Massive MIMO multi-user scheduling method based on improved convolutional neural network

Abstract

The application relates to a Massive MIMO multi-user scheduling method based on an improved convolutional neural network. The method comprises the following steps: a base station sequentially scans an analog precoding DFT codebook, and in combination with channel state information fed back by a user, selects an optimal codebook from each codebook matrix to determine an optimal channel matrix of the user under a 5G communication system; an improved particle swarm algorithm is used to analyze network structure parameters of the convolutional neural network, the network structure parameters of the convolutional neural network used for Massive MIMO multi-user scheduling are determined, an optimal convolutional neural network model is constructed for training, a Massive MIMO multi-user scheduling model based on the convolutional neural network is obtained, and in an industrial Internet of Things scene under the 5G communication system, the optimal channel matrix of a current user to be scheduled is input into the Massive MIMO multi-user scheduling model to predict a current optimal scheduling user set, and the optimal scheduling user set is obtained. Therefore, the calculation load is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a Massive MIMO multi-user scheduling method based on an improved convolutional neural network. BACKGROUND

[0002] At present, 5G communication systems have been widely used, but due to the massive nodes of the Internet of Things, the 5G and B5G (super 5th generation mobile communication) network is required to support a large number of access users. The Massive Multiple Input Multiple Output (Massive MIMO) technology is one of the effective ways to solve the above problems, the core idea of which is to configure a large number of antennas at the base station end and support multiple user communication at the same time by dividing the spatial resources, which makes the system can serve multiple users on the same time-frequency resource, but the actual number of access network users is often greater than the number of users that the system can serve at the same time, so it is necessary to efficiently select the best users from the access users for resource allocation (i.e. multi-user scheduling), and in the case of massive nodes in industrial Internet of Things, the complexity of the scheduling algorithm is challenged.

[0003] Another link that affects the complexity of the scheduling algorithm is the method of selecting users. Since in multi-user MIMO, users simultaneously occupy frequency band resources. It is necessary to use precoding to eliminate inter-user interference, and using ZF precoding will completely eliminate inter-user interference, but the user precoding matrix changes with the user combination, that is, the achievable rate of the user will affect the result of user selection, and the selected user set will in turn affect the precoding matrix, and then affect the achievable rate of the user. When using the MMSE precoding matrix to minimize the interference, inter-user interference still exists, so the interference of the selected user will affect the achievable rate of the user. It is this problem that prevents the user selection method from directly selecting the "best" user to achieve the highest throughput. The optimal method is to exhaustively search and compare the system throughput under different user combinations to find the best user combination. But it is obvious that the exhaustive method has too large a calculation amount. Therefore, the existing method basically uses a greedy algorithm, however, the greedy algorithm still needs to compare the throughput of different user combinations, and needs a large number of singular value decomposition (SVD) processes to calculate the achievable rate of the user under different user combinations, thereby generating a huge calculation load. SUMMARY

[0004] Therefore, it is necessary to provide a Massive MIMO multi-user scheduling method based on an improved convolutional neural network, which can reduce the calculation load.

[0005] A Massive MIMO multi-user scheduling method based on an improved convolutional neural network, the method comprising:

[0006] Step 1, the base station scans the analog precoding DFT codebook in order, and selects the best codebook from each codebook matrix in combination with the channel state information fed back by the user, to construct the best channel matrix of the user under the 5G communication system;

[0007] Step 2, the network structure parameters of the convolutional neural network are analyzed by using the improved particle swarm algorithm, and the network structure parameters of the convolutional neural network used for Massive MIMO multi-user scheduling are determined, to construct the best convolutional neural network model for Massive MIMO multi-user scheduling;

[0008] Step 3, the best convolutional neural network model is trained to obtain a Massive MIMO multi-user scheduling model based on convolutional neural network;

[0009] Step 4, in the industrial Internet of Things scene under the 5G communication system, the best channel matrix of the current user to be scheduled is input into the Massive MIMO multi-user scheduling model to predict the current best scheduling user set, and the best scheduling user set is obtained.

