Millimeter wave channel prediction method and system for mobile scene

Abstract

The invention relates to a wireless communication technology, in particular to a millimeter wave channel prediction method and system for a mobile scene, and the method comprises the steps: building a single-cell time division duplex MIMO channel model which comprises a base station and a user; the method comprises the following steps: constructing a millimeter wave channel prediction network oriented to a mobile scene on the basis of a transformer neural network, improving a position coding layer into a learnable position coding layer, and introducing a convolution module into a multi-head self-attention layer to improve the multi-head self-attention layer into a multi-head convolution self-attention layer; training the millimeter wave channel prediction network to enable the millimeter wave channel prediction network to adapt to a mobile scene; and establishing a channel prediction model comprising a channel estimator and a millimeter wave channel prediction network, the channel estimator obtaining known channel state information according to the pilot signal, and the millimeter wave channel prediction network predicting future channel state information according to the known channel state information. According to the invention, for a mobile scene under a millimeter wave channel, future channel state information is obtained through channel prediction, and the capacity of a communication system can be effectively improved.

Description

Technical Field

[0001] The present invention relates to wireless communication technology, and in particular to a millimeter wave channel prediction method and system for mobile scenarios. Background Art

[0002] Millimeter-wave MIMO technology is a key technology in modern wireless communications. By adding more antennas to both the transmitter and receiver, system capacity can be increased. In a MIMO system, additional antennas also help extend signal transmission distance and coverage, while reducing the impact of multipath fading and signal interference on the communication system.

[0003] However, when using millimeter-wave MIMO technology for communications, pilot signals are typically transmitted to obtain channel state information. However, millimeter-wave channels are complex and volatile, and this channel state information can easily become outdated. In millimeter-wave MIMO, the quality of the precoding used to transmit the transmitted signal is closely linked to the instantaneous channel state information. If the channel state information used for precoding becomes outdated, the precoding quality degrades, resulting in a loss of system reachability and data rate.

[0004] Most existing channel prediction methods are not suitable for millimeter wave channels, especially those in mobile scenarios. Therefore, how to improve the system capacity and communication quality of millimeter wave MIMO channels through channel prediction is an urgent problem that needs to be solved. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems existing in the prior art and provide a millimeter wave channel prediction method and system for mobile scenarios. A millimeter wave channel prediction network is constructed based on a neural network to predict future channel states in advance, which can effectively improve the precoding quality of the millimeter wave MIMO channel, achieve higher system capacity, and enhance communication quality.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a millimeter wave channel prediction method for mobile scenarios, comprising the following steps:

[0008] Establish a single-cell time-division duplex MIMO channel model, including one base station and one user;

[0009] Based on the transformer neural network, a millimeter wave channel prediction network for mobile scenarios was constructed. The position encoding layer was improved to a learnable position encoding layer, and a convolutional module was introduced into the multi-head self-attention layer to improve it into a multi-head convolutional self-attention layer.

[0010] Train the millimeter wave channel prediction network to adapt it to mobile scenarios;

[0011] A channel prediction model is established, including a connected channel estimator and a millimeter wave channel prediction network. The channel estimator is used to obtain known channel state information based on the pilot signal, and the millimeter wave channel prediction network predicts future channel state information based on the known channel state information.

[0012] In a second aspect, an embodiment of the present invention provides a millimeter wave channel prediction system for mobile scenarios, including the following modules:

[0013] MIMO channel model building module, used to build a single-cell time-division duplex MIMO channel model, including one base station and one user;

[0014] The millimeter wave channel prediction network construction module is used to build a millimeter wave channel prediction network for mobile scenarios based on the transformer neural network. It improves the position encoding layer into a learnable position encoding layer and introduces a convolution module to the multi-head self-attention layer to improve it into a multi-head convolutional self-attention layer.

