WO2022206747A1

WO2022206747A1 - High-efficiency mimo channel feedback method and device based on binarized neural network

Info

Publication number: WO2022206747A1
Application number: PCT/CN2022/083630
Authority: WO
Inventors: 王劲涛; 陆智麟; 张彧; 阳辉
Original assignee: 清华大学
Priority date: 2021-04-01
Filing date: 2022-03-29
Publication date: 2022-10-06
Also published as: CN113098805B; CN113098805A

Abstract

The present application discloses a high-efficiency MIMO channel feedback method and device based on a binarized neural network. The method comprises: determining the dimension of a space-frequency domain downlink channel matrix of a user-side channel according to the number of OFDM subcarriers in a communication system and the number of antennas of a base station side; inputting the space-frequency domain downlink channel matrix into a self-encoder based on the binarized neural network for compression to obtain a channel feature vector, and sending the channel feature vector from a user side to the base station side by means of an uplink; and decoding the received channel feature vector by means of a neural network-based self-decoder of the base station side to obtain the space-frequency domain downlink channel matrix. In the solution, self-encoder neural network deployment requiring low overhead can be carried out at the user side having limited resources, and a more practical channel compression feedback solution is realized.

Description

Efficient MIMO channel feedback method and device based on binary neural network

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims the priority of the Chinese patent application number "202110356946.1" filed by Tsinghua University on April 1, 2021 with the title of "Efficient MIMO Channel Feedback Method and Device Based on Binarized Neural Network".

technical field

The present application relates to the field of communication technologies, and in particular, to an efficient MIMO channel feedback method and device based on a binarized neural network.

Background technique

The channel feedback accuracy of the traditional codebook-based MIMO (Multiple-Input Multiple-Output, multiple-input multiple-output system) system is limited, especially when the MIMO scale is expanded to massive MIMO, the feedback accuracy will be further attenuated.

The full-channel feedback or sub-full-channel feedback of the MIMO system requires a large amount of information to be fed back, which brings unbearable feedback overhead. In addition, it is difficult for traditional compressed sensing technology to achieve sufficiently low compression ratio and sufficiently low loss of compressed information.

The channel compression feedback technology of MIMO system based on deep learning achieves lower compression information loss at a lower compression rate through the learning of channel characteristics. It is too cumbersome in terms of storage and calculation.

SUMMARY OF THE INVENTION

The present application aims to solve one of the technical problems in the related art at least to a certain extent.

To this end, one purpose of this application is to propose an efficient MIMO channel feedback method based on a binary neural network, which can implement low-overhead autoencoder neural network deployment at the resource-constrained user end, and achieve a more practical Channel compression feedback scheme.

Another object of the present application is to propose an efficient MIMO channel feedback device based on a binarized neural network.

To achieve the above purpose, an embodiment of the present application proposes an efficient MIMO channel feedback method based on a binarized neural network, including:

S1, the dimension of the space-frequency domain downlink channel matrix of the channel at the user end is determined according to the number of OFDM subcarriers in the communication system and the number of antennas at the base station end;

S2, inputting the space-frequency domain downlink channel matrix into an autoencoder based on a binarized neural network for compression to obtain a channel eigenvector, and sending the channel eigenvector from the user terminal to the base station through the uplink;

S3: Decode the received channel feature vector through a neural network-based self-decoder at the base station to obtain the space-frequency domain downlink channel matrix.

In order to achieve the above purpose, another embodiment of the present application provides an efficient MIMO channel feedback device based on a binarized neural network, including:

The user-end preamble module is used to determine the dimension of the space-frequency domain downlink channel matrix of the user-end channel according to the number of OFDM subcarriers and the number of antennas at the base station in the communication system;

The first compression module is used for inputting the space-frequency domain downlink channel matrix into an autoencoder based on a binary neural network for compression to obtain a channel eigenvector, and compressing the channel eigenvector from the user terminal through the uplink sent to the base station;

The first decompression module is configured to decode the received channel feature vector through a neural network-based self-decoder at the base station to obtain the space-frequency domain downlink channel matrix.

The high-efficiency MIMO channel feedback method and device based on the binarized neural network in the embodiments of the present application quantize the floating-point fully-connected layer in the original autoencoder into a binarized fully-connected layer by introducing the network binarization technology. In terms of storage, the binarized auto-encoder is one-thirtieth of the original floating-point auto-encoder; in terms of computing speed, the binarized auto-encoder is twice the original floating-point auto-encoder. The ultra-lightweight MIMO channel compression feedback auto-encoder designed through the above-mentioned network binarization technology is very beneficial to network deployment in a variety of resource-constrained user equipment vendors.