[0010] In one embodiment, the base station scans the analog precoding DFT codebook in order, and selects the best codebook from each codebook matrix in combination with the channel state information fed back by the user, to determine the best channel matrix of the user under the 5G communication system, comprising:

[0011] The space is divided according to the number of base station antennas N r and the number of user antennas N t , N r and N t different angle array response vectors are written into a matrix respectively, to construct an analog precoding DFT codebook;

[0012] The base station scans the analog precoding DFT codebook in order, and selects the angle from the analog precoding DFT codebook in combination with the channel state information fed back by the user, to realize the best beamforming and determine the best channel matrix of the user under the 5G communication system.

[0013] In one embodiment, the network structure parameters of the convolutional neural network are analyzed by using the improved particle swarm algorithm, and the network structure parameters of the convolutional neural network used for Massive MIMO multi-user scheduling are determined, to construct the best convolutional neural network model for Massive MIMO multi-user scheduling, comprising:

[0014] The parameters are set in advance, the value range of the network structure of the convolutional neural network is set, N sets of network structure parameter combinations are randomly generated according to the value range, and are used as N particles in the initial population P;

[0015] N network models are constructed by using the N sets of network structure parameter combinations, the N network models are trained and tested by using sample data, the prediction accuracy of the N network models is obtained, the prediction accuracy of the N network models is used as the fitness function, and the fitness value of each particle is calculated according to the fitness function;

[0016] The fitness value of the current position is compared with the fitness value corresponding to the historical best position, if the fitness value of the current position is higher, the current position is used to update the historical best position;

[0017] The fitness value of the current position is compared with the fitness value corresponding to the global best position, if the fitness value of the current position is higher, the current position is used to update the global best position;

[0018] The speed and position of each particle are updated according to the adaptive speed updating and position updating method;

[0019] Until the termination condition is reached, the best network structure parameter combination of the convolutional neural network is output, otherwise the step of calculating the fitness value of each particle according to the fitness function is returned;

[0020] The output best network structure parameter combination of the convolutional neural network is used as the network structure parameter of the convolutional neural network for Massive MIMO multi-user scheduling, to construct the best convolutional neural network model for Massive MIMO multi-user scheduling.

[0021] In one embodiment, the expression of the adaptive speed updating and position updating method is:

[0022] r(t+1)=4r(t)(1-r(t)),r(0)=rand

[0023]

[0024] Wherein, ω is the inertia weight, ω max = 0.9, ω min = 0.4, k is the current iteration number, M is the total iteration number, rand is a random number between 0 and 1, r(t) and r(t+1) represent the corresponding values of chaotic mapping t, t+1 times respectively;

[0025] When :

[0026]

[0027] When Time:

[0028]

[0029] where k is the current iteration number, M is the total iteration number, is the current velocity, is the current position, d = 1, 2, …, D, i = 1, 2, … n, c1 and c2 are the social learning factor and self-learning factor of the particles in the population respectively, r1 and r2 are random parameters between 0 and 1, is the current individual extreme value, is the current group extreme value, is the current group best position.

[0030] In one embodiment, the optimal channel matrix of the user under the 5G communication system is used as a training sample to train the optimal convolutional neural network model, and a Massive MIMO multi-user scheduling model based on a convolutional neural network is obtained, which includes:

[0031] The optimal channel matrix of the user under the 5G communication system is obtained as a training sample set through an exhaustive algorithm;

[0032] The training sample set is used to train the optimal convolutional neural network model offline until the optimal convolutional neural network model meets the convergence condition, and a Massive MIMO multi-user scheduling model based on a convolutional neural network is obtained.

[0033] The above-mentioned Massive MIMO multi-user scheduling method based on an improved convolutional neural network determines the optimal channel matrix of the user under the 5G communication system by scanning the analog precoding DFT codebook in order at the base station and selecting the optimal codebook from each codebook matrix in combination with the channel state information fed back by the user, analyzes the network structure parameters of the convolutional neural network using the improved particle swarm algorithm, determines the network structure parameters of the convolutional neural network for Massive MIMO multi-user scheduling, constructs the optimal convolutional neural network model for Massive MIMO multi-user scheduling, trains the optimal convolutional neural network model, and obtains a Massive MIMO multi-user scheduling model based on a convolutional neural network. In the industrial Internet of Things scenario under the 5G communication system, the optimal channel matrix of the current user to be scheduled is input into the Massive MIMO multi-user scheduling model to predict the current best scheduling user set, and the best scheduling user set is obtained. Therefore, using the deep learning method based on the convolutional neural network can greatly reduce the calculation amount of the scheduling algorithm, reduce the complexity of the traditional user scheduling, and reduce the calculation load. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 1 is a flow chart of a Massive MIMO multi-user scheduling method based on an improved convolutional neural network in one embodiment;

