for public use.Simulation results show that the proposed method achieves satisfying performance at varying sampling frequencies and symbol rates with the least degradation on source domain data.The proposed method provides a new idea for modulation classification on unlabeled signals,pushing the deep learning deployment toward real practical scenarios.