[0015] A millimeter wave channel prediction network training module is used to train the millimeter wave channel prediction network to adapt it to mobile scenarios;

[0016] The channel prediction model establishment module is used to establish a channel prediction model, including a connected channel estimator and a millimeter wave channel prediction network. The channel estimator is used to obtain known channel state information based on the pilot signal, and the millimeter wave channel prediction network predicts future channel state information based on the known channel state information.

[0017] Compared with the prior art, the effective effects brought about by the technical solution provided by the present invention include at least:

[0018] The millimeter wave channel prediction method and system for mobile scenarios provided by the present invention can provide accurate channel state information in complex and changeable millimeter wave channels, thereby improving precoding quality and achieving higher system and rate; it can effectively alleviate the problem of channel state information of millimeter wave channels in mobile scenarios being easily outdated, achieve higher system capacity, and enhance wireless communication quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a flow chart of a millimeter wave channel prediction method for mobile scenarios provided by an embodiment of the present invention;

[0020] Figure 2 A schematic diagram of the structure of a millimeter wave channel prediction network according to an embodiment of the present invention;

[0021] Figure 3Schematic diagram of the structure of the multi-head convolutional self-attention layer in an embodiment of the present invention;

[0022] Figure 4 Schematic diagram of the structure of the channel prediction model established in an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram comparing the normalized mean square error of predicted future channel state information between the millimeter wave channel prediction network according to an embodiment of the present invention and the channel predictors based on the RNN model, the LSTM model, and the transformer model;

[0024] Figure 6 A schematic diagram comparing the normalized mean square error of predicted future channel state information between the millimeter wave channel prediction network of the present invention and a transformer-based model, a transformer model using a learnable position encoding layer, and a transformer model using a multi-head convolutional self-attention layer;

[0025] Figure 7 A schematic diagram comparing the cumulative distribution functions of the millimeter wave channel prediction network of the present invention and the channel predictors based on the RNN model, LSTM model, and transformer model;

[0026] Figure 8 This is a structural block diagram of a millimeter wave channel prediction system for mobile scenarios provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0028] Example 1

[0029] See also Figure 1 In this embodiment, a millimeter wave channel prediction method for mobile scenarios is provided, including the following steps:

[0030] S1. Establish a single-cell time division duplex MIMO channel model, where the MIMO channel model includes one base station and one user.

[0031] In the single-cell time-division duplex MIMO channel model, both the base station and the user use uniform linear array antenna configurations. In this embodiment, the base station is configured with N b antennas, the user configured Nu antennas, the channel state information of the i-th frame between the base station and the user can be expressed as:

[0032]

[0033] Where L is the number of propagation paths, α l is the attenuation factor of the lth path, θ b,l is the arrival angle of the base station on the lth path, θ u,l is the departure angle of the user on the lth path, f d is the Doppler frequency of the lth path, T s is the frame length of the lth path, a b (·) and a u (·) is the array steering vector, [·] H Represents the conjugate transpose of a vector.

[0034] The channel state information is considered to be constant within the coherence time of the channel, but changes between two different coherence times.

[0035] Array steering vector a b (·) and a u (·) are:

[0036]

[0037] in, is the wave number of the carrier, λ is the wavelength of the carrier, and d is the antenna spacing; [·] T Represents the transpose of a vector.

[0038] S2. Construct a millimeter wave channel prediction network for mobile scenarios. The millimeter wave channel prediction network is built on the basis of the transformer neural network. The position encoding layer is improved into a learnable position encoding layer, and the multi-head self-attention layer is improved into a multi-head convolutional self-attention layer by introducing a convolution module.

[0039] like Figure 2 As shown, the millimeter wave channel prediction network of this embodiment is improved based on the transformer network model, and specifically includes a first input layer, a second input layer, a first position coding layer, a second position coding layer, an encoder, a decoder, and an output layer. The first input layer is connected to the encoder via the first position coding layer, the second input layer is connected to the decoder via the second position coding layer, the output end of the encoder is connected to the decoder, and the output layer is provided at the output end of the decoder.

