Disclosure of Invention
The application aims to provide a codebook-free near-field beamforming method, a codebook-free near-field beamforming device, codebook-free near-field beamforming equipment, codebook-free near-field beamforming medium and codebook-free near-field beamforming product, which can be used for efficiently extracting and estimating a near-field channel and reducing the cost of beam training.
In order to achieve the above object, the present application provides the following solutions:
in a first aspect, the present application provides a codebook-free near-field beamforming method based on deep learning, including:
the method comprises the steps of constructing and training a convolutional neural network model, wherein the convolutional neural network model comprises a plurality of feature extraction modules, each feature extraction module comprises two convolutional layers, two regularization layers and two activation layers, wherein the filling is set to be 1 in the first layer of convolution, the size of a convolution kernel is set to be 2 multiplied by 2 in the second layer of convolution, the filling is set to be 0 in the second layer of convolution, and the size of the convolution kernel is set to be 2 multiplied by 2;
Inputting the current channel state information into a trained convolutional neural network model, and determining an optimal beamforming vector;
And adjusting the transmitting signals of the antenna array based on the optimal shaping vector to form a directional beam so as to finish near-field beam shaping.
In a second aspect, the present application provides a codebook-free near-field beamforming apparatus based on deep learning, including:
The model building and training module is used for building and training a convolutional neural network model, wherein the convolutional neural network model comprises a plurality of feature extraction modules, each feature extraction module comprises two convolutional layers, two regularization layers and two activation layers, wherein the size of a convolution kernel is set to be 1 when the first layer is convolved, the size of the convolution kernel is set to be 2 multiplied by 2 when the second layer is convolved, the size of the convolution kernel is set to be 0, and the size of the convolution kernel is set to be 2 multiplied by 2;
The optimal beamforming vector determining module is used for inputting the current channel state information into the trained convolutional neural network model to determine an optimal beamforming vector;
and the near-field beam forming module is used for adjusting the transmitting signals of the antenna array based on the optimal forming vector to form directional beams so as to finish near-field beam forming.
In a third aspect, the present application provides a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the computer program to implement the steps of the depth learning based codebook-less near field beamforming method as described in any of the above.
In a fourth aspect, the present application provides a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the deep learning based codebook-less near field beamforming method of any one of the above.
In a fifth aspect, the present application provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the codebook-less near-field beamforming method based on deep learning as defined in any one of the above.
According to the specific embodiment provided by the application, the application discloses the following technical effects:
The application provides a codebook-free near-field beamforming method, a codebook-free near-field beamforming device, codebook-free near-field beamforming equipment, codebook-free near-field beamforming medium and codebook-free near-field beamforming products, which are used for effectively extracting characteristics of channel state information by a convolutional neural network model through ingenious design of a convolutional kernel and a filling size in the convolutional neural network model, so that an optimal beamforming vector is derived. The application relies on the powerful data processing capacity and the high-efficiency parallel computing capacity of the convolutional neural network model, which not only can effectively reduce the cost brought by beam training, but also can obtain higher beam forming precision.
Detailed Description
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
With the rapid development of mobile communication technology, the 6G communication technology is becoming a research hotspot for the next generation wireless communication system. The 6G technology aims at realizing higher spectral efficiency, lower power consumption and higher antenna number and density, so as to meet the requirements of the fields of future intellectualization, internet of things and the like for high speed, low delay and high reliability. To achieve these objects, a very large-scale antenna array technology is considered as one of the important technologies of 6G.
In conventional wireless communication systems, antenna array technology is mainly focused on far field communication, i.e., where the signal propagation distance is long. In this case, the signal wavefront can be approximated as a plane wave, simplifying the process of channel modeling and beamforming. However, as the communication band shifts to higher frequency bands (such as millimeter waves and terahertz), the communication distance becomes shorter, the number of antennas increases significantly, and the rayleigh distance increases gradually, which results in a transition of signal propagation from far field to near field characteristics. In a near-field environment, the signal wave front shows spherical wave characteristics, and the traditional far-field channel model and the wave beam forming method are not applicable any more. However, the near field communication has the problems of complex channel modeling, complex channel state information, large beam training overhead and the like.
