A neural network design method and device for channel estimation

Abstract

The present invention provides a neural network design method and device for channel estimation, which relates to the field of wireless communication technology. It includes: calculating the current estimated value CSI entropy through the historical estimated value of the channel state; modeling the change of CSI as a first-order Markov model to obtain the amount of information required for the CSI matrix of the channel state information at the current moment; constructing a new network structure using the amount of information required for the CSI matrix at the current moment, training the new neural network offline, determining the weight parameters and bias parameters, and completing the neural network design for channel estimation. The present invention proposes a neural network structure for channel estimation that does not rely on pilots, and uses information entropy theory to evaluate the strong time correlation characteristics of wireless channel data to improve the performance of the channel estimation algorithm based on the neural network. The neural network structure for channel estimation proposed by the present invention can improve the accuracy of channel estimation, reduce the delay of channel estimation, and meet the real-time requirements of wireless communication systems.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a neural network design method and device for channel estimation. Background Art

[0002] The method and accuracy of channel estimation affect the transmission quality of mobile communication systems. Channel estimation algorithms obtain CSI (Channel State Information) by mining and revealing the internal characteristics of wireless transmission signals, such as inserting pilot signals, approximating minimum error and nonlinear mapping. At present, commonly used channel estimation methods include LS (Least Square), MMSE (Minimum Mean Square Error), LMMSE (Linear Minimum Mean Square Error), compressed sensing channel estimation algorithm, neural network algorithm, etc., which cannot obtain accurate results when pilot signals are limited. At the same time, the rapid change of the channel requires the insertion of more pilot signals to ensure the estimation performance, resulting in a decrease in spectrum efficiency. Among them, neural networks have been widely used in channel estimation in different scenarios due to their excellent nonlinear mapping ability and supervised learning using training sets with obvious data features. They can fit and approximate any complex function.

[0003] Most estimation algorithms based on neural networks use existing network structures oriented to image data features, such as convolutional layers, pooling layers, and batch normalization (BN) layers. Although good channel estimation performance is achieved, there is still a lack of neural network structures designed for the time correlation of wireless channels to further improve channel estimation accuracy and reduce the time loss of channel estimation.

[0004] The estimation algorithms based on neural networks in the existing technology have certain limitations. Most neural network structures originate from the field of image processing. There is a lack of neural network structures designed for the characteristics of wireless channel data, which often leads to the failure to achieve the expected accuracy in channel estimation. It also increases the time consumption in the neural network training and deployment stages. Summary of the invention

[0005] In view of the problem that the prior art lacks a neural network structure designed for wireless channel data characteristics and cannot achieve the expected accuracy, the present invention proposes a neural network design method and device for channel estimation.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] On the one hand, a neural network design method for channel estimation is provided, comprising:

[0008] S1: Obtain the CSI change of the current channel state estimate value through the historical estimate value of the channel state, and model the CSI change as a first-order Markov model;

[0009] S2: The CSI conditional entropy is calculated by the first-order Markov model. The CSI conditional entropy represents the amount of information required to obtain the CSI matrix of the channel state information at the current moment.

[0010] S3: construct a new recurrent neural network structure based on the obtained CSI conditional entropy, and the new recurrent neural network structure is used for channel estimation;

[0011] S4: Perform offline training on the new recurrent neural network, determine weight parameters and bias parameters, and complete the neural network design for channel estimation.

[0012] Optionally, step S1 further includes:

[0013] Based on the historical estimated value, the amount of information required for estimating the CSI entropy at the current moment is obtained according to the following formula (1):

[0014]

[0015] Among them, e t represents the amount of information required to estimate the CSI entropy at time t; represents CSI conditional entropy; H(·) represents entropy operation; represents the CSI matrix estimated at time t; W c represents the weight parameter, b c Represents the bias parameter.

[0016] Optionally, in step S1, the change of CSI is modeled as a first-order Markov model, including:

[0017] According to the following formula (2), the change of CSI is modeled as a first-order Markov model:

[0018] p(H t |H t-1 ,...,H1)=p(H t |H t-1 ) (2)

[0019] Among them, H t represents the CSI matrix at time t.

