CN111901024B - MIMO channel state information feedback method based on fitting depth learning resistance - Google Patents

Abstract

The invention discloses a MIMO channel state information feedback method based on anti-fitting deep learning, which belongs to the field of communication and includes the following steps: firstly, an Anti-overfitting CSI net model is constructed, and a channel matrix is divided into a real part and an imaginary part for input respectively The encoder that enters the user side includes a convolutional layer and a fully connected layer. The data is encoded and passed through the feedback link to reach the receiving end. The decoder at the receiving end includes an anti-fitting layer, a fully connected layer, a RefineNet layer, and a convolutional layer. , and finally output the predicted channel matrix. After the model is built, the model is trained offline. First, the model parameters are initialized, and the model is saved after the error converges. Finally, the trained and saved model is used to predict the channel state information online. The invention can further improve the recovery precision of the information matrix, ensure that the system transmitter obtains accurate channel state information, and improve the communication quality of the system.

Description

MIMO channel state information feedback method based on anti-fitting deep learning

technical field

The invention relates to the field of communication, in particular to a massive MIMO channel state information feedback method based on anti-fitting deep learning.

Background technique

As the key technology of the fifth generation (5G) communication system, massive multiple-input multiple-output (MIMO) technology has the advantages of high spectral efficiency, large system capacity, and strong system robustness. With accurate feedback to the transmitting end, the MIMO system has a higher data transmission rate than the OFDM system, and improves the reliability of the system. Therefore, massive MIMO technology has attracted more and more attention from industry and academia. However, the significant advantages of massive MIMO technology depend to a large extent on the transmitter's availability of channel state information for the downlink. In a frequency division duplex massive MIMO system, the base station needs to obtain downlink CSI through feedback from the receiver. However, the use of large-scale antenna arrays leads to a dramatic increase in channel feedback overhead.

In a MIMO wireless communication system, the traditional channel state information feedback method has serious shortcomings. At present, the traditional research methods in MIMO channel state information feedback have been widely used in communication, but there are many shortcomings. First, they rely heavily on the assumption that the channel is sparse. However, channels are not completely sparse on any basis, and may even have no interpretable structure. Second, compressed sensing algorithms use random projections and do not fully exploit the channel structure. In addition, most of the existing signal reconstruction algorithms are iterative methods, and the reconstruction speed is relatively slow.

In order to realize the channel feedback method with high accuracy and high efficiency, a feedback scheme based on deep learning is proposed. The application of deep learning in channel feedback has a good feedback effect, which makes the communication have good performance and ensures the stability of the system. Using deep learning theory to conduct offline training data for channel information at the receiving end of the communication system, and restore data online, greatly reduce the burden of the feedback link and improve the accuracy of restoring channel state information. However, at present, these studies have not considered the problem of over-fitting in the channel information feedback of deep learning methods. Over-fitting will lead to problems such as reduced prediction performance and model training results, resulting in insufficient accuracy of the final predicted channel state information, and communication cannot be guaranteed. quality. Therefore, how to find a deep learning algorithm that can improve the recovery speed and accuracy is the key to the channel information feedback scheme.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a MIMO channel state information feedback method based on anti-fitting deep learning, which solves the problem of insufficient accuracy of predicted channel state information due to overfitting in the current deep learning method.

For solving the above-mentioned technical problems, the technical scheme adopted in the present invention is:

The MIMO channel state information feedback method based on anti-fitting deep learning includes the following steps:

(1) Constructing an Anti-overfitting CSI net model, using convolutional neural networks as encoders and decoders, convolutional neural networks can utilize spatial local correlation by strengthening the local connection patterns between neurons in adjacent layers, and the channel matrix The real and imaginary parts of H are used as the input of the Anti-overfitting CSI net model;

(2) The channel matrix H data enters the encoder, the encoder is located at the user end sending the data, and the channel matrix H is encoded into a low-dimensional data, and the encoder includes a convolution layer and a fully connected layer;

(3) After the data is encoded, it enters the feedback link and reaches the receiving end;

(4) At the base station at the receiving end, the decoder starts decoding, and rebuilds the Anti-overfitting CSI net model from the low-dimensional data at the encoding end; the decoder at the receiving end includes an anti-fitting layer, a fully connected layer, a RefineNet layer, and a volume Multilayer, output the predicted channel matrix;

(5) After the construction of the Anti-overfitting CSI net model is completed, the model is trained offline. First, the model parameters are initialized, and the model is saved after the error converges. Finally, the trained and saved Anti-overfitting CSI net model is used to predict the channel state information online.