Additional aspects and advantages of the present application will be set forth, in part, in the following description, and in part will be apparent from the following description, or learned by practice of the present application.

Description of drawings

The above and/or additional aspects and advantages of the present application will become apparent and readily understood from the following description of embodiments taken in conjunction with the accompanying drawings, wherein:

FIG. 1 is a flowchart of an efficient MIMO channel feedback method based on a binarized neural network according to an embodiment of the present application;

2 is a schematic diagram of an efficient MIMO channel feedback method based on a binarized neural network according to an embodiment of the present application;

3 is a schematic diagram of an end-to-end training of a binarized autoencoder according to an embodiment of the present application;

4 is a schematic diagram of a visualization of a binarized dimensionally compressed fully-connected layer in an autoencoder and a dimensionally expanded fully-connected layer in a self-decoder according to an embodiment of the present application;

FIG. 5 is a schematic structural diagram of an efficient MIMO channel feedback device based on a binarized neural network according to an embodiment of the present application.

Detailed ways

The following describes in detail the embodiments of the present application, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals refer to the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary, and are intended to be used to explain the present application, but should not be construed as a limitation to the present application.

The following describes an efficient MIMO channel feedback method and device based on a binarized neural network according to the embodiments of the present application with reference to the accompanying drawings.

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

FIG. 1 is a flowchart of an efficient MIMO channel feedback method based on a binarized neural network according to an embodiment of the present application.

As shown in Figure 1, the efficient MIMO channel feedback method based on binarized neural network includes the following steps:

Step S1: Determine the dimension of the space-frequency domain downlink channel matrix of the channel at the user end according to the number of OFDM subcarriers and the number of antennas at the base station in the communication system.

Determine the number of OFDM subcarriers N _c and the number of base station antennas N _t according to preset communication requirements or preset communication systems, thereby determining the space-frequency domain downlink channel matrix

dimension.

Specifically, the embodiments of the present application are mainly applied in a frequency division duplex (FDD) system to reduce the storage and computing overhead of a resource-constrained user end, where the base station end is a multi-antenna array, and the user end is a single antenna. However, when channel feedback needs to be performed in a time division duplex (TDD) system, the same system structure can also be used to perform compression feedback to reduce the storage and computing overhead of the user end.

Further, after determining the dimension of the space-frequency domain downlink channel matrix of the user terminal channel, the space-frequency domain downlink channel matrix can also be processed, and the space-frequency domain downlink channel matrix can be transformed by two discrete Fourier transforms into The angle-delay domain downlink channel matrix, and the non-zero sub-matrix in the angle-delay domain downlink channel matrix is intercepted by non-zero sub-matrix cutting. Take a nonzero subarray as the input to the autoencoder.

It can be understood that the discrete Fourier transform operation can be used for any number of OFDM subcarriers N _c and the number of antennas N _t at the base station. The method of intercepting the non-zero sub-arrays is determined by the actual channel conditions. When the channel delay is high or complex, the intercepted non-zero sub-arrays are larger, otherwise, only smaller non-zero sub-arrays are required.

In step S2, the space-frequency domain downlink channel matrix is input into an autoencoder based on a binarized neural network for compression to obtain a channel eigenvector, and the channel eigenvector is sent from the user end to the base station through the uplink.

After compression by the self-encoder, the user end feeds back the compressed eigenvectors through the uplink instead of directly feeding back the original channel matrix, thereby greatly reducing the overhead of channel feedback.

Specifically, an autoencoder based on a binarized neural network can be an unprocessed space-frequency domain downlink channel matrix, and directly compress the space-frequency domain downlink channel matrix to obtain a channel eigenvector. If the downlink channel matrix in the space-frequency domain is processed, the non-zero sub-matrix is used as the input of the self-encoder to compress, and the channel eigenvector is obtained.

It can be understood that, in the embodiments of the present application, it is a feasible solution whether to adopt discrete Fourier transform for the downlink channel matrix in the space-frequency domain.

Further, the autoencoder based on the binarized neural network is composed of a convolutional layer in the pre-order and a fully connected layer in the post-order, and the fully connected layer in the autoencoder based on the binarized neural network is binarized. , the convolutional layers are binarized or not.

