Large-scale MIMO simulation domain beam selection method, apparatus and device, and storage medium
By using a convolutional neural network model and differential evolution algorithm to construct the network structure in a large-scale MIMO system, the target simulation precoding matrix is quickly selected, solving the problem of low beam selection efficiency in the existing technology and achieving efficient beam selection and improved accuracy.
Patent Information
- Application Number
- CN202410823229.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-24
- Publication Date
- 2024-11-26
AI Technical Summary
In existing technologies, beam selection methods based on exhaustive search are inefficient in large-scale MIMO systems, especially in millimeter-wave large-scale antenna systems where the complexity is too high, resulting in low beam selection efficiency.
By determining multiple sets of simulated precoding matrices and equivalent channel matrices, a network structure is constructed using a convolutional neural network model based on the differential evolution algorithm to quickly select the target simulated precoding matrix. Combined with the circular arc function to control the mutation operator and crossover operator, the beam selection efficiency is improved.
This method enables rapid determination of the target simulation precoding matrix in large-scale MIMO systems, improves beam selection efficiency, enhances the prediction accuracy of convolutional neural network models, and reduces system complexity.
Smart Images

Figure CN119030573A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present application relate to the field of communication technology, and in particular to a large-scale MIMO analog domain beam selection method, apparatus, device and storage medium. Background Art
[0002] The analog domain precoding technology based on fixed codebook is mainly implemented by using multi-antenna arrays. Multi-antenna array technology forms different analog domain precoding matrices by controlling the phase weight of each antenna element, thereby generating analog domain beams of different angles. However, in actual communication systems, as the user position is fixed, the transmission direction between the base station and the user becomes certain, which means that the transmitter only needs to use a beam in a specific direction for transmission, and the receiver uses a beam in a specific direction for reception to meet the communication needs of the system. Therefore, when multiple analog domain beams are formed at both the transmitting and receiving ends, in order to reduce the power loss of the transmitting end and improve the transmission efficiency of the system, it is particularly important to select the best analog domain beam pair (that is, determine a pair of optimal analog domain precoding matrices) for transmission.
[0003] In the prior art, the main beam selection method is to scan the analog domain codebook, that is, to determine the best beam pair by exhaustive search. Exhaustive search is a brute force beam search mechanism that determines the optimal transmission beam pair by global search of all beam pair combinations. Figure 1 is a schematic diagram of the exhaustive search process in the prior art, such as Figure 1 As shown in , assuming that the serving base station b contains M beams and the user i has N receiving beams, then the beam pair combinations are M*N combinations. The exhaustive search process is as follows Figure 1 As shown, the base station uses its M transmission beams in turn to repeatedly send reference signals to user i; for each transmission beam, base station b will send reference signals to each receiving beam of user i in different time slots in turn, and then user i selects the best receiving beam under the current transmission beam according to a certain beam selection criterion (such as the maximum signal-to-noise ratio criterion) and feeds it back to its serving base station; after scanning all the transmission beams, that is, after traversing all the beam pairs, an optimal beam pair can be determined for transmission.
[0004] For the full-traversal exhaustive beam search algorithm, its performance is seriously affected by the number of antenna elements and the number of analog domain beams; when the antenna elements are small and the number of analog domain beams is small, this search algorithm has a faster search speed, lower complexity, and can achieve the best system performance. However, for millimeter-wave massive antenna systems, due to the large number of analog domain beam pairs at both ends of the transmitter and receiver, the full-traversal beam search is bound to bring higher complexity, making the beam selection inefficient. Summary of the invention
[0005] The embodiments of the present application provide a large-scale MIMO analog domain beam selection method, apparatus, device and storage medium, which are helpful to improve the beam selection efficiency in a large-scale MIMO system.
[0006] In order to solve the above problems, in a first aspect, an embodiment of the present application provides a large-scale MIMO analog domain beam selection method, including:
[0007] Determine multiple groups of simulated precoding matrices according to the transmit antenna array information of the target base station and the receive antenna array information of the user, each group of simulated precoding matrices including a receive-end simulated precoding matrix and a transmit-end simulated precoding matrix;
[0008] Determine a plurality of equivalent channel matrices according to the plurality of groups of simulated precoding matrices and the target channel model, each of the equivalent channel matrices corresponding to a group of the simulated precoding matrices;
[0009] The multiple equivalent channel matrices are input into a trained convolutional neural network model, and the convolutional neural network model selects a target simulated precoding matrix from the multiple equivalent channel matrices based on a target beam selection criterion. The network structure of the convolutional neural network model is determined based on a differential evolution algorithm, and the mutation operator in the differential evolution algorithm is controlled based on an arc function.
[0010] In a second aspect, an embodiment of the present application provides a large-scale MIMO analog domain beam selection device, including:
[0011] The analog codebook determination module is used to determine multiple groups of analog precoding matrices according to the transmitting antenna array information of the target base station and the receiving antenna array information of the user, each group of the analog precoding matrices includes a receiving end analog precoding matrix and a transmitting end analog precoding matrix;
[0012] An equivalent channel determination module, used to determine a plurality of equivalent channel matrices according to a plurality of groups of simulated precoding matrices and a target channel model, each of the equivalent channel matrices corresponding to a group of the simulated precoding matrices;
[0013] A beam selection module is used to input the multiple equivalent channel matrices into a trained convolutional neural network model, and select a target simulated precoding matrix from the multiple equivalent channel matrices based on a target beam selection criterion through the convolutional neural network model, wherein the network structure of the convolutional neural network model is determined based on a differential evolution algorithm, and the mutation operator in the differential evolution algorithm is controlled based on an arc function.
[0014] In a third aspect, an embodiment of the present application further provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the large-scale MIMO analog domain beam selection method described in the embodiment of the present application is implemented.
[0015] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the large-scale MIMO analog domain beam selection method disclosed in an embodiment of the present application.
[0016] The large-scale MIMO analog domain beam selection method, apparatus, device and storage medium provided in the embodiments of the present application determine multiple groups of analog precoding matrices based on the transmitting antenna array information of the target base station and the receiving antenna array information of the user, determine multiple equivalent channel matrices based on the multiple groups of analog precoding matrices and the target channel model, and select the target analog precoding matrix from the multiple equivalent channel matrices based on the target beam selection criterion through a convolutional neural network model. For millimeter-wave large-scale antenna systems, the target analog precoding matrix can also be quickly determined through the convolutional neural network model, thereby improving the beam selection efficiency in the large-scale MIMO system. Moreover, since the network structure of the convolutional neural network model is determined based on the differential evolution algorithm, the construction of the convolutional neural network structure is more intelligent, and the prediction accuracy of the convolutional neural network model can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0018] Figure 1 is a schematic diagram of an exhaustive search process in the prior art;
[0019] Figure 2 It is a flowchart of a large-scale MIMO analog domain beam selection method provided in an embodiment of the present application;
[0020] Figure 3 This is a schematic diagram of a service process from a base station to a user in a 5G communication scenario in an embodiment of the present application;
[0021] Figure 4 This is a schematic diagram of the arrangement of antenna elements in an embodiment of the present application;
[0022] Figure 5It is a flowchart for determining the network structure and training of a convolutional neural network in an embodiment of the present application;
[0023] Figure 6 It is a schematic diagram of the structure of a convolutional neural network in an embodiment of the present application;
[0024] Figure 7 is a flow chart of selecting the best beam pair through a convolutional neural network model in an embodiment of the present application;
[0025] Figure 8 It is a structural schematic diagram of a large-scale MIMO analog domain beam selection device provided in an embodiment of the present application;
[0026] Fig. 9 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0028] Figure 2 is a flowchart of a large-scale MIMO simulation domain beam selection method provided in an embodiment of the present application, such as Figure 2 As shown, the method includes: step 210 to step 230.
