Non-cellular large-scale MIMO precoding method and system

Abstract

The invention relates to the technical field of wireless communication, and discloses a cellular-free large-scale MIMO precoding method and system, and the method comprises the steps: constructing a cellular-free large-scale MIMO system; the access points randomly generate Rayleigh-distributed channel state information, and according to the channel state information, an FP precoding algorithm is deeply expanded into a neural network so as to generate a local model based on each access point; constructing a federated learning loss function according to the rate function, and setting a convergence target; federated learning training is carried out on the local model, and the channel state information is used as a data set input by federated learning training; and the local model parameters are aggregated until the objective function reaches convergence, and optimal precoding output is obtained. The distributed federated learning and the FP precoding algorithm are combined, the limited backhaul link capacity is better utilized, the downlink data transmission performance of the cellular-free large-scale MIMO system is further optimized, the total spectrum efficiency of the limited backhaul link capacity is improved, and the forward transmission burden is relieved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a non-cellular large-scale MIMO precoding method and system. Background Art

[0002] As a new generation of wireless communication technology, the cell-free massive MIMO system combines the characteristics of massive MIMO and distributed antennas, which can effectively solve the contradiction between the coverage and speed of traditional cellular networks, thereby significantly promoting breakthroughs in key network performance of future mobile communications.

[0003] Although non-cellular massive MIMO has many excellent features, in actual implementation, considering the hardware cost, the backhaul link capacity between AP (Access Point) and CPU (Central Processing Unit) is very limited. According to relevant research, the system transmission rate will be limited by the capacity of the backhaul link. Therefore, it is necessary to find a way to make full use of the limited backhaul link capacity. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a non-cellular massive MIMO precoding method and system to solve the problem that the capacity of the backhaul link between the AP and the CPU is very limited and the system transmission rate is limited by the capacity of the backhaul link.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a non-cellular massive MIMO precoding method, comprising: acquiring a video stream in a monitoring area;

[0008] Construct a cell-free massive MIMO system consisting of M access points with L antennas and K single-antenna users;

[0009] The access point randomly generates Rayleigh distributed channel state information, and according to the channel state information, deeply expands the FP precoding algorithm into a neural network to generate a local model based on each access point;

[0010] Construct the loss function of federated learning based on the rate function and set the convergence target;

[0011] Performing federated learning training on the local model, using the channel state information as a data set input for the federated learning training;

[0012] The local model parameters are aggregated until the objective function converges to obtain the optimal precoding output.

[0013] As a preferred solution of the non-cellular massive MIMO precoding method of the present invention, wherein: the construction of a non-cellular massive MIMO system including M access points with L antennas and K single-antenna users includes:

[0014] 1 CPU, M access points with L antennas, and K single-antenna users;

[0015] Among them, each access point and each user are evenly distributed in a square area with a side length of D, and their distribution is independent of each other;

[0016] Each access point is equipped with a single antenna, and the large-scale fading coefficient of the channel between the mth access point and the kth user is β mk It means that all access points are connected to the central processor via a backhaul link with limited capacity.

[0017] As a preferred solution of the non-cellular massive MIMO precoding method of the present invention, wherein: the access point randomly generates Rayleigh distributed channel state information, including:

[0018] The channel state information between all access points and all users includes the fast fading vector h mk and the large-scale fading coefficient β mk ;

[0019] Each access point obtains the local channel state information g mk , where g mk is the mth row and kth column element of the channel state information matrix G, expressed as:

[0020]

[0021] Take g mk The real part of The imaginary part is The complex field g mk Mapping to the real number field gives:

[0022] G′ of all users mk Integrated into G′ m As the input dataset of the mth access point.

[0023] As a preferred solution of the non-cellular massive MIMO precoding method of the present invention, the FP precoding algorithm is deeply expanded into a neural network according to the channel state information to generate a local model based on each access point, including:

[0024] Replace the optimization method step in the FP algorithm with a T-layer stochastic gradient descent step;

[0025] The update step size of each layer of gradient descent is set as a learnable parameter, and the replaced FP algorithm is deeply expanded into a neural network;

[0026] Combine the channel state information matrix G to obtain the precoding matrix V after deep expansion;

[0027] The deep unfolding neural network of the FP algorithm is deployed on all access points, and the CPU uniformly sends the learnable parameters of each access point to calculate the initial precoding.

