CN113381950A

CN113381950A - Efficient MIMO channel feedback method and device based on network aggregation strategy

Info

Publication number: CN113381950A
Application number: CN202110447922.7A
Authority: CN
Inventors: 王劲涛; 陆智麟; 张彧; 张超
Original assignee: Tsinghua University
Current assignee: Tsinghua University
Priority date: 2021-04-25
Filing date: 2021-04-25
Publication date: 2021-09-10
Anticipated expiration: 2041-04-25
Also published as: CN113381950B

Abstract

The invention discloses a high-efficiency MIMO channel feedback method and a device based on a network aggregation strategy, comprising the following steps: training the aggregation feedback network, deploying a self-encoder and a quantization module in the trained aggregation feedback network to a client, and deploying a self-decoder and a quantization module in the trained aggregation feedback network to the client; performing channel estimation through a user side to obtain a downlink channel matrix, performing Fourier transform twice on the downlink channel matrix to transform the downlink channel matrix from a space-frequency domain to an angle-time delay domain, and intercepting a non-zero sub-matrix of the transformed matrix; compressing the non-zero sub-array through a self-encoder and a quantization module to obtain a characteristic vector, and sending the characteristic vector to a base station end; decoding the characteristic vector through a quantization module and a self-decoder of a base station end; and carrying out zero filling and inverse discrete Fourier transform on the decoded matrix to obtain a downlink channel matrix. The method has the advantages of high feedback precision, good network elasticity and flexible deployment, and realizes the MIMO channel compression feedback with high precision and low expense.

Description

Efficient MIMO channel feedback method and device based on network aggregation strategy

Technical Field

The invention relates to the technical field of data transmission, in particular to a high-efficiency MIMO channel feedback method and device based on a network aggregation strategy.

Background

With the rapid development of 5G, the massive MIMO technology is gradually becoming a competitive mainstream technical solution. The feedback precision of the MIMO system based on the codebook is seriously limited, and the performance can be further reduced under the condition of large-scale MIMO; the full-channel feedback has better performance, but the feedback overhead is too large, and the overhead of the full-channel feedback is further increased along with the expansion of the scale of the MIMO system. Therefore, full-channel compressed feedback is a promising feedback technique for MIMO systems.

However, due to the insufficient sparsity of the MIMO downlink channel, especially the MIMO downlink channel in the angle-time delay domain, it is difficult to implement higher-precision channel compression feedback for the conventional compressed sensing technology. And a compression feedback network composed of an auto-encoder and an auto-decoder based on deep learning can well learn the characteristics of a channel, so that high-precision practical compression feedback is realized.

The existing compressed feedback network is generally rigid in design, cannot realize simple expansion and good deployment, and the feedback performance of the existing compressed feedback network needs to be further improved.

Disclosure of Invention

The present invention is directed to solving, at least to some extent, one of the technical problems in the related art.

Therefore, one objective of the present invention is to provide an efficient MIMO channel feedback method based on a network aggregation strategy, which has high feedback accuracy, good network elasticity, flexible deployment, and implements MIMO channel compression feedback with high accuracy and low overhead.

Another objective of the present invention is to provide an efficient MIMO channel feedback apparatus based on a network aggregation strategy.

In order to achieve the above object, an embodiment of the present invention provides an efficient MIMO channel feedback method based on a network aggregation policy, including:

training an aggregation feedback network, deploying a self-encoder and a quantization module in the trained aggregation feedback network to a user side, and deploying a self-decoder and a quantization module in the trained aggregation feedback network to the client side;

performing channel estimation through a user side to obtain a downlink channel matrix, performing Fourier transform twice on the downlink channel matrix to transform the downlink channel matrix from a space-frequency domain to an angle-time delay domain, and intercepting a non-zero sub-matrix of the transformed matrix;

compressing the non-zero subarray through a self-encoder and a quantization module to obtain a characteristic vector, and sending the characteristic vector to a base station end;

decoding the characteristic vector through a quantization module and a self-decoder of a base station end;

and carrying out zero filling and inverse discrete Fourier transform on the decoded matrix to obtain the downlink channel matrix.

In order to achieve the above object, an embodiment of another aspect of the present invention provides an efficient MIMO channel feedback apparatus based on a network aggregation policy, including:

the training module is used for training the aggregation feedback network, deploying a self-encoder and a quantization module in the trained aggregation feedback network to a user side, and deploying a self-decoder and a quantization module in the trained aggregation feedback network to the client side;

the system comprises a user side preorder module, a data acquisition module and a data acquisition module, wherein the user side preorder module is used for performing channel estimation through a user side to obtain a downlink channel matrix, performing Fourier transform twice on the downlink channel matrix to transform the downlink channel matrix from a space-frequency domain to an angle-time delay domain, and intercepting a non-zero sub-matrix of the transformed matrix;

the compression module is used for compressing the non-zero subarray through a self-encoder and a quantization module to obtain a characteristic vector, and the characteristic vector is sent to a base station end;

the decompression module is used for decoding the characteristic vector through a quantization module and a self-decoder of the base station end;

and the base station end follow-up module is used for carrying out zero filling and inverse discrete Fourier transform on the decoded matrix to obtain the downlink channel matrix.