[0029] Step 210, determining multiple groups of simulated precoding matrices according to the transmitting antenna array information of the target base station and the receiving antenna array information of the user, each group of the simulated precoding matrices includes a receiving end simulated precoding matrix and a transmitting end simulated precoding matrix.
[0030] Each group of simulated precoding matrices corresponds to one user.
[0031] Construct a Massive MIMO application scenario model from the target base station to the user. The Massive MIMO application scenario model from the target base station to the user includes a scenario where one base station serves multiple users. The base station and the user end use hybrid precoding to send and receive data streams, respectively, to form Massive MIMO multi-user hybrid precoding. Among them, Massive MIMO (Massive Multiple-Input Multiple-Output, large-scale dense multiple-input multiple-output antenna array) is a core key technology of 5G. Compared with traditional MIMO, the spatial resolution of Massive MIMO is significantly enhanced compared with the existing MIMO, and it can deeply mine spatial dimension resources, so that multiple users in the network can use the spatial freedom provided by Massive MIMO to communicate with the base station at the same time and frequency resources, thereby greatly improving the spectrum efficiency without increasing the base station density and bandwidth. Massive MIMO can use beamforming technology to provide precise service range for serving users.
[0032] For a communication system model, such as a 5G communication system model, a communication scenario is established from the base station side to the user. Figure 3 This is a schematic diagram of a service process between a base station and a user in a 5G communication scenario in an embodiment of the present application, such as Figure 3 As shown in the figure, the base station first sends the data stream to the antenna through the digital precoding matrix and the analog precoding matrix to realize beamforming, and then transmits it to the user through millimeter wave for service. Figure 3 In the system model, the downlink based on Massive MIMO multi-user hybrid precoding system is considered, and the base station is configured with N t antennas, each user terminal is equipped with N r antennas, the number of transmission streams is N s Consider the case of multi-user precoding, where each user receives only one data stream, i.e. The RF front end of the base station adopts a hybrid connection mode and is equipped with RF links are connected to a uniform linear array (ULA). Assume that there are N u candidate users, of which only N can be selected k The purpose of beam selection in the embodiment of the present application is to select N u Select N from the candidate users k Serving users.
[0033] Since Massive MIMO has too many antennas, it is often impossible to deploy an RF chain for each antenna in actual deployment. Therefore, Massive MIMO often uses hybrid precoding for beamforming operations.
[0034] In the embodiment of the present application, when beam selection is performed, beam selection can be performed in the analog domain, so it is necessary to determine the analog precoding matrix.
[0035] For the hybrid precoding structure in the 5G communication system model, Figure 4 Schematic diagram of antenna array arrangement in the embodiment of the present application. Figure 4 As shown, θ represents the antenna element (such as Figure 4 Where elm0, elm1, …, elmN) are the transmission beam angles, and d is the distance between adjacent antenna elements. The time difference between adjacent elements can be deduced as:
[0036]
[0037] Among them, τ is the time difference between adjacent arrays, and c is the speed of light.
[0038] The phase difference between adjacent antenna elements is:
[0039]
[0040] in, Represents the phase difference between adjacent antenna elements.
[0041] Let ω = 2πf, Substituting into the above formula, the phase difference between adjacent arrays is obtained as:
[0042]
[0043] Wherein, λ represents the wavelength of the RF signal, and f represents the frequency of the RF signal.
[0044] When the antenna array is a uniform linear array (ULA), the receiving antenna array a r (θ i,l ) and the transmitting antenna array They can be expressed as:
[0045]
[0046] Among them, θ i,l represents the receiving beam angle of the antenna element in the receiving antenna array of the i-th user, represents the transmit beam angle of the antenna element in the transmit antenna array for the i-th user.
[0047] According to the above analysis, the analog precoding DFT codebook, i.e., a set of analog precoding matrices, can be expressed as:
[0048]
[0049] Among them, δ in the codebook i satisfy in N r Indicates the number of antennas at the receiving end. i satisfy in N t Indicates the number of transmitting antennas. Each element in the codebook represents the range covered by the beam. represents the simulated precoding matrix at the receiving end, Indicates the number of RF links at the receiving end, represents the simulated precoding matrix at the transmitter, Indicates the number of RF links on the transmitting side.
[0050] Step 220: Determine a plurality of equivalent channel matrices according to the plurality of groups of simulated precoding matrices and the target channel model, each of the equivalent channel matrices corresponding to a group of the simulated precoding matrices.
[0051] After obtaining the simulated precoding matrix through the above codebook method, the equivalent channel matrix between the target base station and the m users is determined respectively, as shown in the following formula:
[0052]
[0053] Among them, H eff,i Represents the equivalent channel matrix between the i-th user and the target base station, and its dimension size is represents the simulated precoding matrix of the receiving end of the i-th user, H i represents the channel matrix of the target channel model, represents the simulated precoding matrix at the transmitting end of the i-th user.
[0054] Step 230, input the multiple equivalent channel matrices into the trained convolutional neural network model, and select the target simulated precoding matrix from the multiple equivalent channel matrices based on the target selection criterion through the convolutional neural network model, the network structure of the convolutional neural network model is determined based on the differential evolution algorithm, and the mutation operator in the differential evolution algorithm is controlled based on the arc function.
