TWI669921B - Feedback method for use as a channel information based on deep learning - Google Patents

Abstract

A feedback method based on deep learning as channel state information can reduce the complexity of reconstructing channel state information, and solve the problem of poor performance of the prior art. The method includes: splitting a channel state information into a real part matrix and an imaginary part matrix by using a coding training model at a receiving end to generate a coded word, and the matrix size of the channel state information is , Expressed as the number of original subcarriers of the adopted OFDM system, and ≧1, Nt denotes the number of base station transmit antennas of the transmitting end, and Nt≧1; the receiving end feeds back the coded word to a transmitting end; the transmitting end obtains the coded word and encodes the code with a decoding training model The word is converted back to the real part matrix and the imaginary part matrix; and the decoding training model converts the real part matrix and the imaginary part matrix of the transmitting end back to the channel state information.

Description

Feedback learning method based on depth learning as channel state information

The invention relates to a feedback method of channel state information, in particular to a method for reducing channel time complexity based on deep learning, as a feedback method of channel state information.

In recent years, with the rapid development of personal communication needs and the rapid increase of multimedia information exchange, the relatively usable spectrum is very limited, making the spectrum a precious resource. Therefore, multiple-input multiple-output (Multiple-Input Multiple-Output, MIMO) technology is highly regarded as one of the key technologies in the field of wireless communication. It has beamforming capability, diversity gain capability and multiplexing gain capability. It can be used at the transmitting end and the receiving end. At the same time, multiple antennas and related communication signal processing technologies are used, so that spatial freedom can be provided without increasing the bandwidth, thereby effectively improving the system capacity and spectrum efficiency of the wireless system.

According to the multi-input and multi-output technology, it is generally possible to use Time-Division Duplexing (TDD) or Frequency-Division Duplexing (FDD), in which duplex of wireless communication is used. (Duplex) technology refers to a method of realizing two-way communication between a transmitting end and a receiving end by means of channel access, so that two communication devices can mutually transmit data. In addition, the large-scale multi-input and multi-output (Massive MIMO) technology extended by MIMO technology is able to increase system capacity and spectrum efficiency to support a larger number of users. It is the main technology of the fifth generation wireless communication system, in which the channel reciprocity of the time division duplex technology relies on a complicated calibration process, and the existing system uses a large number of frequency division duplex technologies, such as most existing mobile phones. The system adopts frequency division duplex technology, which makes the large-scale frequency division duplex multiple input multiple output (FDD Massive MIMO) system an important development direction of multi-input and multi-output technology.

However, for the existing large-scale frequency division duplex multiple-input multiple-output system, when the user equipment (User Equipment, UE) of a receiving end needs to feed back a channel state information (CSI) in the downlink. When a base station (BS) to which a transmitting end belongs is given, the channel status information is first simplified to make the channel structure exhibit sparse characteristics, and then Compressive Sensing (CS) is used. The signal of the channel status information is compressed, and the channel status information can be restored after multiple iterations on the transmitting end. Because the conventional use of the compressed sensing method requires multiple iterations to restore the channel state information, the system time complexity is increased and the system performance is reduced.

In view of this, the feedback method of the conventional channel state information does still have to be improved.

In order to solve the above problems, an object of the present invention is to provide a feedback method based on deep learning as channel state information, which can utilize depth learning to reduce system time complexity to reconstruct the channel state information.

The deep learning method based on depth learning of the present invention includes the following steps: splitting a channel state information into a real part matrix and an imaginary part matrix by a coding training model at a receiving end to generate a coded word , the matrix size of the channel status information is xNt, Expressed as the number of original subcarriers of the adopted OFDM system, and ≧1, Nt denotes the number of base station transmit antennas of the transmitting end, and Nt≧1; the receiving end feeds back the coded word to a transmitting end; the transmitting end obtains the coded word and encodes the code with a decoding training model The word is converted back to the real part matrix and the imaginary part matrix; and the decoding training model converts the real part matrix and the imaginary part matrix of the transmitting end back to the channel state information.

Accordingly, the deep learning method based on the channel state information of the present invention can greatly compress the channel state information by using deep learning when transmitting channel state information, and reconstruct the channel state information with extremely low complexity. In order to improve the acquisition efficiency of the channel status information, the system can be applied to large-scale multi-input and multi-output technology to fully utilize its advantages.

