Signal Separation Based On Deep Learning

With the rapid development of the economy and technology,the information revolution represented by the Internet of Things,cloud computing,and artificial intelligence is progressively going forward.In the current era,interconnection between a large number of mobile devices has become a remarkable feature.However,the increasing interconnectivity has led to explosion in the number of wireless devices and a large number of non-cooperative communication scenarios,which makes signal separation increasingly important.In complex electromagnetic environments,signal separation plays a critical role in improving communication efficiency.While research on the separation of singlechannel modulation signals has made some progress over the years,conventional methods have been limited by artificially designed feature extraction,which are insufficiently robust.Deep learning,which relies on learnable feature extraction,has been widely used in tasks such as channel estimation,modulation recognition,speech denoising,and speech separation,and has emerged as a powerful tool in addressing signal separation problems.Aiming at the separation of modulated signals that exist in the same frequency band,this thesis proposes two signal separation models based on deep learning:AttTCNNet,which combines temporal convolutional network and attention mechanism,and CRNRVQNet,which is based on a residual vector separator.AttTCNNet consists of three main components:encoder,separator,and decoder.The model aims to separate signals in the time domain and adopts the scale-invariant source-to-noise ratio of the estimated signal as the loss function to update the model parameters.AttTCNNet stacks residual convolutional networks with different downsampling rates in encoder and decoder to extract features and reconstruct waveforms.The separator,which combines temporal convolutional networks and attention mechanism,performs feature mask estimation for signal separation.CRNRVQNet utilizes vector quantization,which retains the crucial information of the input vector,to construct the residual vector separator.The separator filters out the source signal features from the mixed signal features.Based on AttTCNNet,the encoder and decoder in CRNRVQNet adopt an interleaved structure of convolutional neural network and intraframe recurrent neural network.Due to the fact that public modulation signal separation dataset in the related research work can be hardly found,this thesis creates a modulation signal separation dataset based on the RML2016.10a dataset.Based on the dataset,this paper explores the influence of the different configurations in AttTCNNet and CRNRVQNet on the separation performance and conducts a visualization study on the separation process.The experimental results show that AttTCNNet and CRNRVQNet exhibit better separation performance compared to other deep learning models under the similar number of parameters and computational complexity.