Research On Deep Learning Based Multi-User Detection For Massive Machine Type Communication

Different with traditional speech communication,in massive machine type communication(MTC)system,a large number of access nodes sporadically transmit small data packets with a low data rate.The control signaling carrying user information takes up a larger portion of communication overhead in this case,new designs on the PHY layer is needed to cut off the transmission overhead.Compressive Sensing(CS)theory,that is an emerging theory,breaks through the sampling frequency limit given in Nyquist sampling theorem,is able to recover the sampled data with fewer measurements and reduce lots of redundant information generated in the traditional sampling process.Some researchers have used CS in the design of massive MTC systems,and proposed a CS based multi-user detection(CS-MUD)scheme to detect active users through random access by exploiting sparsity to reduce overhead though avoid the control signaling.However,the high computational complexity of conventional sparse reconstruction algorithms prohibits the implementation of CS-MUD in real communication systems.Therefore,based on this system,this paper introduces deep learning(DL)into massive machine type communication.Firstly it discusses the system modeling of the multiuser detection method using the deep learning theory,the multiuser detection problem is converted to a multi label classification problem which can be solved by a deep neural network.The training data is generate by the pilot sequence.And once we get the trained model,a real-time multiuser detection is possible.Different with the typical CS problem that involves a sparse model,in the massive MTC communication,the system has a block structure owing to the multipath wireless channel.In particular,we design a novel sparse neural network,a new block activation layer is proposed to capture the block sparse structure in the multiuser detection problem,for a further improve in the detection accuracy.At the same time,this paper also studies the optimization of the model based on the distribution of user data.Finally,a multiuser detect model based deep learning for MMTC system is built with MXNet and Python.Experimental results show that in comparison to the basic CS reconstruction algorithms and the designed model.Experimental results show that the DL based MUD reduces the detection time by more than a thousand times compared with traditional compressed sensing reconstruction algorithms.Experimental results also show that the proposed block activation neural network model and user-distribution-based optimization method can effectively improve the accuracy of multi-user detection and shorten the time of multi-user access detection.