The convolutional neural network of the coding training model includes at least a first layer and a second layer, the first layer has two channels, and a real matrix of the receiving end and a matrix of the imaginary part are respectively provided in a matrix size. Convolution operation is performed on the core of KxK, and a linear rectification function is executed to generate a two-feature map. The second layer obtains a feedback parameter with a matrix size of Nx1 from the two-feature map, and converts the feedback parameter into an M-dimensional vector. Code word, a data compression ratio of the channel status information Where K ≧ 3, N=NcxNtx2, N represents the product of the length, width and number of the feature map, and Nc represents the number of subcarriers obtained by the OFDM system, and ≧Nc≧1. In this way, compared with the conventional compressed sensing method, the feature value possessed by the channel state information cannot be fully utilized, and the present invention can obtain the feature value of the coded word from the channel state information, and has the identifiable character of the enhanced coded word. Sexual effect.

The decompression training model includes at least a first layer and a second refining network layer, and the first layer of the decoding training model converts the code word back to the feedback parameter to obtain an initial estimate of a matrix size of NcxNt. a matrix and an initial imaginary matrix, each of the refining network layers includes at least one input layer and three convolution layers, and an input layer of the first refining network layer in the second refining network layer is used to input the initial Estimating the real part matrix and the initial estimated imaginary part matrix, the first convolutional layer of the first refining network layer is for the initial estimated real part matrix and the initial estimated imaginary part matrix, and the core of the matrix size KxK is rolled Product operation and perform a linear rectification function to generate 8 feature maps, second The convolution layer convolves the eight feature maps with a kernel of matrix size KxK, and performs a linear rectification function to generate 16 feature maps, and the third convolution layer has a matrix size of KxK for the 16 feature maps. The kernel performs a convolution operation, and performs a linear rectification function to generate two feature maps, and performs zero padding on the two feature maps generated by the third convolution layer, so that the matrix size of the two feature maps is input to the input layer. The initial estimated real part matrix and the matrix of the initial estimated imaginary part matrix are the same, and the matrix of the two characteristic maps is respectively added to the initial estimated real part matrix and the initial estimated imaginary part matrix as the second refining network The second layer in the layer refines the input of the input layer of the network layer, and then performs convolution operations and performs linear rectification functions through the three convolution layers of the second refining network layer, so that the initial estimated real part matrix And the initial estimated imaginary part matrix is transformed back to the real part matrix and the imaginary part matrix. In this way, compared with the conventional compressed sensing method, the present invention can restore the channel state information through deep learning without multiple iterations, which has the effect of reducing time complexity.

The receiving end uses the coding training model to calculate the channel state information by two-dimensional discrete Fourier transform, so that the channel state information is converted from the spatial frequency to the angle and time, and the coding training model retains the channel state information. The first Nc column is a value that is not zero to generate a truncation matrix, and the truncation matrix is split into the real part matrix and the imaginary part matrix to generate the coded word. In this way, the amount of information required for feedback at the receiving end can be reduced, which has the effect of reducing channel state information feedback overhead.

The data compression ratio is 1/4, 1/16, 1/32 or 1/64. In this way, compared with the conventional compressed sensing method, the present invention can greatly compress the channel state information through deep learning, and has the effect of further reducing channel state information feedback overhead.

Among them, the data compression ratio is 1/4 or 1/32. In this way, compared with the conventional compressed sensing method, the present invention can perform compression in the case that the channel state information is not a sparse matrix through deep learning, and can be applied to different antenna structures, and has the effect of improving the use range.

〔this invention〕

S1‧‧‧ coding step

S2‧‧‧ decoding step

S3‧‧‧ Sparse steps

L11‧‧‧ first floor

L12‧‧‧ second floor

L21‧‧‧ first floor

L22‧‧‧Refining the network layer

L221‧‧‧Input layer

L222‧‧‧Convolutional layer

L223‧‧‧Convolutional layer

L224‧‧‧Convolutional layer

Figure 1 is a flow chart of a method in accordance with an embodiment of the present invention.

Figure 2 is a block diagram showing the architecture of an embodiment of the present invention.

The above and other objects, features and advantages of the present invention will become more <RTIgt; The method according to a preferred embodiment of the method for the deep learning of the channel state information feedback method comprises: an encoding step S1 and a decoding step S2.

