CN111901024A

CN111901024A - MIMO channel state information feedback method based on fitting depth learning resistance

Info

Publication number: CN111901024A
Application number: CN202010745080.9A
Authority: CN
Inventors: 李鑫滨; 赵海红; 韩赵星; 于海峰; 骆曦
Original assignee: Yanshan University
Current assignee: Yanshan University
Priority date: 2020-07-29
Filing date: 2020-07-29
Publication date: 2020-11-06
Anticipated expiration: 2040-07-29
Also published as: CN111901024B

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 AOCN model is constructed, and a channel matrix is divided into a real part and an imaginary part, which are respectively input into the coding of the user end The encoder includes a convolutional layer and a fully connected layer. The data is encoded and passed through the feedback link to 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. The final output prediction the channel matrix. After the AOCN model is constructed, 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 AOCN 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 fitting deep learning resistance

Technical Field

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

Background

The large-scale multiple-input multiple-output (MIMO) technology is used as a key technology of a fifth generation (5G) communication system, has the advantages of high spectrum efficiency, large system capacity, strong system robustness and the like, and has higher data transmission rate and improved system reliability compared with an OFDM system in order to ensure that channel state information obtained by channel estimation can be accurately fed back to a transmitting end. Therefore, massive MIMO technology is receiving increasing attention from the industry and academia. However, the significant advantages of massive MIMO techniques depend to a large extent on the channel state information available to the transmitter for the downlink. In a fdd MIMO system, a base station needs to obtain downlink CSI through feedback from a receiving end. However, the use of large-scale antenna arrays results in a drastic increase in channel feedback overhead.

In the MIMO wireless communication system, the conventional channel state information feedback method has a serious disadvantage. At present, traditional research methods in MIMO channel state information feedback are widely applied to communication, but have many defects. First, they rely heavily on the assumption that the channel is sparse. However, the channels are not completely sparse on any basis and may even have no interpretable structure. Secondly, the compressed sensing algorithm uses random projections, and the channel structure is not fully utilized. In addition, most of the existing signal reconstruction algorithms are iterative methods, and the reconstruction speed is low.

In order to realize a high-precision and high-efficiency channel feedback method, a feedback scheme based on deep learning is provided, and the application of the deep learning in channel feedback has a good feedback effect, so that the communication has good performance, and the stability of the system is guaranteed. The deep learning theory is adopted to carry out off-line training data on the channel information at the receiving end of the communication system, and the data is recovered on line, so that the load of a feedback link is greatly reduced, and the accuracy of recovering the channel state information is improved. However, at present, the problem of overfitting in channel information feedback of a deep learning method is not considered in the researches, and the overfitting can cause the problems of reduced prediction performance, reduced model training result and the like, so that the accuracy of finally predicted channel state information is insufficient, and the quality of communication cannot be guaranteed. Therefore, how to find a deep learning algorithm which can improve the recovery speed and accuracy is the key of the channel information feedback scheme.

Disclosure of Invention

The invention provides an anti-fitting deep learning MIMO channel state information feedback method, and solves the problem that the accuracy of predicted channel state information is insufficient due to over-fitting in the conventional deep learning method.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows:

the MIMO channel state information feedback method based on fitting deep learning comprises the following steps:

(1) constructing an AOCN model, and using a convolutional neural network as an encoder and a decoder, wherein the convolutional neural network can use spatial local correlation by enhancing a local connection mode between neurons of adjacent layers, and a real part and an imaginary part of a channel matrix H are used as the input of the AOCN model;

(2) the channel matrix H data enters an encoder, the encoder is positioned at a user side for sending the data and encodes the channel matrix H into low-dimensional data, and the encoder comprises a convolutional layer and a full connection layer;

(3) the data enters a feedback link after being coded and reaches a receiving end;

(4) at a base station of a receiving end, a decoder starts decoding, and the low-dimensional data of an encoding end is subjected to secondary construction of an AOCN model; the decoder at the receiving end comprises an anti-fitting layer, a full connection layer, a RefineNet layer and a convolution layer, and outputs a predicted channel matrix;

(5) after the AOCN model is built, the model is subjected to off-line training, model parameters are initialized, the model is stored after error convergence, and finally the trained and stored AOCN model is subjected to on-line channel state information prediction.

The technical scheme of the invention is further improved as follows: the step (2) comprises the following steps:

at the encoder's convolutional layer, this layer uses a kernel of size 3 × 3 to generate two signatures; after convolutional layer, we reshape the eigenmap into a vector and use a fully-connected layer to generate codeword s, which is a vector of size mx 1; the convolutional layer and the fully-connected layer simulate the projection of the compressed sensing and act as encoders.