The application designs a convolutional neural network model, and can efficiently extract and process near-field channel characteristics through offline training. By means of offline training and online reasoning of the convolutional neural network model, rapid beam forming is achieved, the overhead of beam training is obviously reduced, and the problems existing in the existing near field communication are solved.
In order that the above-recited objects, features and advantages of the present application will become more readily apparent, a more particular description of the application will be rendered by reference to the appended drawings and appended detailed description.
In an exemplary embodiment, as shown in fig. 1, there is provided a codebook-free near-field beamforming method based on deep learning, including the following steps S1 to S3. Wherein:
The method comprises the steps of S1, constructing and training a convolutional neural network model, wherein the convolutional neural network model comprises a plurality of feature extraction modules, each feature extraction module comprises two convolutional layers, two regularization layers and two activation layers, the size of a convolution kernel is set to be 1 when the first layer is convolved, the size of the convolution kernel is set to be 2 multiplied by 2 when the second layer is convolved, the size of the convolution kernel is set to be 0, and the size of the convolution kernel is set to be 2 multiplied by 2.
S2, inputting the current channel state information into a trained convolutional neural network model, and determining an optimal beamforming vector.
And S3, adjusting the transmitting signals of the antenna array based on the optimal shaping vector to form a directional beam so as to complete near-field beam shaping.
Before step S1, the method further comprises the steps of constructing a near field communication model between the base station and the user and collecting channel state information. In fig. 2, rn is a distance between a user and an nth antenna unit, r is a distance between a central antenna and the user, Dn is a distance between the nth antenna unit and the central antenna, θn is an angle between the nth antenna unit and the user, and θ is an angle between the central antenna and the user.
Specific:
Executing the main body, namely computer simulation software.
The input is the number of users K, the signal to noise ratio SNR, the number of antenna units N, and the normalized transmitting power P.
Operation the downstream beam training of a large-scale multiple-input single-output (MISO) system was studied. A Base Station (BS) is equipped with a linear array (ULA) of N antenna elements (i.e. antenna elements) with a spacing d=λ/2, λ representing the carrier wavelength. Whereas millimeter wave communication (i.e., near field communication) relies primarily on line of sight (LOS) links, the present application contemplates a single LOS Radio Frequency (RF) chain.
When the distance between the antenna array and the receiver is smaller than the Rayleigh Distance (RD), the receiver is located in the near field region. The equation for the rayleigh distance is:
where D represents the maximum size of the antenna array, λ represents the carrier wavelength, and f represents the carrier frequency. Let the distance from the center antenna to the user be r and the angle be θ. Considering the characteristics of the near-field spherical wave, the distance rn between the UE and the nth antenna element is:
the downlink channel hn of the subcarrier between the nth antenna element and the UE can be expressed as:
Wherein βn is the nth antenna element path attenuation coefficient.
By stacking the channels of all antenna elements into one vector, channel state information hk of the kth UE can be obtained:
hk＝[h1,h2,…,hn…,hN]
it can be seen that the near field antenna response is a function of angle and distance. Based on this, the near field steering vector can be represented as:
The downlink received signal yk of the kth UE may be expressed as:
Wherein G represents antenna unit power, w represents a beamforming vector, H represents a transpose, Hk is acquired channel state information of the kth user, s is a downlink transmission signal, and η is noise and interference obeying complex gaussian distribution. Assuming that the downlink transmit signal is |s|2 =1, the achievable rate of the kth UE can be expressed as:
wherein hi is the acquired channel state information of the ith user,Is the signal to noise ratio of the user.
In the above formula, other non-serving UE interference is considered noise. Channel state information acquired from the base station and the user according to the near field communication model is used as input of a convolutional neural network model.