[0020] Optionally, based on formula (2), on the basis of the historical estimated value, according to the following formula (3), the amount of information required for the CSI matrix at the current moment is obtained through the CSI conditional entropy:

[0021]

[0022] Optionally, in step S3, a new recurrent neural network structure is constructed using the obtained CSI conditional entropy, including:

[0023] The CSI conditional entropy is embedded into the recurrent neural network using the linear rectification function ReLU to construct a new recurrent neural network structure, which is used for channel estimation.

[0024] Optionally, in step S4, offline training is performed on the new recurrent neural network to determine weight parameters and bias parameters to complete the neural network design for channel estimation, including:

[0025] S41: according to the amount of information required by the CSI matrix at the current moment, input the received signal matrix Y into the new recurrent neural network and output the CSI matrix H;

[0026] S42: Train the new recurrent neural network through the neural network training algorithm until convergence, and determine the final weight parameter W c and the bias parameter b c , complete the design of neural network for channel estimation.

[0027] Optionally, the received signal matrix Y is obtained according to the following formula (4):

[0028] Y=XH+N (4)

[0029] Wherein, X represents the transmitted signal, N represents the additive white Gaussian noise, and H represents the CSI matrix.

[0030] Optionally, the neural network training algorithm includes but is not limited to a gradient descent method, a Newton algorithm, or an adaptive moment estimation Adam algorithm.

[0031] Optionally, the CSI matrix is ​​obtained according to an existing channel model, or using channel data of a measured real environment; the channel model includes but is not limited to a spatial channel SCM model, a Jakes model, a WINNER model, or a 3rd Generation Partnership Project 3GPP standard channel model.

[0032] On the one hand, a neural network design device for channel estimation is provided, the device comprising:

[0033] A CSI entropy calculation module is used to obtain the CSI change of the current channel state estimation value through the historical estimation value of the channel state, and model the CSI change as a first-order Markov model;

[0034] An information amount calculation module is used to obtain CSI conditional entropy through a first-order Markov model calculation, where the CSI conditional entropy represents the amount of information required to obtain the CSI matrix of the channel state information at the current moment;

[0035] A network structure building module is used to construct a new recurrent neural network structure through the obtained CSI conditional entropy, and the new recurrent neural network structure is used for channel estimation;

[0036] The offline training module is used to perform offline training on the new neural network through the amount of information required by the CSI matrix of the current channel state, determine the weight parameters and bias parameters, and complete the design of the neural network for channel estimation.

[0037] The above technical solution of the embodiment of the present invention has at least the following beneficial effects:

[0038] In the above scheme, the present invention proposes a neural network structure for channel estimation that does not rely on pilots, and uses information entropy theory to evaluate the strong time correlation characteristics of wireless channel data to improve the performance of the channel estimation algorithm based on neural networks. The neural network structure for channel estimation proposed by the present invention can improve the accuracy of channel estimation, reduce the delay of channel estimation, and meet the real-time requirements of wireless communication systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 is a flow chart of a neural network design method for channel estimation provided by an embodiment of the present invention;

[0041] Figure 2 is a flow chart of a neural network design method for channel estimation provided by an embodiment of the present invention;

[0042] Figure 3 is a CeRU structure diagram of a neural network design method for channel estimation provided by an embodiment of the present invention;

[0043] Figure 4 It is a CeLSTM channel estimation algorithm architecture diagram of a neural network design method for channel estimation provided by an embodiment of the present invention;

[0044] Figure 5 is a MSE performance comparison diagram of a neural network design method for channel estimation provided by an embodiment of the present invention;

[0045] Figure 6 is a delay comparison diagram of a neural network design method for channel estimation provided by an embodiment of the present invention;

[0046] Figure 7 It is a device block diagram of a neural network design device for channel estimation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, an embodiment of the present invention provides a neural network design method for channel estimation, including:

[0049] S101: obtaining a CSI change of a channel state estimation value at a current moment through a historical estimation value of the channel state, and modeling the CSI change as a first-order Markov model;

[0050] S102: obtaining CSI conditional entropy through calculation using a first-order Markov model, where the CSI conditional entropy represents the amount of information required to obtain the CSI matrix of the channel state information at the current moment;

[0051] S103: constructing a new recurrent neural network structure by using the obtained CSI conditional entropy, and the new recurrent neural network structure is used for channel estimation;

[0052] S104: Perform offline training on the new recurrent neural network, determine weight parameters and bias parameters, and complete the design of the neural network for channel estimation.

[0053] Optionally, step S101 further includes:

[0054] Based on the historical estimated value, the amount of information required for estimating the CSI entropy at the current moment is obtained according to the following formula (1):

[0055]

[0056] Among them, e t represents the amount of information required to estimate the CSI entropy at time t; represents CSI conditional entropy; H(·) represents entropy operation; represents the CSI matrix estimated at time t; W c represents the weight parameter, b c Represents the bias parameter.

[0057] Optionally, in step S1, the change of CSI is modeled as a first-order Markov model, including:

[0058] According to the following formula (2), the change of CSI is modeled as a first-order Markov model:

[0059] p(H t|H t-1 ,...,H1)=p(H t |H t-1 ) (2)

[0060] Among them, H t represents the CSI matrix at time t.

[0061] Optionally, based on formula (2), on the basis of the historical estimated value, according to the following formula (3), the amount of information required for the CSI matrix at the current moment is obtained through the CSI conditional entropy:

[0062]

[0063] Optionally, in step S103, a new recurrent neural network structure is constructed by using the obtained CSI conditional entropy, including:

[0064] The CSI conditional entropy is embedded into the recurrent neural network using the linear rectification function ReLU to construct a new recurrent neural network structure, which is used for channel estimation.

[0065] Optionally, in step S104, offline training is performed on the new recurrent neural network to determine weight parameters and bias parameters to complete the neural network design for channel estimation, including:

[0066] S141: According to the amount of information required by the CSI matrix at the current moment, a received signal matrix Y is input into the new recurrent neural network, and a CSI matrix H is output;

[0067] S142: Train the new recurrent neural network through the neural network training algorithm until convergence, and determine the final weight parameter W c and the bias parameter b c , complete the design of neural network for channel estimation.

[0068] Optionally, the received signal matrix Y is obtained according to the following formula (4):

[0069] Y=XH+N (4)

[0070] Wherein, X represents the transmitted signal, N represents the additive white Gaussian noise, and H represents the CSI matrix.

[0071] Optionally, the neural network training algorithm includes but is not limited to a gradient descent method, a Newton algorithm, or an adaptive moment estimation Adam algorithm.

[0072] Optionally, the CSI matrix is ​​obtained according to an existing channel model, or using channel data of a measured real environment; the channel model includes but is not limited to a spatial channel SCM model, a Jakes model, a WINNER model, or a 3rd Generation Partnership Project 3GPP standard channel model.

[0073] In an embodiment of the present invention, the present invention proposes a neural network structure for channel estimation that does not rely on pilot signals, and uses information entropy theory to evaluate the strong time correlation characteristics of wireless channel data to improve the performance of a neural network-based channel estimation algorithm. The neural network structure for channel estimation proposed in the present invention can improve the accuracy of channel estimation, reduce the delay of channel estimation, and meet the real-time requirements of wireless communication systems.

[0074] like Figure 2 As shown, an embodiment of the present invention provides a neural network design method for channel estimation, including:

[0075] S201: Obtain a CSI change of a channel state estimation value at a current moment through a historical estimation value of the channel state, and model the CSI change as a first-order Markov model.

[0076] In a feasible implementation, the present invention proposes a neural network structure for channel estimation that is independent of pilot signals, uses information entropy theory to evaluate the strong time correlation characteristics of wireless channel data to improve the performance of a neural network-based channel estimation algorithm, and proposes a neural network structure that uses information entropy to achieve channel estimation.