As an embodiment, only the fully connected layer used to adjust the dimension of the channel matrix is binarized. The pre-order convolutional layers can be binarized or not. To ensure performance, it is generally recommended not to perform binarization. In fact, the fully connected layer usually occupies more than 90% of the weight of the autoencoder network parameters and calculations, so the binarized fully connected layer basically means the binarized fully autoencoder.

Step S3: Decode the received channel eigenvectors through a neural network-based self-decoder at the base station to obtain a downlink channel matrix in the space-frequency domain, thereby completing downlink channel feedback in a low-overhead MIMO system.

Specifically, if the discrete Fourier transform is not performed on the space-frequency domain downlink channel matrix transformation, the space-frequency domain downlink channel matrix is obtained by directly decoding the channel vector through the self-decoder. If discrete Fourier transform and non-zero sub-array cutting are performed on the space-frequency domain downlink channel matrix, the received channel eigenvectors are decoded by the neural network-based self-decoder at the base station to obtain non-zero angle-time Delay domain downlink channel sub-array, and then perform zero-padding and two inverse discrete Fourier transform on the non-zero angle-delay domain downlink channel sub-array to obtain the space-frequency domain downlink channel matrix, and complete the complete downlink channel compression feedback process.

There is a certain difference between the angle-delay domain non-zero channel sub-array and the real angle-delay domain non-zero channel sub-array after decompression. Such a difference is the information loss of the compressed feedback. The present application can make such information loss sufficiently small through adequate and proper training of the self-encoder and self-decoder, that is, the recovered channel sub-array is sufficiently accurate and usable. In addition, the use of more powerful autoencoders and autodecoders can further reduce such information loss.

Further, neural network-based self-decoders can have various designs, such as utilizing residual networks, utilizing multi-resolution networks, utilizing attention mechanism augmentation networks, and so on.

The application is for several extensions of the binarized autoencoder of the MIMO channel compression network, mainly aimed at expanding and strengthening the existing convolutional neural network in the autoencoder, so as to obtain better feedback performance. Such as adding additional convolutional layers, using additional residual links, expanding the size of the convolution kernel, etc. The extension of the self-decoder of the MIMO channel compression network mainly adds redundant decoding units to the self-decoder to obtain better feedback performance.

Further, in the embodiments of the present application, it is necessary to perform end-to-end training on the binarized self-encoder and self-decoder.

First get the training dataset and test dataset. Channel data can be generated in batches by using the existing mature channel model, or the device can be used to collect channel data from the actual environment. The former is less expensive but may degrade in performance when deployed; the latter is more expensive but the trained network performs better in real-world environments. After the obtained channel matrix is preprocessed by discrete Fourier transform and non-zero sub-array cutting, a data set that can be used for end-to-end network training is generated. After the data is obtained, it is divided into training set and test set. A collection requires no duplicate data.

The autoencoder and autodecoder are then connected together for end-to-end training, where the optimizer can use ADAM and the loss function uses mean squared error (MSE). Specifically, the input of the self-encoder is directly used as the input of the self-decoder; and the mean squared error (MSE) is calculated from the angle-delay domain non-zero channel sub-array output by the self-decoder and its true value. Therefore, the forward propagation of the network will eventually calculate the value of the loss function MSE, and then backpropagate through the loss function value, that is, automatically derive the gradient according to the chain rule to obtain the gradient of each layer. After the gradient is obtained, the adaptive learning rate is given by using the initialization and optimization rules of the ADAM optimizer, and then the parameters of each layer can be updated.

In end-to-end training, binarization modules in autoencoders (such as binarized fully connected layers) are not differentiable. In order to achieve normal end-to-end training and iteration, the gradient low-pass approximation method is adopted, that is, for all parameters whose absolute value is less than the threshold, the gradient is back-transmitted as is, and the gradient is zero-forcing for the parameters whose absolute value is greater than the threshold. Therefore, the entire feedback network can achieve complete gradient backpropagation, enabling end-to-end training. As shown in Figure 3, the specific process is as follows:

a) Find all binarization layers in the autoencoder. By default, only the last fully connected layer belongs to the binarization layer. Calculates equivalent weights based on the current floating point parameters of the binarization layer.

b) Binarize the floating-point parameters of the binarization layer, and the method of binarization is to take its sign, that is, positive numbers and zeros are binarized to 1; negative numbers are binarized to -1.

c) Based on the above binarized parameters, perform forward propagation on the entire feedback network (self-encoder plus cascaded decoder) to obtain the loss function MSE.