[0055] The network structure of the convolutional neural network can be determined by the differential evolution algorithm. When the network structure of the convolutional neural network is determined by the differential evolution algorithm, the mutation operator adopts the mutation operator controlled by the arc function. The mutation operator and the crossover operator are controlled by the arc function, so that the mutation operator and the crossover operator can be dynamically adjusted with the iteration process. Among them, the differential evolution algorithm (Differential Evolution Algorithm, DE algorithm) is one of the intelligent optimization algorithms. Like the genetic algorithm, the differential evolution algorithm is also an optimization algorithm based on modern intelligence theory. It guides the direction of optimization search through the group intelligence generated by the mutual cooperation and competition between individuals in the group. The basic idea of the differential evolution algorithm is: starting from a randomly generated initial population, a new individual is generated by summing the vector difference between any two individuals in the population and the third individual, and then the new individual is compared with the corresponding individual in the contemporary population. If the fitness of the new individual is better than the fitness of the current individual, the new individual is used to replace the old individual in the next generation, otherwise the old individual is still preserved. Through continuous evolution, excellent individuals are retained and inferior individuals are eliminated, guiding the search to approach the optimal solution. In view of the characteristics of the DE algorithm, the CNN (Convolutional Neural Network) network structure selection can be modeled as an optimization problem, and then solved by the DE algorithm to obtain the best CNN model structure in the same depth. CNN is a type of feedforward neural network that includes convolution calculations and has a deep structure. It is commonly used in image processing and face recognition. Convolutional neural networks have representation learning capabilities and can classify input information in a translation-invariant manner according to their hierarchical structure. The embodiment of the present application improves CNN based on the DE algorithm to obtain ICNN (Improved Convolutional Neural Network). In terms of model prediction accuracy, ICNN has higher prediction accuracy than neural networks of the same depth.
[0056] When the convolutional neural network is trained, an exhaustive search method can be used to determine the optimal beam pair based on the target beam selection criterion, and then used as a data set for convolutional neural network training, so that the trained convolutional neural network model can predict the optimal analog precoding matrix from multiple sets of equivalent channel matrices.
[0057] The target base station and the multiple equivalent channel matrices of the corresponding users are respectively input into the trained convolutional neural network model, and the equivalent channel matrices are processed by the convolutional neural network model based on the target beam selection criterion to predict the matching probability corresponding to each equivalent channel matrix, and a preset number of equivalent channel matrices with higher matching probability are selected from multiple equivalent channel matrices, and the analog precoding matrices corresponding to these equivalent channel matrices are used as target analog precoding matrices, and a group of target analog precoding matrices is a target beam pair. Among them, the target beam selection criterion can be, for example, a maximum signal-to-noise ratio criterion or a maximum signal-to-interference-plus-noise ratio criterion, and the appropriate beam selection criterion can be selected according to different scenarios.
[0058] The large-scale MIMO analog domain beam selection method provided in the embodiment of the present application determines multiple groups of analog precoding matrices based on the transmitting antenna array information of the target base station and the receiving antenna array information of the user, determines multiple equivalent channel matrices based on the multiple groups of analog precoding matrices and the target channel model, and selects the target analog precoding matrix from the multiple equivalent channel matrices based on the target beam selection criterion through the convolutional neural network model. For the millimeter wave large-scale antenna system, the target analog precoding matrix can also be quickly determined through the convolutional neural network model, thereby improving the beam selection efficiency in the large-scale MIMO system. Moreover, since the network structure of the convolutional neural network model is determined based on the differential evolution algorithm, the construction of the convolutional neural network structure is more intelligent, and the prediction accuracy of the convolutional neural network model can be improved.
[0059] Before using the convolutional neural network model to perform beam selection, it is necessary to determine the network structure of the convolutional neural network based on the differential evolution operator and train the convolutional neural network after the network structure is determined. Figure 5 is a flowchart for determining the network structure and training of a convolutional neural network in an embodiment of the present application, such as Figure 5 As shown, the process of determining the network structure and training of the convolutional neural network includes steps 510 to 570.
[0060] Step 510, generating an initial population according to the value range of the hyperparameters of the convolutional neural network, and using the initial population as the current population.
[0061] Each individual in the initial population is a hyperparameter combination of the convolutional neural network, and the hyperparameter is a structural parameter of the convolutional neural network. The hyperparameter combination may include convolution kernel size, number of convolution layers, number of fully connected layers, etc. The hyperparameter value range is the value range of each hyperparameter included in a preset hyperparameter combination.
[0062] Get the relevant parameters of the differential evolution algorithm, including the maximum number of iterations L, population size N, mutation operator F, crossover operator CR, etc. The population size is the total number of individuals in the population.
[0063] According to the hyperparameter value range of the convolutional neural network, a hyperparameter is randomly determined from the value range of each hyperparameter to obtain a hyperparameter combination. In this way, multiple hyperparameter combinations are determined respectively. When the number of determined hyperparameter combinations reaches the set population size, these hyperparameter combinations form the initial population. The initial population is first used as the current population, and the following steps are iterated.
[0064] Step 520, determining the fitness function value of each individual in the current population, and determining the individual with the largest fitness function value as the optimal individual of the current population.
[0065] Among them, the fitness function value is the prediction accuracy of the convolutional neural network determined by the individual for the selection of a sample simulated precoding matrix, and the sample simulated precoding matrix includes multiple groups of simulated precoding matrices, each group of simulated precoding matrices includes a receiving end simulated precoding matrix and a transmitting end simulated precoding matrix.
[0066] Convolutional neural networks are constructed for each individual, and the convolutional neural network is trained based on a training data set consisting of multiple sets of simulated precoding matrices and corresponding optimal beam pairs. After the training is completed, a test data set (consisting of sample simulated precoding matrices, and the test data set also includes the labeled data corresponding to each sample simulated precoding matrix, and the labeled data is the optimal beam pair in the sample simulated precoding matrix) is used to determine the prediction accuracy of the convolutional neural network. The prediction accuracy is to determine the fitness function value of the individual of the convolutional neural network, and the fitness function value of each individual in the current population is determined in this way. The fitness function value of each individual in the initial population is compared, and the individual with the largest fitness function value is taken as the optimal individual of the current population.
[0067] Step 530: According to the optimal individual, a mutation operation is performed on the current population based on a mutation operator controlled by an arc function to obtain a temporary solution.
[0068] Combined with the best individual in the current population, the individuals in the current population are mutated based on the mutation operator to obtain a temporary solution. The temporary solution is also a population, and the population size is the set population size.
[0069] In one embodiment of the present application, the mutation operation is performed on the current population based on the optimal individual and a mutation operator controlled by an arc function to obtain a temporary solution, including: determining the mutation operator based on the current number of iterations and the maximum number of iterations; randomly selecting an odd number of individuals from the current population, and taking one of the odd number of individuals as a target individual; performing mutation operations on the target individual and the optimal individual based on the mutation operator and the odd number of individuals to obtain an individual in the temporary solution; and cyclically executing the operations of randomly selecting an odd number of individuals from the current population and generating an individual in the temporary solution until the number of individuals in the temporary solution reaches the population size.
[0070] The embodiment of the present application combines the common mutation mechanisms DE / best / 2 and DE / rand / 2 of the differential evolution algorithm, so that the mutation mechanism focuses more on random individuals x in the early stage of the search. r1 (g) Maintain population diversity and focus on the current optimal individual x in the later stage of the search best (g) Accelerate the convergence of the algorithm. At the same time, the mutation operator is dynamically adjusted based on the increase of the current iteration number, so that the mutation operator decreases with the increase of the current iteration number, changing in an arc shape.