[0077] In a feasible implementation, the recurrent neural network includes but is not limited to LSTM, BiLSTM, etc.

[0078] In a feasible implementation, based on the historical estimated value, the amount of information required for estimating the CSI entropy at the current moment is obtained according to the following formula (1):

[0079]

[0080] Among them, e t represents the amount of information required to estimate the CSI entropy at time t; represents CSI conditional entropy, H(·) represents entropy operation; represents the CSI matrix estimated at time t; W c represents the weight parameter, b c Represents the bias parameter.

[0081] In a feasible implementation mode, firstly, According to the following formula (2), the change of CSI is modeled as a first-order Markov model:

[0082] p(H t |H t-1 ,...,H1)=p(H t |H t-1 ) (2)

[0083] Among them, H t represents the CSI matrix at time t.

[0084] S202: Obtain CSI conditional entropy through first-order Markov model calculation, where the CSI conditional entropy represents the amount of information required to obtain the CSI matrix of the channel state information at the current moment.

[0085] In a feasible implementation, based on formula (2), on the basis of historical estimated values, according to the following formula (3), the amount of information required by the CSI matrix at the current moment is obtained through the CSI conditional entropy:

[0086]

[0087] S203: construct a new recurrent neural network structure through the obtained CSI conditional entropy, and the new recurrent neural network structure is used for channel estimation.

[0088] In a feasible implementation, Figure 3 As shown, the CSI conditional entropy is embedded into the recurrent neural network using the linear rectification function ReLU to construct a new recurrent neural network structure, the new recurrent neural network structure is CeRU (CSIEntropy-based ReLU, a linear rectification function based on channel state information); the new recurrent neural network structure is used for channel estimation.

[0089] S204: Based on the amount of information required by the CSI matrix at the current moment, the received signal matrix Y is input into the new recurrent neural network, and the CSI matrix H is output.

[0090] In a feasible implementation, during the training phase of the model, the input of the new recurrent neural network is the received signal matrix Y, and the output is the channel state matrix H, according to the amount of information required by the CSI matrix at the current moment.

[0091] In a feasible implementation manner, the received signal matrix Y is obtained according to the following formula (4):

[0092] Y=XH+N (4)

[0093] Wherein, X represents the transmitted signal, N represents the additive white Gaussian noise, and H represents the CSI matrix.

[0094] In a feasible implementation, the CSI matrix is ​​obtained based on an existing channel model, or using channel data of a measured real environment; the channel model includes but is not limited to a spatial channel SCM model, a Jakes model, a WINNER model, or a 3GPP standard channel model.

[0095] S205: Train the new recurrent neural network through a neural network training algorithm until convergence, and determine the final weight parameter W c and the bias parameter b c , complete the design of neural network for channel estimation.

[0096] In a feasible implementation, the neural network training algorithm includes but is not limited to a gradient descent method, a Newton algorithm, or an adaptive moment estimation Adam (Adaptive Moment Estimation) algorithm.

[0097] In a feasible implementation, Figure 4 The above method is simulated and implemented as shown. The LSTM (Long Short-Term Memory) based on the CeRU structure, namely CeLSTM, is used to estimate the channel of the wireless communication system in the time-varying non-stationary scenario. The channel state matrix is ​​generated by the WINNER II model, with a frequency of 2.6GHz and a moving speed of 80km / h. The Adam algorithm is used to train CeLSTM offline until convergence, and the weight parameter W is determined. c and the bias parameter b c In the online estimation stage, CeLSTM takes the received signal matrix Y as input and outputs the estimated channel state matrix H.

[0098] In a feasible implementation, Figure 5 As shown in FIG, the MSE performance comparison of the CeLSTM channel estimation algorithm of the present invention with the classic channel estimation algorithm and the LSTM algorithm is shown in FIG. Figure 6 The figure shows the delay comparison between the CeLSTM channel estimation algorithm of the present invention and the LSTM algorithm. It can be seen that the neural network structure for channel estimation proposed in the present invention can improve the accuracy of channel estimation, reduce the delay of channel estimation, and meet the real-time requirements of wireless communication systems.