[0040] The encoder consists of a multi-head convolutional self-attention layer, a feedforward layer, a residual connection, and a normalization layer. The decoder also consists of a multi-head convolutional self-attention layer, a multi-head cross self-attention layer, a feedforward layer, a residual connection, and a normalization layer. Unlike the multi-head convolutional self-attention layer, the feedforward layer, and the multi-head cross self-attention layer, the residual connection and normalization layers have two inputs, while the multi-head convolutional self-attention layer has only one input. The encoder and decoder use residual connections and layer normalization operations through the residual connection and normalization layers.

[0041] The first input layer, the second input layer, and the output layer are all linear layers. The input layer maps the low-dimensional channel data to a high-dimensional space through linear transformation to obtain high-dimensional data; the output layer maps the high-dimensional data back to a low-dimensional space through linear transformation and outputs the future channel state information, which can be expressed as:

[0042] Linear(x)=xW+b

[0043] Among them, x is the input data of the linear layer, Linear(x) is the output data of the linear layer, W is the matrix used for linear transformation, and b is the added bias.

[0044] The first position coding layer and the second position coding layer are both learnable position coding layers. This embodiment improves the position coding layer so that the position coding can be continuously optimized during the channel predictor training process. Specifically, the learnable position coding process is expressed as:

[0045] LPE (pos,i) =fθ(·)

[0046] Among them, pos is the position of the channel data, i is the dimension index, θ is the neural network parameter of the millimeter wave channel prediction network, and f θ (·) is the change of position encoding during the training process of the mmWave channel prediction network.

[0047] The multi-head convolutional self-attention layer is improved on the basis of the multi-head self-attention layer, such as Figure 3 shown.

[0048] The self-attention mechanism is a correlation-finding process that requires three different linear transformations on the input data: query transformation, key transformation, and value transformation. The self-attention mechanism first finds the correlation between the query-transformed and key-transformed data, and then multiplies the obtained correlation with the value-transformed data to obtain the output of the self-attention, which can be expressed as:

[0049]

[0050] Among them, Q, K, and V are linear transformation matrices used for query transformation, key transformation, and value transformation, respectively. is a scaling factor used to avoid variance increase; softmax(·) is the activation function, and the expression can be

[0051] The multi-head self-attention layer allows the channel predictor to obtain information from multiple different places, which can be expressed as:

[0052] multihead(Q,K,V)=concat(head1,…,head h )W

[0053] Among them, head i represents the self-attention result of the i-th head, concat(·) represents the concatenation operation of all self-attention results, and W is a linear transformation of the result.

[0054] This embodiment introduces a convolution module into the multi-head self-attention layer, thereby improving it into a multi-head convolutional self-attention layer. Compared to the multi-head self-attention layer, before the data of the multi-head convolutional self-attention layer enters the query transformation and key transformation, it needs to be processed using a causal convolution kernel with a convolution kernel size of k and a stride of 1. Among them, the causal convolution kernel can prevent the millimeter wave channel prediction network from paying attention to future information when processing data, making the millimeter wave channel prediction network process data in a way that is consistent with actual usage scenarios.

[0055] The multi-head cross-self-attention layer is similar to the multi-head convolutional self-attention layer, being constructed by introducing a convolutional module into the multi-head self-attention layer. The difference lies in the input. The multi-head cross-self-attention layer takes two inputs: the encoder output and the output of the previous layer (i.e., the multi-head convolutional self-attention layer). The multi-head cross-self-attention layer performs key and value transformations on the encoder output and query transformation on the output of the previous layer before performing self-attention calculations.

[0056] The feedforward layer uses two linear transformations and one nonlinear transformation to perform deeper feature extraction on the channel data, which can be expressed as:

[0057] FFN(x)=max(0,xW1+b1)W2+b2

[0058] Where x is the input data of the feedforward layer, FFN(x) is the output of the feedforward layer, W1 and W2 are matrices used for linear transformation, b1 and b2 are added biases, and max(·) represents the maximum value operation.