[0028] As a preferred solution of the non-cellular massive MIMO precoding method of the present invention, wherein: constructing a loss function of federated learning according to a rate function includes:

[0029] The loss function of federated learning is set to the negative of the sum rate, expressed as:

[0030]

[0031] Where X l For the input data set, For the dataset Length;

[0032] As a preferred solution of the non-cellular massive MIMO precoding method of the present invention, the convergence target is optimized by federated learning, and the objective function is expressed as:

[0033]

[0034] Among them, f represents the functional relationship between the model input and output, Indicates that the loss function relationship takes the negative of the result of f, θ l is the weight parameter of the model at the lth AP, ω is the weight parameter of the model at the central processor, and st is represented as a constraint condition.

[0035] As a preferred solution of the non-cellular massive MIMO precoding method of the present invention, the local model parameters are aggregated until the objective function converges, including:

[0036] The local model of each access point is learned using mini-batch gradient descent;

[0037] Train the local models of each access point separately. When the training reaches the set number of training times, each local model will update the model parameters Upload to the CPU;

[0038] Aggregate the model parameters of each access point to obtain the new global model parameter ω t , expressed as:

[0039] The updated global model parameter ω t Downlink to each access point;

[0040] Each access point uses the received global model parameters as local model parameters, namely: And let t=t+1, and repeat the above training process of the local model of each access point until the system objective function converges.

[0041] In a second aspect, the present invention provides a non-cellular massive MIMO precoding system, comprising:

[0042] A system building module for building a cell-free massive MIMO system including M access points with L antennas and K single-antenna users;

[0043] A network model generation module, configured to randomly generate Rayleigh distributed channel state information for the access point, and to deeply expand the FP precoding algorithm into a neural network according to the channel state information, so as to generate a local model based on each access point;

[0044] The loss function generation module is used to construct the loss function of federated learning based on the rate function and set the convergence target;

[0045] A training module, used for performing federated learning training on the local model, wherein the channel state information is used as a data set for federated learning training input;

[0046] The convergence output module is used to aggregate the local model parameters until the objective function converges to obtain the optimal precoding output.

[0047] In a third aspect, the present invention provides a computing device, comprising:

[0048] Memory and processor;

[0049] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the cellular-free massive MIMO precoding method are implemented.

[0050] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the non-cellular massive MIMO precoding method.

[0051] Compared with the prior art, the present invention has the following beneficial effects: the present invention obtains channel state information for each access point, calculates local channel state information for each access point, and performs original precoding calculation for all access points according to the FP precoding algorithm combined with the channel state information; expands the complex optimization solution method in the FP algorithm into a neural network, and sets learnable parameters; deploys a local network of distributed federated learning for each access point, calculates the spectrum efficiency and the total spectrum efficiency by receiving signals; takes maximization and rate as the goal, and obtains the optimal precoding information under the constraints of total transmission power and backhaul link capacity.

[0052] The present invention combines distributed federated learning and FP precoding algorithms to improve the overall spectrum efficiency of non-cellular massive MIMO systems with limited backhaul link capacity; and only local information and model trainable parameters need to be transferred from the CPU to the AP during training, greatly reducing the burden of the fronthaul. In addition, after the training is completed, each AP can independently generate effective precoding using the locally trained model, eliminating the need to transmit any further information to the CPU, greatly reducing the fronthaul load, and achieving significant rate improvements and computation time reductions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0054] Figure 1 This is a schematic diagram of the overall process of the non-cellular massive MIMO precoding method according to the first embodiment of the present invention;

[0055] Figure 2 A schematic diagram of a model structure and parameter transfer in a federated learning process in a cellular-free massive MIMO precoding method according to a first embodiment of the present invention;

[0056] Figure 3 A neutralization rate curve diagram of the non-cellular massive MIMO precoding method according to the second embodiment of the present invention;

[0057] Figure 4 This is a comparison chart of the running time of three schemes in the non-cellular massive MIMO precoding method described in the second embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0059] Example 1

[0060] Reference Figure 1-Figure 2 , as an embodiment of the present invention, provides a non-cellular massive MIMO precoding method, comprising:

[0061] S100: constructing a cell-free massive MIMO system including M access points with L antennas and K single-antenna users;

[0062] In an embodiment of the present application, a non-cellular massive MIMO system including M access points with L antennas and K single-antenna users is constructed, including:

[0063] 1 CPU, M access points with L antennas, and K single-antenna users;

[0064] Among them, each access point and each user are evenly distributed in a square area with a side length of D, and their distribution is independent of each other;

[0065] Each access point is equipped with a single antenna, and the large-scale fading coefficient of the channel between the mth access point and the kth user is β mk It means that all access points are connected to the central processor via a backhaul link with limited capacity.