[0099] like Figure 7 As shown, an embodiment of the present invention provides a neural network design device 300 for channel estimation, the device comprising:

[0100] The CSI entropy calculation module 310 obtains a change estimate value through a historical estimate value of the channel state, wherein the change estimate value is a channel state information CSI conditional entropy; constructs a new recurrent neural network structure through the CSI conditional entropy, wherein the new recurrent neural network structure is used for channel estimation;

[0101] An information amount calculation module 320 is used to model the CSI conditional entropy as a first-order Markov model to obtain the information amount required by the CSI matrix of the channel state at the current moment;

[0102] The offline training module 330 is used to perform offline training on the new neural network through the amount of information required by the CSI matrix of the current channel state, determine the weight parameters and bias parameters, and complete the design of the neural network for channel estimation.

[0103] Preferably, the CSI entropy calculation module 310 is used to embed the CSI conditional entropy into a recurrent neural network using a linear rectification function ReLU to construct a new recurrent neural network structure, and the new recurrent neural network structure is used for channel estimation.

[0104] Preferably, the CSI entropy calculation module 310 is further used to obtain the amount of information required for estimating the CSI entropy at the current moment according to the following formula (1) based on the historical estimated value:

[0105]

[0106] Among them, e t represents the amount of information required to estimate the CSI entropy at time t; represents CSI conditional entropy; H(·) represents entropy operation; represents the CSI matrix estimated at time t; W c represents the weight parameter, b c Represents the bias parameter.

[0107] Preferably, the information amount calculation module 320 is used to model the CSI conditional entropy as a first-order Markov model according to the following formula (2):

[0108] p(H t |H t-1 ,...,H1)=p(H t |H t-1 ) (2)

[0109] Among them, H t represents the CSI matrix at time t;

[0110] Based on formula (2), on the basis of historical estimated values, according to the following formula (3), the amount of information required by the CSI matrix at the current moment is obtained through the CSI conditional entropy:

[0111]

[0112] Preferably, the offline training module 330 is further used to input the received signal matrix Y into the new recurrent neural network and output the CSI matrix H according to the amount of information required by the CSI matrix at the current moment;

[0113] The new recurrent neural network is trained by the neural network training algorithm until convergence, and the final weight parameter W is determined. c and the bias parameter b c , complete the design of neural network for channel estimation.

[0114] Preferably, the received signal matrix Y is obtained according to the following formula (4):

[0115] Y=XH+N (4)

[0116] Wherein, X represents the transmitted signal, N represents the additive white Gaussian noise, and H represents the CSI matrix.

[0117] Preferably, the neural network training algorithm includes but is not limited to the gradient descent method, the Newton algorithm or the adaptive moment estimation Adam algorithm.

[0118] Preferably, the CSI matrix is ​​obtained according to an existing channel model, or using channel data of a measured real environment; the channel model includes but is not limited to a spatial channel SCM model, a Jakes model, a WINNER model, or a 3GPP standard channel model.

[0119] In a feasible implementation, the present invention proposes a neural network structure for channel estimation that does not rely on pilot signals, and uses information entropy theory to evaluate the strong time correlation characteristics of wireless channel data to improve the performance of the channel estimation algorithm based on the neural network. The neural network structure for channel estimation proposed in the present invention can improve the accuracy of channel estimation, reduce the delay of channel estimation, and meet the real-time requirements of wireless communication systems.