[0059] The residual connection operation adds the output of one layer directly to the output of the next layer, which can effectively alleviate the gradient vanishing and gradient exploding problems in neural networks. The layer normalization operation normalizes the input data in each layer and maps the input data to a preset numerical range or distribution. This can accelerate the convergence of the neural network and improve the stability of the neural network during training. The residual connection operation and layer normalization operation can be expressed as:

[0060] y=Layernorm(x+F(x))

[0061] Among them, y represents the output after residual connection and layer normalization operations, x represents the input data of the neural network; F(·) represents the sub-layer operation process that requires residual connection, such as multi-head convolution self-attention operation or feedforward layer operation; Layernorm(·) represents the layer normalization operation.

[0062] S3. Train the millimeter wave channel prediction network so that it can adapt to mobile scenarios.

[0063] The data used to train the millimeter wave channel prediction network is millimeter wave MIMO channel data in mobile scenarios. The optimization function used in the training process is the normalized mean square error loss function, which can be expressed as:

[0064]

[0065] Where NMSE is the normalized mean square error, n is the total number of channel state information, h i is the true value of the channel state information, Channel state information predicted by the millimeter wave channel prediction network.

[0066] S4. Establish a channel prediction model, including a connected channel estimator and a millimeter wave channel prediction network, where the channel estimator is used to obtain known channel state information based on the pilot signal, and the millimeter wave channel prediction network predicts future channel state information based on the known channel state information.

[0067] During the channel prediction process, for a trained millimeter-wave channel prediction network, known channel state information obtained through pilot signals is input into the input end of the millimeter-wave channel prediction network, and future channel state information can be obtained at the output end of the millimeter-wave channel prediction network. Using this future channel state information can help wireless communication systems communicate.

[0068] like Figure 4As shown, in the channel prediction model of this embodiment, first, known channel state information is obtained based on the pilot signal, then p known channel state information is input into the trained millimeter wave channel prediction network, and finally q future channel state information (i.e., channel state information to be predicted) is output.

[0069] S41. Determine known channel state information through a channel estimator according to a pilot sequence sent by a user to a base station.

[0070] Specifically: the pilot sequence received in the i-th frame for:

[0071] y i =H i s i +w i

[0072] Among them, N p is the length of the pilot sequence, is the pilot sequence transmitted orthogonally between different antennas, is the known channel state information of the i-th frame, is zero-mean additive white Gaussian noise.

[0073] S42. Predict q future channel state information through a millimeter wave channel prediction network based on p known channel state information.

[0074] The expression for predicting q future channel state information can be expressed as:

[0075]

[0076] Where f(·) is the millimeter wave channel prediction network, is the channel state information to be predicted for the i-th frame, H i is the known channel state information of the i-th frame.

[0077] In this embodiment, in the millimeter wave channel prediction network, the input of the encoder is p known channel state information; the input of the decoder is q channel state information input taken from the end of the p known channel state information.

[0078] In this embodiment, mathematical modeling and simulation of the system are carried out. The simulation uses CDL-A channel data generated by Matlab for experiments. The base station uses a uniform linear array configured with 4 antennas, and the user uses a uniform linear array configured with 2 antennas. The carrier frequency is 30GHz, the user moving speed is 60km / h, and the causal convolution kernel size of the multi-head convolution self-attention layer is 3. The simulation uses 25 known channel state information to predict 5 unknown channel state information, such as Figure 5-Figure 7 shown.

[0079] from Figure 5 From the comparison, it can be seen that the normalized mean square error of the millimeter wave channel prediction network of the present invention at various prediction lengths is smaller than the normalized mean square error of the channel predictor based on the RNN model, LSTM model, and transformer model, which can better realize the prediction of future channel conditions.

[0080] from Figure 6 A comparison shows that compared to the transformer model channel predictor, the transformer model channel predictor that adopts a learnable position coding layer or a multi-head convolutional self-attention layer can improve the prediction accuracy of future channel state information. Moreover, the millimeter wave channel prediction network of the present invention can achieve even better channel prediction results because it adopts both a learnable position coding layer and a multi-head convolutional self-attention layer. Therefore, the channel prediction method of the present invention has better performance in predicting future channel state information.