The further improvement of the technical solution of the present invention is: step (2) includes:

In the convolutional layer of the encoder, this layer uses a kernel of size 3×3 to generate two feature maps; after the convolutional layer, we reshape the feature map into a vector and use a fully connected layer to generate codeword s, which is a vector of size M × 1; convolutional and fully connected layers simulate the projection of compressed sensing and act as encoders.

The further improvement of the technical solution of the present invention is: the decoder workflow described in step (4) is expressed as:

The codeword s is obtained at the receiving end, and then the neural network layer (as a decoder) is used to map it back to the channel matrix H; the first layer of the decoder is an anti-fitting unit, and we add a random dropout (Dropout) algorithm, its method In order to randomly lose a certain proportion of node information in each training cycle, that is, to change a certain proportion of the output of the previous layer to zero in this training phase, and let the nodes of the next layer determine the value according to the remaining information; the second layer is a fully connected layer that takes the randomly deactivated s as input and outputs two matrices of size N _c ×N _t as the initial estimates of the real and imaginary parts of H; then, the initial estimates are input to In several subdivision network units that are continuously refined and reconstructed; the RefineNet layer includes multiple RefineNet units, and each RefineNet unit consists of four layers. In the RefineNet unit, the first layer is the input layer, and all the remaining 3 layers use 3 ×3 kernels; the second and third layers generate 8 and 16 feature maps, respectively, and the last layer generates the final reconstruction of H; with appropriate zero-padding, the feature maps generated by the three convolutional layers are set to The input channel matrix size is the same as N _c ×N _t ; ReLU(x)=max(x,0) is selected as the activation function, and batch normalization is performed on each layer;

After the channel matrix is refined through a series of RefineNet units, the channel matrix is input to the final convolutional layer, and the values are scaled to the [0, 1] range using the sigmoid function.

The further improvement of the technical scheme of the present invention is: step (5) comprises the following content:

值得注意的是，Anti-overfitting CSI net的输入和输出都是归一化的通道矩阵，其元素在[0,1]范围内缩放；与自动编码器类似，Anti-overfitting CSI net是一种无监督学习算法；损失函数为均方误差(mean squared error,MSE)，计算方法如下：To train the Anti-overfitting CSI net, we use end-to-end learning for all kernel and bias values of the encoder and decoder; the parameter set is denoted as Θ={Θ _en ,Θ _de }; the input to the Anti-overfitting CSI net is H _i , the reconstructed channel matrix is

It is worth noting that both the input and output of Anti-overfitting CSI net are normalized channel matrices whose elements are scaled in the range [0, 1]; similar to autoencoders, Anti-overfitting CSI net is a Supervised learning algorithm; the loss function is mean squared error (MSE), and the calculation method is as follows:

where ||·|| ₂ is the Euclidean norm and T is the total number of samples in the training set.

The further improvement of the technical solution of the present invention is: the principle data of the random deactivation algorithm passes through the input layer, the multi-layer hidden layer and the output layer; in the normal data transmission process, the neurons of each layer are all connected to the neurons of the next layer. , so that when a large amount of data is used for training, it often causes overfitting. Therefore, a random deactivation algorithm is proposed, which randomly deactivates the neurons connected to each layer and the next layer, so that the neural The purpose of this is to reduce the training data and prevent overfitting; the random deactivation algorithm just randomly deactivates the neurons during training. During the test, all neurons will be reconnected together, and after the calculation of the neural network, the final output of high-precision calculation results, so that it can obtain complete channel state information.

Owing to having adopted the above-mentioned technical scheme, the technical progress that the present invention obtains is:

The invention provides a massive MIMO channel state information feedback method based on anti-fitting deep learning, solves the problem of insufficient accuracy of predicted channel state information due to overfitting in the current deep learning method, further improves the restoration accuracy of the information matrix, and ensures the system The transmitting end obtains accurate channel state information to improve the communication quality of the system.

Description of drawings

Fig. 1 is the Anti-overfitting CSI net model construction and training flow chart of the present invention;

Fig. 2 is the Anti-overfitting CSI net model network architecture diagram of the present invention;

3 is a schematic diagram of the normal training of the neural network of the present invention;

FIG. 4 is a schematic diagram of adding an anti-fitting unit for training according to the present invention.