The high-efficiency MIMO channel feedback method and device based on the network aggregation strategy have the following advantages that:

1) through a specially designed neural network self-encoder and a specially designed neural network self-decoder, the characteristic learning and the characteristic compression of the MIMO channel, especially a large-scale MIMO channel are realized, so that the feedback overhead is greatly reduced.

2) Compared with the design based on a naive convolutional neural network, the feedback network design based on the network aggregation strategy has the characteristics of high feedback precision, good network elasticity and flexible deployment, and the proposed aggregation compression feedback network can provide a high-efficiency adaptive scheme aiming at the resource limitations of different user sides and base station sides, thereby realizing the MIMO channel compression feedback with high precision and low cost.

3) From the application example, the aggregation feedback network in the invention can be used for channel compression feedback of the MIMO system under various conditions. Particularly for Frequency Division Duplex (FDD) systems, the aggregation network provided by the invention can keep the accuracy of the fed-back MIMO channel under extremely high compression rate.

Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

The foregoing and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

fig. 1 is a flow chart of an efficient MIMO channel feedback method based on a network aggregation strategy according to an embodiment of the present invention;

fig. 2 is a flow chart of an efficient MIMO channel feedback method based on a network aggregation strategy according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of an aggregate feedback network setup according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of an implementation of an aggregated feedback network according to an embodiment of the present invention;

fig. 5 is a schematic structural diagram of an efficient MIMO channel feedback apparatus based on a network aggregation strategy according to an embodiment of the present invention.

Detailed Description

Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.

The following describes an efficient MIMO channel feedback method and apparatus based on a network aggregation policy according to an embodiment of the present invention with reference to the accompanying drawings.

First, an efficient MIMO channel feedback method based on a network aggregation policy proposed according to an embodiment of the present invention will be described with reference to the accompanying drawings.

Fig. 1 is a flowchart of an efficient MIMO channel feedback method based on a network aggregation strategy according to an embodiment of the present invention.

As shown in fig. 1, the efficient MIMO channel feedback method based on the network aggregation strategy includes the following steps:

and step S1, training the aggregation feedback network, deploying the self-encoder and the quantization module in the trained aggregation feedback network to a user side, and deploying the self-decoder and the quantization module in the trained aggregation feedback network to the client side.

Further, training the aggregated feedback network further comprises:

collecting and training a downlink channel matrix;

transforming the training downlink channel matrix from a space-frequency domain to an angle-time delay domain through two discrete Fourier transforms, and intercepting non-zero sub-matrices of the transformed matrix to form a training data set;

and (3) according to the resource limitations of the actual user side and the base station side and the requirements on time delay and feedback precision, stretching the aggregation feedback network, selecting a proper quantization scheme, and training the aggregation feedback network based on the training data set.

Further, acquiring a training downlink channel matrix, comprising:

and splicing downlink channels on all the subcarriers into a complete training downlink channel matrix according to the number of the OFDM subcarriers and the number of the base station-side antennas.

Further, acquiring a training downlink channel matrix, comprising:

and training a downlink channel matrix through model simulation or acquisition in a physical environment.

Specifically, the number N of OFDM subcarriers is determined according to system conditions_cAnd the number of base station side antennas N_tSplicing the downlink channels on all the sub-carriers into a complete downlink channel matrix

Through two discrete Fourier transforms

And after the space-frequency domain is converted into the angle-time delay domain, intercepting the non-zero submatrix H as the input of the compression network. Raw data

According to actual needs, the model can be generated through simulation of the model or directly collected in a physical environment.

And combining the resource limitations of the actual user side and the base station side, comprehensively considering the requirements on time delay and feedback precision, properly extending and contracting the aggregation feedback network, and selecting a proper quantization scheme. And finishing the training of the aggregation feedback network based on the data set generated above. Deploying the trained self-encoder to a user side; and deploying the trained self-decoder to the base station side.

Further, the aggregation feedback network comprises group convolution, and channel feature extraction is carried out through the group convolution.

Specifically, in the embodiment of the present invention, a group convolution operation is introduced in the MIMO compression feedback based on the neural network, and the aggregation network design mainly utilizes group convolution (group convolution) to complete higher-level and more flexible channel feature extraction. On one hand, diversified channel characteristics can be learned in parallel through packet convolution, and higher-precision channel feedback is completed; on the other hand, the complexity of the feedback network can be conveniently and effectively adjusted by adjusting the group number of the grouping convolution, so that the method is suitable for different hardware resource limitations.

The aggregated feedback network may include other conventional neural network components besides group convolution (network aggregation operations), including but not limited to fully connected layers, normal convolutional layers, various normalization layers (e.g., batch normalization), various activation function layers (e.g., ReLU, leakyreu, Sigmoid, etc.), various attention mechanism layers, and the like.

It will be appreciated that the present invention can be adapted to FDD systems as well as TDD systems.