[0035] Figure 2 A schematic diagram of a scheduling scenario for an industrial IoT user of a 5G communication system in one embodiment;

[0036] Figure 3 A schematic diagram of an antenna array in one embodiment;

[0037] Figure 4 Schematic diagram of a convolutional neural network structure in one embodiment;

[0038] Figure 5 This is a flowchart of optimizing the network structure parameters of a convolutional neural network based on the improved particle swarm algorithm in one embodiment. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0040] In one embodiment, Figure 1 As shown, a Massive MIMO multi-user scheduling method based on an improved convolutional neural network is provided. The method is applied to a terminal as an example and includes the following steps:

[0041] In step 1, the base station sequentially scans the simulated precoding DFT codebook and selects the best codebook from each codebook matrix based on the channel state information fed back by the user to construct the optimal channel matrix for the user in the 5G communication system.

[0042] Step 2: Use the improved particle swarm algorithm to analyze the network structure parameters of the convolutional neural network, determine the network structure parameters of the convolutional neural network for Massive MIMO multi-user scheduling, and build the optimal convolutional neural network model for Massive MIMO multi-user scheduling.

[0043] The improved particle swarm algorithm may be an improved particle swarm optimization (PSO) algorithm.

[0044] Among them, particle swarm optimization algorithm (PSO) is a kind of random search algorithm based on group cooperation by simulating the foraging behavior of bird swarm, which is generally considered as a kind of swarm intelligence (SI). Its own limitations are also obvious. With the increase of iteration times, other particles will accelerate to the temporary optimal solution, at this time, the speed and position updating operation will be difficult to increase the diversity of the population, causing the difference between the individuals in the population to gradually decrease, if the temporary optimal solution is a local optimal solution of the function problem, then the population individuals will be difficult to jump out of the local range affected by the temporary optimal solution, so as to cause the algorithm to fall into local optimum, and premature convergence phenomenon appears. Moreover, in the iteration process, it is also easy to appear that a certain individual is a local optimal solution, leading to slow convergence, and then may lead to premature algorithm. In actual research and application, the function problems encountered are mostly high-dimensional and multi-extremum, so the particle swarm algorithm is easy to fall into the local optimum circle and cannot search the global optimal solution.

[0045] Among them, the improved particle swarm algorithm can balance the global search ability and the local search ability, and improve the search efficiency of the particle swarm algorithm.

[0046] The main parameters of the improved particle swarm algorithm include: population size N, particle dimension D, iteration number M, inertia weight ω and learning factor c1, c2. For the convergence and optimization ability of the algorithm, the population size and the iteration number are certainly better. But when the population size and the iteration number increase, the algorithm will certainly increase the operation time. How to reasonably set the inertia weight ω and the learning factor c1, c2 within a certain time. The inertia weight ω of the present application is set as a fixed value or linearly decreases to optimize the performance of the PSO algorithm. Therefore, the inertia weight ω is adopted to realize performance optimization by non-linearly decreasing with the iteration number and combining with chaotic mapping, and the corresponding calculation expression of the inertia weight ω is:

[0047] r (t+1) = 4r (t) (1-r (t)), r (0) = rand

[0048]

[0049] Among them, ω max = 0.9, ω min = 0.4, k is the current iteration number, M is the total iteration number, rand is a random number between 0 and 1, r (t) and r (t+1) represent the corresponding values of chaotic mapping t and t+1 times respectively.

[0050] Wherein, for adaptive speed update and position update method, a population position update method is adopted according to the number of iterations to adaptively change, in the unimproved particle swarm optimization algorithm (PSO), since the iteration later stage, the particle changes the position constantly, thus the particle becomes more and more similar, easily leading to the algorithm falling into local optimum. And the main reason is: in the population update, the position of the next particle is obtained by directly adding the position of the previous particle and the particle update speed, thereby leading to falling into local optimum in the iteration later stage. The particle swarm optimization algorithm (PSO) is improved into an adaptive speed update and position update method based on a nonlinear function, which can balance the global search ability and the local search ability, improve the search efficiency of the particle swarm algorithm, prevent falling into local optimum in the iteration later stage, and the expression of the adaptive speed update and position update method is:

[0051] When ,

[0052]

[0053] When ,

[0054]

[0055]

[0056] Wherein, k is the current iteration number, M is the total iteration number, is the current speed, is the current position, d=1, 2, …, D, i=1, 2, …n, c1 and c2 are the social learning factor and self-learning factor of the population particle respectively, r1 and r2 are random parameters between 0-1, is the current individual extreme value, is the current group extreme value, is the current group best position.

