Smart channel state information feedback method for intelligent metasurface wireless communication

Abstract

The application discloses a kind of intelligent channel state information feedback methods for intelligent metasurface wireless communication.First, construct the model RIS-CsiNet based on neural network, wherein the two encoders located at user end respectively compress and channel G between base station-intelligent metasurface and channel h between intelligent metasurface-user, two network modules are deployed at base station end, the first module input is the feedback information of channel h, output is the beamforming vector of base station and the phase shift of each unit in intelligent metasurface, in order to introduce the information of channel G into the design of base station end, its corresponding feedback information is input to super network, and the output is the network parameter of the last layer of the first network;Then, the model RIS-CsiNet is trained, and the user reachable rate is maximized;Finally, the trained RIS-CsiNet is used in communication system.The application can reduce the feedback overhead of intelligent metasurface assisted communication system, jointly carry out channel feedback and active / passive beamforming design, and reduce complexity.

Description

Technical Field

[0001] This invention relates to an intelligent channel state information feedback method for intelligent metasurface wireless communication, belonging to the field of communication technology. Background Technology

[0002] Reconfigurable Intelligent Surface (RIS)-assisted wireless communication is widely considered a core enabling technology for next-generation mobile communications. Unlike traditional communication systems that passively adapt to the channel environment, RIS modifies the channel by carefully adjusting programmable artificial electromagnetic units, showing great promise in scenarios such as coverage enhancement. Similar to traditional technologies, the prerequisite for achieving various gains is that the base station obtains downlink channel state information (CSI). Using the acquired CSI, the base station designs its beamforming vector and the phase shift matrix at the RIS end, thereby improving the user's received signal-to-noise ratio and user data rate. In Frequency-Division Duplex (FDD) systems, uplink and downlink transmissions operate on different frequency bands, and channel reciprocity no longer exists. Therefore, in RIS-assisted FDD systems, users need to feed back their estimated downlink CSI obtained through the channel. Since the base station is equipped with a large number of antennas, and the RIS end has hundreds of electromagnetic units, the CSI dimension in RIS-assisted wireless communication systems is extremely high. Directly feeding it back would consume a significant amount of uplink control channel resources. Existing RIS-assisted wireless communication CSI feedback techniques primarily utilize the sparsity of CSI to explore the impact of RIS introduction on the sparse structure of the channel, and then use compressed sensing technology to compress and reconstruct the CSI. However, due to the high complexity of the reconstruction algorithm in compressed sensing and its heavy reliance on sparse prior assumptions, it cannot fully utilize environmental knowledge within the channel, thus its accuracy and speed cannot meet the requirements of practical systems. Therefore, designing an efficient and high-precision CSI feedback technique is crucial for RIS wireless communication. Summary of the Invention

[0003] Technical Problem: The technical problem to be solved by this invention is to propose an efficient and high-precision channel state information feedback method in an intelligent metasurface communication system based on artificial intelligence, thereby solving the problems of high channel state information feedback overhead, low feedback accuracy, and high algorithm complexity in intelligent metasurfaces.

[0004] Technical Solution: This invention is an intelligent channel state information feedback method for intelligent metasurface wireless communication, the method comprising the following steps:

[0005] Step 1: Construct a feedback model RIS-CsiNet composed of neural networks. The user end includes two encoders, Encoder 1 and Encoder 2, and a quantization module that does not require training. The two encoders at the user end are used to generate low-dimensional representations of the two channels, namely channel G and channel h, i.e., feedback bitstreams SG and Sh, for the base station-smart metasurface and smart metasurface-user channels. The neural network at the base station uses the received feedback information, i.e., S... G and S h The beamforming vector v at the base station and the discrete phase shift [θ1,…,θ] of the smart metasurface are directly generated via neural networks. N ];

[0006] Step 2: Train the feedback model RIS-CsiNet, with the optimization objective being to maximize the user reachability rate, and obtain the model parameters;

[0007] Step 3: The trained feedback model RIS-CsiNet is used in the intelligent metasurface-assisted wireless communication scenario, where two encoders are deployed at the user end and other neural network modules are deployed at the base station end to achieve efficient channel state information feedback.

[0008] The encoder 1 is used to compress the channel h between the smart metasurface and the user. It consists of three fully connected layers. The parameters of each layer are randomly initialized. The real and imaginary parts of the complex form of the channel h are separated and concatenated as the input of the encoder 1. The number of neurons in the last fully connected layer is the same as the number of feedback bits. The activation function is sigmoid. The output of this layer is normalized to (0, 1). After 1-bit quantization, the output generates the feedback bitstream S. h The data is transmitted to the base station via the uplink control channel.