[0081] exist Figure 7 In , the cumulative distribution function reveals the distribution of the difference between the predicted future channel state information and the true value. Figure 7 From the comparison, it can be seen that the amplitude error of the channel state information predicted by the millimeter wave channel prediction network of the present invention is about 80% within 0.02, while the amplitude error of the channel state information predicted by the channel predictor based on the RNN model and the LSTM model is only about 60% within 0.02; compared with the channel predictor based on the RNN model and the LSTM model, the prediction error of the channel prediction method of the present invention is smaller.

[0082] Example 2

[0083] Based on the same concept as the millimeter wave channel prediction method for mobile scenarios in the above-mentioned embodiment 1, this embodiment provides a millimeter wave channel prediction system for mobile scenarios, which can be used to execute each step of the above-mentioned millimeter wave channel prediction method for mobile scenarios.

[0084] For ease of explanation, the structural diagram of the millimeter wave channel prediction system embodiment for mobile scenarios only shows the parts related to the embodiment of the present invention. Those skilled in the art will understand that the illustrated structure does not constitute a limitation on the device, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0085] See also Figure 8 In this embodiment, a millimeter wave channel prediction system 100 for mobile scenarios is provided. The system includes the following modules:

[0086] MIMO channel model building module 101 is used to build a single-cell time-division duplex MIMO channel model including one base station and one user; wherein both the base station and the user adopt a uniform linear array antenna configuration;

[0087] Millimeter wave channel prediction network construction module 102 is used to build a millimeter wave channel prediction network for mobile scenarios based on the transformer neural network, improve the position encoding layer into a learnable position encoding layer, and introduce a convolution module into the multi-head self-attention layer to improve it into a multi-head convolutional self-attention layer;

[0088] The millimeter wave channel prediction network training module 103 is used to train the millimeter wave channel prediction network to adapt the millimeter wave channel prediction network to mobile scenarios; the training process uses channel data in mobile scenarios, and the loss function used in the training process is a normalized mean square error loss function;

[0089] The channel prediction model establishment module 104 is used to establish a channel prediction model, including a connected channel estimator and a millimeter wave channel prediction network. The channel estimator is used to obtain known channel state information based on the pilot signal, and the millimeter wave channel prediction network predicts future channel state information based on the known channel state information.

[0090] It should be noted that the millimeter wave channel prediction system for mobile scenarios of the present invention corresponds one-to-one to the millimeter wave channel prediction method for mobile scenarios of the present invention. The technical features and beneficial effects described in the above-mentioned embodiment of the millimeter wave channel prediction method for mobile scenarios are applicable to the embodiment of the millimeter wave channel prediction system for mobile scenarios. For specific contents, please refer to the description in the embodiment of the method of the present invention. No further details will be given here. This is hereby declared.

[0091] In addition, in the implementation of the millimeter wave channel prediction system for mobile scenarios in the above-mentioned embodiment, the logical division of each program module is only an example. In actual applications, the above-mentioned functions can be assigned to different program modules as needed, for example, for the configuration requirements of the corresponding hardware or the convenience of software implementation. That is, the internal structure of the millimeter wave channel prediction system for mobile scenarios is divided into different program modules to complete all or part of the functions described above.

[0092] In summary, the millimeter wave channel prediction method and system for mobile scenarios provided by the present invention can provide accurate channel state information under complex and changeable millimeter wave channels. The millimeter wave channel prediction method of the present invention can effectively alleviate the problem that the channel state information of the millimeter wave channel in mobile scenarios is easily outdated, achieve higher system capacity, and improve the quality of wireless communication.