[0057] Step 3, training the best convolutional neural network model to obtain a Massive MIMO multi-user scheduling model based on convolutional neural network.

[0058] Step 4, in the industrial Internet of Things scene under the 5G communication system, input the best channel matrix of the current user to be scheduled into the Massive MIMO multi-user scheduling model to predict the current best scheduling user set and obtain the best scheduling user set.

[0059] Wherein, the Massive MIMO multi-user scheduling model is used for the base station end, the base station end includes a base station and multiple industrial Internet of Things user nodes, and the base station selects the best user set according to the maximum throughput. For example Figure 2The following is a schematic diagram of the scheduling scenario of industrial IoT users in the 5G communication system. In the scheduling scenario of industrial IoT users in the 5G communication system, the downlink of the Massive MIMO multi-user Hybrid Beam Forming (HBF) system is considered. The base station is configured with N t antennas, each user terminal is equipped with N r The base station RF front end adopts a fully connected mode and is equipped with N RF Each RF link is connected to a uniform planar array (UPA) via a phase shifter network. When serving IIoT users, the base station first implements beamforming for the data stream using analog and digital precoding. The base station then uses the Massive MIMO multi-user scheduling model to predict the user scheduling set and select the optimal user set for service.

[0060] In one example, for a single cell, assuming there are N u Candidate IoT users (currently waiting to be scheduled), of which only N can be selected RF The Massive MIMO multi-user scheduling model on the base station side is used to predict the user scheduling set. u Select N out of candidate IoT users RF users (i.e., the optimal scheduling user set) for service.

[0061] In the Hybrid Beam Forming (HBF) system, the hybrid precoding is performed by the analog precoding. and digital precoder Composition, among which F RF The matrix structure is N t ×N RF Matrix, F BB The matrix structure is N RF ×N s Matrix. F RF It is defined as constant amplitude and adjustable phase. BB Its phase and amplitude parameters can be adjusted. t is the number of transmitting antennas, N RF is the number of RF radio links, N s is the number of data streams in the Massive MIMO multi-user system. is the analog precoder F RF Nth RF The value of the column, is the digital precoder F BB Chinese N sThe values of the columns. After the precoding described above, the transmitted signal can be expressed as:

[0062] x = F RF F BB s

[0063] where F RF is an analog precoder, F BB is a digital precoder, is an N S x 1 transmitted signal vector, where represents the N S th data stream, and T is the matrix transpose. Assuming that the user terminals are equipped with linear receivers, the received signal can be expressed as:

[0064]

[0065] where H n represents the user channel matrix of the n th user, and by transforming, the received signal can be expressed as:

[0066]

[0067] where s n is the transmitted signal, is the useful signal, is the inter-user interference, and q is the channel noise interference. The signal-to-interference ratio of the k th user and the n th beam can be expressed as:

[0068]

[0069] where H k is the user channel matrix between the k th user and the base station, P is the transmitted signal power, and σ 2 is the noise power. F RF and respectively represent the N t x N RF order analog precoding matrix of the n th user, and the N RF x N s order digital precoding matrix. By using the MMSE algorithm to minimize the interference, the precoding matrix can be obtained.

[0070] Because the millimeter wave has a short wavelength and a high free space path loss, in order to reflect the limited scattering and sparse low rank of the millimeter wave channel, the millimeter wave propagation environment is usually described by a cluster channel model, that is, the S-V model. The model describes the millimeter wave channel matrix as

[0071]

[0072] where L is the total number of multipaths, and L idenotes the i-th propagation cluster, g i,l denotes the complex gain of the i-th multipath in the i-th propagation cluster, a r (θ i,l ) and denote the receive and transmit array response vectors, respectively, denote the azimuth and elevation angles of arrival and departure, respectively.