Further, training the aggregated feedback network further comprises:

selecting the expansion degree of the elastic aggregation feedback network, and selecting a compression ratio and a feature vector quantization ratio combination according to the feedback overhead limit;

determining the aggregation design of the aggregation feedback network according to the determined expansion degree and the compression quantization rate;

training is performed using an Adam optimizer and MSE loss function along with a training data set.

Further, a learnable adaptive activation function parametricalrlu is included in the aggregated feedback network.

The invention comprises the effective application of a new activation function in MIMO channel compression feedback, and each branch in a polymerization network can obtain greater characteristic learning autonomy by learning the default negative half-axis gradient of the LEAKYRELU, thereby obtaining higher feedback precision.

The strategy for resisting overfitting in the MIMO channel compression feedback training based on the aggregation network can resist overfitting pressure brought by expansion of the aggregation compression network by setting a smoother learning rate strategy and increasing network regularization.

Several extensions to the aggregate autoencoder are directed to the MIMO channel compression feedback for the aggregation network. Including using parallel fully-connected layers to enforce dense feature learning capabilities, using lowlevelskipconnection to implement multi-level feature combinations, etc. These enhancements can achieve higher feedback accuracy at a lower cost and can be used when the user side resources are more abundant.

And step S2, performing channel estimation through the user terminal to obtain a downlink channel matrix, performing Fourier transform twice on the downlink channel matrix to transform the downlink channel matrix from a space-frequency domain to an angle-time delay domain, and intercepting a non-zero sub-matrix of the transformed matrix.

Specifically, as shown in fig. 2, during actual feedback, the ue obtains the downlink channel through channel estimation

And transforming to an angle-time delay domain and intercepting a non-zero sub-array H. And sending the H into a self-encoder based on a polymerization network to obtain a compressed feature vector v. The base station only needs to feed back v but not H, so that the feedback overhead can be obviously reduced; after receiving the characteristic vector v at the base station end, sending the characteristic vector v to a self-decoder based on a polymerization network, thereby recovering an original non-zero angle-time delay domain subarray H, and acquiring an original downlink channel through zero filling and inverse discrete Fourier transform

The MIMO channel is transformed from the space-frequency domain to the angle-time delay domain by a discrete fourier transform. Since the time delay of multipath is in a certain range, the transformed angle-time delay domain MIMO channel is highly sparse in the time delay dimension. Since most elements in the angle-time delay domain channel matrix are zero, only the non-zero submatrices of the angle-time delay domain channel matrix need to be intercepted and fed back. The aggregation feedback network can utilize the self-encoder to compress the non-zero subarray at the user end to further reduce the feedback capacity, and then utilize the self-decoder to recover the original non-zero subarray of the degree-time delay domain channel at the base station end. And finally, obtaining an original channel matrix through zero filling and two times of inverse discrete Fourier transform.

And step S3, compressing the non-zero sub-array through the self-encoder and the quantization module to obtain a characteristic vector, and sending the characteristic vector to the base station.

Specifically, the feature vector v compressed by the self-encoder is directly and uniformly quantized from the original 32-bit floating point number to the k-bit fixed point number. The value-taking strategy of k is explained in the invention. The actual feedback bit number of the quantized feature vector is greatly reduced, and the feedback overhead can be further reduced.

It will be appreciated that by extending the number of group convolutional groups from the decoder side, an efficient extension to the aggregate self-decoder is achieved. The cost of the user side sensitive to the expanded resources is unchanged; the base station terminal with relatively strong hardware resources gets more accurate feedback effect through higher calculation power and storage consumption. The advantages of the extension strategy proposed in the present invention are the simplicity of the extension scheme (only the number of groups of group convolutions needs to be adjusted) and the high cost-performance of extension (by a reasonable extra cost to trade for a more significant feedback effect gain).

Step S4, the feature vector is decoded by the quantization module and the self-decoder of the base station.

And step S5, carrying out zero filling and inverse discrete Fourier transform on the decoded matrix to obtain a downlink channel matrix.

The efficient MIMO channel feedback method based on the network aggregation strategy according to the embodiment of the present invention is described in detail below.

The first step, determining the original channel needing feedback, Massive MIMO FDD system, with N at the base station end_tAnd (4) setting the user side to only have 1 antenna. Let the number of OFDM subcarriers be N_cThen the user terminal estimates the obtained downlink channel

Dimension of (A) is N_c×N_tSince each channel element is complex, the total size of the feedback required is 2N_cN_t。

And secondly, transforming the original channel to an angle-time delay domain, and intercepting a non-zero sub-array. The original space-frequency domain channel matrix can be formed by two DFTs

And transforming to an angle time delay domain. After transformation, due to the limited multipath complexity, the angle-time delay domain channel matrix H shows high sparsity in the time delay domain. Most elements in H are zero elements or near zero elements, feedback is not needed, and therefore only non-zero submatrices H of H need to be intercepted_aAnd (6) carrying out feedback.