Due to the close correlation of the near field channel with the angle and distance information, near field codebooks are typically large in size, introducing significant computational complexity when codebook-based beam planning is employed. In a practical scenario, the main focus is on beam gain towards the target UE location, rather than optimizing the beam gain per grid for the whole space. The application does not need a predefined codebook, and realizes the effective control of the beam forming shape through deep learning so as to maximize the achievable data transmission rate. Considering the normal mode constraint |wi|2 =1, for i=1..the optimization problem for the N, beamforming vector w is expressed as:
s.t.wi∈w,|wi|2=1
The channel state information in XL-MIMO systems is complex, such that the above-mentioned problems have significant nonlinearities and non-convexities. The present application introduces deep learning as a powerful tool to solve these problems. An unsupervised learning convolutional neural network model is designed, so that the beamforming vector can be flexibly adapted and learned under the complex channel condition. The solution aims at finding the mapping function Φξ of the parameters ζ, which uses the channel state information to predict the optimal beamforming vector w'. The mapping function can be expressed as:
Φξ{h}→{w'}
Where h represents channel state information of the user.
In a specific embodiment, the training process of the convolutional neural network model in the step S1 comprises the steps of inputting acquired channel state information into the convolutional neural network model, outputting equivalent phases of beamforming coefficients, converting the equivalent phases into beamforming vectors by adopting an Euler formula, calculating user reachable rates based on the beamforming vectors and the acquired channel state information, constructing a loss function based on the user reachable rates, and adjusting parameters of the convolutional neural network model through the loss function to complete training of the convolutional neural network model. Specific:
() Design convolutional neural network model
Execution subject, computer
The operation is that the dimension of input data is 2 XN, and the complex signal characteristics are extracted by adding the additional dimension to be converted into 1 X2 XN and then input into the convolutional neural network of the upper graph.
As shown in fig. 3, each feature extraction module includes two convolutional layers, two regularization layers (BatchNorm), and two activation layers (ReLU). The padding is set to 1 in the first layer convolution, the convolution kernel size is 2 x2, and the padding is set to 0 in the second layer convolution, the convolution kernel size is 2 x 2. By skillfully designing the convolution kernel and the filling size, the convolution neural network can effectively extract real part information, imaginary part information and equivalent phases (combination of real parts and imaginary parts) of channel state information, and meanwhile, the dimension of the data is ensured to be unchanged before and after the feature extraction module processes. After the data passes through a feature extraction module, carrying out averaging pooling to carry out downsampling, repeating the process twice, changing the data size into BxC x2 x N/4, then carrying out upsampling twice and feature extraction module processing twice, and restoring the final data size into Bx1 x2 x N which is the same as the original input data size. And then flattening the data into a one-dimensional vector, and mapping the data to an equivalent phase by using a 'Tanh' layer as an output after passing through the full connection layer.
(2) Design loss function
Execution subject, computer
The operation is that an Euler formula is applied to convert the equivalent phase alpha of the convolutional neural network output into the required beamforming vector, and the complex output is given by the following formula:
w'=exp(j·α·π)=cos(πα)+j·sin(πα)
wherein, theIt can be seen that pi alpha has a definite physical meaning, the elements of which correspond to the equivalent phases of the beamforming coefficients in w'. Because the application adopts a deep learning method without codebook, the neural network training is guided by the loss function directly related to the beam forming target. The loss function of a task is defined as the negative of the achievable rate:
Where Q represents the total number of training samples, i.e. the total number of acquired channel state information, Q is the number of acquired channel state information, K represents the total number of users,And w' is a beamforming vector obtained based on a convolutional neural network model in the training process.