[0071] When performing a mutation operation, randomly select an odd number of individuals from the current population (such as 3 individuals or 5 individuals, etc.), and use one of them as a target individual, which is used to determine a basis vector in the mutation operation; determine a basis vector in the mutation operation based on the target individual and the optimal individual, determine a differential vector based on the individuals other than the target individual among the odd number of individuals, and determine an individual in a temporary solution based on the basis vector and the differential vector; repeat the above operations of randomly selecting an odd number of individuals from the current population and generating an individual in a temporary solution until the number of individuals in the temporary solution reaches the population size.
[0072] The mutation operation is controlled by the mutation operator based on arc control, so that the mutation mechanism focuses more on random individuals in the early stage of the search to maintain population diversity, and focuses on the current optimal individual in the later stage of the search to accelerate the convergence of the algorithm.
[0073] In one embodiment of the present application, the mutation operation is performed on the target individual and the optimal individual according to the mutation operator and the odd number of individuals to obtain an individual in the temporary solution, including: determining a first weight of the target individual and a second weight of the optimal individual according to the mutation operator, the sum of the first weight and the second weight being 1; performing weighted summation of the target individual and the optimal individual according to the first weight and the second weight to obtain a basis vector of the mutation operation; determining the differential vectors of the individuals other than the target individual in the odd number of individuals according to the mutation operator; and taking the sum of the basis vector and the differential vector as an individual in the temporary solution.
[0074] 1 / 2 of the mutation operator can be determined as the first weight of the target individual, and the second weight can be determined based on the first weight; weighted sum is performed on the target individual and the optimal individual to obtain a basis vector for the mutation operation; every two individuals other than the target individual in an odd number of individuals are taken as a group, and differences are made on each group of other individuals to obtain a difference vector of the other individuals in the group; the mutation operator is used as the weight of each difference vector, and weighted sum is performed on each difference vector to obtain a difference vector; the basis vector and the difference vector are summed to obtain an individual in the temporary solution.
[0075] In summary, taking the odd number of individuals as 5 individuals as an example, the mutation operation can be expressed as:
[0076] v i (g) = γ·x r1 (g)+(1-γ)·x best (g)+F·[x r2 (g)-x r3 (g)]+F·[x r4 (g)-x r5 (g)]
[0077] Among them, g represents the current iteration number, v i (g) represents an individual in the temporary solution obtained after performing the mutation operation, γ represents the first weight, F represents the mutation operator, x r1 (g), x r2 (g), x r3 (g), x r4 (g), x r5 (g) is a random individual in the current population, x best (g) is the optimal individual of the current population, x best (g), x r1 (g), x r2 (g), x r3 (g), x r4 (g), xr5 (g) are different from each other.
[0078] In one embodiment of the present application, determining the mutation operator according to the current number of iterations and the maximum number of iterations includes:
[0079] According to the current number of iterations and the maximum number of iterations, the mutation operator is determined according to the following formula:
[0080]
[0081] Wherein, F represents the mutation operator, g represents the current iteration number, and G represents the maximum iteration number. F∈[0,2]. The maximum iteration number is the set termination iteration number.
[0082] Step 540: Perform a crossover operation on the temporary solution and the current population to obtain a new solution.
[0083] A crossover operation is performed on the individuals in the temporary solution and the individuals in the current population, so that part of the hyperparameters of an individual in the temporary solution and another part of the hyperparameters of an individual in the current population form a new individual, which can be processed separately in the order of the individuals to obtain new individuals of the population size, and the new individuals of this population size constitute a new solution.
[0084] In one embodiment of the present application, performing a crossover operation on the temporary solution and the current population to obtain a new solution includes:
[0085] Determine the crossover operator according to the current number of iterations and the maximum number of iterations;
[0086] According to the following formula, a crossover operation is performed on the temporary solution and the current population based on the crossover operator to obtain a new solution:
[0087]
[0088] Among them, g represents the current iteration number, u ij (g) represents the jth hyperparameter in the i-th individual in the new solution, v ij (g) represents the jth hyperparameter in the i-th individual in the temporary solution, x ij (g) represents the jth hyperparameter in the i-th individual in the current population, randb(j) represents the j-th random number generated between [0, 1], rnbr(i)∈(1, 2, 3…d) represents a randomly selected sequence, d represents the dimension of each individual, and CR represents the crossover operator.
[0089] From the above crossover operation formula, it can be seen that when a crossover operation is performed on an individual in the temporary solution and an individual in the current population, for example, a crossover operation is performed on the i-th individual, multiple random numbers between [0, 1] are generated, and the number of random numbers generated is the dimension of each individual, and the dimension of an individual is the number of hyperparameters in a hyperparameter combination; if the j-th random number among the multiple random numbers is less than or equal to the crossover operator or is equal to the value selected in the random selection sequence, then the j-th hyperparameter u of the i-th individual in the new solution is ij (g) Take the jth hyperparameter v in the i-th individual in the temporary solution ij (g); If the jth random number among the multiple random numbers is greater than the crossover operator or is not equal to the value selected in the random selection sequence, then the jth hyperparameter u of the i-th individual in the new solution ij (g) Take the jth hyperparameter x of the i-th individual in the current population ij (g).
[0090] By randomly selecting the sequence rnbr(i)∈(1, 2, 3…d), we can ensure that the individual u in the new solution ij (g) from v ij (g) At least one parameter is obtained. The crossover operator CR∈[0,1]. From the crossover operation formula, we can see that the value of CR affects the diversity and convergence speed of the population. The larger the CR, the greater the probability that randb(j)≤CR. The individual u in the new solution i The parameters in (g) are determined by the individual v in the temporary solution i (g) The more it provides, the more conducive it is to local search and accelerate the convergence rate. The smaller the CR is, the opposite is true, that is, the individual u in the new solution i The parameters in (g) are determined by the individual v in the temporary solution i (g) provides less, and the individual x in the current population i (g) The more it provides, the greater the population diversity will be, which is more conducive to global search.
[0091] In one embodiment of the present application, determining the crossover operator according to the current number of iterations and the maximum number of iterations includes:
[0092] According to the current number of iterations and the maximum number of iterations, the crossover operator is determined according to the following formula:
[0093]
[0094] Wherein, CR represents the crossover operator, g represents the current number of iterations, and G represents the maximum number of iterations.
[0095] The crossover operator is also controlled by an arc function. The crossover operator is dynamically adjusted based on the increase in the current number of iterations, so that the crossover operator increases with the increase in the current number of iterations, changing in an arc shape. The maximum value of the crossover operator is 1, that is, CR∈[0,1]. Through this crossover operator, the CR is large in the early iteration of the differential evolution algorithm, and the individual u in the new solution is i The parameters in (g) are determined by the individual v in the temporary solution i (g) The more it provides, the more conducive it is to local search and accelerate the convergence rate. In the later iteration of the differential evolution algorithm, the CR is small, and the individual u in the new solution i The parameters in (g) are determined by the individual v in the temporary solution i (g) provides less, and the current population of individuals x i (g) The more it provides, the greater the population diversity will be, which is more conducive to global search.
[0096] Step 550: Determine the next generation population based on the current population and the new solution.