And thirdly, compressing the non-zero sub-array by the user side by using an auto-encoder. Will intercept good nonZero subarray H_aInputting the self-encoder based on the aggregation network structure which is trained by the user side. Self-encoder extractable and compressed non-zero sub-array H based on aggregation network structure_aThereby generating a lower-dimensional feature vector v. Dimension of the feature vector v compared to the original non-zero sub-matrix H_aCan be reduced by several times to dozens of times, thereby greatly reducing the cost of actual feedback.

And fourthly, losslessly feeding back the characteristic vector v to the base station end through digital uplink feedback.

Fifthly, the base station side recovers the non-zero submatrix H from the characteristic vector v by using a self-decoder_a. And the base station sends the received feature vector v to a trained self-decoder based on a polymerization network structure to complete the dimension recovery and feature recovery of the original subarray. Through proper training, the self-decoder based on the aggregation network structure enables the restored non-zero subarrays to be as close as possible to the original non-zero subarrays, and high-precision channel compression feedback is achieved.

And sixthly, the base station end obtains an original channel needing feedback through two times of inverse discrete Fourier transform. Based on the sparsity of the original angle-time delay domain channel, the restored non-zero subarrays are subjected to zero filling and inverse discrete Fourier transform, and new information loss is basically not introduced.

In the intelligent channel compression feedback method provided by the invention, the core is that the self-encoder is designed based on an aggregation network structure, or the self-decoder is designed based on an aggregation network structure, or both the self-encoder and the self-decoder are designed based on an aggregation network structure. Compared with the traditional visual image, the CSI image has rich characteristic patterns, and particularly shows complex diversity in spatial characteristics under the conditions of change of position distribution of a user relative to a base station, change of the number of multi-paths, change of channel scenes and the like. The traditional single convolutional neural network is difficult to learn spatial domain characteristics with too strong diversity, so that the network performance is poor. The aggregation network design proposed by the invention is based on the group convolution idea and can be equivalent to the parallel structure of a plurality of or even dozens of single convolution branches. Each single convolution branch can learn relatively independent CSI spatial domain characteristics, so that the diversity learning capability of the spatial domain characteristics is greatly enriched. And finally, overlapping the diversified spatial domain features to restore feature dimensions, thereby allowing the construction of a deep neural network.

The network aggregation structure provided by the invention also has the following specific design advantages: convenient feedback network elastic expansion and contraction can be realized by simply increasing or decreasing the number of branches in the aggregation structure in the feedback self-encoder or self-decoder. Specifically, when resources (including storage resources and computing resources) of the ue/bs are more abundant than the existing aggregation feedback network, the number of branches of the aggregation structure is increased to use additional resources for obtaining higher feedback accuracy; when resources (including storage resources and computing resources) of the user terminal/base station terminal are insufficient relative to the existing aggregation feedback network, the number of branches of the aggregation structure is reduced so as to provide the best feedback precision under the current resource limitation. The design in the invention is very beneficial to realizing the high-efficiency deployment of the feedback network on the basis of the variable base station resource allocation condition and the diversified user equipment types.

The network aggregation structure provided by the invention also has the following specific design advantages: a learnable adaptive activation function parametricalrlu (PReLU) is introduced into the feedback task and used in combination with the aforementioned aggregation network structure to obtain better feedback performance. In particular, conventional feedback networks mostly employ fixed activation functions, such as ReLU, leakyreu, etc. The form of the fixed activation function is fixed during network design and is not changeable during network training. In the invention, a learnable adaptive activation function PReLU is introduced, and the activation amplitude is obtained through learning and has the capability of adapting to different tasks. In particular, when used in conjunction with a converged network, the PReLU activation functions on different convergence branches are learned independently. In other words, the technology of the invention allows each aggregation branch to learn the own activation function, thereby further strengthening the characteristic diversity learned by the whole aggregation feedback network, and further improving the feedback precision.

The network aggregation structure provided by the invention is also proved to be capable of being used in cooperation with a feature vector quantization technology. By quantizing and de-quantizing the low-dimensional features learned by the aggregation network structure, the feedback network provided by the invention can further greatly reduce the feedback overhead under the same feedback precision. Specifically, the user side needs to quantize the compressed feature vector output from the encoder to form a feature bit stream; after the digital uplink feedback losslessly feeds back the feature bit stream to the base station end, the base station end needs to perform corresponding dequantization on the feature bit stream firstly, and then sends the dequantized feature vector to a self-decoder for dimension and feature recovery.

The network aggregation structure provided by the invention also comprises a plurality of variant designs, such as a design of combining a feature pyramid with a multi-scale feature, and the like, and the variant designs can further improve the feedback performance of the whole aggregation feedback network.

Further, in the embodiment of the present invention, the key point includes building of an aggregation feedback network, and the building process of the aggregation feedback network is described in detail below as shown in fig. 3.

