Research On Communication Signal Modulation Recognition Based On Compressive Sensing

Automatic modulation recognition(AMR)is an intelligent signal analysis technology,which automatically determines the modulation scheme from the received waveform of communication signal.As well known to us,AMR plays a key role in the fields closely related to people's livelihood and national security.However,recent AMR deeply relies on the signal acquirement system with high sampling rate and a large observation window,which apparently causes a mass of data to be acquired,transmitted,stored,and processed.In current conditions,the above issues are indeed challenges to hardware design and implementation.The emerging theory of compressive sensing breaks through the limitation of Shannon/Nyquist sampling theorem by replacing signal sampling with information sampling,and adopts the compressed signal instead of Nyquist-sampled signal for signal processing tasks.As a result,the bottleneck resulted from large-scale data processing could be relieved.In this thesis,AMR is considered in the framework of compressive sensing,and researches concentrate to develop novel AMR algorithms with low complexity and stable classification performance.The main work are summarized as follows:1.A novel AMR approach is proposed by exploiting the sparsity of higher order spectrum.Since the locations,strength,number,and existence of higher order spectrum lines can be treated as the signatures of considered modulation scheme,AMR is available by capturing the spectrum lines.In this work,because of the sparsity,the so-called zeropadding fast Fourier transform is taken on the 4th power transformation of compressed signal.And then,through a sorting procedure,features of spectrum lines can be detected,and a well designed feature vector can be formed.To solve the recognition problem between different modulations with same constellation,such as OQPSK vs.QPSK,/4-DQPSK vs.8PSK,the weaker spectrum lines are also considered,and their features are also included by the feature vector.2.A sparse model of temporal higher-order cumulant(HOC)sequence is presented.Since the higher-order cyclic cumulants(CC)can be regarded as the Fourier coefficients of the corresponding HOC sequence,which locate at the cyclic frequencies,CCs are sparsely distributed in the Fourier domain.In light of above analysis,the estimation of CCs can be considered as a CS reconstruction problem.Thus,a rough reconstruction is introduced for pattern recognition,and features are extracted directly from compressive measurements in absence of accurate but cumbersome reconstruction algorithms.3.Motivated by the idea of machine learning,an AMR algorithm based on 6)-sparse autoencoder is proposed.In the 6)-sparse autoencoder,an over-complete redundant dictionary is learnt from the training samples,and the atoms associated to the considered patterns are also selected.Thus,the sparse representations of recognition features can be derived through the learnt dictionary.Actually,the feature extraction is implemented as a matrix multiplication.The whole procedure is performed without any prior knowledge of received signal,and the blind feature extraction is achieved.4.The application of sparse learning is also investigated.Due to the ability of feature preservation,sensing matrix is used as a tool for dimension reduction in feature extraction algorithm.Compressive learning asserts that a reliable linear classifier can be learnt in the compressive measurement domain.In the presented work,a linear support vector machine(SVM)is trained among compressive samples,and an acceptable recognition performance is obtained.In the proposed algorithm,the raw signal just need to go through a simple pre-processing stage,and the whole flow also runs independent of prior knowledge of received signals.