The technical scheme of the invention is further improved as follows: the decoder workflow in step (4) is expressed as:

obtaining the codeword s at the receiving end, and then mapping it back to the channel matrix H using the neural network layer (as a decoder); the first layer of the decoder is an anti-fitting unit, a random deactivation (Dropout) algorithm is added, and the method is that a certain proportion of node information is lost randomly in each training period, namely, a certain proportion of the output of the previous layer is changed into zero in the training stage, and the node of the next layer determines a value according to the remaining information; the second layer is a layer with s subjected to random inactivation treatment as input and two outputs with the size of N_c×N_tThe fully connected layer of the matrix of (a) as an initial estimate of the real and imaginary parts of H; then, the initial estimation number is input into a plurality of subdivided network units which are continuously refined and reconstructed; the reflexet layer includes a plurality of reflexet units, each consisting of four layers, of which the first layer is an input layer and all remaining 3 layers use 3 × 3 kernels; the second layer and the third layer generate 8 and 16 feature maps respectively, and the last layer generates a final reconstruction of H; the profile generated by the three convolutional layers is set to the input channel matrix size N by appropriate zero padding_c×N_tThe same size; selecting ReLU (x) max (x,0) as an activation function, and carrying out batch normalization processing on each layer;

after the channel matrix is refined by a series of RefineNet units, the channel matrix is input to the final convolutional layer, and the sigmoid function is used to scale the values to the [0,1] range.

The technical scheme of the invention is further improved as follows: the step (5) comprises the following steps:

to train the AOCN, we use end-to-end learning for all the kernels and bias values of the encoder and decoder; the parameter set is denoted as Θ ═ Θ_en,Θ_de}; the input of AOCN is H_iThe reconstructed channel matrix is

Notably, the input and output of the AOCN are normalized channel matrices with elements at [0,1]]Zooming in a range; similar to the autoencoder, AOCN is an unsupervised learning algorithm; the loss function is Mean Squared Error (MSE), and is calculated as follows:

wherein | · | purple₂Is the euclidean norm and T is the total number of samples in the training set.

The technical scheme of the invention is further improved as follows: the principle data of the random inactivation algorithm passes through an input layer, a plurality of hidden layers and an output layer; in the normal data transmission process, the neurons of each layer are all connected with the neurons of the next layer, so that overfitting is often caused when a large amount of data are trained, therefore, a random inactivation algorithm is provided, the neurons connected with the neurons of each layer and the next layer are randomly inactivated, the neurons are disconnected with the neurons of the next layer, the training data are reduced, and the purpose of preventing overfitting is achieved; the random inactivation algorithm only inactivates the neurons at random during training to obtain a trained model, and then all the neurons are reconnected together during testing, and a high-precision calculation result is finally output through calculation of the neural network, so that complete channel state information is obtained.

Due to the adoption of the technical scheme, the invention has the technical progress that:

the invention provides a large-scale MIMO channel state information feedback method based on fitting-resistant deep learning, which solves the problem of insufficient accuracy of predicted channel state information caused by over-fitting in the conventional deep learning method, further improves the recovery accuracy of an information matrix, ensures that a system transmitting terminal obtains accurate channel state information, and improves the communication quality of the system.

Drawings

FIG. 1 is a flow chart of AOCN model construction and training of the present invention;

FIG. 2 is a diagram of an AOCN model network architecture according to the present invention;

FIG. 3 is a schematic diagram of normal training of a neural network according to the present invention;

FIG. 4 is a schematic diagram of training with the addition of an anti-fit unit according to the present invention.

Detailed Description

The present invention will be described in further detail with reference to the following examples:

as shown in fig. 1, a MIMO channel state information feedback method based on fitting-resistant deep learning includes the following steps:

(1) constructing an AOCN (Anti-overlapping CSI net, AOCN) model, and using a convolutional neural network as an encoder and a decoder, wherein the convolutional neural network can use spatial local correlation by reinforcing a local connection mode between neurons of adjacent layers, and a real part and an imaginary part of a channel matrix H are used as input of the convolutional neural network;

(2) the channel matrix H data enters an encoder, the encoder is positioned at a user end for sending the data and encodes the channel matrix H into low-dimensional data, and the encoder comprises a convolutional layer and a full connection layer. The method comprises the following steps: at the encoder's convolutional layer, this layer uses a kernel of size 3 × 3 to generate two signatures. After convolutional layers, we reshape the eigenmap into a vector and use a fully-connected layer to generate the codeword s, which is a vector of size mx 1. The first two layers simulate the projection of compressed sensing and act as encoders;

(4) at the base station of the receiving end, the decoder starts decoding, and low-dimensional data of the encoding end is newly constructed. The decoder at the receiving end comprises an anti-fitting layer, a full connection layer, a RefineNet layer and a convolution layer, and outputs a predicted channel matrix. The decoder workflow is represented as: at the receiving end, the codeword s is obtained and then mapped back to the channel matrix H using the neural network layer (as a decoder). The first layer of the decoder we add a random deactivation (Dropout) algorithm.