[0009] The encoder 2 is used to compress the channel G between the base station and the smart metasurface. It consists of two fully connected layers. The parameters of each layer are randomly initialized, and the complex form of the channel G is stretched into a one-dimensional vector. The real and imaginary parts of this vector are separated and concatenated as the input of the encoder 2. The number of neurons in the last fully connected layer is 1 / 4 of the number of feedback bits. Its output is quantized by 4 bits to generate the feedback bit stream S. G It is transmitted to the base station via the uplink control channel.

[0010] In the aforementioned base station-smart metasurface, since the environment between the base station and the smart metasurface is relatively fixed, channel G is considered to remain unchanged over a long period of time. However, due to user movement, the environment between the smart metasurface and the user changes rapidly, and channel h changes rapidly. Therefore, the feedback intervals of channels G and h are different, with G having a larger feedback interval and h having a smaller feedback interval.

[0011] The base station includes an inverse quantization module and two neural network branches: neural network branch 1 and neural network branch 2.

[0012] The inverse quantization module first restores the bit stream transmitted through the uplink control channel into quantized codewords;

[0013] Neural network branch 1 consists of a feature extraction module and two parallel fully connected layers; the feature extraction module consists of four fully connected layers, and its input is the feedback bitstream corresponding to channel h, i.e., S. h The output is a feature vector for phase shift design and beamforming design. The dimensions of the two feature vectors are the same as the dimensions of the phase shift vector and the complex beamforming vector. The parameters of each fully connected layer are randomly initialized. Then, the two extracted feature vectors are input into the fully connected layer to generate phase shift and beamforming vectors. The output of the fully connected layer is processed by a uniform quantizer and a power normalization operation. Here, the parameters of the two fully connected layers are determined by the second branch of the neural network.

[0014] Neural network branch 2, designed based on the concept of a supernetwork, consists of two parallel neural network parameter generation modules; both generation modules are composed of fully connected layers, and their input is the feedback bitstream S corresponding to channel G. G The output is the parameters of the last two fully connected layers in branch 1. The last fully connected layer does not use an activation function. Before training, the parameters of each fully connected layer are randomly initialized.

[0015] The parameters of the model in step two mainly include the weights and biases of the fully connected layers.

[0016] Step two involves training the feedback model RIS-CsiNet using an end-to-end learning approach. This involves jointly training the parameters of the two encoders at the user end and the two neural network branches at the base station to minimize the cost function. Since the gradient of the quantization operation is not differentiable, it is set to a constant 1 during training. The cost function is described as follows:

[0017]

[0018] Where ||·||² is the Euclidean norm, λ is a constant that adjusts the magnitude of the cost function, and Φ = diag(φ₁,...,φ₂). N ) represents the reflection matrix corresponding to the intelligent metasurface, where θ n Let v be the phase shift corresponding to the nth unit of the smart metasurface, and v be the beamforming vector at the base station.

[0019] Beneficial effects: By adopting the above technical solution, the present invention can produce the following technical effects:

[0020] This invention considers the cascaded channel characteristics in smart metasurface wireless communication, providing feedback at both ends of the channel: between the base station and the smart metasurface, and between the smart metasurface and the user. Given the different variation intervals of the two channels, the first channel provides feedback at a lower frequency, while the latter provides feedback at a higher frequency. Through an autoencoder at the user end, the channel dimensionality is significantly reduced. Furthermore, at the base station, the feedback information is directly utilized via a neural network to design the beamforming vector and the phase shift matrix of the smart metasurface, achieving a joint design of feedback and active / passive beamforming, reducing complexity and improving system performance. Specifically, by introducing slowly varying channel information between the base station and the smart metasurface through a hypernetwork architecture, frequent input to this channel is avoided, further reducing the complexity of the neural network.

[0021] In summary, this invention achieves high-precision, low-overhead, and low-complexity channel state information feedback in intelligent metasurface wireless communication, which has significant practical value. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the intelligent metasurface wireless communication system considered in this invention;

[0023] Figure 2 This is a schematic diagram of the intervals between channel G, channel h feedback, and data transmission in this invention;

[0024] Figure 3 This is a schematic diagram of the RIS-CsiNet model proposed in this invention. Detailed Implementation

[0025] The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0026] This invention designs an intelligent channel state information feedback method for smart metasurface wireless communication. Based on the RIS-CsiNet neural network architecture, this method uses two encoders at the user end to compress and encode two channel segments: one between the base station and the smart metasurface, and the other between the smart metasurface and the user. These segments are transmitted via the uplink control channel at different feedback intervals. At the base station, the received feedback information is directly used by the neural network to design the base station beamforming vector and the smart metasurface phase shift. This design achieves efficient, high-precision, and low-complexity channel state information feedback. The method specifically includes the following steps:

[0027] 1: For example Figure 1 As shown, a smart metasurface-assisted large-scale multiple-input multiple-output narrowband wireless communication system is considered, with the base station equipped with N... t=64 transmitting antennas, and the user end is equipped with a single receiving antenna. Due to obstructions from buildings or other objects, there is no signal transmission path between the base station and the user. In this case, a signal transmission path is established through a smart metasurface equipped with N = 16 × 16 reflective elements. In this scenario, the channel consists of two parts: channel G between the base station and the smart metasurface, and channel h between the smart metasurface and the user. The dimensions of channel G and channel h are 64 × 256 and 256 × 1, respectively. Direct feedback would incur significant feedback overhead.

