A channel estimation method for MIMO communication system assisted by intelligent reflector

Abstract

The present invention provides a channel estimation method for a MIMO communication system assisted by an intelligent reflecting surface, applies deep learning technology to the channel estimation of the intelligent reflecting surface, and introduces a conditional generative adversarial network with a received signal as conditional information, thereby effectively improving the accuracy of channel estimation and improving robustness performance by reducing information loss during neural network training; adopts a U‑Net structure and a patching structure in the generator and the discriminator respectively to improve the ability of the generator to generate more realistic data and the ability to identify input detailed information; proposes an optimization loss function for updating the parameters of the generator to maintain the correct optimization direction during the training process, thereby further improving the channel estimation performance of the conditional generative adversarial network in the adversarial training process of the generator and the discriminator. The channel estimation method for a MIMO communication system assisted by an intelligent reflecting surface based on the conditional generative adversarial network of the present invention effectively improves the channel estimation performance of the MIMO communication system assisted by the intelligent reflecting surface.

Description

Technical Field

[0001] The invention relates to a channel estimation method, and is particularly applicable to a channel estimation method of an intelligent reflecting surface assisted MIMO communication system used in the technical field of intelligent reflecting surface assisted MIMO communication. Background Art

[0002] In recent years, millimeter wave communication has become a promising technology for 5G and even future communication systems because it can meet the high-speed requirements, while large-scale MIMO communication systems can reduce the high propagation loss of millimeter wave transmission. However, due to the dense placement of antennas in a limited physical space, millimeter wave large-scale MIMO systems still lead to high energy consumption and hardware complexity even with the use of hybrid beamforming technology. For this reason, smart reflective surfaces have recently been widely used in wireless communications because they can support high-speed communications with high energy efficiency and low cost for future wireless communications. Generally speaking, smart reflective surfaces contain a large number of low-cost passive reflective elements, and these elements can be configured according to the dynamic wireless environment to provide high beamforming gain. In wireless communication systems assisted by smart reflective surfaces, it is particularly important to obtain accurate channel state information to achieve reliable beamforming.

[0003] However, while the smart reflector provides low-cost high beamforming gain, it also brings severe challenges to channel estimation. There are two main difficulties in the communication system assisted by the smart reflector: 1. The large number of components of the smart reflector leads to high complexity of channel estimation; 2. Each component is passive and therefore has no ability to process signals. In order to solve the problem of channel estimation, many traditional channel estimation methods have been proposed to reduce complexity and reduce pilot overhead. However, these methods are either highly complex or cannot achieve satisfactory performance in complex channel models. In addition, although the existing deep learning-based channel estimation methods have improved the channel estimation performance and achieved good robustness, the channel estimation performance is still insufficient due to the fixed and single loss function. In order to further improve the channel estimation performance, the channel estimation method based on deep learning still needs to be studied in depth. Summary of the invention

[0004] Purpose of the invention: To address the deficiencies of the above-mentioned technologies, a channel estimation method for a reflection surface-assisted MIMO communication system is provided, which effectively improves the accuracy of channel estimation, uses received messages as conditional information, conducts adversarial network training, designs a new target loss function, and has good robustness.

[0005] Technical solution: To achieve the above technical objectives, the present invention provides a channel estimation method for a MIMO communication system assisted by an intelligent reflective surface, and the steps are as follows:

[0006] Step 1: Generate the equivalent baseband channels h from the base station to the kth user, from the base station to the smart reflection surface, and from the smart reflection surface to the kth user respectively. D，k , G and h R，k , and obtain the cascade channel H k samples, and at the same time, generate the corresponding received signal y k Samples of

[0007] Step 2: Construct the input-output pair {Y k , H k}, receiving signal matrix Cascade Channel M represents the number of antennas in the base station, and N represents the number of reflective elements;

[0008] Step 3: Based on the received signal as the conditional information of conditional GAN, design the target loss function for training the generator G and discriminator D in conditional GAN;

[0009] Step 4: The real and imaginary parts of each received signal matrix and cascade channel are fed into the conditional GAN. The real and imaginary parts of the received signal matrix of the kth user are represented as Re{Y k} and Im{Y k}, the real and imaginary parts of the concatenated channel of the kth user are represented as Re{H k} and Im{H k};

[0010] Step 5: Use the target loss function to update the parameters of the generator G and the discriminator D during the training process. When training one network, the parameters of the other network must be fixed.