[0073] The above-mentioned Massive MIMO multi-user scheduling method based on an improved convolutional neural network, by the base station, scans the analog precoding DFT codebook in order, and combines the channel state information fed back by the user, to select the best codebook from each codebook matrix, to determine the best channel matrix of the user under the 5G communication system, the improved particle swarm algorithm is used to analyze the network structure parameters of the convolutional neural network, to determine the network structure parameters of the convolutional neural network used for Massive MIMO multi-user scheduling, to construct the best convolutional neural network model used for Massive MIMO multi-user scheduling, the best convolutional neural network model is trained, and the Massive MIMO multi-user scheduling model based on the convolutional neural network is obtained, in the industrial Internet of Things scene under the 5G communication system, the best channel matrix of the current user to be scheduled is input into the Massive MIMO multi-user scheduling model to predict the current best scheduling user set, and the best scheduling user set is obtained. Therefore, using the deep learning method based on the convolutional neural network can greatly reduce the calculation amount of the scheduling algorithm, reduce the complexity of the traditional user scheduling, and reduce the calculation load.

[0074] In one embodiment, the base station scans the analog precoding DFT codebook in order, and combines the channel state information fed back by the user, to select the best codebook from each codebook matrix, to determine the best channel matrix of the user under the 5G communication system, comprising:

[0075] The space is divided according to the number of base station antennas N r and the number of user antennas N t , N r and N t array response vectors of different angles are written into a matrix respectively, and an analog precoding DFT codebook is constructed; the base station scans the analog precoding DFT codebook in order, and combines the channel state information fed back by the user, to select the angle from the analog precoding DFT codebook, to realize the best beamforming, to determine the best channel matrix of the user under the 5G communication system.

[0076] The space is divided according to the number of base station antennas N r and the number of user antennas N t , N r and N tThe array response vectors at different angles are written into a matrix to construct the analog precoding DFT codebook, such as Figure 3 As shown in the schematic diagram of the antenna array, in the beamforming problem, the distance between antenna elements determines the size and direction of the beam. The arrangement of the antenna elements and the beam transmission angle can be used to calculate the simulated precoding DFT codebook for beamforming at the base station. Then, based on user feedback, the optimal angle can be selected from the codebook for beamforming.

[0077] In one example, if Figure 3 As shown, Figure 3 Where θ represents the transmission beam angle of the antenna element, and d represents the distance between antenna elements. The time difference τ between adjacent elements can be deduced as:

[0078]

[0079] Where c is the speed of light, then the phase difference between adjacent arrays is for:

[0080]

[0081] Substituting ω=2πf, f is the frequency, λ is the wavelength, then the phase difference between adjacent arrays is for:

[0082]

[0083] So when considering a uniform plane array, a r (θ i,l )and They can be expressed as:

[0084]

[0085]

[0086] Among them, e is a natural constant.

[0087] According to the above analysis, the analog precoding DFT codebook can be expressed as:

[0088]

[0089]

[0090] Among them, δ in the simulated precoding DFT codebook i With ζ i Should satisfy in in An angle of a transmit codebook beam, An angle of a receive analog precoding DFT codebook beam, each element of the analog precoding DFT codebook represents a range covered by a beam.

[0091] In one embodiment, the improved particle swarm algorithm is used to analyze the network structure parameters of the convolutional neural network, determine the network structure parameters of the convolutional neural network for Massive MIMO multi-user scheduling, and construct the optimal convolutional neural network model for Massive MIMO multi-user scheduling, comprising:

[0092] Each parameter is set in advance, the value range of the network structure of the convolutional neural network is set, and N sets of network structure parameter combinations are randomly generated as N particles in the initial population P according to the value range;

[0093] N sets of network structure parameter combinations are used to construct N network models, sample data is used to train and test the N network models, and the prediction accuracy of the N network models is obtained, the prediction accuracy of the N network models is used as the fitness function, and the fitness value of each particle is calculated according to the fitness function;

[0094] The fitness value of the current position is compared with the fitness value corresponding to the historical best position, if the fitness value of the current position is higher, the current position is updated to the historical best position;

[0095] The fitness value of the current position is compared with the fitness value corresponding to the global best position, if the fitness value of the current position is higher, the current position is updated to the global best position;

[0096] The speed and position of each particle are updated according to the adaptive speed update and position update method;

[0097] Until the termination condition is reached, the optimal network structure parameter combination of the convolutional neural network is output, otherwise the step of calculating the fitness value of each particle according to the fitness function is returned;

[0098] The output convolutional neural network optimal network structure parameter combination is used as the network structure parameter of the convolutional neural network for Massive MIMO multi-user scheduling, and the optimal convolutional neural network model for Massive MIMO multi-user scheduling is constructed.