d) Backpropagation is performed based on the loss function MSE to automatically derive the gradient of each layer.

e) When the backpropagation in the previous step passes through the binarization layer in the autoencoder, the automatic derivation cannot be continued, and the derivative of the binarization function needs to be manually given. Here, the gradient low-pass filter is used as the approximate derivative of the binarization function. Specifically, when the absolute value of the original parameter is greater than 1, the gradient of the parameter is considered to be 0; when the absolute value of the original parameter is less than 1, the parameter is considered to be The gradient of is consistent with the normal floating point gradient.

f) The gradient of all layers of the feedback network can be obtained by the operation defined in the previous step; the update of the floating point parameters of the binarization layer can be calculated from the equivalent weight of the binarization layer and the original gradient of the binarization layer gradient.

g) Combine the update gradient of the binarization layer calculated in f) and the gradients of other layers obtained in d), use the optimizer (ADAM) and the corresponding learning rate to the overall feedback network (auto-encoder cascade auto-decoder ) to update the parameters.

Keeping the end-to-end training and gradually decreasing the learning rate until the network converges, the autoencoder design and autodecoder design of the compressed feedback network can be obtained.

It can be understood that different self-decoder designs, different optimizer designs, and different learning rate descent methods can all achieve end-to-end training of MIMO channel compression binarized self-encoders through gradient low-pass approximation strategies.

The embodiments of the present application include a learning rate reduction strategy that is effective for MIMO channel compression networks (including self-encoders and self-decoders). That is to say, the learning firstly rises rapidly from 0 to the highest point, and then decreases from the highest point to 0 in a cosine curve. The lowest point of the learning rate can be not 0 but a small amount close to zero; the length of the interval between which the learning rate rises and the interval during which it falls can be different, usually the former is much shorter.

End-to-end training of the binarized neural network-based autoencoder and the neural network-based self-decoder, where the key difficulty in end-to-end training is the binarization module in the autoencoder (such as binary fully connected layer) is non-steerable. In order to achieve normal end-to-end training and iteration, the present application adopts the gradient low-pass approximation method, that is, for all parameters whose absolute value is less than the threshold, the gradient is back-transmitted as it is, and the gradient is zero-forcing for the parameters whose absolute value is greater than the threshold. . In this way, the entire feedback network can achieve complete gradient back-propagation, enabling end-to-end training. The core of this is to use the gradient low-pass approximation to complete the end-to-end training of the binary feedback network under the intelligent feedback communication system, rather than the technical details of general feedback network training such as data acquisition and Adam optimizer.

The efficient MIMO channel feedback method based on the binarized neural network of the present application will be described in detail below.

S101: For the FDD system, according to the existing communication system hardware and system software, determine the number N _c of OFDM subcarriers and the number N _t of antennas at the base station, so as to give an overall downlink channel combining each subcarrier

The dimension of , that is, N _c ×N _t . It is transformed by two discrete Fourier transforms

After transforming to the angle-delay domain, the matrix is obtained

As shown in the following formula:

where X and Y are discrete Fourier transform matrices of corresponding dimensions, respectively. Due to the sparsity of the FDD MIMO channel in the delay domain,

Only a limited and concentrated few rows are non-zero. There is no need for feedback for most of the rows that are approximately zero, so the non-zero sub-matrix H composed of the first N _s rows is intercepted at this time as the object of channel feedback, as shown in the following formula:

S102: At the user end, the matrix H will pass through the autoencoder E based on the binarized neural network. The autoencoder consists of a pre-order convolutional layer and a subsequent fully-connected layer, which are responsible for channel spatial feature abstraction and channel dimension compression, respectively, and finally convert the input matrix H into a compressed feature vector v, as shown in the following formula:

E(H)= _EFC(Econv ₍ H))=v

It is worth noting that the dimension of the feature vector v is from a fraction to several tenths of the original matrix H. Therefore, the user end only needs to feed back v but does not need to feed back H, which will significantly reduce the feedback overhead. It is worth noting that the convolutional layer here can have various designs, one of which is shown in the first specific embodiment. The fully connected layer is constrained by the dimensions of the matrix H and the feature vector v. The dimension of the matrix H is 2×N _s ×N _t , where 2 represents the real part and the imaginary part (the element at each position of the channel matrix H is a complex number); and the dimension of the feature vector v is

Where λ is the compression factor, ranging from several to dozens. Therefore, the dimension of the fully connected layer is

After the eigenvector v is ideally fed back through the uplink, the base station will receive the lossless eigenvector v. Note that it is assumed that the uplink is lossless digital transmission, and the possible information loss of the eigenvector itself in transmission is not considered, but focuses on how to recover the original channel matrix as accurate as possible from the compressed eigenvectors.