[0066] It should be noted that the access points are evenly distributed in the area and connected to the central processor via a backhaul link to serve all users in the area.

[0067] S200: The access point randomly generates Rayleigh distributed channel state information, and according to the channel state information, the FP precoding algorithm is deeply expanded into a neural network to generate a local model based on each access point;

[0068] In an embodiment of the present application, the access point randomly generates Rayleigh distributed channel state information, including:

[0069] The channel state information between all access points and all users includes the fast fading vector h mk and the large-scale fading coefficient β mk ;

[0070] Each access point obtains the local channel state information g mk , where g mkis the mth row and kth column element of the channel state information matrix G, expressed as:

[0071]

[0072] Take g mk The real part of The imaginary part is The complex field g mk Mapping to the real number field gives:

[0073] G′ of all users mk Integrated into G′ m As the input dataset of the mth access point.

[0074] In an embodiment of the present application, the FP precoding algorithm is deeply expanded into a neural network according to the channel state information to generate a local model based on each access point, including:

[0075] Replace the optimization method step in the FP algorithm with a T-layer stochastic gradient descent step;

[0076] The update step size of each layer of gradient descent is set as a learnable parameter, and the replaced FP algorithm is deeply expanded into a neural network;

[0077] Combine the channel state information matrix G to obtain the precoding matrix V after deep expansion;

[0078] Specifically, the precoding matrix V output by the gradient descent of the tth layer is t It is expressed as:

[0079]

[0080] Among them, α (t) is the update step size of the gradient descent of the tth layer;

[0081] Furthermore, the unfolded T-layer gradient descent is mapped to the neural network in machine learning, and the update step size of the gradient descent is set as a learnable parameter;

[0082] The processed data signal is sent to each access point through the limited capacity backhaul link. The signal y received by the kth user k Expressed as:

[0083]

[0084] Furthermore, the additive Gaussian white noise of the received wireless channel is normalized to calculate the individual rate R of the kth user in the downlink transmission system. k , the formula is:

[0085] Rk =log2(1+SINR k )

[0086]

[0087] Furthermore, the individual rates R of all users are k Add them together to get the total spectrum efficiency WSR, which is expressed as:

[0088]

[0089] where ω k is the weighting coefficient of the kth user.

[0090] The deep unfolding neural network of the FP algorithm is deployed on all access points, and the CPU uniformly sends the learnable parameters of each access point to calculate the initial precoding.

[0091] Specifically, in this step, the unfolding of the neural network is processed by central processing.

[0092] S300: construct the loss function of federated learning according to the rate function and set the convergence target;

[0093] In the embodiment of the present application, the loss function of federated learning is constructed according to the rate function, including:

[0094] The loss function of federated learning is set to the negative of the sum rate, expressed as:

[0095]

[0096] Where X l For the input data set, For the dataset Length.

[0097] It should be noted that the negative of the sum rate means that when the loss function is minimized, the sum rate is maximized.

[0098] In the embodiment of the present application, the convergence target is optimized by federated learning, and the objective function is expressed as:

[0099]

[0100] Among them, f represents the functional relationship between the model input and output, Indicates that the loss function relationship takes the negative of the result of f, θ l is the weight parameter of the model at the lth AP, ω is the weight parameter of the model at the central processor, and st is represented as a constraint condition.

[0101] It should be noted that the convergence target is set before the first training, and the entire training process relies on the convergence of the objective function to determine the specific situation of the reverse update of the parameters.

[0102] S400: performing federated learning training on the local model, and using the channel state information as a data set for federated learning training input;

[0103] It should be noted that after the training is completed, each AP can independently generate effective precoding using the locally trained model, thereby eliminating the need to transmit any further information to the CPU, greatly reducing the fronthaul load.

[0104] S500: Aggregate local model parameters until the objective function converges to obtain an optimal precoding output.