[0099] Wherein, as Figure 4The illustrated convolutional neural network structure diagram mainly comprises input images, convolution with a plurality of trainable filters and an addable bias, a plurality of feature maps generated after the convolution at the G1 layer, and then each pixel in each group of the feature maps is summed, weighted, and biased through an activation function to obtain a plurality of feature maps of the S2 layer. These maps are filtered to obtain the G3 layer. This hierarchical structure generates S4 as the S2 layer. Finally, the pixel values are rasterized and connected into a vector to input into a traditional neural network to obtain an output.

[0100] wherein, as shown in Figure 5 The flow chart for realizing the parameter optimization of the convolutional neural network structure based on the improved particle swarm algorithm is shown. Due to the complex network structure and numerous threshold parameters of the convolutional neural network (CNN), the optimization process of the convolutional neural network using the improved particle swarm algorithm can be classified into two types: structure type optimization and parameter type optimization. The structure type optimization scheme is essentially to regard the structure parameters of the convolutional neural network as different dimensions of a particle. For example, the number of convolution kernels, the size of the convolution kernel, the type of the activation function of the G1 and G3 layers in Figure 4 , and the pooling type of the S2 and S4 layers are represented by different dimensions of a particle. If the initial population is N particles, the modeling of the particle swarm algorithm should be to randomly initialize N 8-dimensional particles as a population, and the random initialization range is shown in Table 1. The prediction accuracy of the convolutional neural network after M times of training is used as the fitness function, and the speed and position are updated iteratively through the speed and position update formula to find the optimal convolutional neural network structure in the application scenario. The specific algorithm steps are as follows:

[0101] 41) Construct an N 8-dimensional particle swarm, initialize the particle swarm related parameters; randomly generate N particles according to the range in Table 1. And the mean square error value of the training results of the N convolutional neural networks is used as the adaptive function.

[0102] 42) Calculate the particle fitness function value, compare the adaptive value of its current position with the adaptive value corresponding to its historical best position, if the adaptive value of the current position is higher, update the historical best position with the current position.

[0103] 43) Calculate the particle fitness function value, compare the adaptive value of its current position with the adaptive value corresponding to its global best position, if the adaptive value of the current position is higher, update the global best position with the current position.

[0104] 44) Update the speed and position of the particle.

[0105] 45) If the algorithm meets the stopping condition, that is, the mean square error value is lower than a certain required value or the number of iterations reaches the maximum number of iterations, the particle with the global best position is the optimal network structure parameter combination. Otherwise, return to step 42 to continue searching.

[0106] Table 1. Convolutional neural network structure parameters and ranges based on the improved particle swarm optimization algorithm

[0107]

[0108]

[0109] The method of using the unimproved particle swarm optimization algorithm to optimize the network structure parameters of the convolutional neural network is mainly to represent the actual thresholds and biases of all convolution kernels in the G1 and G3 layers through different dimensions of particles. Then the particle swarm algorithm is used to search for the optimal threshold. At that time, this unimproved particle swarm optimization algorithm had a major drawback. As shown in Table 1, if the number of convolution kernels is selected to be 15 and the size of the convolution kernel is 5x5, then Figure 1 The number of thresholds and biases in the convolutional neural network shown will reach 2*15*5*5+2*15=780. Therefore, the particle dimension will also reach 780. With such a high-dimensional search, the performance of the particle swarm algorithm will be greatly reduced. As a result, it will be impossible to obtain the global optimal particle before the number of iterations reaches a certain limit, and the optimization of the convolutional neural network will be minimal. Therefore, this application chooses a structural optimization solution to optimize the convolutional neural network, which can prevent the problem of falling into local optimality in the late iterations.