Scheme selection

In the first step, the expansion degree K of the elastic aggregation feedback network is selected. In the aggregation feedback network, the aggregation branch number g is defined as the expansion degree K of the whole aggregation network, and a series of aggregation feedback networks can be generated by adjusting the expansion degree K. When K increases, resource consumption of the feedback network increases, and better feedback performance is obtained at the same time. Therefore, the expansion degree K of the aggregation feedback network needs to be adjusted according to the resource limitations of the user side and the base station side in the actual system, so as to improve the feedback performance as much as possible on the premise of meeting the resource and delay constraints.

And secondly, selecting a compression rate and eigenvector quantization rate combination according to feedback overhead limitation. The aggregation feedback network provided by the invention comprises quantization and dequantization modules, and the compression rate can be freely selected by changing the dimension of the full connection layer. Obviously, the higher the compression rate and the lower the quantization accuracy, the lower the final feedback overhead. However, lower feedback overhead generally means lower feedback accuracy, so that the limitation of the feedback overhead needs to be given according to the feedback accuracy required by the system. In addition, the combination of compression rate and quantization precision jointly determines the feedback overhead, in other words, there are a plurality of different combinations of compression rate and quantization precision under the unified feedback overhead. Therefore, after the feedback overhead limit is determined, an optimal combination of compression rate and quantization precision needs to be further given, so as to achieve a better feedback effect.

Off-line training

And thirdly, determining the aggregation design of the aggregation feedback network according to the expansion degree K and the compression quantization rate determined in the first step and the second step. The general self-encoder is composed of a simple convolution, an aggregation convolution and a full connection layer, and then a dequantization module; the self-decoder is composed of a full connection layer, a simple convolution and an aggregation convolution, and a front connection dequantization module.

And fourthly, training by using an Adam optimizer and an MSE loss function. Wherein the data can be generated from a channel model and can also be collected by the device in an actual communication scene. The latter is more cost and effective than the former. Where the data in the training set and the test set need to be generated/collected independently.

Online deployment

And fifthly, deploying the user side. The trained self-encoder and the corresponding quantizer are deployed at a user end, wherein the quantizer itself is trained in cooperation with the self-encoder, and belongs to the category of perceptual Quantization Training (QAT).

And sixthly, deploying the base station terminal. And deploying the trained self-decoder to cooperate with a corresponding dequantizer at the base station, wherein the dequantizer and the self-decoder are also trained cooperatively and belong to the category of QAT. In addition, TRT deployment acceleration, INT8 network fixed point and other technologies can be introduced during deployment, so that the cost and delay after deployment are further reduced.

The efficient MIMO channel feedback method based on the network aggregation strategy of the present invention is described below with a specific embodiment.

S101: according to the number N of OFDM subcarriers_cAnd the number of base station side antennas N_tDetermining a downlink channel after system combining

Then, two DFTs are performed to channel the space-frequency domain

Transforming to the angle-delay domain, as shown in the following equation:

wherein

And

are DFT transform matrices. As introduced above, the multipath complexity of the system is limited and therefore the delay falls within a certain range. In other words, in the time delay domain, the channel to be fed back has sparsity, only the non-zero part in the channel needs to be fed back, the part which is approximate to zero can be discarded in advance, and the channel spatial structure can be restored by simply filling zero at the base station end. Based on the above consideration, the first N of the angle-time delay domain matrix H is intercepted_aThe line is used as a channel needing explicit feedback, and the following formula is shown:

H_a＝H[:N_a]

it is worth noting that the original spatial-frequency domain downlink channel

There are different ways of obtaining in different embodiments. A data set is established in a mode of acquiring actual channel information in a scene with high requirements on the online stability of the system; under the condition of limited COST or inconvenient channel acquisition, channel state information can also be directly generated by using a channel model, such as a COST2100, a 3GPP TR.38.901UMi NLOS model, a Saleh-Valenzella model and the like.

S102: based on the requirements of the system on time delay and feedback precision, the design of the compression aggregation feedback network is determined by comprehensively considering the limitations of calculation, storage and power consumption resources. The reduction of the feedback network polymerization degree K can reduce the system delay and the deployment overhead, but can simultaneously reduce the feedback precision. The specific situation of each implementation case needs to be adjusted so as to obtain the optimal balance of the two.

In addition, the limitation of the overhead of the uplink feedback itself in the communication system, that is, how many bits are allowed to be used to describe the original channel in a unit time, needs to be considered. Obviously, the stricter the limitation of the system on the feedback overhead is, the more the feedback precision is damaged, and the more the larger polymerization degree K needs to be introduced to compensate the loss of precision. On the other hand, after the number of uplink feedback bits is limited, each implementation case further needs to specifically determine the compression ratio λ of the feedback network and the quantization precision B of the quantizer. Feedback overhead is proportional to λ B, and it is obvious that there are multiple sets of compression rate and quantization rate combinations under the same feedback overhead, so each specific embodiment needs to determine an optimal combination, thereby giving the best feedback accuracy under the limitation of the system on feedback overhead.

S103: after the system is deployed and on-line, the user end can obtain a downlink channel matrix through channel estimation of a preamble

(during training

Directly obtained from data set sampling) and processed in S101 to obtain the angle-time delay domain non-zero sub-array H_aAs is the channel actually needed for feedback.