S103: Input the received feature vector v into the neural network-based self-decoder D at the base station side. The structure of the self-decoder and the self-encoder are dual, and consist of a fully connected layer with dual dimension and concatenated subsequent convolution layers. The fully connected layer of dimension duality restores the compressed feature vector v to the dimension before compression; the subsequent convolutional layer extracts and restores the spatial information of the original channel matrix H, as shown in the following formula:

A properly trained self-decoder can recover the original non-zero angle-delay domain subarray H, that is,

is sufficiently close to the original channel matrix H. It is worth noting that the dimension of the fully connected layer here is opposite to that of the fully connected layer in the autoencoder, that is

Due to the relatively strong storage and computing capabilities of the base station, the subsequent convolutional layer D _conv can be constructed relatively complex, so as to better restore the spatial characteristics of the channel matrix and achieve better feedback performance.

S104: Finally, the base station can recover the value from the non-zero angle-delay domain sub-array through zero-padding and two inverse discrete Fourier transforms.

Get the recovered value of the original downlink channel

Complete the complete downlink channel compression feedback process, as shown in the following formula:

in

is the angle-delay domain matrix after zero-padded, and is a complex matrix with dimension N _t ×N _c ; and

and

is the corresponding inverse discrete Fourier transform matrix.

The key of this application is to binarize the fully connected layer EFC which occupies most of the storage/computation complexity of the _autoencoder . By simplifying the 32-bit floating point parameters into 1-bit binarization parameters, the storage complexity of this layer is reduced to the original

Since the multiplication operation is not required after binarization, only the addition operation is required, and the calculation cost is also greatly reduced. It is conservatively estimated that the operation speed is increased by more than 2 times. It is worth noting that in most current autoencoder designs, the amount of computation and parameters occupied by the convolutional layer is very small compared to the fully connected layer. Therefore, the loss of binarizing the convolutional layer of the self-encoder is greater than the gain, so the binarization of the convolutional layer is not adopted in the example of this application (but such an operation is practically feasible).

_{Let EBFC} represent the binarized autoencoder fully connected layer. Then the low-complexity MIMO channel compression feedback method based on network binarization proposed in this application can be described by the following formula:

Through the above process, the present application greatly reduces the cost of deploying resource-sensitive user-end self-encoders, and improves the feasibility of the MIMO channel feedback method based on neural network self-encoder-self-decoder in practical systems. The above process can be implemented by the low-complexity MIMO channel compression feedback device based on network binarization depicted in FIG. 2 .

In addition, the above networks all need to obtain the channel compression/decompression capability through end-to-end training, and the parts that need to be trained include E _conv , E _BFC , D _FC and D _conv . The network described in the above embodiment can be trained on the collected/generated data through the end-to-end training strategy described in the content of the application, and the key is to perform the full-connection layer _EBFC on the non-derivable binarized autoencoder. For end-to-end processing, the equivalent weights of its parameters are obtained as shown in Figure 3, and their update gradients are calculated.

Assuming that the floating-point parameters of the E _BFC layer are represented by _FBFC , and the binarization parameters are represented by _BBFC , an instantiated calculation method of the equivalent weight β of the parameters of the binarization layer is as follows:

The corresponding calculation method of the update gradient of its floating-point parameters is as follows:

where _{flow_pass} (F _BFC ) is the aforementioned gradient low-pass filter, which is used to process the approximate derivation of the binarization function.

The low-complexity MIMO channel compression feedback method based on network binarization will be further described below by way of specific embodiments.

1) For an FDD massive MIMO system, the number of OFDM subcarriers is N _c =1024, and the size of the antenna array at the base station is N _t =32. Then its downlink channel matrix

The dimension of is 1024×32. For the downlink channel matrix

Do two discrete Fourier transforms, which can be transformed from the space-frequency domain to the angle-delay domain to obtain the angle-delay domain channel matrix

After that, it is intercepted, and the first 32 lines (usually 16 lines are sufficient, 32 lines can further improve the accuracy) are taken as the complex sub-array H that needs to be fed back, then its dimension is reduced to 32 × 32, and the dimension after conversion to a floating-point number vector That is 2×32×32.