[0110] In one embodiment, the adaptive speed update and position update method is expressed as:

[0111] r(t+1)=4r(t)(1-r(t)),r(0)=rand

[0112]

[0113] where ω max =0.9,ω min = 0.4, k is the current iteration number, M is the total iteration number; when hour:

[0114]

[0115] when hour:

[0116]

[0117]

[0118] Among them, k is the current iteration number, M is the total iteration number, is the current speed, is the current position, d = 1, 2, ..., D, i = 1, 2, ...n, c1 and c2 are the social learning factor and self-learning factor of the particles in the population respectively, and r1 and r2 are random parameters between 0 and 1.

[0119] In one embodiment, the optimal channel matrix of a user in a 5G communication system is used as a training sample to train the optimal convolutional neural network model to obtain a Massive MIMO multi-user scheduling model based on a convolutional neural network, including:

[0120] Through an exhaustive algorithm, the optimal channel matrix of users in the 5G communication system is obtained as a training sample set. The optimal convolutional neural network model is trained offline using the training sample set until the optimal convolutional neural network model meets the convergence conditions, thereby obtaining a Massive MIMO multi-user scheduling model based on the convolutional neural network.

[0121] The core of this application is to improve the network structure design of the convolutional neural network and use the convolutional neural network to establish a Massive MIMO multi-user scheduling model based on the convolutional neural network at the base station end based on the improved convolutional neural network. By inputting the optimal channel matrix into the Massive MIMO multi-user scheduling model, the optimal scheduling user set is predicted. In order to solve the prediction accuracy problem of the Massive MIMO multi-user scheduling model, the improved particle swarm algorithm (IPSO) can be used to optimize the network result parameters. The Massive MIMO multi-user scheduling model is used as the particle dimension in the improved particle swarm algorithm. Through continuous position update and population iteration of particles, the optimal population, that is, the optimal convolutional neural network structure, is obtained, thereby realizing a Massive MIMO multi-user scheduling model based on the improved particle swarm algorithm. In order to solve the problem that the particle swarm optimization algorithm is prone to falling into local optimality, an adaptive inertia weight and a position update scheme that changes with the number of iterations are designed to prevent the particles from becoming more and more similar in the late convergence stage of the algorithm, resulting in falling into local optimality.

[0122] The Massive MIMO multi-user scheduling model based on the improved particle swarm optimization algorithm established by the above method is suitable for predicting the optimal scheduling set in Massive MIMO industrial Internet of Things multi-user scheduling scenarios. Compared with traditional scheduling schemes, this method predicts the optimal scheduling set entirely based on the matrix characteristics of the user channel matrix, without the need for complex SVD calculations and achievable rate calculations. While getting rid of the coupling between user rate and user set, it reduces the computational complexity of multi-user scheduling and improves the speed of multi-user scheduling in industrial Internet of Things scenarios.

[0123] The above-mentioned Massive MIMO multi-user scheduling method based on the improved convolutional neural network processes the characteristics of the matrix, takes the user channel matrix in the massive MIMO as the input of the convolutional neural network, and takes the optimal scheduling user set as the output. Since the convolutional neural network (CNN) can predict the optimal user combination through the channel state matrix, the coupling of the user rate and the user set can be broken, and therefore the use of the deep learning method based on the convolutional neural network can greatly reduce the calculation amount of the scheduling algorithm and reduce the complexity of the traditional user scheduling.

[0124] Further, the network structure parameters of the convolutional neural network are taken as the dimensions of the particles in the Particle Swarm Optimization (PSO) algorithm, and the optimal network parameter model is obtained through the iteration optimization of the improved particle swarm algorithm, so as to improve the prediction accuracy and training speed of the network. The improved particle swarm algorithm is based on the self-adaptation of the particle swarm algorithm weight of the nonlinear function, can balance the global search ability and the local search ability, and improves the search efficiency of the particle swarm algorithm; a new population updating method prevents falling into local optimum in the late iteration.

[0125] It should be understood that, although Figure 1 The steps in the flowchart of the above-mentioned method are displayed in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this article, the execution of these steps has no strict order limitation, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the above-mentioned method can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or sub-steps or stages of other steps.

[0126] The technical features of the above-mentioned embodiments can be combined in any way. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the description.