H_aThe user end passes through a self-encoder epsilon and a quantization module based on a polymerization network

The self-encoder is composed of a convolution network and a full connection layer. The convolutional network part is mainly responsible for extracting spatial domain characteristics of the CSI matrix. When the resource is allowed, the network design can be aggregated in the convolutional network so as to enhance the performance, and when the resource is not allowed, the simple convolutional layer and residual error structure design can be adopted so as to reduce the overhead.

F_conv＝ε_conv(H_a)

The full connection layer part is mainly responsible for compressing the spatial domain characteristics of the CSI matrix. In a specific example, it is desirable to adjust the dimensions of the matrix according to the designed compression ratio, for example when the compression ratio is

Time, convolution output characteristic F_convDimension of 2 XN_a×N_t(ii) a And a full connection layer epsilon_FCOutput characteristic F_FCDimension of

F_FC＝ε_FC(F_conv)

Finally, a Bbit quantization module is given according to the designed quantization rate

And inputting the data into a quantization module to obtain a characteristic bit stream required by the final uplink feedback.

The ue-specific bitstream generation in the whole embodiment can be summarized by the following formula:

it can be seen that the channel is derived from the original space-frequency domain

Characteristic bit stream v to final digital upstream feedback_BThe information compression in the invention goes through three stages, which are respectively: sparse compression after DFT transformation, feature extraction and compression based on a neural network self-encoder, and bit-level precision compression based on quantization. After three-stage compression, the number of bit streams requiring feedback is reduced

The overall compression rate is determined by the scheme selection in S102.

It is noted that in both the present invention and its specific implementation, ideal digital upstream feedback is assumed. Therefore, the possible interference or noise in uplink feedback does not need to be considered, and only the loss in the process of compressing the CSI original information into the characteristic bit stream information needs to be reduced.

S104: after the system is deployed and on-line, the base station end receives the characteristic bit stream v from the ideal digital uplink feedback in S103_BAnd input it to a dequantizer of corresponding precision

Thereby obtaining a feature vector v of full precision.

And then inputting the feature vector v with full precision into an aggregation network self-decoder deployed at a base station end. The self-decoder is mainly composed of a fully-connected layer and a subsequent convolutional network. Wherein the full connection layer

Mainly responsible for recovering the compressed dimension of the CSI matrix from the feature vector v, and in each specific example, the input and output dimensions thereof need to be equal to epsilon in S103_FCThe opposite is true.

While convolutional networks

It is composed of several aggregation network modules (one possible design of aggregation network module is shown in fig. 4), which contain tens of equivalent parallel convolution branches, and can well learn and recover the original oneChannel spatial domain information.

The channel recovery whole flow at the base station end can be described by the following formula:

after the intelligent recovery of the non-zero subarray in the angle-time delay domain is completed, the base station end can obtain the initial spatial-frequency domain downlink channel matrix through post-processing. Firstly, the non-zero subarrays obtained by recovery are correspondingly zero-filled, thereby recovering N_c×N_tChannel dimension of (c):

then, performing inverse DFT operation twice on the complete angle-time delay domain channel matrix, and finishing post-processing to obtain a recovery value of the original space-frequency domain channel matrix:

wherein

And

is the corresponding inverse discrete fourier transform matrix.

Any specific embodiment of the present invention includes four overall steps S101 to S104 as described above, where S101 and S102 are offline scheme selection and network training; s103 and S104 are online post-deployment network reasoning, i.e., actual channel compression and decompression processes. The online reasoning section can be characterized in its entirety by the following set of summary formulas.

The core design of the invention is that a convolution network structure which is more suitable for completing CSI characteristic extraction and dimension compression is designed through a network aggregation structure. Greater epsilon_convAnd

higher feedback accuracy can be provided. More specific application and design details are set forth in the following detailed description.

The efficient MIMO channel feedback method based on the network aggregation strategy will be further described below by way of specific embodiments.

In a first embodiment of the invention:

s1: consider the following system settings:

for an FDD massive MIMO system, the number of OFDM subcarriers is N_c1024, the base station side antenna array size is N_t32. Its downlink channel matrix

Is 1024 × 32. A channel is generated based on an indoor scene in a COST2100 model, a linear antenna array is adopted at a base station end, and default values are adopted for all other parameters. When the channel sparsity of the angle-time delay domain can support N_a16, non-zero sub-array truncated after discrete fourier transform

The dimension is 2 × 16 × 32 after the floating-point number matrix is converted.

S2: consider the following aggregate feedback network design:

for the Indor scene, the channel complexity is low, the channel information entropy is small, and therefore, the excessively complex network design is not needed. The self-encoder considers a cascade simple residual convolution structure and a full connection layer; the self-decoder considers cascading fully-connected layers and two aggregation network structures. The aggregate expansion K is taken to be 40 in conjunction with base station end deployment constraints. All convolution layers are followed by batch normalization layers and activation function layers. The activation function layer adopts the adaptive activation function PReLU:

where α is a learnable negative activation gradient, when α is fixed, the PReLU degenerates to the generic LeakyReLU. It is worth noting here that different independent alpha values can be learned at each convolution channel, thereby maximally learning independent rich CSI features.