[0097] Individuals of the same population size are selected from the current population and the new solution to form the next generation population.
[0098] In one embodiment of the present application, determining the next generation population based on the current population and the new solution includes: determining the fitness function value of each individual in the new solution; generating a temporary next generation population based on the fitness function value of each individual in the current population and the fitness function value of each individual in the new solution; when the individual fitness variance of the temporary next generation population is less than or equal to the variance threshold, randomly generating a preset number of new individuals, and using the new individuals to replace a preset number of individuals in the temporary next generation population to obtain the next generation population.
[0099] Among them, the individual fitness variance is the variance of the fitness function values of all individuals in a population. Assume that the population size, that is, the total number of individuals in the population, is N, and the fitness function value of the i-th individual is f i (g), the average fitness function value of the population is f average (g), then define the individual fitness variance δ of the population 2 for:
[0100]
[0101] Where N is the population size, f i (g) is the fitness function value of the i-th individual, f average (g) is the average fitness function value of the population.
[0102] A convolutional neural network is constructed with each individual in the new solution respectively, and the constructed convolutional neural network is trained based on a training data set consisting of multiple groups of simulated precoding matrices and corresponding optimal beam pairs. After the training is completed, a test data set (consisting of sample simulated precoding matrices, and the test data set also includes labeled data corresponding to each sample simulated precoding matrix, and the labeled data is the optimal beam pair in the sample simulated precoding matrix) is used to determine the prediction accuracy of the convolutional neural network. The prediction accuracy is used to determine the fitness function value of the individual of the convolutional neural network. The fitness function value of each individual in the new solution is determined in this way.
[0103] According to the fitness function value of each individual in the current population and the fitness function value of each individual in the new solution, individuals of the population size are selected from the current population and the new solution as the temporary next-generation population; the individual fitness variance of the temporary next-generation population is determined. When the individual fitness variance of the temporary next-generation population is less than or equal to the variance threshold, it means that the temporary next-generation population has fallen into a local optimum. At this time, a preset number of new individuals can be randomly generated based on the hyperparameter value range, and the randomly generated new individuals are used to replace the preset number of individuals in the temporary next-generation population, and the replaced temporary next-generation population is used as the next-generation population; if the individual fitness variance of the temporary next-generation population is greater than the variance threshold, it means that the temporary next-generation population has not fallen into a local optimum. At this time, the temporary next-generation population can be used as the next-generation population.
[0104] By replacing individuals in the temporary next-generation population with randomly generated individuals when the fitness variance of the temporary next-generation population is less than or equal to the variance threshold, it is possible to avoid falling into the local optimum during the iteration process, thereby greatly reducing the probability of falling into the local optimal solution. Even after falling into the local optimum, it is possible to jump out of the local optimum by replacing new individuals and further find the global optimal solution.
[0105] In one embodiment of the present application, generating a temporary next generation population according to the fitness function value of each individual in the current population and the fitness function value of each individual in the new solution includes:
[0106] In the order of fitness function values from large to small, individuals of the population size are selected from the current population and the new solution as the temporary next generation population.
[0107] According to the fitness function value of each individual from large to small, the current population x i The individuals and new solutions u in (g) i (g) are sorted, and the individuals of the population size in the front position are selected, that is, the individuals of the population size with the largest fitness function value are selected to form the temporary next generation population t i (g+1).
[0108] By selecting individuals from the current population and the new solution in order of fitness function values from large to small to form a temporary next generation population, the iterative process is brought closer to the optimal solution, facilitating rapid convergence.
[0109] Step 560, determining whether the next generation population satisfies the iteration termination condition.
[0110] When the next generation population does not satisfy the iteration termination condition, the next generation population is used as the current population, and steps 520 to 560 are iteratively executed; when the next generation population satisfies the iteration termination condition, step 570 is executed.
[0111] The iteration termination condition may be that there is an individual whose fitness function value is greater than a fitness threshold, or that the number of iterations reaches a maximum number of iterations.
[0112] Step 570, training the convolutional neural network determined by the optimal individual in the next generation population according to the data set consisting of multiple groups of simulated precoding matrices and corresponding optimal beam pairs, to obtain the trained convolutional neural network model.
[0113] In the final next generation population, the individual with the largest fitness function value is determined as the optimal individual in the next generation population. A convolutional neural network is constructed using the hyperparameter combination in the optimal individual in the next generation population, and the convolutional neural network is the best convolutional neural network obtained by the differential evolution algorithm. Figure 6 is a schematic diagram of the structure of a convolutional neural network in an embodiment of the present application, such as Figure 6 As shown, the convolutional neural network may include an input layer, a feature layer, a fully connected layer, and an output layer. During the training process, the input layer is used to input samples, and during the prediction process, the input layer is used to input an equivalent channel matrix to be selected for beam selection.
[0114] In the embodiment of the present application, the convolutional neural network model is obtained by optimizing the hyperparameters by the differential evolution algorithm, which is an improved convolutional neural network (Improved Convolutional Neural Network, ICNN). ICNN is a type of feedforward neural network that includes convolution calculations and has a deep structure, commonly used in image processing, face recognition, and convolutional neural networks have representation learning capabilities and can perform translation-invariant classification of input information according to their hierarchical structure. ICNN is a convolutional neural network improved by an intelligent algorithm (differential evolution algorithm) in the embodiment of the present application. In terms of model prediction accuracy, it has higher prediction accuracy than neural networks of equal depth.
[0115] By using a repeated exhaustive algorithm to obtain multiple groups of optimal beam pairs as training data sets for the convolutional neural network model, the convolutional neural network model is trained offline using the training data sets. The following describes the process of obtaining the optimal beam pair using the exhaustive algorithm.
[0116] Since the number of antennas in Massive MIMO is too large, it is often impossible to deploy an RF chain for each antenna in actual deployment. Therefore, Massive MIMO often uses hybrid precoding for beamforming operations. In the fully digital precoding structure, calculating the digital precoding matrix only requires obtaining the null space of the joint channel matrix of all users by performing SVD decomposition, and allowing the precoding of one user to fall within the null space of the channel matrix of other users to eliminate interference between users. In the hybrid precoding structure of Massive MIMO, since the system adopts a combination of digital precoding and analog precoding, when using the SVD algorithm to calculate the digital precoding matrix, the analog precoding F RF The multi-user joint channel matrix formed by multiplying the matrix with the channel matrix H is subjected to SVD decomposition.
[0117] Therefore, before solving the digital precoding matrix, it is necessary to obtain the analog precoding and construct an equivalent matrix. Specifically, after determining the equivalent channel matrix of each user in the above manner, the equivalent channel matrices of all users are combined to obtain a dimension of The joint equivalent channel matrix H eff :
[0118] H eff =[H eff,1 ,H eff,2 ,H eff,3 ...,H eff,m ]
[0119] Among them, m represents the number of users corresponding to the base station, Indicates the number of RF links at the receiving end, Indicates the number of RF links on the transmitting side.