[0028] 2: Since base stations and smart metasurfaces are mostly deployed at high altitudes, the propagation environment between them is relatively stable. Therefore, if... Figure 2 As shown, channel G does not require frequent feedback, and its feedback interval is relatively large. However, because the user is in motion, their surrounding environment is constantly changing, resulting in significant variations in channel h, such as... Figure 2 As shown, channel h requires frequent feedback. Therefore, the feedback periods of the two channels are different.

[0029] 3: Figure 2 This demonstrates the overall architecture of the RIS-CsiNet model designed in this embodiment. The user end includes two encoders and a quantization module that does not require training.

[0030] Encoder 1 is used to compress the channel h between the smart metasurface and the user. It consists of three fully connected layers. The parameters of each layer are randomly initialized. The real and imaginary parts of the complex channel vector h are separated and concatenated as the input to encoder 1. The first two fully connected layers have 2N and N neurons respectively, and the activation function is ReLU. The last fully connected layer has the same number of neurons as the number of feedback bits, and the activation function is Sigmoid. The output of this layer is normalized to (0, 1), and its output is quantized by 1 bit to generate the feedback bitstream S. h It is transmitted to the base station via the uplink control channel;

[0031] Encoder 2 is used to compress the channel G between the base station and the smart metasurface. It consists of two fully connected layers. The parameters of each layer are randomly initialized. The complex channel G is stretched into a one-dimensional vector. The real and imaginary parts of the vector are separated and concatenated as the input of encoder 2. The number of neurons in both fully connected layers is 1 / 4 of the number of feedback bits. The activation functions are ReLU and Sigmoid, respectively. The output is quantized by 4 bits to generate a feedback bit stream, which is transmitted back to the base station through the uplink control channel.

[0032] 4. The base station includes an inverse quantization module and two neural network modules (branch). The inverse quantization module first recovers the quantized codewords from the bit stream transmitted through the uplink control channel;

[0033] Neural network branch 1 consists of a feature extraction module and two parallel fully connected layers. The feature extraction module comprises four fully connected layers, with the input being the feedback bitstream corresponding to channel h, i.e., S. h The output is a feature vector for phase shift and beamforming design. The dimensions of the two feature vectors are the same as those of the phase shift and complex beamforming vectors. The parameters of each fully connected layer are randomly initialized. The number of neurons in the first three fully connected layers is 4N. t +N, 4N t +2N, 2N t +N, with ReLU as the activation function; then, the two extracted feature vectors are input into the fully connected layer to generate phase shift and beamforming vectors, and the output of the fully connected layer is processed by a uniform quantizer and power normalization operation, where the two fully connected parameters are determined by branch 2.

[0034] Neural network branch 2, designed based on the concept of a supernetwork, consists of two parallel neural network parameter generation modules. Both generation modules are composed of fully connected layers, and their input is the feedback codeword S corresponding to the channel G. G The output is the parameters of the last two fully connected layers in branch 1. The last fully connected layer does not use an activation function. Before training, the parameters of each fully connected layer are randomly initialized.

[0035] 5. The neural network is trained using channels obtained through collection or simulation. Specifically, the Adam optimization algorithm and an end-to-end learning approach are employed to jointly train the parameters of two encoders at the user end and two neural network branches at the base station, minimizing the cost function. Since the gradient of the quantization operation is not differentiable, it is set to a constant of 1 during training. The cost function is described as follows:

[0036]

[0037] Where ||·||² is the Euclidean norm, λ is a constant (adjusting the order of magnitude of the cost function), and Φ = diag(φ₁, ..., φ₂). N ) represents the reflection matrix corresponding to the intelligent metasurface, where θ n Let v be the phase shift corresponding to the nth unit of the smart metasurface, and v be the beamforming vector at the base station.

[0038] 6: The trained RIS-CsiNet model is used in the channel feedback of the intelligent metasurface-assisted FDD system.

[0039] In summary, this invention considers the cascaded channel characteristics in smart metasurface wireless communication, providing feedback at both ends of the channel between the base station and the smart metasurface, and between the smart metasurface and the user. Through an autoencoder at the user end, the channel dimensionality is significantly reduced. Furthermore, at the base station, the feedback information is directly utilized via a neural network to design the beamforming vector and the phase shift matrix of the smart metasurface, achieving a joint design of feedback and active / passive beamforming, reducing complexity and improving system performance. Specifically, by introducing slowly varying channel information between the base station and the smart metasurface through a hypernetwork architecture, frequent input to this channel is avoided, further reducing the complexity of the neural network.