The encoding step S1 is based on a channel status information at a receiving end Simplified into a codeword, the receiver then feeds the codeword back to a transmitting end. In this embodiment, the channel status information It is transmitted by the base station of the transmitting end to the user equipment of the receiving end, that is, the channel status information of the downlink. Specifically, the receiving end can use the coded training model to display the channel status information. Splitting into a real part matrix and an imaginary part matrix to generate a coded word, the receiving end feeds back the coded word to the transmitting end. Where the channel status information Based on the spatial frequency, and the matrix size can be xNt, Indicates the number of original subcarriers of the OFDM system used in the present invention, and ≧1, Nt is expressed as the number of transmitting antennas of the base station, and Nt≧1.

Among them, the coding training model is trained by deep learning Convolutional Neural Networks (CNNs). For example, the convolutional neural network may include at least a first layer L11 and a second layer L12. The first layer L11 and the second layer L12 may each be a convolutional layer and a pool. In the embodiment, the first layer L11 may be a convolution layer, and the second layer L12 may be a fully connected layer. The first layer L11 has two channels, and the two channels respectively provide a real matrix of the receiving end and the imaginary matrix is convoluted by a kernel of a matrix size KxK, and performs a linear rectification function (Rectified Linear Unit) , ReLU) to generate a Feature Map, where K≧3. The second layer L12 obtains a feedback parameter having a matrix size of Nx1 from the two feature maps, and converts the feedback parameter into an encoded word of an M-dimensional vector. Channel status information Data compression ratio Where N is expressed as the product of the length, width and number of the feature map, and N = NcxNtx2, Nc is expressed as the number of subcarriers taken by the OFDM system, and ≧Nc≧1; M is expressed as the dimension of the codeword, and M N.

The decoding step S2 is capable of acquiring the encoded word by the transmitting end, and converting the encoded word back to the channel status information. . Specifically, the transmitting end can convert the coded word back to the real part matrix and the imaginary part matrix by using a decoding training model, and the decoding training model converts the real part matrix and the imaginary part matrix back to the channel state information. To complete the decoding step S2. In this embodiment, the decoding training model may include at least a first layer L21 and a second refining network layer (RefineNet) L22, and the first layer L21 of the decoding training model may convert the coded word back to the feedback parameter. To obtain an initial estimated real part matrix with a matrix size of NcxNt and an initial estimated imaginary part matrix. The second layer L22 of the decoding training model converts the initial estimated real part matrix and the initial estimated imaginary part matrix back to the real part matrix and the imaginary part matrix. In this embodiment, the first layer L21 may be A fully connected layer.

Specifically, each of the refinement network layers L22 may include at least one input layer L221 and three convolution layers L222 to L224. The input layer L221 of the first refining network layer in the second refining network layer L22 is used to input the initial estimated real part matrix and the initial estimated imaginary part matrix. Subsequently, the first convolution layer L222 of the first refining network layer L22 convolves the initially estimated real part matrix and the initial estimated imaginary part matrix with a matrix having a matrix size of 3x3, and performs a linear rectification function. To generate eight feature maps, the second convolution layer L223 convolves the eight feature maps with a kernel of matrix size KxK, and performs a linear rectification function to generate 16 feature maps, and a third convolution layer L224 pair The 16 feature maps are convoluted by a kernel of matrix size KxK, and a linear rectification function is performed to generate two feature maps, and two special features generated by the third convolution layer L224 are generated. The sign spectrum performs Zero Padding so that the matrix size of the two feature maps is the same as the matrix size of the initial estimated real part matrix input to the input layer L221 and the initial estimated imaginary part matrix. Then, the sum of the two feature maps and the initial estimated real part matrix and the initial estimated imaginary part matrix are used as input of the input layer L221 of the second refining network layer in the second refining network layer L22, And performing convolution operation and performing a linear rectification function through the three convolutional layers L222~L224 of the second refining network layer, so that the initial estimated real part matrix and the initial estimated imaginary part matrix are transformed back to the real part matrix And the imaginary matrix.