[0120] Then, for H eff Performing SVD decomposition yields:
[0121]
[0122] in, for The unitary matrix of ∑ is a diagonal matrix whose non-zero elements on the diagonal are the joint equivalent channel matrix H eff The singular values of for The unitary matrix of .
[0123] At this time, the received signal of the i-th user can be expressed as:
[0124]
[0125] Among them, s i Represents the transmitted signal, n i Indicates noise.
[0126] If you use the matrix For the transmitted signal s i If precoding is performed, the received signal of the i-th user can be expressed as:
[0127]
[0128] Multiply both sides of the equation by have to:
[0129]
[0130] Slightly transform the above formula to get:
[0131] y′ i =∑s i +n′ i
[0132] As shown in the above formula, since ∑ is a diagonal matrix, after SVD transformation, the original channel model becomes multiple unrelated channels, and the system capacity becomes the sum of these channels:
[0133] y′ i,j =λ j s i,j +n′ i
[0134] Among them, λ j is the jth singular value in the joint equivalent channel matrix, j = 1, 2, 3, ..., N s , N s is the rank of the joint equivalent channel matrix, that is, the number of data streams that the system can transmit.
[0135] Therefore, through the above derivation, the optimal digital precoding matrix should be Matrix first N s The matrix consists of columns corresponding to singular values. Write in blocks:
[0136]
[0137] The structure is Its form is as follows:
[0138]
[0139] in, It represents the digital precoding matrix of the transmitter of the i-th user obtained by SVD, and its dimension size is S i Indicates the number of transmission streams corresponding to the i-th user.
[0140] Through the above steps, we can get the simulated precoding matrix F at the transmitting end. RF , the receiving end simulates the precoding matrix W RF , the digital precoding matrix F at the transmitting end BB , the receiving end digital precoding matrix W BB (The digital precoding matrix at the receiving end is generally in a fixed form), where F RF The dimension size is W RF The dimension size is F BB The dimension size is W BB The dimension size is F RF It is defined as constant amplitude and adjustable phase. BB Its phase and amplitude parameters can be adjusted. After the above precoding, the transmitted signal of the i-th user can be expressed as:
[0141]
[0142] in, YesN s ×1 transmission signal vector.
[0143] Then the received signal of the i-th user is:
[0144]
[0145] The first part represents the useful signal received by the i-th user, and the second part represents the interference caused by other users to the i-th user. i represents the Gaussian white noise of the channel.
[0146] The signal-to-interference-noise ratio of the i-th user can be expressed as:
[0147]
[0148] Among them, H i is the channel matrix between the i-th user and the base station, p is the transmitted signal power, σ 2 is the noise power.
[0149] 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 clustered channel model, namely the Saleh-Valenzuela model. At this time, the channel model is as follows:
[0150]
[0151] When the target beam selection criterion is the maximum signal-to-noise ratio, the beam pair with the maximum signal-to-noise ratio can be determined as the optimal beam pair (simulating the precoding matrix) according to the above formula, thereby constructing a training data set.
[0152] The constructed training data set is used to train the convolutional neural network determined by the optimal individual in the final next-generation population (i.e., the best population) to obtain a trained convolutional neural network model.
[0153] Figure 7 is a flowchart of selecting the best beam pair through a convolutional neural network model in an embodiment of the present application, such as Figure 7 As shown, in the offline training stage, a simulated precoding codebook is first constructed, and then an exhaustive algorithm is used to obtain multiple groups of simulated precoding codebooks (i.e., the best beam pairs), a training data set is constructed, and finally the training data set is used to train the convolutional neural network model to obtain a trained convolutional neural network model; in the online prediction stage, for the target base station, a simulated precoding codebook is first constructed, and then the equivalent channel matrix of each user in the simulated precoding codebook is input into the trained convolutional neural network model, and the best simulated precoding codebook is predicted by the convolutional neural network model.
[0154] The embodiment of the present application adopts a differential evolution algorithm to determine the hyperparameters of a convolutional neural network, improves the crossover and mutation process of the differential evolution algorithm, and newly adds a population individual replacement mechanism, so that when the differential evolution algorithm is solving the optimization problem, when the algorithm falls into a local optimum, the algorithm can jump out of the local optimum and further find the global optimal solution after increasing the diversity of the population through the population replacement mechanism; in solving the optimization problem, the improved differential evolution algorithm has a greatly reduced probability of falling into a local optimal solution because the algorithm has newly added an arc function (dynamically adjusting the mutation operator F and the crossover operator CR with the iteration process) and a new population replacement mechanism (even after falling into the local optimum, the algorithm can also be replaced by the population replacement mechanism). Jump out), resulting in the algorithm having better optimization ability than other intelligent algorithms (such as PSO algorithm, ant colony algorithm); and when the differential evolution algorithm is used in the convolutional neural network model, it is convenient to set different population ranges for different hyperparameters (such as convolution kernel size, number of convolution layers, etc.), so as to make the parameter optimization of the convolutional neural network model more accurate in different usage scenarios; in the embodiment of the present application, the construction of the convolutional neural network model is more intelligent, and there is no need to select through the empirical formula of the neural network model, which avoids repeated attempts, and the optimal model structure can be found through the improved differential evolution algorithm to achieve a more accurate convolutional neural network model.
[0155] When performing beam selection, the embodiment of the present application performs beam selection on the obtained analog precoding codebook, inputs all codebook matrices into the convolutional neural network model, performs beam selection based on the target beam selection criterion (such as the maximum signal-to-noise ratio criterion, etc.), and outputs the optimal analog precoding matrix (beam pair); compared with the prior art, the embodiment of the present application no longer needs to exhaustively calculate all beams when performing beam selection, and directly uses the convolutional neural network model for prediction and selection. At the same time, a variety of different selection criteria can be selected for prediction in different scenarios, such as the maximum signal-to-noise ratio criterion, the maximum signal-to-interference-plus-noise ratio criterion, etc., and the actual scope of use is wider; by using the high-precision convolutional neural network model improved based on the differential evolution algorithm, the surge in the number of antennas and the excessive complexity of the exhaustive calculation in the Massive MIMO scenario, which leads to excessive system delay, is avoided, so as to achieve a beam selection with lower complexity.
[0156] Figure 8 is a structural diagram of a large-scale MIMO analog domain beam selection device provided in an embodiment of the present application, such as Figure 8 As shown, the device comprises:
[0157] The simulated codebook determination module 810 is used to determine multiple groups of simulated precoding matrices according to the transmitting antenna array information of the target base station and the receiving antenna array information of the user, each group of the simulated precoding matrices includes a receiving end simulated precoding matrix and a transmitting end simulated precoding matrix;
[0158] An equivalent channel determination module 820 is used to determine a plurality of equivalent channel matrices according to a plurality of groups of simulated precoding matrices and a target channel model, each of the equivalent channel matrices corresponding to a group of the simulated precoding matrices;
[0159] The beam selection module 830 is used to input the multiple equivalent channel matrices into the trained convolutional neural network model, and select the target simulated precoding matrix from the multiple equivalent channel matrices based on the target beam selection criterion through the convolutional neural network model. The network structure of the convolutional neural network model is determined based on the differential evolution algorithm, and the mutation operator in the differential evolution algorithm is controlled based on the arc function.