(3) Model training
Execution body high performance computer or GPU server
The operation is that the collected channel state information is input into a convolutional neural network model, and the real part and the imaginary part are stacked as input data of the convolutional neural network model to be trained after regularization. The architecture of the convolutional neural network model comprises a plurality of convolutional layers and a full-connection layer, so as to capture complex channel characteristics, and an output result is converted into an equivalent phase through a 'Tanh' layer. And converting the equivalent phase into a beam forming vector by using an Euler formula, calculating the loss by adopting a formula of a loss function, and updating the convolutional neural network model parameters by back propagation. As training proceeds, the loss function decreases and the user achievable rate increases, and when the convolutional neural network model converges, the loss function tends to stabilize and the user achievable rate is maximized. The convolutional neural network model parameters are saved for subsequent deployment.
In a specific embodiment, the trained convolutional neural network is deployed online. Specifically:
execution subject base station
The operation is that as shown in figure 4, the trained convolutional neural network model is deployed in a base station, the current channel state information is input in real time in the communication process, and the base station can rapidly output the optimal beamforming vector reaching the maximum sum rate according to the trained convolutional neural network model. And the base station adjusts the transmitting signals according to the output to form directional beams so as to complete beam forming.
The technical effect of the method is verified through simulation experiments, and an equidistant half-wave linear array with n=256 is arranged at the BS. The application is compared with an exhaustive beam forming scheme with the near field layering, the far field layering and the calculation cost limited to 256, and the accessibility is used as an evaluation index.
Fig. 5 shows the change in the achievable rate with signal-to-noise ratio for different beam training schemes, where the signal-to-noise ratio increases from-20 dB to 20dB. Fig. 5 clearly shows that the proposed solution is always superior to the existing far-field and near-field beam training solutions over a wide range of signal-to-noise ratio values. When the signal-to-noise ratio is 20dB, the reachability of the scheme provided by the application is respectively improved by about 18%, 30% and 120% compared with the near-field hierarchical beam forming scheme, the far-field hierarchical beam forming scheme and the detailed searching scheme. Notably, when the signal-to-noise ratio is > 5dB, the scheme has a significant performance gain compared to the comparison scheme.
The carrier frequency is a key factor affecting the communication rate, and fig. 6 shows the achievable rate curves of the proposed scheme and the comparison scheme under different carrier frequencies. For all carrier frequencies, the scheme proposed by the present application is always superior to other schemes in terms of the achievable rate. For far field layering schemes, near field propagation is increasingly dominant as the carrier frequency increases, resulting in substantially no increase in the average achievable rate of the far field scheme when the carrier is greater than 40 GHz. The best performance of all solutions was observed at a carrier frequency of 50 GHz. This is because the free space propagation loss generated during transmission of the high frequency signal increases, and as the carrier frequency increases, the performance gain obtained by increasing the number of antennas will not be sufficient to cancel the increasing free space propagation loss. When fc >50GHz, the performance of the comparison scheme is degraded, while the performance of the proposed solution remains relatively stable, confirming the robustness of the proposed solution.
Fig. 7 shows the dependence of the achievable rate on the UE-to-BS distance. When the UE approaches the BS, the scheme provided by the application has superior rate performance compared with the comparison scheme. For far field schemes, the advantage of near field propagation results in a rapid decrease in the average transmission rate of the far field scheme as distance decreases. Due to the limitation of the maximum training cost, the exhaustive solution is severely degraded in the near field region, preventing effective search of near field locations. Furthermore, the near field layered beam training scheme may moderately mitigate the drop in average achievable rate, but gradually degrades when the distance exceeds 30 m. Furthermore, when the distance is less than 20 meters, the performance thereof is significantly delayed from the proposal of the application. In contrast, the present application demonstrates the ability to search for the best beamforming vector with minimal pilot overhead. This demonstrates the robustness of the proposed solution in near-field and far-field communication scenarios.