[0120] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A neural network design method for channel estimation, characterized in that: include: S1: Obtain the CSI change of the current channel state estimation value through the historical estimation value of the channel state, and model the CSI change as a first-order Markov model; S2: Obtain CSI conditional entropy by calculating the first-order Markov model, where the CSI conditional entropy represents the amount of information required to obtain the CSI matrix of the channel state information at the current moment; S3: construct a new recurrent neural network structure by using the obtained CSI conditional entropy, where the new recurrent neural network structure is used for channel estimation; In step S3, a new recurrent neural network structure is constructed by using the obtained CSI conditional entropy, including: The CSI conditional entropy is embedded into a recurrent neural network using a linear rectification function ReLU to construct a new recurrent neural network structure, and the new recurrent neural network structure is used for channel estimation; S4: Offline training is performed on the new recurrent neural network to determine weight parameters and bias parameters, thereby completing the design of a neural network for channel estimation.

2. The neural network design method for channel estimation according to claim 1, characterized in that: The step S1 further includes: based on the historical estimated value, obtaining the amount of information required for estimating the CSI entropy at the current moment according to the following formula (1): Among them, e t represents the amount of information required to estimate the CSI entropy at time t; represents CSI conditional entropy; H(·) represents entropy operation; represents the CSI matrix estimated at time t; W c represents the weight parameter, b c Represents the bias parameter.

3. The neural network design method for channel estimation according to claim 2, characterized in that: In step S1, the change of the CSI is modeled as a first-order Markov model, including: According to the following formula (2), the change of CSI is modeled as a first-order Markov model: p(H t |H t-1 ,...,H1)=p(H t |H t-1 ) (2) Among them, H t represents the CSI matrix at time t.

4. The neural network design method for channel estimation according to claim 2, characterized in that: In the step S2, the CSI conditional entropy is calculated by the first-order Markov model, and the CSI conditional entropy represents the amount of information required to obtain the CSI matrix of the channel state information at the current moment, including: based on formula (2), on the basis of the historical estimated value, according to the following formula (3), the amount of information required to obtain the CSI matrix at the current moment is obtained by the CSI conditional entropy, 。 5. The neural network design method for channel estimation according to claim 4, characterized in that: In step S4, the new recurrent neural network is trained offline to determine weight parameters and bias parameters, and the neural network design for channel estimation is completed, including: S41: according to the amount of information required by the CSI matrix at the current moment, input the received signal matrix Y into the new recurrent neural network and output the CSI matrix H; S42: Train the new recurrent neural network through a neural network training algorithm until convergence, and determine the final weight parameter W c and the bias parameter b c , complete the design of neural network for channel estimation.

6. The neural network design method for channel estimation according to claim 5, characterized in that: The received signal matrix Y is obtained according to the following formula (4): Y=XH+N (4) Wherein, X represents the transmitted signal, N represents the additive white Gaussian noise, and H represents the CSI matrix.

7. The neural network design method for channel estimation according to claim 6, characterized in that: The neural network training algorithm includes but is not limited to the gradient descent method, the Newton algorithm or the adaptive moment estimation Adam algorithm.

8. The neural network design method for channel estimation according to claim 7, characterized in that: The CSI matrix is ​​obtained according to an existing channel model, or using channel data of a measured real environment; the channel model includes but is not limited to a spatial channel SCM model, a Jakes model, a WINNER model, or a 3rd Generation Partnership Project 3GPP standard channel model.

9. A neural network design device for channel estimation, characterized in that: The device comprises: A CSI entropy calculation module, used to obtain a CSI change of a channel state estimation value at a current moment through a historical estimation value of the channel state, and model the CSI change as a first-order Markov model; An information amount calculation module, used to calculate and obtain CSI conditional entropy through the first-order Markov model, where the CSI conditional entropy represents the amount of information required to obtain the CSI matrix of the channel state information at the current moment; A network structure building module, used for constructing a new recurrent neural network structure by using the obtained CSI conditional entropy, wherein the new recurrent neural network structure is used for channel estimation; The network structure building module is used to embed the CSI conditional entropy into a recurrent neural network using a linear rectification function ReLU to construct a new recurrent neural network structure, and the new recurrent neural network structure is used for channel estimation; The offline training module is used to perform offline training on the new neural network through the amount of information required by the CSI matrix of the current channel state, determine the weight parameters and bias parameters, and complete the design of the neural network for channel estimation.

Images