[0160] Optionally, the device further comprises:
[0161] A population initialization module, used to generate an initial population according to a value range of a hyperparameter of a convolutional neural network, and use the initial population as a current population, wherein each individual in the initial population is a hyperparameter combination of the convolutional neural network, and the hyperparameter is a structural parameter of the convolutional neural network;
[0162] A fitness determination module, used to determine the fitness function value of each individual in the current population, and determine the individual with the largest fitness function value as the optimal individual of the current population, wherein the fitness function value is the prediction accuracy of the convolutional neural network determined by the individual for the selection of a sample simulated precoding matrix, and the sample simulated precoding matrix includes multiple groups of simulated precoding matrices;
[0163] A mutation operation module, used for performing a mutation operation on the current population according to the optimal individual and based on a mutation operator controlled by an arc function to obtain a temporary solution;
[0164] A crossover operation module, used for performing a crossover operation on the temporary solution and the current population to obtain a new solution;
[0165] A next generation population determination module, used to determine the next generation population according to the current population and the new solution;
[0166] An iteration control module, used for, when the next generation population does not meet the iteration termination condition, using the next generation population as the current population, and iteratively performing operations of determining the fitness function value of each individual in the current population and determining the next generation population, until the next generation population meets the iteration termination condition;
[0167] The convolutional network training module is used to train the convolutional neural network determined by the optimal individual in the next generation population according to a data set consisting of multiple groups of simulated precoding matrices and corresponding optimal beam pairs, so as to obtain the trained convolutional neural network model.
[0168] Optionally, the mutation operation module includes:
[0169] A mutation operator determination unit, used to determine the mutation operator according to the current number of iterations and the maximum number of iterations;
[0170] An individual selection unit, used for randomly selecting an odd number of individuals from the current population, and taking one of the odd number of individuals as a target individual;
[0171] A mutation operation unit, configured to perform a mutation operation on the target individual and the optimal individual according to the mutation operator and the odd number of individuals, to obtain an individual in the temporary solution;
[0172] A loop control unit is used to loop through the operations of randomly selecting an odd number of individuals from the current population and generating an individual in a temporary solution until the number of individuals in the temporary solution reaches the population size.
[0173] Optionally, the mutation operation unit is specifically used for:
[0174] Determining, according to the mutation operator, a first weight of the target individual and a second weight of the optimal individual, wherein the sum of the first weight and the second weight is 1;
[0175] According to the first weight and the second weight, performing weighted summation on the target individual and the optimal individual to obtain a basis vector of a mutation operation;
[0176] According to the mutation operator, respectively determine the difference vectors of the individuals other than the target individual in the odd number of individuals;
[0177] The sum of the basis vector and the difference vector is taken as an individual in the temporary solution.
[0178] Optionally, the mutation operator determination unit is specifically used to:
[0179] According to the current number of iterations and the maximum number of iterations, the mutation operator is determined according to the following formula:
[0180]
[0181] Wherein, F represents the mutation operator, g represents the current number of iterations, and G represents the maximum number of iterations.
[0182] Optionally, the cross-operation module includes:
[0183] A crossover operator determination unit, used to determine a crossover operator according to a current number of iterations and a maximum number of iterations;
[0184] A crossover operation unit is used to perform a crossover operation on the temporary solution and the current population based on the crossover operator according to the following formula to obtain a new solution:
[0185]
[0186] Among them, g represents the current iteration number, u ij (g) represents the jth hyperparameter in the i-th individual in the new solution, v ij (g) represents the jth hyperparameter in the i-th individual in the temporary solution, x ij (g) represents the jth hyperparameter in the i-th individual in the current population, rnadb(j) represents the j-th random number generated between [0,1], rnbr(i)∈(1,2,3…d) represents a randomly selected sequence, d represents the dimension of each individual, and CR represents the crossover operator.
[0187] Optionally, the crossover operator determination unit is specifically used to:
[0188] According to the current number of iterations and the maximum number of iterations, the crossover operator is determined according to the following formula:
[0189]
[0190] Wherein, CR represents the crossover operator, g represents the current number of iterations, and G represents the maximum number of iterations.
[0191] Optionally, the next generation population determination module includes:
[0192] An individual fitness determination unit, used to determine the fitness function value of each individual in the new solution;
[0193] A temporary next-generation population generation unit, used to generate a temporary next-generation population according to the fitness function value of each individual in the current population and the fitness function value of each individual in the new solution;
[0194] The next generation population determination unit is used to randomly generate a preset number of new individuals when the individual fitness variance of the temporary next generation population is less than or equal to the variance threshold, and use the new individuals to replace the preset number of individuals in the temporary next generation population to obtain the next generation population.
[0195] Optionally, the temporary next generation population generation unit is specifically used for:
[0196] In the order of fitness function values from large to small, individuals of the population size are selected from the current population and the new solution as the temporary next generation population.
[0197] The large-scale MIMO analog domain beam selection device provided in the embodiment of the present application is used to implement the various steps of the large-scale MIMO analog domain beam selection method described in the embodiment of the present application. The specific implementation methods of each module of the device refer to the corresponding steps and will not be repeated here.
[0198] The large-scale MIMO analog domain beam selection device provided in the embodiment of the present application determines multiple groups of analog precoding matrices according to the transmitting antenna array information of the target base station and the receiving antenna array information of the user, determines multiple equivalent channel matrices according to the multiple groups of analog precoding matrices and the target channel model, and selects the target analog precoding matrix from the multiple equivalent channel matrices based on the target beam selection criterion through the convolutional neural network model. For the millimeter wave large-scale antenna system, the target analog precoding matrix can also be quickly determined through the convolutional neural network model, thereby improving the beam selection efficiency in the large-scale MIMO system. Moreover, since the network structure of the convolutional neural network model is determined based on the differential evolution algorithm, the construction of the convolutional neural network structure is more intelligent, and the prediction accuracy of the convolutional neural network model can be improved.
[0199] Fig. 9 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, such as Fig. 9 As shown, the electronic device 900 may include one or more processors 910 and one or more memories 920 connected to the processor 910. The electronic device 900 may also include an input interface 930 and an output interface 940 for communicating with another device or system. The program code executed by the processor 910 may be stored in the memory 920.
[0200] The processor 910 in the electronic device 900 calls the program code stored in the memory 920 to execute the massive MIMO analog domain beam selection method in the above embodiment.
[0201] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the large-scale MIMO analog domain beam selection method as described in the embodiment of the present application are implemented.
[0202] The embodiment of the present application also provides a computer program product, which, when executed by a processor, implements the steps of the large-scale MIMO analog domain beam selection method as described in the embodiment of the present application.