2) Input the sub-array H into the self-encoder E at the user end, where the self-encoder E consists of a convolutional layer and a fully connected layer. Set the target compression factor to 8 times, set the convolution layer as a cascade of two 3×3 convolutions, each convolution is followed by a batch normalization layer and an activation function layer; while the fully connected layer realizes the channel dimension Compression, its parameter dimension is 2048×256, so the dimension of the feature vector v output from the encoder E is 1×256. The fully connected layer of the autoencoder in this application is binarized, so in fact, the parameter amount here is only 16K (the original floating-point parameter amount is 512K), that is, the parameter amount is reduced to the original

In addition, there is no need to perform multiplication during network calculation here, because multiplying by ±1 is equivalent to changing the sign of the current element, which eventually degenerates into addition and subtraction between vector elements. The operation speed of addition is faster than multiplication on most hardware, so the forward transmission speed of the binarized fully-connected layer here is more than twice that of the original floating-point parameter fully-connected layer.

3) Perform ideal digital feedback on the feature vector v, and the base station receives the lossless feature vector v and inputs it into the neural network-based self-decoder. The self-decoder consists of a trained one-dimensionally expanded fully connected layer and a more powerful convolutional neural network. The dimension of the dimension-expanded fully-connected layer is dual with the dimension-compressed fully-connected layer in the autoencoder, and its parameter dimension is 256×2048. It is worth noting that the self-decoder parameters in this application are normal 32-bit floating point numbers. After the fully connected layer is expanded by dimension, the dimension of the feature vector v is expanded back to the original 2048=2×32×32. The subsequent convolution sub-network is mainly composed of a simple convolution layer and a residual structure. Each residual structure includes three convolution layers, and the convolution kernel size is 3; the three convolution kernels expand the number of channels respectively. Increase to 8, expand to 16 and compress back to 2, to achieve feature recovery by scaling the channel. In this example, two such residual structures are cascaded, and a simple convolutional layer with kernel size 3 is added at the end to form the final self-decoder convolutional network. As can be seen from the above structure, the matrix finally recovered from the decoder

The dimension is still 2×32×32, which is 32×32 in terms of complex dimensions.

4) pair matrix

Zero-padding restores its dimensions from 32×32 to 1024×32. Perform two inverse discrete Fourier transforms on the matrix that restores the OFDM dimension to obtain the final MIMO channel feedback result.

In the above specific example, the training data is collected from the actual environment, or generated by the channel model (COST2100 model, 3GPPTR.38.901UMiNLOS model or simple Saleh-Valenzuela model, etc.). Taking the high-frequency indoor scene of COST2100 as an example, after training batch data based on ADAM optimizer, the channel recovery normalization below -10dB can be obtained through the cascaded compression feedback network of the binarized auto-encoder and self-decoder. The normalized mean square error provides sufficiently accurate downlink channel information for subsequent beamforming.

In this specific example, the visualization of the binarized dimensionally compressed fully-connected layer in the autoencoder and the dimensionally expanded fully-connected layer in the self-decoder obtained after training is shown in Figure 4 . It can be seen that the parameters of the fully connected layer after training and learning describe some spatial characteristics of the channel.

It can be seen from the above introduction that the solution of the embodiment of the present application is to binarize the FC layer in the self-encoder, instead of quantizing or binarizing the compressed feature vector output by the FC layer. As long as the final FC layer uses network binarization, whether discrete Fourier transform is used in the compression process or not, it belongs to the concept of the present application.

The purpose of this application is to reduce the cost of deploying the neural network on the client device, so it is necessary to compress the neural network itself, rather than reducing the feedback cost, and further quantify the CSI feature vector that needs to be fed back. The overhead reduced by the present application includes storage overhead and calculation overhead, and the overhead of channel feedback before and after compression is unchanged.

The embodiments of the present application do not consider the compression of feature vectors, but focus on the compression of the network itself and the reduction of deployment costs. The binarization mentioned in this application is network binarization, and specifically, binarization is performed on the fully connected layer of the feedback self-encoder, rather than the binarization of the compressed feature vector.

In this application, regardless of whether other layers in the pre-order autoencoder are binarized (other layers may be convolutional layers or fully connected layers), as long as network binarization is used at the last fully connected layer, it is included in this application. within the scope of the application concept of the application.