Fig. 8 shows the achievable rate as a function of angle. Here, θ gradually increases from-60 ° to 60 °. For far field stratification schemes, performance drops significantly as θ approaches 0 degrees. This degradation is due to the enhancement of near field effects near zero angle. Furthermore, the near field layered beam training scheme shows significant fluctuations, because the near field hybrid scheme creates a near field codebook by uniformly sampling angles and distances in a cartesian coordinate system, which is challenging to achieve stable beamforming performance throughout the near field environment. In contrast, the proposed solution of the present application always achieves the highest average achievable rate over all considered angular ranges. This result highlights the robustness and effectiveness of the application in handling challenges brought by angle changes in near field communication scenarios.
Based on the same inventive concept, the embodiment of the application also provides a device for realizing the codebook-free near-field beamforming method based on the deep learning. The implementation of the solution provided by the device is similar to the implementation described in the above method, so the specific limitation in the embodiments of one or more codebook-free near-field beamforming devices based on deep learning provided below may be referred to above for the limitation of the codebook-free near-field beamforming method based on deep learning, which is not repeated here.
In one exemplary embodiment, there is provided a codebook-free near field beamforming apparatus based on deep learning, including:
The model building and training module is used for building and training a convolutional neural network model, the convolutional neural network model comprises a plurality of feature extraction modules, each feature extraction module comprises two convolutional layers, two regularization layers and two activation layers, the size of a convolution kernel is set to be 1 when the first layer is convolved, the size of the convolution kernel is set to be 2 multiplied by 2 when the second layer is convolved, the size of the convolution kernel is set to be 0, and the size of the convolution kernel is set to be 2 multiplied by 2.
And the optimal beamforming vector determining module is used for inputting the current channel state information into the trained convolutional neural network model to determine an optimal beamforming vector.
And the near-field beam forming module is used for adjusting the transmitting signals of the antenna array based on the optimal forming vector to form directional beams so as to finish near-field beam forming.
In an exemplary embodiment, a computer device is provided, comprising a memory and a processor, the memory having stored therein a computer program, the processor performing the steps of the method embodiments described above when the computer program is executed. The computer device may be a server or a terminal, and its internal structure may be as shown in fig. 9. The computer device includes a processor, a memory, an Input/Output interface (I/O) and a communication interface. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface is connected to the system bus through the input/output interface. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device is used for storing data to be processed. The input/output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used for communicating with an external terminal through a network connection. The computer program, when executed by a processor, implements a codebook-less near-field beamforming method based on deep learning.
It will be appreciated by persons skilled in the art that the architecture shown in fig. 9 is merely a block diagram of some of the architecture relevant to the present inventive arrangements and is not limiting as to the computer device to which the present inventive arrangements are applicable, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, comprising a memory and a processor, the memory having stored therein a computer program, the processor performing the steps of the method embodiments described above when the computer program is executed.
In an exemplary embodiment, a computer-readable storage medium is provided, in which a computer program is stored which, when being executed by a processor, carries out the steps of the method embodiments described above.
In an exemplary embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of the method embodiments described above.
It should be noted that, the user information (including but not limited to user information, user personal information, etc.) and the data (including but not limited to data for analysis, stored data, presented data, etc.) related to the present application are both information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data are required to meet the related regulations.
Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magneto-resistive random access Memory (Magnetoresistive RandomAccess Memory, MRAM), ferroelectric Memory (Ferroelectric RandomAccess Memory, FRAM), phase change Memory (PHASE CHANGE Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (RandomAccess Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in various forms such as static random access memory (Static RandomAccess Memory, SRAM) or dynamic random access memory (Dynamic RandomAccess Memory, DRAM), etc.
The databases referred to in the embodiments provided herein may include at least one of a relational database and a non-relational database. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processor referred to in the embodiments provided in the present application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, or the like, but is not limited thereto.
The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.
The principles and embodiments of the present application have been described herein with reference to specific examples, which are intended to facilitate an understanding of the principles and concepts of the application and are to be varied in scope and detail by persons of ordinary skill in the art based on the teachings herein. In view of the foregoing, this description should not be construed as limiting the application.