[0127] The above-mentioned embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A Massive MIMO multi-user scheduling method based on improved convolutional neural network, characterized in that, The method comprises: Step 1, the base station scans the analog precoding DFT codebook in sequence, and selects the best codebook from each codebook matrix in combination with the channel state information fed back by the user, to determine the best channel matrix of the user under the 5G communication system; Step 2, the improved particle swarm algorithm is used to analyze the network structure parameters of the convolutional neural network, determine the network structure parameters of the convolutional neural network for Massive MIMO multi-user scheduling, and construct the best convolutional neural network model for Massive MIMO multi-user scheduling; Step 3, the best channel matrix of the user under the 5G communication system is used as a training sample to train the best convolutional neural network model, and a Massive MIMO multi-user scheduling model based on convolutional neural network is obtained; Step 4, in the industrial Internet of Things scene under the 5G communication system, the best channel matrix of the current user to be scheduled is obtained by using the method of step 1, and the current best scheduling user set is predicted by inputting the Massive MIMO multi-user scheduling model, to obtain the best scheduling user set.

2. The method of claim 1, wherein, The base station scans the analog precoding DFT codebook in sequence, and selects the best codebook from each codebook matrix in combination with the channel state information fed back by the user, to determine the best channel matrix of the user under the 5G communication system, comprising: The space is divided according to the number N of base station antennas r and the number N of user antennas t , and N r and N t array response vectors of different angles are written into a matrix respectively to construct a simulation precoding DFT codebook; The base station scans the analog precoding DFT codebook in sequence, and selects the best codebook from each codebook matrix in combination with the channel state information fed back by the user, to determine the best channel matrix of the user under the 5G communication system, comprising:

3. The method of claim 1, wherein, The improved particle swarm algorithm is used to analyze the network structure parameters of the convolutional neural network, determine the network structure parameters of the convolutional neural network for Massive MIMO multi-user scheduling, and construct the best convolutional neural network model for Massive MIMO multi-user scheduling, comprising: Each parameter is set in advance, the value range of the network structure of the convolutional neural network is set, and N sets of network structure parameter combinations are randomly generated as N particles in the initial population P according to the value range; N network models are constructed by using N sets of network structure parameter combinations, sample data is used to train and test N network models, the prediction accuracy of N network models is obtained, the prediction accuracy of N network models is used as the fitness function, and the fitness value of each particle is calculated according to the fitness function; The fitness value of the current position is compared with the fitness value corresponding to the historical best position, if the fitness value of the current position is higher, the current position is used to update the historical best position; The fitness value of the current position is compared with the fitness value corresponding to the global best position, if the fitness value of the current position is higher, the current position is used to update the global best position; The speed and position of each particle are updated according to the adaptive speed update and position update method; Until the termination condition is reached, the best network structure parameter combination of the convolutional neural network is output, otherwise the step of calculating the fitness value of each particle according to the fitness function is returned. The output convolutional neural network optimal network structure parameter combination is taken as a network structure parameter of a convolutional neural network for massive MIMO multi-user scheduling, so as to construct an optimal convolutional neural network model for massive MIMO multi-user scheduling.

4. The method of claim 3, wherein, The expression of the adaptive speed and position updating method is: r(t+1) = 4r(t)(1-r(t)), r(0) = rand where ω is the inertia weight, ω max = 0.9, ω min = 0.4, k is the current iteration number, M is the total iteration number, rand is a random number between 0 and 1, and r(t) and r(t+1) represent the corresponding values of the chaotic map at times t and t+1, respectively. When time: When Time: wherein, is the current velocity, is the current position, d = 1, 2,..., D, i = 1, 2,..., N, D is the dimension of the particle, c1 and c2 are the social learning factor and self-learning factor of the particle in the population, respectively, and r1 and r2 are random parameters between 0 and 1, is the current individual extremum, is the current group extremum, is the current group best position.

5. The method of claim 1, characterized in that, The optimal channel matrix of the user under the 5G communication system is taken as a training sample to train the optimal convolutional neural network model, so as to obtain a massive MIMO multi-user scheduling model based on a convolutional neural network, which comprises the following steps: The optimal channel matrix of the user under the 5G communication system is obtained as a training sample set through an exhaustive algorithm; The training sample set is used to perform offline training on the optimal convolutional neural network model until the optimal convolutional neural network model meets a convergence condition, so as to obtain a massive MIMO multi-user scheduling model based on a convolutional neural network.

Images