The feedback precision and the deployment requirement are integrated, and the compression ratio is taken as

And taking the quantization precision as B-2. Whereby the dimension of the fully connected layer of the self-encoder is

Dimension of the self-decoder full connection layer is

Overall compression ratio of

From the original 2 × 1024 × 32 × 32 bits to 256 bits. Experiments show that the scheme can obtain the feedback precision of more than-10 dB after being properly trained.

S3: consider the following aggregate feedback network training scheme:

and (3) performing feedback network training by adopting an Adam optimizer and an MSE loss function. To combat overfitting, the weight decay was taken to be 1 e-5. And a smooth cosine preheating learning rate adjustment mode is adopted, so that the network can be converged and fitted better. During training, the batch size is about 100-400 according to the training equipment.

During training, the quantization and dequantization modules need to be integrated into the network for collaborative training, so that the loss of quantization on precision is reduced as much as possible, a certain degree of regularization effect is achieved, and the overfitting of the network is resisted.

S4: the online reasoning specifically comprises the following steps:

firstly, the self-encoder and the self-decoder trained in S3 are respectively deployed to a user terminal and a base station terminal, and then a downlink channel matrix obtained by channel estimation is arranged at the user terminal

Obtaining the data to be compressed by two DFT and non-zero sub-array interception

Finally, H is put_aInputting the deployed self-encoder and quantizer to obtain a characteristic bit stream v_B∈{0,1}²⁵⁶。

Characteristic bit stream v_BThe ideal digital uplink feedback is transmitted to a base station end, and then the ideal digital uplink feedback is sent to a corresponding de-quantizer and a self-decoder to recover the non-zero sub-array of the angle-time delay domain

Finally, the channel of the original space-frequency domain can be obtained through the corresponding zero filling operation and two times of inverse DFT

In a second embodiment of the invention:

s1: consider the following system settings:

Is 1024 × 32. Generating a channel based on an outdoor scene in a COST2100 model, wherein a base station end adopts a linear antenna array and all other parametersThe number assumes a default value. At the moment, the channel sparsity of the angle-time delay domain intelligently supports N_a32, non-zero sub-array truncated after discrete fourier transform

The dimension is 2 × 32 × 32 after the floating-point number matrix is converted.

S2: consider the following aggregate feedback network design:

for an Outdoor scene, the channel complexity is high, the channel information entropy is large, and the feature extraction compression is difficult. Therefore, there is a need for an enhanced network design to maintain higher accuracy of feedback. The self-encoder considers the aggregation network module with narrower cascade connection and the full connection layer; the self-decoder considers cascading fully-connected layers and three aggregation network modules. In addition, in order to enhance the feature extraction capability of the self-encoder of the outdor scene, the full connection layer is divided into two parts, the original channel feature and the channel feature after the abstraction of the convolution network are respectively received, and the two parts are added to serve as the final feature to be output to the quantizer.

In combination with base station end deployment constraints, we derive from the encoder aggregate expansion K₁8; self-decoder aggregate expansion K₂80. Similarly, all convolution layers are followed by batch normalization and activation function layers. The activation function layer adopts the adaptive activation function PReLU:

By integrating the feedback precision and the deployment requirement, the compression ratio is taken as

And taking the quantization precision as B-4. Whereby the dimension of the fully connected layer of the self-encoder is

Dimension of the self-decoder full connection layer is

Overall compression ratio of

Compressed from the original 2 × 1024 × 32 × 32 bits to 2048 bits. Experiments show that the scheme can obtain the feedback precision of more than-10 dB after being properly trained.

S3: consider the following aggregate feedback network training scheme:

S4: the online reasoning specifically comprises the following steps:

Finally, H is put_aInputting the deployed self-encoder and quantizer to obtain a characteristic bit stream v_B∈{0,1}²⁰⁴⁸。

Characteristic bit stream v_BFirst transmitted by ideal digital uplink feedbackTo the base station end, then sent to the corresponding de-quantizer and self-decoder to recover the non-zero sub-array of the angle-time delay domain