According to the efficient MIMO channel feedback method based on the binarized neural network proposed in the embodiment of the present application, by introducing the network binarization technology, the floating-point fully-connected layer in the original autoencoder is quantized into a binarized fully-connected layer. In terms of storage, the binarized auto-encoder is one-thirtieth of the original floating-point auto-encoder; in terms of computing speed, the binarized auto-encoder is twice the original floating-point auto-encoder. The ultra-lightweight MIMO channel compression feedback auto-encoder designed through the above-mentioned network binarization technology is very beneficial to network deployment in a variety of resource-constrained user equipment vendors.

Next, an efficient MIMO channel feedback device based on a binarized neural network proposed according to an embodiment of the present application will be described with reference to the accompanying drawings.

As shown in FIG. 5 , the high-efficiency MIMO channel feedback device based on the binarized neural network includes: a user-end preamble module 501 , a first compression module 502 and a first decompression module.

The UE preamble module 501 is configured to determine the dimension of the space-frequency domain downlink channel matrix of the UE channel according to the number of OFDM subcarriers and the number of base station antennas in the communication system.

The first compression module 502 is configured to input the space-frequency domain downlink channel matrix into an autoencoder based on a binarized neural network for compression to obtain a channel eigenvector, and send the channel eigenvector from the user end to the base station through the uplink end.

The first decompression module 503 is configured to decode the received channel feature vector through a neural network-based self-decoder at the base station to obtain a space-frequency domain downlink channel matrix.

Further, in the embodiments of this application, it also includes:

The transformation module 504 is configured to transform the space-frequency domain downlink channel matrix into an angle-delay domain downlink channel matrix through two discrete Fourier transforms, and intercept the angle-delay domain downlink channel matrix through non-zero sub-array cutting. nonzero subarray.

The second compression module 505 is used for inputting the non-zero sub-array into the autoencoder based on the binarized neural network for compression to obtain the channel feature vector, and sending the channel feature vector from the user end to the base station end through the uplink.

The second decompression module 506 is configured to decode the received channel feature vector through the neural network-based self-decoder at the base station to obtain a non-zero angle-delay domain downlink channel sub-array, and for the non-zero angle-delay domain The downlink channel sub-array performs zero-padding and two inverse discrete Fourier transforms to obtain the downlink channel matrix in the space-frequency domain.

The base station finally obtains the downlink channel of the space-frequency domain MIMO system

Can be used as input for other subsequent modules, such as beamforming modules, etc.

It should be noted that the foregoing explanations of the method embodiment are also applicable to the apparatus of this embodiment, and details are not repeated here.

According to the high-efficiency MIMO channel feedback device based on the binarized neural network proposed in the embodiment of the present application, by introducing the network binarization technology, the floating-point fully-connected layer in the original autoencoder is quantized into a binarized fully-connected layer. In terms of storage, the binarized auto-encoder is one-thirtieth of the original floating-point auto-encoder; in terms of computing speed, the binarized auto-encoder is twice the original floating-point auto-encoder. The ultra-lightweight MIMO channel compression feedback auto-encoder designed through the above-mentioned network binarization technology is very beneficial to network deployment in a variety of resource-constrained user equipment vendors.

In addition, the terms "first" and "second" are only used for descriptive purposes, and should not be construed as indicating or implying relative importance or implying the number of indicated technical features. Thus, a feature delimited with "first", "second" may expressly or implicitly include at least one of that feature. In the description of the present application, "plurality" means at least two, such as two, three, etc., unless expressly and specifically defined otherwise.

In the description of this specification, description with reference to the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples", etc., mean specific features described in connection with the embodiment or example , structure, material or feature is included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms are not necessarily directed 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, those skilled in the art may combine and combine the different embodiments or examples described in this specification, as well as the features of the different embodiments or examples, without conflicting each other.

Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limitations to the present application. Embodiments are subject to variations, modifications, substitutions and variations.