Preferably, the present invention may also include a sparse step S3. The sparse step S3 is performed by the receiving end to the channel state information by using the coding training model. Calculated by two-dimensional discrete Fourier transform to make the channel status information Converted from spatial frequency to angle and time based on the channel status information According to the spatial frequency dependence of orthogonal frequency division multiplexing multiple input multiple output (MIMO-OFDM) and channel, it can exhibit sparse characteristics in the angular delay domain (Angular-Delay Domain).

In detail, in the dimension of time delay, since the time delay between the arrival of multipath is within a certain period of time, the channel status information may not be needed. All the time points in the middle. The coding training model retains the channel status information The value of the middle Nc column is not zero to generate a truncation matrix H, and the truncation matrix H is split into the real part matrix and the imaginary part matrix to generate the coded word. In this way, the amount of information that the receiver needs to feed back can be reduced, which has the effect of reducing the channel state information feedback overhead.

For example, the present invention is based on deep learning as a channel state information feedback method, which can be applied to a large-scale frequency division duplex multiple input multiple output (FDD Massive MIMO) system, and combined with orthogonal frequency division multiplexing Technical system. First, in order to train the coding training model of the encoding step S1 and the decoding training model of the decoding step S2, the present invention can pass the The COST 2100 Channel Model (see L. Liu et al., "The COST 2100 Channel Model" IEEE Wireless Commun., vol. 19, no. 6, pp. 92-99, Dec.2012.) Establish two channels, one of which is the indoor environment of the 5.3 GHz band, and the other of the two channels is the outdoor environment of the 300 MHz band, each of which has 100,000 sample data and 30,000 samples. Data and 20,000 sample data are used as the training training model and the training set, the verification set and the test set of the decoding training model. Each of the two channels is provided with 32 transmitting antennas at one transmitting end, and the OFDM system divides the frequency band into 1024 subcarriers. Quantity.

In this embodiment, the channel status information is transmitted through the thinning step S3. After converting to angle and time as the basis, retain the channel status information In the first 32 lines whose value is not zero, a two-feature map of 32x32 matrix size is generated. The second layer L12 of the coding training model generates a feedback parameter with a matrix size of 2048×1 according to the two feature maps, and converts the feedback parameter into an M-dimensional vector coding word to complete the training learning of the coding training model. In addition, the decoding training model converts the coded word into the feedback parameter to obtain an initial estimated real part matrix with a matrix size of 32x32 and an initial estimated imaginary part matrix, the initial estimated real part matrix and the initial estimate The imaginary part matrix is further subjected to a convolution operation through the second refining network layer L22, so that the initial estimated real part matrix and the initial estimated imaginary part matrix are converted back to the real part matrix and the imaginary part matrix. The decoding training model then converts the real part matrix and the imaginary part matrix back to the channel state information .

The invention compares the feedback method based on deep learning as channel state information with the conventional methods such as LASSO, TVAL3, BM3D-AMP and CsiRecovNet. In the present invention, the channel matrix is The reconstructed channel matrix is measured by the normalized mean square error (NMSE) to measure the difference between the pre-encoding and decoded channel state information, and the channel state information of the feedback is compared by the cosine similarity ρ (Consine Similarity), and the comparison result can be as Table 1 shows. Wherein, the formula of the normalized mean square error and the cosine similarity can be respectively expressed as follows: Where H is the original truncation matrix, Is the truncated matrix after reconstruction.

among them, Representing a channel vector on the nth subcarrier; Is the reconstructed channel vector on the nth subcarrier.

Table 1 above describes the present invention and the conventional methods, respectively, in the channel status information The normalized mean squared error and cosine similarity of the data compression ratios are 1/4, 1/16, 1/32, and 1/64, and the best results are shown in bold. For example, when the compression ratio is 1/4, the channel state information is converted into a codeword of a 512-dimensional vector. As can be seen from Table 1, the present invention achieves the lowest normalized mean square error and the highest cosine similarity, so the present invention is significantly superior to the above conventional methods. Furthermore, the average run time of the present invention is 0.0035 s, which is also superior to the conventional LASSO of 0.1828 s, the TVAL3 of 0.5717 s and the BM3D-AMP of 0.3155 s. In addition, when in an outdoor environment, and the base station transmitting antennas of the transmitting end are 16 and 48 respectively, the comparison result can be as shown in Table 2. It can be seen that the present invention is also superior to the above conventional methods.