[0040] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for intelligent channel state information feedback in smart metasurface wireless communication, characterized in that, The method includes the following steps: Step 1: Construct a feedback model RIS-CsiNet composed of neural networks. The user end includes two encoders, Encoder 1 and Encoder 2, and a quantization module that does not require training. The two encoders at the user end are used to generate low-dimensional representations of the two channels, namely channel G and channel h, namely base station-smart metasurface and smart metasurface-user, which are the feedback bit streams S. G and S h The neural network at the base station utilizes the received feedback information, i.e., S G and S h The beamforming vector v and discrete phase shift of the smart metasurface are directly generated at the base station via neural networks. Step 2: Train the feedback model RIS-CsiNet, with the optimization objective being to maximize the user reachability rate, and obtain the model parameters; Step 3: The trained feedback model RIS-CsiNet is used in the intelligent metasurface-assisted wireless communication scenario, where two encoders are deployed at the user end and other neural network modules are deployed at the base station end to achieve efficient channel state information feedback. The encoder 1 is used to compress the channel h between the smart metasurface and the user. It consists of three fully connected layers. The parameters of each layer are randomly initialized. The real and imaginary parts of the complex form of the channel h are separated and concatenated as the input of the encoder 1. The number of neurons in the last fully connected layer is the same as the number of feedback bits. The activation function is sigmoid. The output of this layer is normalized to (0, 1). After 1-bit quantization, the output generates the feedback bitstream S. h The data is transmitted to the base station via the uplink control channel. The encoder 2 is used to compress the channel G between the base station and the smart metasurface. It consists of two fully connected layers. The parameters of each layer are randomly initialized, and the complex form of the channel G is stretched into a one-dimensional vector. The real and imaginary parts of this vector are separated and concatenated as the input of the encoder 2. The number of neurons in the last fully connected layer is 1 / 4 of the number of feedback bits. Its output is quantized by 4 bits to generate the feedback bit stream S. G It is transmitted to the base station via the uplink control channel.

2. The intelligent channel state information feedback method for intelligent metasurface wireless communication according to claim 1, characterized in that, In the aforementioned base station-smart metasurface, since the environment between the base station and the smart metasurface is relatively fixed, channel G is considered to remain unchanged over a long period of time. However, due to user movement, the environment between the smart metasurface and the user changes rapidly, and channel h changes rapidly. Therefore, the feedback intervals of channels G and h are different, with G having a larger feedback interval and h having a smaller feedback interval.

3. The intelligent channel state information feedback method for intelligent metasurface wireless communication according to claim 1, characterized in that, The base station includes an inverse quantization module and two neural network paths: neural network branch 1 and neural network path 2. The inverse quantization module first restores the bit stream transmitted through the uplink control channel into quantized codewords; Neural network branch 1 consists of a feature extraction module and two parallel fully connected layers; the feature extraction module consists of four fully connected layers, and its input is the feedback bitstream corresponding to channel h, i.e., S. h The output is a feature vector for phase shift design and beamforming design. The dimensions of the two feature vectors are the same as the dimensions of the phase shift vector and the complex beamforming vector. The parameters of each fully connected layer are randomly initialized. Then, the two extracted feature vectors are input into the fully connected layer to generate phase shift and beamforming vectors. The output of the fully connected layer is processed by a uniform quantizer and a power normalization operation. Here, the parameters of the two fully connected layers are determined by the second branch of the neural network. Neural network branch 2, designed based on the concept of a supernetwork, consists of two parallel neural network parameter generation modules; both generation modules are composed of fully connected layers, and their input is the feedback bitstream S corresponding to channel G. G The output is the parameters of the last two fully connected layers in branch 1. The last fully connected layer does not use an activation function. Before training, the parameters of each fully connected layer are randomly initialized.

4. The intelligent channel state information feedback method for intelligent metasurface wireless communication according to claim 1, characterized in that: The parameters of the model in step two mainly include the weights and biases of the fully connected layers.

5. The intelligent channel state information feedback method for intelligent metasurface wireless communication according to claim 1 or 3, characterized in that: Step two involves training the feedback model RIS-CsiNet using an end-to-end learning approach. This involves jointly training the parameters of the two encoders at the user end and the two neural network branches at the base station to minimize the cost function. Since the gradient of the quantization operation is not differentiable, it is set to a constant 1 during training. The cost function is described as follows: in, Let be the Euclidean norm, and λ be a constant that adjusts the order of magnitude of the cost function. Let be the reflection matrix corresponding to the intelligent metasurface, where . θ n Let v be the phase shift corresponding to the nth unit of the smart metasurface, and v be the beamforming vector at the base station.

Images