Research Methodology:methods Of Device Activity Detection For Massive Machine-type Communication Via Neural Networks

As one of the three key scenarios of the fifth-generation wireless systems(5G),massive machine type communication(m MTC)is widely used in intelligent transport systems,city brain and health monitoring etc.,which has provided great convenience for human production and life.At the same time,however,huge numbers of users,large amounts of data and complex operational scenarios also pose serious technology challenges in the communications field,including the problem of radio access control.Grant-free non-orthogonal user access schemes provide a promising method towards massive number of devices and sporadic device activity patterns in the network.Since the sparsity of user activity pattern presents a compressed sensing(CS)problem,many algorithms of sparse linear inversion can be leveraged to detect active devices and estimate channels.Based on this scheme,we propose a novel neural network(NN).Firstly,we briefly introduce the development history of modern communication technology and the main application scenarios of next-generation wireless technologies,and highlight some practical applications of the m MTC scenario.Then we introduce the current research status of sparse signal recovery algorithms and deep unfolding net(DUN),which play key roles in our following research.Secondly,based on the pratical scenarios of m MTC,a mathematical model is established,which aims to abstract the user activity detection as a solution to sparse inverse problems.We suppose that each user is equipped with one antenna.When the base station(BS)has multiple antennas,the solution to this problem can be regarded as a high-dimensional sparse matrix recovery problem with similar inter-column sparse patterns.When the BS has a single antenna,the problem is reduced to a classic sparse linear inverse problem.At the same time,we introduce and re-deriving two classic CS algorithms: iterative shrinkage thresholding algorithm(ISTA)and approximate message passing(AMP)algorithm.These two algorithms can be unfolded and optimized in the DUN.Then,we introduce two types of NN,including data-driven NN and model-driven NN,and explain the advantages of applying model-driven networks to communication problems.As a new type of interpretable NN,its internal structure is based on a rigorous mathematically proven algorithm.Due to the fixed structure,it has very few trainable parameters and low training complexity,which is more suitable for communication systems with strict delay requirements.We propose a novel network called the Stein 's unbiased risk estimate AMP net(SUREAMPNet),SANet for short.This is a new network based on the AMP algorithm,with divergence-free adaptive denoiser as the core and DUN as the framework.Starting with single measurement vector(SMV),we derive the calculation method for the coefficient of the adaptive denoiser and show the structure of the SANet-SMV network.Simulation results show the superiority of the SANet-SMV network compared with other algorithms in terms of convergence speed,convergence results and ill-condition countermeasures.We also illustrate the no prior needed and low training complexity of SANet-SMV.Finally,we extend the SMV into a high-dimensional multiple measurement vector(MMV)network.With transmission power and scale fading considered,this network can convert complex signals into real-valued calculation.The strategies of offline training and online usage are also detailed.Simulation results show the superiority of SANet-MMV in the performance of error rate compared to the mainstream method vector approximate message passing(VAMP).In addition,when the number of antennas increases,SANet-MMV can maintain its own training complexity and reduce error rate.We also study the feasibility of SANet-MMV in practical,propose schemes of choosing the number of layers,and show how can it detect device activity without knowing the access probability of users and how long it takes to calculate in a trained network.SANet outperforms many other algorithms in performance and has strong robustness.It can be well applied in high-dimensional problems to solve the multi-antenna problem.As an alternative of NN,its no prior needed,low training complexity and efficient computing capabilities provide the possibility for its application in practical scenarios.