Detailed ways

Below in conjunction with embodiment, the present invention is described in further detail:

As shown in Figure 1, a MIMO channel state information feedback method based on anti-fitting deep learning includes the following steps:

(1) Construct AOCN (Anti-overfitting CSI net, AOCN) model, using convolutional neural network as encoder and decoder, convolutional neural network can utilize spatial localization by strengthening the local connection pattern between neurons in adjacent layers correlation, the real and imaginary parts of the channel matrix H as its input;

(2) The channel matrix H data enters the encoder. The encoder is located at the user end that sends the data, and encodes the channel matrix H into a low-dimensional data. The encoder includes a convolution layer and a fully connected layer. It consists of the following: In the convolutional layer of the encoder, this layer uses a kernel of size 3×3 to generate two feature maps. After the convolutional layer, we reshape the feature map into a vector and use a fully connected layer to generate the codeword s, which is a vector of size M × 1. The first two layers simulate the projection of compressed sensing and act as encoders;

(3) After the data is encoded, it enters the feedback link and reaches the receiving end;

(4) At the base station at the receiving end, the decoder starts decoding, and reconstructs the low-dimensional data at the encoding end. The decoder at the receiving end includes an anti-fitting layer, a fully connected layer, a RefineNet layer, a convolutional layer, and outputs the predicted channel matrix. The decoder workflow is expressed as: the codeword s is obtained at the receiver and subsequently mapped back to the channel matrix H using a neural network layer (as the decoder). In the first layer of the decoder, we add a random dropout algorithm.

The principle of the random deactivation algorithm is shown in Figure 3 and Figure 4. The principle of the random deactivation algorithm The data passes through the input layer, the multi-layer hidden layer, and the output layer. In the normal data transmission process, the data is transmitted as shown in Figure 3, and the neurons of each layer are all connected to the neurons of the next layer, so when a large amount of data is used for training, it will often cause overfitting. Therefore, The random deactivation algorithm is proposed, which randomly deactivates the neurons connected to the next layer in each layer, so that the neurons are disconnected from the neurons in the next layer. The purpose of this is to reduce the training data and achieve The purpose of preventing overfitting. The random deactivation algorithm just randomly deactivates neurons during training. After the trained model is obtained, all neurons will be reconnected during testing. After the calculation of the neural network, the final output of high-precision calculation results, so that to obtain complete channel state information.

The second layer is a fully connected layer that takes the randomly deactivated s as input and outputs two matrices of size N _c ×N _t as initial estimates of the real and imaginary parts of H. The initial estimates are then fed into several subdivision network units that are continuously refined and reconstructed. Each RefineNet unit consists of four layers, in a RefineNet unit, the first layer is the input layer, and all the remaining 3 layers use 3×3 kernels. The second and third layers generate 8 and 16 feature maps, respectively, and the last layer generates the final reconstruction of H. The feature maps generated by the three convolutional layers are set to the same size as the input channel matrix size _Nc × _Nt with appropriate zero-padding. Select ReLU(x)=max(x,0) as the activation function, and batch normalize each layer.

After the channel matrix is refined through a series of RefineNet units, the channel matrix is input to the final convolutional layer, and the values are scaled to the [0, 1] range using the sigmoid function.

值得注意的是，Anti-overfitting CSI net的输入和输出都是归一化的通道矩阵，其元素在[0,1]范围内缩放。与自动编码器类似，Anti-overfitting CSI net是一种无监督学习算法。损失函数为均方误差(mean squared error,MSE)，计算方法如下：(5) The construction of the Anti-overfitting CSI net model is completed, as shown in Figure 2, and then the model is trained offline. First, the model parameters are initialized, and the model is saved after the error converges. Finally, the trained and saved Anti-overfitting CSI net model is predicted online. Channel state information. To train the Anti-overfitting CSI net, we use end-to-end learning for all kernel and bias values of the encoder and decoder. The parameter set is denoted as Θ = {Θ _en , Θ _de }. The input of Anti-overfitting CSI net is H _i , and the reconstructed channel matrix is

It is worth noting that both the input and output of the Anti-overfitting CSI net are normalized channel matrices whose elements are scaled in the range [0, 1]. Similar to autoencoders, Anti-overfitting CSI net is an unsupervised learning algorithm. The loss function is mean squared error (MSE), and the calculation method is as follows:

where ||·|| ₂ is the Euclidean norm and T is the total number of samples in the training set.

The above-mentioned embodiments are only to describe the preferred embodiments of the present invention, and do not limit the scope of the present invention. On the premise of not departing from the design spirit of the present invention, those of ordinary skill in the art can Such deformations and improvements shall fall within the protection scope determined by the claims of the present invention.

Claims (3)

1. The MIMO channel state information feedback method based on anti-fitting deep learning is characterized by comprising the following steps: (1) Construct an Anti-overfitting CSI net model, using convolutional neural network as encoder and decoder, convolutional neural network utilizes spatial local correlation by strengthening the local connection pattern between neurons in adjacent layers, the channel matrix H The real and imaginary parts are used as the input of the Anti-overfitting CSI net model; (2) The channel matrix H data enters the encoder, the encoder is located at the user end sending the data, and the channel matrix H is encoded into a low-dimensional data, and the encoder includes a convolution layer and a fully connected layer; (3) After the data is encoded, it enters the feedback link and reaches the receiving end; (4) At the base station at the receiving end, the decoder starts to decode and reconstructs the low-dimensional data at the encoding end; the decoder at the receiving end includes an anti-fitting layer, a fully connected layer, a RefineNet layer, and a convolution layer, and outputs the predicted channel matrix; (5) After the construction of the Anti-overfitting CSI net model is completed, the model is trained offline. First, the model parameters are initialized, and the model is saved after the error converges. Finally, the trained and saved Anti-overfitting CSI net model is used to predict the channel state information online; The decoder workflow described in step (4) is expressed as: The codeword s is obtained at the receiving end, and then mapped back to the channel matrix H using a convolutional neural network; the first layer of the decoder is an anti-fitting layer, and we add a random deactivation algorithm, which is randomly selected in each training cycle. Losing a certain proportion of node information, that is, changing a certain proportion of the output of the previous layer to zero in this training phase, and letting the nodes of the next layer determine the value according to the remaining information; the second layer is a random deactivation method. The processed s is the input and outputs a fully connected layer of two matrices of size N _c ×N _t as the initial estimates of the real and imaginary parts of H, where: N _c is the number of receive antennas and N _t is the number of transmit antennas ; Then, the initial estimates are input into several subdivision network units that are continuously refined and reconstructed; the RefineNet layer includes multiple RefineNet units, each RefineNet unit consists of four layers, in the RefineNet unit, the first layer is the input layer , all remaining 3 layers use 3 × 3 kernels; the second and third layers generate 8 and 16 feature maps, respectively, and the last layer generates the final reconstruction of H; with appropriate zero-padding, the three volumes The feature map generated by the stacked layer is set to the same size as the input channel matrix size N _c ×N _t ; ReLU(x)=max(x, 0) is selected as the activation function, and batch normalization is performed on each layer; After the channel matrix is refined through a series of RefineNet units, the channel matrix is input to the final convolutional layer, and the values are scaled to the [0,1] range using the sigmoid function; Step (5) includes the following: To train the Anti-overfitting CSI net, we use end-to-end learning for all kernel and bias values of the encoder and decoder; the parameter set is denoted as Θ={Θ _en ,Θ _de }, where Θ _en is the encoding parameter set, Θ _de is the decoding parameter set; the input of Anti-overfitting CSI net is H _i , and the reconstructed channel matrix is

Among them, H _i is the real channel matrix, f _de is the decoding function, f _en is the encoding function, and si is the compressed codeword of the _i -th channel matrix; it is worth noting that the input and The outputs are all normalized channel matrices whose elements are scaled in the range [0,1]; the loss function is mean squared error (MSE), which is calculated as follows:

where ||·|| ₂ is the Euclidean norm and T is the total number of samples in the training set. 2. The MIMO channel state information feedback method based on anti-fitting deep learning according to claim 1, wherein step (2) comprises: In the convolutional layer of the encoder, this layer uses a kernel of size to generate two feature maps; after the convolutional layer, we reshape the feature map into a vector and use a fully connected layer to generate the codeword s , which is a vector of size ; the convolutional and fully connected layers simulate the projection of compressed sensing and act as encoders. 3. The MIMO channel state information feedback method based on anti-fitting deep learning according to claim 1, is characterized in that: the principle data of random deactivation algorithm passes through input layer, multi-layer hidden layer, output layer; in normal data transmission process The neurons of each layer are all connected to the neurons of the next layer, and the random deactivation algorithm randomly deactivates the neurons that are connected to the next layer, so that the neurons of the next layer are connected to the neurons of the next layer. Unit disconnection; the random deactivation algorithm randomly deactivates neurons during training, and after the trained model is obtained, all neurons will be reconnected during testing.

Images