[0093] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0094] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A millimeter wave channel prediction method for mobile scenarios, characterized in that: The steps include: Establish a single-cell time-division duplex MIMO channel model, including one base station and one user; Based on the transformer neural network, a millimeter wave channel prediction network for mobile scenarios is constructed. The position encoding layer is improved to a learnable position encoding layer, and the multi-head self-attention layer is improved to a multi-head convolutional self-attention layer by introducing a convolution module. Train the millimeter wave channel prediction network to adapt it to mobile scenarios; A channel prediction model is established, including a connected channel estimator and a millimeter wave channel prediction network, the channel estimator is used to obtain known channel state information according to a pilot signal, and the millimeter wave channel prediction network predicts future channel state information according to the known channel state information.

2. The millimeter wave channel prediction method according to claim 1, characterized in that: In the single-cell time-division duplex MIMO channel model, both the base station and the user use uniform linear array antenna configurations. The base station is configured with N b Antennas, user-configured N u root antenna; the channel state information of the i-th frame between the base station and the user is expressed as: Where L is the number of propagation paths, α l is the attenuation factor of the lth path, θ b,l is the arrival angle of the base station on the lth path, θ u,l is the departure angle of the user on the lth path, f d is the Doppler frequency of the lth path, T s is the frame length of the lth path, a b (·) and a u (·) is the array steering vector, [·] H Represents the conjugate transpose of a vector.

3. The millimeter wave channel prediction method according to claim 2, characterized in that: Array steering vector a b (·) and a u (·) are: in, is the wave number of the carrier, λ is the wavelength of the carrier, and d is the antenna spacing; [·] T Represents the transpose of a vector.

4. The millimeter wave channel prediction method according to claim 2, characterized in that: The process of building a channel prediction model includes: According to the pilot sequence sent by the user to the base station, the known channel state information is determined by the channel estimator; the pilot sequence received in the i-th frame for: y i =H i s i +w i Among them, N p is the length of the pilot sequence, is the orthogonal transmission pilot sequence between different antennas, is the known channel state information of the i-th frame, is zero-mean additive white Gaussian noise; According to p known channel state information, q future channel state information is predicted through the millimeter wave channel prediction network, which is expressed as: Where f(·) is the millimeter wave channel prediction network, is the future channel state information of the i-th frame, H i is the known channel state information of the i-th frame.

5. The millimeter wave channel prediction method according to claim 1 or 4, characterized in that: The millimeter wave channel prediction network includes a first input layer, a second input layer, a first position coding layer, a second position coding layer, an encoder, a decoder and an output layer; the first input layer is connected to the encoder via the first position coding layer, the second input layer is connected to the decoder via the second position coding layer, the output end of the encoder is connected to the decoder, and the output layer is arranged at the output end of the decoder; The first position encoding layer and the second position encoding layer are both learnable position encoding layers. The learnable position encoding process is expressed as: LPE (pos,i) =f θ (·) Among them, pos is the position of the channel data, i is the dimension index, θ is the neural network parameter of the millimeter wave channel prediction network, and f θ (·) Changes to position encoding during the mmWave channel prediction network training process; The encoder includes a multi-head convolutional self-attention layer, a feedforward layer, a residual connection, and a normalization layer; the decoder includes a multi-head convolutional self-attention layer, a multi-head cross self-attention layer, a feedforward layer, a residual connection, and a normalization layer; The multi-head convolutional self-attention layer is obtained by introducing a convolution module into the multi-head self-attention layer. Before the data of the multi-head convolutional self-attention layer enters the query transformation and key transformation, the causal convolution kernel is used to process the input data. The multi-head cross self-attention layer is obtained by introducing the convolution module into the multi-head self-attention layer; the input of the multi-head cross self-attention layer includes the output of the encoder and the output of the multi-head convolutional self-attention layer; the multi-head cross self-attention layer performs key transformation and value transformation on the output of the encoder, performs query transformation on the output of the multi-head convolutional self-attention layer, and then performs self-attention calculation; The first input layer, the second input layer and the output layer are all linear layers. The first input layer and the second input layer map the low-dimensional channel data to the high-dimensional space through linear transformation to obtain high-dimensional data; the output layer maps the high-dimensional data back to the low-dimensional space through linear transformation and outputs the future channel state information.