The principle of the random inactivation algorithm is shown in fig. 3 and 4, and the principle data of the random inactivation algorithm passes through an input layer, a plurality of hidden layers and an output layer. In the normal data transmission process, data is transmitted as shown in fig. 3, all neurons in each layer are connected with neurons in the next layer, so that an overfitting phenomenon is often caused when a large amount of data is trained, therefore, a random inactivation algorithm is provided, which carries out random inactivation on the neurons in each layer connected with the next layer to disconnect the neurons from the neurons in the next layer, so that the training data is reduced, and the purpose of preventing overfitting is achieved. The random inactivation algorithm only inactivates the neurons at random during training to obtain a trained model, and then all the neurons are reconnected together during testing, and a high-precision calculation result is finally output through calculation of the neural network, so that complete channel state information is obtained.

The second layer is a layer with s subjected to random inactivation treatment as input and two outputs with the size of N_c×N_tAs an initial estimate of the real and imaginary parts of H. The initial estimated number is then input into several subdivided network elements that are progressively refined and reconstructed. Each reflonenet unit consists of four layers, of which 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 profile generated by the three convolutional layers is set to the input channel matrix size N by appropriate zero padding_c×N_tThe same size. And selecting ReLU (x) max (x,0) as an activation function, and performing batch normalization processing on each layer.

(5) The AOCN model is constructed, as shown in fig. 2, then the model is trained offline, the model parameters are initialized, the model is stored after error convergence, and finally the trained and stored AOCN model is used to predict the channel state information online. To train the AOCN, we use end-to-end learning for all the kernels and bias values of the encoder and decoder. The parameter set is denoted as Θ ═ Θ_en，Θ_de}. The input of AOCN is H_iThe reconstructed channel matrix is

Notably, the input and output of the AOCN are normalized channel matrices with elements at [0,1]]And (4) zooming in and out. Similar to an autoencoder, AOCN is an unsupervised learning algorithm. The loss function is Mean Squared Error (MSE), and is calculated as follows:

The above-mentioned embodiments are merely illustrative of the preferred embodiments of the present invention, and do not limit the scope of the present invention, and various modifications and improvements of the technical solution of the present invention by those skilled in the art should fall within the protection scope defined by the claims of the present invention without departing from the spirit of the present invention.

Claims

1. The MIMO channel state information feedback method based on anti-fitting deep learning is characterized by comprising the following steps:

(1) Construct the AOCN model, using the convolutional neural network as the encoder and decoder, the convolutional neural network can take advantage of the spatial local correlation by strengthening the local connection pattern between neurons in adjacent layers, the real part of the channel matrix H and the imaginary part as the input of the AOCN 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 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 convolutional layer. channel matrix;

(5) After the AOCN model is constructed, 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 AOCN model is used to predict the channel state information online.

2. The MIMO channel state information feedback method based on anti-fitting deep learning according to claim 1, wherein the step (2) comprises:

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.

3. The MIMO channel state information feedback method based on anti-fitting deep learning according to claim 1, is characterized in that: 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 the decoder) is used to map it back to the channel matrix H; the first layer of the decoder is the anti-fitting layer, and we add a random deactivation algorithm, which is In a training cycle, a certain proportion of node information is randomly lost, that is, a certain proportion of the output of the previous layer is changed to zero in this training phase, and the nodes in the next layer can determine the value according to the remaining information; the second layer is a The s that has undergone random deactivation 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; then, the initial estimates are input to several continuous In the subdivision network unit of the refinement reconstruction; 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 kernel; 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 match the input channel matrix The 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.

4. The MIMO channel state information feedback method based on anti-fitting deep learning according to claim 1, is characterized in that: step (5) comprises the following content:

To train AOCN, we use end-to-end learning for all kernel and bias values of encoder and decoder; the parameter set is denoted as Θ={Θ _en ,Θ _de }; the input of _AOCN is Hi, the reconstructed channel matrix for

It is worth noting that both the input and output of AOCN are normalized channel matrices whose elements are scaled in the range [0, 1]; similar to autoencoders, AOCN is an unsupervised learning algorithm; the loss function is The mean squared error (MSE) is calculated as follows:

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

5. The MIMO channel state information feedback method based on anti-fitting deep learning according to claim 3, characterized in that: the random deactivation algorithm principle data 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 when a large amount of data is used for training, it will often cause overfitting. Therefore, the random deactivation algorithm is proposed, which will The neurons connected to the next layer are randomly deactivated, so that the neurons are disconnected from the neurons in the next layer. The purpose of this is to reduce training data and prevent overfitting; the random deactivation algorithm is only The neurons are randomly deactivated during training, and after the trained model is obtained, all neurons will be reconnected during testing. After the calculation of the neural network, high-precision calculation results are finally output, so that a complete channel can be obtained. status information.