The high-efficiency MIMO channel feedback method based on the network aggregation strategy is mainly applied to intelligent channel compression feedback in an MIMO system, and realizes characteristic learning and characteristic compression of MIMO channels, especially large-scale MIMO channels, through a specially designed neural network self-encoder and a specially designed neural network self-decoder, so that the feedback overhead is greatly reduced. In the intelligent channel compression feedback technology, the design of a feedback network is critical. Compared with the design based on a naive convolutional neural network, the feedback network design based on the network aggregation strategy disclosed by the invention has the characteristics of high feedback precision, good network elasticity and flexible deployment, and aiming at the resource limitations of different user sides and base station sides, the aggregation compression feedback network provided by the invention can provide a high-efficiency adaptive scheme, thereby realizing the MIMO channel compression feedback with high precision and low cost. From the application example, the aggregation feedback network in the invention can be used for channel compression feedback of the MIMO system under various conditions. Particularly for Frequency Division Duplex (FDD) systems, the aggregation network provided by the invention can keep the accuracy of the fed-back MIMO channel under extremely high compression rate. Specifically, in an FDD system, the MIMO channel may first be transformed from the space-frequency domain to the angle-time delay domain by discrete fourier transformation. Since the time delay of multipath is in a certain range, the transformed angle-time delay domain MIMO channel is highly sparse in the time delay dimension. Since most elements in the angle-time delay domain channel matrix are zero, only the non-zero submatrices of the angle-time delay domain channel matrix need to be intercepted and fed back. The aggregation feedback network provided by the invention can utilize the self-encoder to compress the non-zero subarray at the user end to realize the further reduction of the feedback capacity, and then utilize the self-decoder to recover the original degree-time delay domain channel non-zero subarray at the base station end. And finally, obtaining an original channel matrix through zero filling and two times of inverse discrete Fourier transform.

Next, an efficient MIMO channel feedback apparatus based on a network aggregation strategy proposed according to an embodiment of the present invention is described with reference to the accompanying drawings.

As shown in fig. 5, the efficient MIMO channel feedback apparatus based on the network aggregation strategy includes: a training module 501, a user-side preamble module 502, a compression module 503, a decompression module 504, and a base-side follow-up module 505.

The training module 501 is configured to train the aggregation feedback network, deploy a self-encoder and a quantization module in the trained aggregation feedback network to a user side, and deploy a self-decoder and a quantization module in the trained aggregation feedback network to the client side.

The user end preamble module 502 is configured to obtain a downlink channel matrix through channel estimation performed by a user end, perform two fourier transforms on the downlink channel matrix to transform the downlink channel matrix from a space-frequency domain to an angle-time delay domain, and intercept a non-zero sub-matrix of the transformed matrix.

And a compressing module 503, configured to compress the non-zero sub-array through the self-encoder and the quantizing module to obtain a feature vector, and send the feature vector to the base station.

A decompression module 504, configured to decode the feature vector through a quantization module and a self-decoder at the base station.

And a base station subsequent module 505, configured to perform zero padding and inverse discrete fourier transform on the decoded matrix to obtain a downlink channel matrix.

Further, training the aggregated feedback network further comprises:

collecting and training a downlink channel matrix;

Further, training the aggregated feedback network further comprises:

It should be noted that the foregoing explanation of the method embodiment is also applicable to the apparatus of this embodiment, and is not repeated herein.

According to the efficient MIMO channel feedback device based on the network aggregation strategy, provided by the embodiment of the invention, the self-encoder and the quantization module in the trained aggregation feedback network are deployed to the user side and the self-decoder and the quantization module in the trained aggregation feedback network are deployed to the client side by training the aggregation feedback network; performing channel estimation through a user side to obtain a downlink channel matrix, performing Fourier transform twice on the downlink channel matrix to transform the downlink channel matrix from a space-frequency domain to an angle-time delay domain, and intercepting a non-zero sub-matrix of the transformed matrix; compressing the non-zero sub-array through a self-encoder and a quantization module to obtain a characteristic vector, and sending the characteristic vector to a base station end; decoding the characteristic vector through a quantization module and a self-decoder of a base station end; and carrying out zero filling and inverse discrete Fourier transform on the decoded matrix to obtain a downlink channel matrix. The method has the advantages of high feedback precision, good network elasticity and flexible deployment, and realizes the MIMO channel compression feedback with high precision and low expense.

Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.

In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims

1. A high-efficiency MIMO channel feedback method based on a network aggregation strategy is characterized by comprising the following steps:

2. The method of claim 1, wherein training the aggregated feedback network further comprises:

collecting and training a downlink channel matrix;

transforming the training downlink channel matrix from a space-frequency domain to an angle-time delay domain through two discrete Fourier transforms, and intercepting a non-zero sub-matrix of the transformed matrix to form a training data set;

and stretching the aggregation feedback network according to the resource limitations of the actual user side and the base station side and the requirements on time delay and feedback precision, selecting a proper quantization scheme, and training the aggregation feedback network based on the training data set.

3. The method of claim 2, wherein acquiring the training downlink channel matrix comprises:

4. The method of claim 2, wherein acquiring the training downlink channel matrix comprises:

and acquiring the training downlink channel matrix through model simulation or in a physical environment.

5. The method of claim 1, wherein the aggregate feedback network comprises a group convolution through which channel feature extraction is performed.

6. The method of claim 2, wherein training the aggregated feedback network further comprises:

training is performed using an Adam optimizer and MSE loss function and the training data set.

7. The method of claim 1, wherein a learnable adaptive activation function Parametric ReLU is included in the aggregate feedback network.

8. An efficient MIMO channel feedback apparatus based on a network aggregation policy, comprising:

9. The apparatus of claim 8, wherein training the aggregated feedback network further comprises:

collecting and training a downlink channel matrix;

10. The apparatus of claim 9, wherein training the aggregated feedback network further comprises: