Lightweight neural network channel estimation method based on intelligent reflecting surface

Abstract

The application discloses a kind of light weight neural network channel estimation methods based on intelligent reflecting surface, comprising: setting simulation scene;Get the signal estimation data accepted by signal receiving end;Take the first M channel estimation data collected by each user and the cascade channel between base station b and user equipment k, and obtain channel estimation data set by merging;Set the loss function of teacher network, input DNN neural network to channel estimation data set, and iteratively train DNN neural network;Set the loss function of student network, input DNN neural network to channel estimation data set, and iteratively train DNN neural network;Estimate the channel using the student network that passes verification.The application does not need a large number of pilot output, while reducing the operation time, significantly improves the efficiency of channel estimation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication network optimization, and in particular to a light neural network channel estimation method based on intelligent reflecting surface. BACKGROUND

[0002] With the promotion of the global new round of industrial iteration and upgrading, the global technology development presents an acceleration and superimposed development trend. The information communication industry, which has the characteristics of technology-intensive and value-intensive, has gradually become a key factor that determines the technology evolution trend, influences the frontier industry ecology, and even the national core competitiveness and innovation ability. 5G communication networks are widely deployed globally, and new technologies for 6G are emerging.

[0003] Reconfigurable intelligent surface technology, as a potential new technology for 6G, is a hot research topic in 6G in recent years. Reconfigurable intelligent surface is composed of a large-scale device array and an array control module. Its principle is to control the on-off of the PIN diode on the surface of the array unit, change the unit reflection coefficient, and dynamically control the beam propagation characteristics of the reflected signal of the intelligent reflecting surface as a whole. Because it can replace traditional communication device radio frequency devices and significantly reduce the cost of setting up communication systems, it has been widely concerned by the industry and academia.

[0004] Currently, there is a problem of difficulty in obtaining channel information in the application of reconfigurable intelligent surface. For complex channels, current channel estimation methods such as least squares, minimum mean square error, matrix decomposition, and compressed sensing have great potential, but their performance is greatly dependent on high-quality antenna array output, and they have a great dependence on pilots, which will lead to an increase in computational complexity. In the actual scenario of wireless communication networks containing RIS, due to the immaturity of RIS phase control technology, the channel estimation process needs to be more lightweight to meet the actual needs of its carrying, and to reduce the time overhead of deep neural network channel estimation.

[0005] The document "Power of Deep Learning for Channel Estimation and Signal Detection in OFDM Systems" proposes a knowledge distillation method based on attention transfer mechanism, however, this research does not involve the attention transfer mechanism and knowledge distillation method of neural network channel estimation such as non-image classification problem. The document "PAYING MORE ATTENTION TO ATTENTION: IMPROVING THE PERFORMANCE OF CONVOLUTIONAL NEURAL NETWORKS VIA ATTENTION TRANSFER" proposes a convolutional neural network channel estimation method for wireless communication direct link channel estimation process, but this research does not involve its channel estimation method in the indirect communication link containing intelligent metasurface. SUMMARY

[0006] The purpose of the present application is to propose a lightweight neural network channel estimation method based on intelligent reflecting surface, which improves the wireless communication network channel estimation process containing RIS, optimizes the channel information expression process of the indirect link and the input mode of the neural network and the attention transfer mechanism. The present application does not require a large number of pilot outputs, and can significantly improve the efficiency of channel estimation while reducing the operation time, solving the problems of large dependence on pilots, high computational complexity of wireless communication MIMO network channel estimation, and large time overhead of neural network channel estimation technology, paving the way for the practical application of reconfigurable intelligent metasurface.

[0007] To achieve the above technical purpose, the technical scheme adopted by the present application is:

[0008] The present application discloses a lightweight neural network channel estimation method based on intelligent reflecting surface, which comprises the following steps:

[0009] S1, setting a simulation scene, the simulation scene includes a plurality of base stations, 1 RIS and a plurality of user equipment, each base station is equipped with antennas, the RIS has elements;

[0010] S2, using DeepMIMO data set, deducing the channel from the base station b to the RIS and the channel from the RIS to the user equipment , multiplying the collected channel samples left The signal data received by the user equipment is vectorized and added with Gaussian white noise, and finally multiplied by the pilot signal transmitted by the base station on the left to obtain the signal data received by the signal receiving end , The signal data received by the user equipment is vectorized and added with Gaussian white noise, and finally multiplied by the pilot signal transmitted by the base station on the left to obtain the signal data received by the signal receiving end

[0011] S3, taking the first M signal data collected by each user And the cascade channel between the base station b and the user equipment k , and merging to obtain the channel estimation data set

[0012] S4, setting the loss function of the teacher network, inputting part of the channel estimation data set into the DNN neural network, and iteratively training the DNN neural network; in each iteration, each user processes multiple local samples to calculate the gradient of the weight, and updates using the Adam optimizer; after the iteration is completed, the trained DNN neural network is output as the teacher network

[0013] S5, setting the loss function of the student network, inputting the remaining part of the channel estimation data set into the DNN neural network, and iteratively training the DNN neural network to train and verify the student network

[0014] S6, using the verified student network to estimate the channel.

[0015] In step S3, the signal estimation data The calculation formula is: ,

[0016] Wherein, The phase shift matrix of IRS is represented as , The diagonal element is represented by the exponential coordinate, The phase angle is represented by the exponential coordinate, The phase angle is represented by the exponential coordinate, The base station b forms a transmission beam forming vector for the user k The noise vector at the user k The pilot signal is The diagonal element is taken out The cascade channel between the base station b and the user equipment k

[0017] The data required for channel estimation is calculated by the foregoing formula , which is combined with to obtain the channel estimation data set , The size of the training data set of the user equipment served by the base station is represented by the first The size of the training data set of the user equipment served by the base station is represented by the first The size of the training data set of the user equipment served by the base station is represented by the first

[0018] In step S4, the first 200 data points out of the 2000 data points collected by each user are taken and divided into training and validation sets in a 9:1 ratio.

[0019] Step S5 further includes:

[0020] S51, set up a DNN neural network with 2 input channels, 32 output channels, 3 convolutional kernel dimensions, 1 padding factor, 3 convolutional layers, 64 samples per batch, and a learning rate of... The optimizer is Adam, and the iterator iteration count is 200.

[0021] S52, Define the loss function for the teacher network: ;

[0022] In the formula, For weight parameters, Indicates by the first The first base station service The size of the training dataset for each user device This is the estimated value of the cascaded channel between base station b and user equipment k;

[0023] S53, define the initial weight matrix, calculate the distance between the user and RIS, set it to zero if it exceeds 2m, and take the inverse ratio according to the distance within 2m, and normalize it, thereby setting the initial weights of the DNN network;

[0024] S54. According to the initial weights set in S53, the channel estimation dataset is fed into the training dataset for multiple users one by one, and the DNN neural network is trained to convergence.

[0025] S55, Using the above DNN neural network as the teacher network, extract its feature matrix, and sum the squares of the extracted feature matrix along the channel dimension to obtain its attention map. ,in This represents the attention map of the j-th layer.

[0026] Step S54 further includes:

[0027] Calculated and estimated Compared with actual data loss function ;

[0028] By optimizing the weight parameters To minimize the loss function of the teacher network: Apply gradient descent to the weight parameters Adjustments are made, and the formula is as follows: ,in, The weight parameters for the next iteration. The weight parameter generated for this iteration, The learning step, Indicates the calculation gradient.

[0029] Step S6 further includes:

[0030] The student matrix is set, the last 1800 of the 2000 data collected by each user of the data set is divided into a training set and a validation set according to 9:1, the number of convolution layers is set to 2 layers, the attention map is calculated, and is denoted as The loss function formula of the student network is defined as follows: , wherein, The mean square error of the student network, The Euclidean distance of the normalized attention map of the teacher network and the student network, The KL divergence between the teacher network and the student network.

[0031] Compared with the prior art, the beneficial effects of the present application are as follows:

[0032] First, the lightweight neural network channel estimation method based on the intelligent reflecting surface of the present application uses an intelligent channel estimation algorithm based on a DNN algorithm to realize the acquisition of CSI for the RIS indirect signal transmission link in RIS-assisted wireless communication.

[0033] Second, the lightweight neural network channel estimation method based on the intelligent reflecting surface of the present application uses an attention transfer-based knowledge distillation algorithm, and the lightweight neural network channel estimation model greatly reduces the actual application difficulty of the existing RIS system based on FPGA driving;

[0034] Third, the lightweight neural network channel estimation method based on the intelligent reflecting surface of the present application realizes the training of the scene data set on the reconfigurable intelligent super surface, expands the actual RIS communication network CSI acquisition, and provides a key basis for the overall optimization design of the communication system, and promotes the landing of RIS-related applications. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The wireless communication system model diagram of the present application;

[0036] Figure 2 The user distribution diagram of the system of the present application is 4 base stations, 1 RIS and 20 users;

[0037] Figure 3 The 20-user internal distribution diagram of the present application;

[0038] Figure 4 The algorithm iteration convergence diagram of the present application;

[0039] Figure 5 The channel estimation accuracy of the present application under a signal-to-noise ratio of 0-14 dB is compared with other typical methods, such as the MMSE and DML methods, in a graph;

[0040] Figure 6 A flowchart of the light-weight deep neural network channel estimation method based on the intelligent reflecting surface of the present application. DETAILED DESCRIPTION

[0041] The embodiments of the present application are further described in detail below with reference to the accompanying drawings.

[0042] Reference Figure 6 The present application discloses a light-weight deep neural network channel estimation method based on an intelligent reflecting surface, which comprises the following steps:

[0043] S1, setting a simulation scene, the simulation scene comprising a plurality of base stations, one RIS and a plurality of user equipment, each base station being equipped with antennas, the RIS having elements;

[0044] S2, using the DeepMIMO data set to derive the channel from the base station b to the RIS and the channel from the RIS to the user equipment Collecting channel samples Left-multiplying , then vectorizing and adding Gaussian white noise, and finally left-multiplying the pilot transmitted by the base station to obtain the signal data accepted by the signal receiving end , representing the signal data of the base station b received by the user equipment;

[0045] S3, taking the first M signal data collected by each user and the cascaded channel between the base station b and the user equipment k to obtain a channel estimation data set;

[0046] S4, setting the loss function of the teacher network, inputting part of the channel estimation data set into the DNN neural network, and iteratively training the DNN neural network; in each iteration, each user processes multiple local samples to calculate the gradient of the weight, and updates using the Adam optimizer; after the iteration is completed, the trained DNN neural network is output as the teacher network;

[0047] S5, setting the loss function of the student network, inputting the remaining part of the channel estimation data set into the DNN neural network, and iteratively training the DNN neural network to train and verify the student network;

[0048] S6, the student network with verification passes estimates the channel.

[0049] As Figure 1 shown, the corresponding system model includes: base station, RIS and user. Neglecting the direct link from the base station to the user, the present application focuses on the indirect link from the base station through the RIS to the user, and realizes the channel estimation for the indirect link.

[0050] In terms of specific implementation, the present application includes DeepMIMO dataset generation, pilot-aided cascaded channel operation, DNN model training and attention transfer mechanism knowledge distillation, which can be subdivided into the following steps:

[0051] Step 1, define the simulation scene, including 4 base stations (BS), 1 RIS and 20 user devices. Each base station is equipped with antennas, and the RIS has elements. Figure 2 Distribution diagram of 4 base stations, 1 RIS and 20 users of the system of the present application. Figure 3 Distribution diagram of 20 users of the present application.

[0052] Step 2, define the channel from base station b to RIS as , and the channel from RIS to user equipment k as . Using the channel samples generated by the DeepMIMO dataset, base station 5 is defined as RIS, and 2000 samples are collected for each user. The collected normalized and . All channel data used in this embodiment analysis come from the DeepMIMO dataset. Using the DeepMIMO dataset, the channel from base station b to RIS, and the channel from RIS to user equipment k are derived. The DeepMIMO dataset generates channel samples through ray tracing technology, taking into account key environmental variables including geometry and frequency. This embodiment focuses on the 60GHz outdoor ray tracing scene '01', in which base station 5 is designated as RIS. Each DeepMIMO-generated channel sample is normalized to ensure , , and 2000 samples are collected for each user.

[0053] Step 3, the data generated by the DeepMIMO dataset is not directly put into DNN training, but the collected channel samples are left multiplied by Then vectorize it and add Gaussian white noise, and finally left multiply the pilot signal transmitted by the base station to get the signal received by the signal receiving end . Define the channel estimation data set, the calculation formula is: ;

[0054] Where, is the phase shift matrix of IRS, which can be expressed as , represents the element diagonalization operation; represents the transmit beamforming vector formed by the base station b for the user k; is the noise vector at the user k; is the pilot signal; represents taking out the diagonal elements; represents the cascade channel between the base station b and the user equipment k. In the channel estimation process, the phase shift matrix of RIS and the beamforming vector of the base station and the pilot are all predetermined and known to both communication participants. Therefore, we can calculate the data required for channel estimation by formula (1) , combine it with into a data set , where, represents the size of the training data set of the kth user equipment served by the bth base station.

[0055] Step 4, take the first 200 of the 2000 data collected by each user, divide the training set and the validation set, and the ratio is 9:1 (training set 180, validation set 20).

[0056] Step 5, set the DNN neural network, the input channel number is 2, the output channel number is 32, the convolution kernel dimension is 3, the expansion coefficient (padding) is 1, the convolution layer number is 3, 64 samples are a Batch, the learning rate is , the optimizer is Adam, and the iterator number is 200 times.

[0057] Step 6, define the loss function, and use the mean square error function to calculate the error, the formula is as follows:

[0058] .

[0059] Step 7, define the initial weight matrix, calculate the distance between the user and the RIS, and set it to zero if it exceeds 2m, and take the inverse ratio according to the distance within 2m and normalize it.

[0060] Step 8, according to the initial weight, put the data set The model was trained on each of the 20 users individually until training convergence was achieved. The DNN neural network is fed with input data. In each iteration, each user processes 64 local samples (batches) to compute the gradient of the weights and updates them using the Adam optimizer. This process continues for 100 iterations; during this time, the most recently trained network is used for performance testing. The learning rate is initialized to... The value is halved every 30 iterations. The iterative process is specifically characterized by:

[0061] Calculated and estimated Compared with actual data loss function By appropriately optimizing the weight parameters To minimize the loss function mentioned earlier, that is: ;

[0062] Apply gradient descent to the weights Adjustments are made, and the formula is as follows: .

[0063] Figure 4 This is the iterative convergence graph of the algorithm of this invention.

[0064] Step 9: For the above DNN network, it is considered a teacher network. Its feature matrix is ​​extracted, and the sum of the squares of the feature matrix along the channel dimension is used as its attention map. ,in This represents the attention map of the j-th layer.

[0065] Step 10: Following steps 4-8 above, set up the student matrix. Take the last 1800 data points from each user's 2000 data points in the dataset, and divide them into training and validation sets in a 9:1 ratio. Set the number of convolutional layers to 2, and calculate their attention map, denoted as... The loss function formula for the student network is defined as follows:

[0066] ,

[0067] in, This represents the mean squared error of the student network. The Euclidean distance represents the normalized teacher and student online attention maps. This represents the KL divergence between the teacher network and the student network.

[0068] Step 11: Conduct training and verification on the above student network.

[0069] Step 12, final verification, use the model after 200 iterations to estimate the data under different signal-to-noise ratios of 0-14dB, and calculate the error with the perfect data.

[0070] In summary, the application realizes a light-weight RIS-assisted communication network channel estimation algorithm based on a deep neural network, successfully realizes the convergence of the loss function, and estimates the effective channel data according to the received signal, providing a good foundation for the design and optimization of RIS-assisted wireless communication networks. The algorithm is lightened by using knowledge distillation, which can cope with various complex channel environments and be carried in actual RIS scenarios. Referring to Figure 5 , the simulation and comparison with various algorithms prove the good characteristics of the application.

[0071] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming languages Java and interpreted scripting language JavaScript, etc.

[0072] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The devices that implement the functions specified in one or more flows and / or blocks.

[0073] These computer program instructions can also be stored in a computer readable storage medium that can guide the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The devices that implement the functions specified in one or more flows and / or blocks.

[0074] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks.

[0075] Although preferred embodiments of the application have been described herein, it will be apparent to those skilled in the art that various modifications and changes can be made to the embodiments without departing from the spirit and scope of the application. Accordingly, it is intended that all such modifications and changes be included within the scope of the application as set forth in the following claims and their equivalents.

[0076] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described herein.