[0203] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
[0204] The above is a detailed introduction to a large-scale MIMO analog domain beam selection method, device, equipment and storage medium provided in an embodiment of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
[0205] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.
Claims
1. A massive MIMO analog domain beam selection method, characterized in that: include: Determine multiple groups of simulated precoding matrices according to the transmit antenna array information of the target base station and the receive antenna array information of the user, each group of simulated precoding matrices including a receive-end simulated precoding matrix and a transmit-end simulated precoding matrix; Determine a plurality of equivalent channel matrices according to the plurality of groups of simulated precoding matrices and the target channel model, each of the equivalent channel matrices corresponding to a group of the simulated precoding matrices; The multiple equivalent channel matrices are input into a trained convolutional neural network model, and the convolutional neural network model selects a target simulated precoding matrix from the multiple equivalent channel matrices based on a target beam selection criterion. The network structure of the convolutional neural network model is determined based on a differential evolution algorithm, and the mutation operator in the differential evolution algorithm is controlled based on an arc function.
2. The method according to claim 1, characterized in that Before inputting the multiple equivalent channel matrices into the trained convolutional neural network model, the method further includes: Generate an initial population according to the value range of the hyperparameter of the convolutional neural network, and use the initial population as the current population, each individual in the initial population is a hyperparameter combination of the convolutional neural network, and the hyperparameter is a structural parameter of the convolutional neural network; Determine the fitness function value of each individual in the current population, and determine the individual with the largest fitness function value as the optimal individual of the current population, wherein the fitness function value is the prediction accuracy of the convolutional neural network determined by the individual for the selection of a sample simulated precoding matrix, and the sample simulated precoding matrix includes multiple groups of simulated precoding matrices; According to the optimal individual, a mutation operator controlled by an arc function is used to perform a mutation operation on the current population to obtain a temporary solution; Performing a crossover operation on the temporary solution and the current population to obtain a new solution; Determining a next generation population according to the current population and the new solution; When the next generation population does not meet the iteration termination condition, the next generation population is used as the current population, and operations of determining the fitness function value of each individual in the current population and determining the next generation population are iteratively performed until the next generation population meets the iteration termination condition; According to a data set consisting of multiple groups of simulated precoding matrices and corresponding optimal beam pairs, a convolutional neural network determined by the optimal individual in the next generation population is trained to obtain the trained convolutional neural network model.
3. The method according to claim 2, characterized in that The step of performing a mutation operation on the current population based on the optimal individual and a mutation operator controlled by an arc function to obtain a temporary solution includes: Determine the mutation operator according to the current number of iterations and the maximum number of iterations; Randomly selecting an odd number of individuals from the current population, and taking one of the odd number of individuals as a target individual; According to the mutation operator and the odd number of individuals, performing mutation operations on the target individual and the optimal individual to obtain an individual in the temporary solution; The operation of randomly selecting an odd number of individuals from the current population and generating an individual in the temporary solution is cyclically executed until the number of individuals in the temporary solution reaches the population size.
4. The method according to claim 3, characterized in that The step of performing mutation operations on the target individual and the optimal individual according to the mutation operator and the odd number of individuals to obtain an individual in the temporary solution includes: Determining, according to the mutation operator, a first weight of the target individual and a second weight of the optimal individual, wherein the sum of the first weight and the second weight is 1; According to the first weight and the second weight, performing weighted summation on the target individual and the optimal individual to obtain a basis vector of a mutation operation; According to the mutation operator, respectively determine the difference vectors of the individuals other than the target individual in the odd number of individuals; The sum of the basis vector and the difference vector is taken as an individual in the temporary solution.
5. The method according to claim 3 or 4, characterized in that: The step of determining the mutation operator according to the current number of iterations and the maximum number of iterations includes: According to the current number of iterations and the maximum number of iterations, the mutation operator is determined according to the following formula: Wherein, F represents the mutation operator, g represents the current number of iterations, and G represents the maximum number of iterations.
6. The method according to any one of claims 2 to 4, characterized in that: The performing a crossover operation on the temporary solution and the current population to obtain a new solution includes: Determine the crossover operator according to the current number of iterations and the maximum number of iterations; According to the following formula, a crossover operation is performed on the temporary solution and the current population based on the crossover operator to obtain a new solution: Among them, g represents the current iteration number, u ij (g) represents the jth hyperparameter in the i-th individual in the new solution, v ij (g) represents the jth hyperparameter in the i-th individual in the temporary solution, x ij (g) represents the jth hyperparameter in the i-th individual in the current population, randb(j) represents the j-th random number generated between [0, 1], rnbr(i)∈(1, 2, 3…d) represents a randomly selected sequence, d represents the dimension of each individual, and CR represents the crossover operator.
7. The method according to claim 6, characterized in that Determining the crossover operator according to the current number of iterations and the maximum number of iterations includes: According to the current number of iterations and the maximum number of iterations, the crossover operator is determined according to the following formula: Wherein, CR represents the crossover operator, g represents the current number of iterations, and G represents the maximum number of iterations.
8. The method according to any one of claims 2 to 4, characterized in that: The step of determining the next generation population according to the current population and the new solution comprises: Determine the fitness function value of each individual in the new solution; Generate a temporary next generation population according to the fitness function value of each individual in the current population and the fitness function value of each individual in the new solution; When the individual fitness variance of the temporary next generation population is less than or equal to the variance threshold, a preset number of new individuals are randomly generated, and the new individuals are used to replace the preset number of individuals in the temporary next generation population to obtain the next generation population.
9. The method according to claim 8, characterized in that The step of generating a temporary next generation population according to the fitness function value of each individual in the current population and the fitness function value of each individual in the new solution comprises: In the order of fitness function values from large to small, individuals of the population size are selected from the current population and the new solution as the temporary next generation population.
10. A massive MIMO analog domain beam selection device, characterized in that: include: The analog codebook determination module is used to determine multiple groups of analog precoding matrices according to the transmitting antenna array information of the target base station and the receiving antenna array information of the user, each group of the analog precoding matrices includes a receiving end analog precoding matrix and a transmitting end analog precoding matrix; An equivalent channel determination module, used to determine a plurality of equivalent channel matrices according to a plurality of groups of simulated precoding matrices and a target channel model, each of the equivalent channel matrices corresponding to a group of the simulated precoding matrices; A beam selection module is used to input the multiple equivalent channel matrices into a trained convolutional neural network model, and select a target simulated precoding matrix from the multiple equivalent channel matrices based on a target beam selection criterion through the convolutional neural network model, wherein the network structure of the convolutional neural network model is determined based on a differential evolution algorithm, and the mutation operator in the differential evolution algorithm is controlled based on an arc function.
11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the large-scale MIMO analog domain beam selection method described in any one of claims 1 to 9 is implemented.
12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the large-scale MIMO analog domain beam selection method described in any one of claims 1 to 9 are implemented.
Citation Information
Cited By
Channel prediction method and device, computer equipment, storage medium and computer program product
CN120785447A