On the other hand, as can be seen from Table 3, relative to the status information of the channel Performing the sparse step S3 to make the channel status information Converting from spatial frequency to angle and time, when the sparse step S3 is not performed, when the channel status information When the data compression ratio is 1/4 and 1/32 respectively, the performance of the present invention is superior to the execution of the thinning step S3, whether in an indoor environment or in an outdoor environment.

While the invention has been described in connection with the preferred embodiments described above, it is not intended to limit the scope of the invention. The technical scope of the invention is protected, and therefore the scope of the invention is defined by the scope of the appended claims.

Claims (6)

A feedback method based on deep learning as channel state information includes the following steps: splitting a channel state information into a real part matrix and an imaginary part matrix by a coding training model at a receiving end to generate a coded word, The matrix size of the channel status information is xNt, Expressed as the number of original subcarriers of the adopted OFDM system, and ≧1, Nt denotes the number of base station transmit antennas of the transmitting end, and Nt≧1; and the receiving end feeds back the coded word to a transmitting end; the transmitting end obtains the coded word and uses a decoding training model to The code word is converted back to the real part matrix and the imaginary part matrix; and the decoding training model converts the real part matrix and the imaginary part matrix of the transmitting end back to the channel state information. The method for rewarding channel state information based on deep learning, as described in claim 1, wherein the convolutional neural network of the coded training model includes at least a first layer and a second layer, the first layer having Two channels, and a real part matrix respectively provided at the receiving end and the imaginary part matrix are convoluted with a core of matrix size KxK, and a linear rectification function is executed to generate a two-characteristic map, and the second layer is composed of the two characteristic maps Obtaining a feedback parameter with a matrix size of Nx1, and converting the feedback parameter into an encoded word of an M-dimensional vector, a data compression ratio of the channel state information Where K≧3, N=NcxNtx2, N is expressed as the product of the length, width and number of the feature map, and Nc is expressed as the number of subcarriers obtained by the OFDM system, and ≧Nc≧1. The method for rewarding channel state information based on deep learning as described in claim 2, wherein the decompression training model includes at least one first layer and two refining network layers, and the first layer of the decoding training model is Converting the coded word back to the feedback parameter to obtain an initial estimated real part matrix having a matrix size of NcxNt and an initial estimated imaginary part matrix, each of the refined network layers comprising at least one input layer and three convolutional layers, the second refining The first refining network in the network layer The input layer of the road layer is used to input the initial estimated real part matrix and the initial estimated imaginary part matrix, the first convolutional layer of the first refining network layer is the initial estimated real part matrix and the initial estimated imaginary part a matrix, performing convolution operations on a kernel of matrix size KxK, and performing a linear rectification function to generate eight feature maps, and a second convolution layer convolving the eight feature maps with a kernel of matrix size KxK, and Performing a linear rectification function to generate 16 feature maps, a third convolution layer convolving the 16 feature maps with a kernel of matrix size KxK, and performing a linear rectification function to generate 2 feature maps, the third The two feature maps generated by the convolutional layer perform zero padding, so that the matrix size of the two feature maps is the same as the matrix size of the initial estimated real part matrix and the initial estimated imaginary part matrix input to the input layer, and then the 2 The matrix of the feature maps is respectively added to the initial estimated real part matrix and the initial estimated imaginary part matrix as input of the input layer of the second refining network layer in the second refining network layer, and then sequentially passed through Three convolutions of two refining network layers The layer performs a convolution operation and performs a linear rectification function to convert the initial estimated real part matrix and the initial estimated imaginary part matrix back to the real part matrix and the imaginary part matrix. The method for rewarding channel state information based on deep learning as described in claim 3, wherein the receiving end calculates the channel state information by two-dimensional discrete Fourier transform using the coded training model, so that the channel state information is The spatial frequency is converted into an angle and time base, and the coded training model retains a value in the channel state information in which the first Nc column is not zero, to generate a truncation matrix, and splits the truncation matrix into the real part matrix and the An imaginary matrix to produce the codeword. The method for feeding back based on deep learning as channel state information according to claim 4, wherein the data compression ratio is 1/4, 1/16, 1/32 or 1/64. The feedback method based on deep learning as channel state information according to item 3 of the patent application scope, wherein the data compression ratio is 1/4 or 1/32.