[0105] In the embodiment of the present application, the local model parameters are aggregated until the objective function reaches convergence, including:

[0106] The local model of each access point is learned using mini-batch gradient descent;

[0107] Train the local models of each access point separately. When the training reaches the set number of training times, that is, the training epoch is equal to epoch_num, each local model will update the model parameters Upload to the CPU;

[0108] Aggregate the model parameters of each access point to obtain the new global model parameter ω t , expressed as:

[0109] The updated global model parameter ω t Downlink to each access point;

[0110] Each access point uses the received global model parameters as local model parameters, namely: And let t=t+1, and repeat the above training process of the local model of each access point until the system objective function converges.

[0111] Specifically, this convergence step can be performed by a central processor in the system.

[0112] Exemplarily, the specific learning steps are as follows:

[0113] By traversing 10^4 different channel state matrices G, the current precoding information V is calculated according to the current learnable parameters, the total weighted sum rate of the system is calculated, and the learnable parameters are updated by back-propagating machine learning gradients. After all APs have undergone a learning process, the CPU aggregates the learnable parameters and performs an average operation. The averaged parameters are then deployed to the local deep learning model of each AP to maximize the weighted sum rate.

[0114] Map the FP algorithm to a neural network, and replace the complex optimization solution method in the algorithm with T-layer stochastic gradient descent to facilitate the mapping of the neural network and simplify the algorithm calculation complexity. The input of the neural network is G, the output is V, and the model-related learning parameters are set;

[0115] It should be noted that the model-related learning parameters are set as follows: the initial learning rate is 0.01, the number of learning data is 64,000, and a total of 1,000 data sets are divided into groups of 64 data.

[0116] The above is a schematic scheme of a non-cellular massive MIMO precoding method of this embodiment. It should be noted that the technical solution of the non-cellular massive MIMO precoding system and the technical solution of the non-cellular massive MIMO precoding method described above belong to the same concept. For details not described in detail in the technical solution of the non-cellular massive MIMO precoding system in this embodiment, please refer to the description of the technical solution of the non-cellular massive MIMO precoding method described above.

[0117] The non-cellular massive MIMO precoding system in this embodiment includes:

[0118] A system building module for building a cell-free massive MIMO system including M access points with L antennas and K single-antenna users;

[0119] A network model generation module is used for randomly generating Rayleigh distributed channel state information at access points, and based on the channel state information, the FP precoding algorithm is deeply expanded into a neural network to generate a local model based on each access point;

[0120] The loss function generation module is used to construct the loss function of federated learning based on the rate function and set the convergence target;

[0121] The training module is used to perform federated learning training on the local model, and the channel state information is used as the input data set of the federated learning training;

[0122] The convergence output module is used to aggregate the local model parameters until the objective function converges to obtain the optimal precoding output.

[0123] This embodiment further provides a computing device, applicable to a case where there is no cellular massive MIMO precoding, including:

[0124] Memory and processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the non-cellular large-scale MIMO precoding method proposed in the above embodiment.

[0125] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for realizing non-cellular massive MIMO precoding as proposed in the above embodiment is implemented.

[0126] The storage medium proposed in this embodiment and the method for implementing non-cellular large-scale MIMO precoding proposed in the above embodiment belong to the same inventive concept. The technical details not described in detail in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0127] Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention 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 a computer floppy disk, read-only memory (ReadOnly, Memory, ROM), random access memory (RandomAccess Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform the methods of various embodiments of the present invention.

[0128] Example 2

[0129] Reference Figure 3 and Figure 4 Based on the previous embodiment, this embodiment provides an application comparison case of a non-cellular large-scale MIMO precoding method to illustrate the feasibility and beneficial effects of our solution.

[0130] Figure 3 The horizontal axis represents the number of users in the system, and the vertical axis represents the sum rate value. The model-driven federated learning precoding refers to the method of the present invention, the local MMSE precoding is the ordinary local precoding, and the local FP algorithm precoding is the FP original algorithm.

[0131] From the figure, we can see that the sum rate value of the non-cellular massive MIMO precoding method based on model-driven federated learning is much larger than that of the local MMSE precoding; although it is slightly smaller than the local FP algorithm precoding, it is the local FP algorithm precoding. Figure 4 It can be seen that its running time is much shorter than the local FP algorithm precoding.

[0132] This shows that the cellular-free massive MIMO precoding method based on model-driven federated learning can significantly improve the system and rate without adding additional payload on the front end; therefore, this method greatly reduces the local computing time while improving the system and rate, proving the feasibility and superiority of the method of the present invention.