6. The millimeter wave channel prediction method according to claim 5, characterized in that: The future channel state information output by the output layer is expressed as: Linear(x)=xW+b Among them, x is the input data of the linear layer, Linear(x) is the output data of the linear layer, W is the matrix used for linear transformation, and b is the added bias.

7. The millimeter wave channel prediction method according to claim 5, characterized in that: The feedforward layer uses two linear transformations and one nonlinear transformation to perform deeper feature extraction on the channel data, which can be expressed as: FFN(x)=max(0,xW1+b1)W2+b2 Among them, x is the input data of the feedforward layer, FFN(x) is the output of the feedforward layer, W1 and W2 are matrices used for linear transformation, b1 and b2 are added biases, and max(·) represents the maximum value operation.

8. The millimeter wave channel prediction method according to claim 1, characterized in that: The data used to train the millimeter wave channel prediction network is the millimeter wave MIMO channel data in the mobile scenario. The optimization function used in the training process is the normalized mean square error loss function, which is expressed as: Where NMSE is the normalized mean square error, n is the total number of channel state information, and h is i is the true value of the channel state information, Channel state information predicted by the millimeter wave channel prediction network.

9. A millimeter wave channel prediction system for mobile scenarios, characterized in that: Includes the following modules: A MIMO channel model building module is used to build a single-cell time-division duplex MIMO channel model, including one base station and one user; The millimeter wave channel prediction network construction module is used to build a millimeter wave channel prediction network for mobile scenarios based on the transformer neural network, improve the position encoding layer into a learnable position encoding layer, and introduce a convolution module into the multi-head self-attention layer to improve it into a multi-head convolutional self-attention layer; A millimeter wave channel prediction network training module is used to train the millimeter wave channel prediction network to adapt the millimeter wave channel prediction network to mobile scenarios; The channel prediction model establishment module is used to establish a channel prediction model, including a connected channel estimator and a millimeter wave channel prediction network. The channel estimator is used to obtain known channel state information based on the pilot signal, and the millimeter wave channel prediction network predicts future channel state information based on the known channel state information.

10. The millimeter wave channel prediction system according to claim 9, characterized in that: The millimeter wave channel prediction network includes a first input layer, a second input layer, a first position coding layer, a second position coding layer, an encoder, a decoder and an output layer; the first input layer is connected to the encoder via the first position coding layer, the second input layer is connected to the decoder via the second position coding layer, the output end of the encoder is connected to the decoder, and the output layer is arranged at the output end of the decoder; The first position encoding layer and the second position encoding layer are both learnable position encoding layers. The learnable position encoding process is expressed as: LPE (pos,i) =f θ (·) Among them, pos is the position of the channel data, i is the dimension index, θ is the neural network parameter of the millimeter wave channel prediction network, and f θ (·) Changes to position encoding during the mmWave channel prediction network training process; The encoder includes a multi-head convolutional self-attention layer, a feedforward layer, a residual connection, and a normalization layer; the decoder includes a multi-head convolutional self-attention layer, a multi-head cross self-attention layer, a feedforward layer, a residual connection, and a normalization layer; The multi-head convolutional self-attention layer is obtained by introducing a convolution module into the multi-head self-attention layer. Before the data of the multi-head convolutional self-attention layer enters the query transformation and key transformation, the causal convolution kernel is used to process the input data. The multi-head cross self-attention layer is obtained by introducing the convolution module into the multi-head self-attention layer; the input of the multi-head cross self-attention layer includes the output of the encoder and the output of the multi-head convolutional self-attention layer; the multi-head cross self-attention layer performs key transformation and value transformation on the output of the encoder, performs query transformation on the output of the multi-head convolutional self-attention layer, and then performs self-attention calculation; The first input layer, the second input layer and the output layer are all linear layers. The input layer maps the low-dimensional channel data to the high-dimensional space through linear transformation to obtain high-dimensional data; the output layer maps the high-dimensional data back to the low-dimensional space through linear transformation and outputs the future channel state information.

Images