Claims

1. A lightweight neural network channel estimation method based on a smart reflective surface, characterized in that, The lightweight neural network channel estimation method includes the following steps: S1, Set up a simulation scenario. The simulation scenario includes multiple base stations, one RIS, and multiple user equipment. Each base station is equipped with N t There are one antenna, and the RIS has N elements; S2, using the DeepMIMO dataset, analyzes the channel G from base station b to RIS. b And the channel from RIS to user equipment k The derivation is performed, and the collected channel samples G are... b Left multiplication Then, the signal is vectorized and Gaussian white noise is added. Finally, it is multiplied by the pilot signal transmitted by the base station to obtain the signal data y received by the signal receiver. b,k y b,k This represents the signal data received by the user equipment from base station b; S3, retrieve the first M signal data y collected from each user. b,k and the cascaded channel h between base station b and user equipment k b,k The training dataset is obtained by merging the datasets. S4. Set the loss function for the teacher network, input a portion of the training dataset into the DNN neural network, and iteratively train the DNN neural network. In each iteration, each user processes multiple local samples to calculate the gradient of the weights and updates them using the Adam optimizer. After the iteration is complete, output the trained DNN neural network as the teacher network. S5, set the loss function of the student network, input the remaining training dataset into the DNN neural network, iteratively train the DNN neural network, train and validate the student network; S6. The channel is estimated using a verified student network; In step S3, the signal data y b,k The calculation formula is: in, Let be the phase shift matrix of the IRS, expressed as: diag(·) represents the element-wise diagonalization operation. Let θ1, θ2...θ be the diagonal elements represented in exponential coordinates. N Phase angle; This indicates that base station b is the transmit beamforming vector formed by user k; n k x is the noise vector at user k; b,k It is a pilot signal; This indicates retrieving the diagonal elements; This represents the cascaded channel between base station b and user equipment k; The cascaded channel h is calculated using the aforementioned formula. b,k , and y b,k Merge into training dataset D b,k This represents the size of the training dataset for the k-th user equipment served by the b-th base station; Step S5 further includes: S51, set up a DNN neural network with 2 input channels, 32 output channels, 3 convolutional kernel dimensions, 1 padding factor, 3 convolutional layers, 64 samples per batch, and a learning rate of 1e. -3 The optimizer is Adam, and the iterator iteration count is 200. S52, Define the loss function for the teacher network: In the formula, ω is the weighting parameter, and D b,k This represents the size of the training dataset for the k-th user equipment served by the b-th base station. This is the estimated value of the cascaded channel between base station b and user equipment k; S53, define the initial weight matrix, calculate the distance between the user and RIS, set it to zero if it exceeds 2m, and take the inverse ratio according to the distance within 2m, and normalize it, thereby setting the initial weights of the DNN network; S54. According to the initial weights set in S53, the training dataset is fed into the training dataset and trained on multiple users one by one until the DNN neural network converges. S55, Using the above DNN neural network as the teacher network, extract its feature matrix, and sum the squares of the extracted feature matrix along the channel dimension to obtain its attention map. Where j represents the attention map of the j-th layer; Step S54 further includes: Calculated and estimated Compared with actual data h b,k The loss function L(ω); The loss function of the teacher network is minimized by optimizing the weight parameter ω. The weight parameter ω is adjusted using the gradient descent method, as shown in the formula: oh t+1 =ω t -or t g(ω t ) Where, ω t+1 ω is the weight parameter for the next iteration. t Let η be the weight parameter generated in this iteration. t To learn the step size, g(·) represents the calculation of the gradient; Step S6 further includes: Set up a student matrix, taking the last 1800 of the 2000 data points collected from each user in the dataset, and divide them into training and validation sets in a 9:1 ratio. Set the number of convolutional layers to 2, and calculate their attention maps, denoted as... The loss function formula for the student network is defined as follows: in, This represents the mean squared error of the student network. D represents the Euclidean distance between the attention maps of the normalized teacher network and student network. KL This represents the KL divergence between the teacher network and the student network.

2. The lightweight neural network channel estimation method based on intelligent reflective surfaces according to claim 1, characterized in that, In step S1, the simulation scenario includes 4 base stations, 1 RIS, and 20 user equipment; each base station is equipped with N t =16 antennas, RIS has N=64 elements.

3. The lightweight neural network channel estimation method based on intelligent reflective surfaces according to claim 1, characterized in that, In step S4, the first 200 data points out of the 2000 data points collected by each user are taken and divided into training and validation sets in a 9:1 ratio.

Images