[0133] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A non-cellular massive MIMO precoding method, characterized in that: include: Construct a cell-free massive MIMO system consisting of M access points with L antennas and K single-antenna users; The access point randomly generates Rayleigh distributed channel state information, and according to the channel state information, deeply expands the FP precoding algorithm into a neural network to generate a local model based on each access point; Construct the loss function of federated learning based on the rate function and set the convergence target; Performing federated learning training on the local model, using the channel state information as a data set input for the federated learning training; The local model parameters are aggregated until the objective function converges to obtain the optimal precoding output.

2. The non-cellular massive MIMO precoding method according to claim 1, wherein: The construction includes a non-cellular massive MIMO system with M access points with L antennas and K single-antenna users, including: 1 CPU, M access points with L antennas, and K single-antenna users; Among them, each access point and each user are evenly distributed in a square area with a side length of D, and their distribution is independent of each other; Each access point is equipped with a single antenna, and the large-scale fading coefficient of the channel between the mth access point and the kth user is β mk It means that all access points are connected to the central processor via a backhaul link with limited capacity.

3. The non-cellular massive MIMO precoding method according to claim 2, wherein: The access point randomly generates Rayleigh distributed channel state information, including: The channel state information between all access points and all users includes the fast fading vector h mk and the large-scale fading coefficient β mk ; Each access point obtains the local channel state information g mk , where g mk is the mth row and kth column element of the channel state information matrix G, expressed as: Take g mk The real part of The imaginary part is The complex field g mk Mapping to the real number field gives: G′ of all users mk Integrated into G′ m As the input dataset of the mth access point.

4. The non-cellular massive MIMO precoding method according to claim 3, wherein: According to the channel state information, the FP precoding algorithm is deeply expanded into a neural network to generate a local model based on each access point, including: Replace the optimization method step in the FP algorithm with a T-layer stochastic gradient descent step; The update step size of each layer of gradient descent is set as a learnable parameter, and the replaced FP algorithm is deeply expanded into a neural network; Combine the channel state information matrix G to obtain the precoding matrix V after deep expansion; The deep unfolding neural network of the FP algorithm is deployed on all access points, and the CPU uniformly sends the learnable parameters of each access point to calculate the initial precoding.

5. The non-cellular massive MIMO precoding method according to claim 4, characterized in that: The loss function of federated learning is constructed based on the rate function, including: The loss function of federated learning is set to the negative of the sum rate, expressed as: Where X l For the input data set, For the dataset Length; 6. The non-cellular massive MIMO precoding method according to claim 5, characterized in that: The convergence target is optimized through federated learning, and the objective function is expressed as: Among them, f represents the functional relationship between the model input and output, Indicates that the loss function relationship takes the negative of the result of f, θ l is the weight parameter of the model at the lth AP, ω is the weight parameter of the model at the central processor, and st is represented as a constraint condition.

7. The non-cellular massive MIMO precoding method according to claim 6, wherein: Aggregate local model parameters until the objective function converges, including: The local model of each access point is learned using mini-batch gradient descent; Train the local model of each access point separately. When the training reaches the set number of training times, each local model will update the model parameters Upload to the CPU; Aggregate the model parameters of each access point to obtain the new global model parameter ω t , expressed as: The updated global model parameter ω t Downlink to each access point; Each access point uses the received global model parameters as local model parameters, namely: And let t=t+1, and repeat the above training process of the local model of each access point until the system objective function converges.

8. A system using the non-cellular massive MIMO precoding method according to claim 1, characterized in that: include: A system building module for building a cell-free massive MIMO system including M access points with L antennas and K single-antenna users; A network model generation module, configured to randomly generate Rayleigh distributed channel state information for the access point, and to deeply expand the FP precoding algorithm into a neural network according to the channel state information, so as to generate a local model based on each access point; The loss function generation module is used to construct the loss function of federated learning based on the rate function and set the convergence target; A training module, used for performing federated learning training on the local model, wherein the channel state information is used as a data set for federated learning training input; The convergence output module is used to aggregate the local model parameters until the objective function converges to obtain the optimal precoding output.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the non-cellular massive MIMO precoding method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the non-cellular massive MIMO precoding method according to any one of claims 1 to 7.

Images