Claims

An efficient MIMO channel feedback method based on a binarized neural network, characterized in that it comprises the following steps:

S1, the dimension of the space-frequency domain downlink channel matrix of the channel at the user end is determined according to the number of OFDM subcarriers in the communication system and the number of antennas at the base station end;

S2, inputting the space-frequency domain downlink channel matrix into an autoencoder based on a binarized neural network for compression to obtain a channel eigenvector, and sending the channel eigenvector from the user terminal to the base station through the uplink;

S3: Decode the received channel feature vector through a neural network-based self-decoder at the base station to obtain the space-frequency domain downlink channel matrix.
The method of claim 1, further comprising:

The space-frequency domain downlink channel matrix is transformed into an angle-delay domain downlink channel matrix through two discrete Fourier transforms, and the non-zero sub-matrix cutting is used to intercept the non-zero downlink channel matrix in the angle-delay domain subarray;

Inputting the non-zero sub-array into an autoencoder based on a binarized neural network for compression to obtain a channel feature vector, and sending the channel feature vector from the user end to the base station end through the uplink;

The received channel feature vector is decoded by the neural network-based self-decoder at the base station to obtain a non-zero angle-delay domain downlink channel sub-array, and the non-zero angle-delay domain downlink channel sub-array is processed The space-frequency domain downlink channel matrix is obtained by zero-padding and inverse discrete Fourier transform twice.
The method of claim 1, wherein:

The communication system is a frequency division duplex system or a time division duplex system.
The method of claim 1, wherein:

The binarized neural network-based autoencoder is composed of a pre-order convolutional layer and a post-order fully connected layer, and the fully connected layer in the binarized neural network-based autoencoder is binarized. , the convolutional layer binarizes or does not binarize.
The method of claim 1, wherein the neural network-based self-decoder comprises a residual network, a multi-resolution network, and an attention mechanism augmentation network.
The method according to claim 1, further comprising: performing end-to-end training on the binarized neural network-based self-encoder and the neural network-based self-decoder, specifically:

Get training datasets and test datasets;

The self-encoder based on the binary neural network and the self-decoder based on the neural network are connected together to perform end-to-end training using the training data set, and the gradient low-pass filter is used in the training. In this way, end-to-end training is maintained, and the learning rate is gradually decreased until the network converges, and the trained binarized neural network-based autoencoder and the neural network-based autodecoder are tested through a test dataset.
The method according to claim 6, wherein a plurality of channel matrices are obtained through an actual channel environment or a channel model describing the actual channel environment, and the plurality of channel matrices are subjected to discrete Fourier transform and non-zero sub-matrix cutting A training data set and a test data set are generated, and the data in the training data set and the test data set are not duplicated.
The method of claim 6, wherein the end-to-end training further comprises:

Determine all binarization layers in the autoencoder. By default, only the last fully connected layer belongs to the binarization layer, and the equivalent weight is calculated based on the current floating-point parameters of the binarization layer;

Binarize the floating-point parameter of the binarization layer, and the way of binarization is to take its sign;

Based on the binarized parameters, forward the entire feedback network to obtain the loss function;

Backpropagation is performed based on the loss function to automatically obtain the gradient of each layer, and the gradient low-pass filter is used as the approximate derivative of the binarization layer in the autoencoder to obtain the gradient of all layers of the feedback network;

From the equivalent weight of the binarization layer and the original gradient of the binarization layer, the update gradient of the floating point parameters of the binarization layer is calculated;

Combined with the update gradient of the binarization layer and the gradients of other layers, the optimizer and the corresponding learning rate are used to update the parameters of the overall feedback network.
An efficient MIMO channel feedback device based on a binarized neural network, characterized in that it includes:

The user-end preamble module is used to determine the dimension of the space-frequency domain downlink channel matrix of the user-end channel according to the number of OFDM subcarriers and the number of antennas at the base station in the communication system;

The first compression module is used for inputting the space-frequency domain downlink channel matrix into an autoencoder based on a binary neural network for compression to obtain a channel eigenvector, and compressing the channel eigenvector from the user terminal through the uplink sent to the base station;

The first decompression module is configured to decode the received channel feature vector through a neural network-based self-decoder at the base station to obtain the space-frequency domain downlink channel matrix.
The device of claim 9, further comprising:

a transformation module, configured to transform the space-frequency domain downlink channel matrix into an angle-delay domain downlink channel matrix through two discrete Fourier transforms, and intercept the angle-delay domain downlink channel through non-zero sub-array cutting a nonzero submatrix in a matrix;

The second compression module is used for inputting the non-zero sub-array into an autoencoder based on a binarized neural network for compression, obtaining a channel feature vector, and sending the channel feature vector from the user end to the base station through the uplink end;

The second decompression module is used for decoding the received channel feature vector through the neural network-based self-decoder at the base station to obtain a non-zero angle-delay domain downlink channel sub-array, and for the non-zero angle- The space-frequency domain downlink channel matrix is obtained by performing zero-padding and two inverse discrete Fourier transforms on the downlink channel sub-array in the delay domain.