[0011] Step 6: Use the cascade channel and the received signal matrix to train the conditional GAN ​​adversarially. After the conditional GAN ​​is trained, the new cascade channel and the received signal matrix are fed into the network, so that the trained conditional GAN ​​outputs the estimated cascade channel

[0012] Furthermore, the intelligent reflector-assisted MIMO communication system is as follows: the base station uses a baseband precoder Send K data symbols The transmitted downlink signal is expressed as in p krepresents the transmission power allocated to the kth user, so the signal received by the kth user is expressed as

[0013]

[0014] in, represents the diagonal reflection coefficient matrix, β n ∈[0, 1] and θ n ∈[0, 2π) represent the reflection amplitude and phase shift of the nth reflection element respectively;

[0015] A mutually orthogonal pilot signal is transmitted within each channel coherence time. and P≤M; the cascade channel matrix between the kth user and the base station can be expressed as where Θ k =diag{h R，k}, H k ψ=GΨh R，k , where Ψ = diag{ψ}, so the signal received by the kth user is expressed as

[0016]

[0017] in, Indicates receiving signal, represents the pilot signal matrix, To obey Gaussian distribution The noise vector of the kth user;

[0018] When P = M pilot signals are sent, only one element of the smart reflector is turned on at a time through the backhaul link, and the reflection beamforming vector of the nth frame can be expressed as ψ n =[0,…,0,ψ n , 0, …, 0] T , then the received signal of the kth user is expressed as

[0019]

[0020] in, and represents a 1×P row vector, h k，n is the cascade channel matrix H k The nth column, that is, h k，n =H k ψ n ;

[0021] Therefore, the estimated concatenated channel matrix for the kth user is Can be constructed as The received signal matrix Y of the kth user k It can be expressed as

[0022] Furthermore, the objective loss function for training the generator G is expressed as:

[0023]

[0024] The new objective loss function for training the discriminator D is expressed as:

[0025]

[0026] Among them, γ G is the parameter set of the generator G, γ D is the parameter set of the discriminator D, z represents the prior distribution p z The noise vector, p data represents the target distribution containing real channel samples, G(z) represents the generator G; the outputs of the generator G and the discriminator D become G(z, Y k ) and D(H k , Y k );η represents the weight of the loss function; during the training process, the generator G learns to generate more realistic channel samples G(z) to disrupt the correct judgment of the discriminator D, while the discriminator D aims to distinguish the real channel from the channel samples generated by the generator G; at the same time, the parameter sets of the generator G and the discriminator D are updated by the loss functions of formulas (4) and (5) respectively. When the parameters of one of the two neural networks are updated, the parameters of the other network are fixed.

[0027] Furthermore, the training steps of the conditional GAN ​​in step 5 include:

[0028] Step 5.1: Update the generator G;

[0029] Step 5.2: Sample the cascaded channel samples of the small batch;

[0030] Step 5.3: Obtain a small batch of received signal matrices as conditional information;

[0031] Step 5.4: Sample the noise vector of the mini-batch;

[0032] Step 5.5: Boost the stochastic gradient of the objective loss function to update the parameters of the generator G;

[0033] Step 5.6: Update the discriminator D;

[0034] Step 5.7: Sample the concatenated channel samples for the mini-batch.

[0035] Step 5.8: Obtain a small batch of received signal matrices as conditional information;

[0036] Step 5.9: Sample the noise vector of the mini-batch;

[0037] Step 5.10: Reduce the stochastic gradient of the objective loss function to update the parameters of the discriminator D.

[0038] Furthermore, the network structure of the conditional generative adversarial network conditional GAN ​​is as follows: conditionalGAN consists of a generator G and a discriminator D, where the {2nd, 3rd, 4th} layers of the generator G are all convolutional layers with 2×2 filters. These three layers are used to change the dimension of the input to match H k To keep it consistent, the {5, 6, 7, 8}th layers of generator G are 4 identical encoding blocks, each of which contains an LReLU activation function, a convolutional layer with a 3×3 filter, and a batch normalization layer; the {9, 10, 11}th layers of generator G are 3 identical decoding blocks, each of which contains a ReLU activation function, a batch normalization layer, and a convolutional layer; the first layer of generator G is the input layer, and the last layer is the hyperbolic tangent activation function, which can scale the output value of the generator to [-1, 1] to avoid mode collapse The {1, 2, 3} layers of the discriminator D are all 3 encoding blocks. The 4th layer of the discriminator D is a convolutional layer with a 3×3 filter and zero padding, followed by an LReLU activation function and a batch normalization layer. The last layer of the discriminator D is a convolutional layer with a Sigmoid activation function, which averages all responses of the receptive field and outputs the average value between [0, 1]. In addition, the U-Net structure and the patching structure are adopted in the generator and the discriminator respectively to improve the generator's ability to generate more realistic data and recognize detailed information of the input.

[0039] Beneficial effects:

[0040] 1) This method uses the conditional GAN ​​structure to solve the problem of designing loss functions in traditional deep learning methods. It can not only learn the mapping from received signals to channels, but also use adaptive loss functions to correctly train the network, effectively improving the channel estimation performance and providing an effective and feasible solution for the channel estimation problem of MIMO communication systems assisted by smart reflectors;

[0041] 2) The channel estimation scheme designed by this method not only has NMSE performance far exceeding that of traditional methods, but also shows superiority over existing deep learning channel estimation methods, and can still achieve satisfactory performance in low signal-to-noise ratio environments;

[0042] 3) This method achieves better NMSE performance under the conditions of pilot contamination and pilot shortage, showing superior robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A diagram of a MIMO communication system assisted by an intelligent reflector constructed in accordance with the present invention; Figure 2 A schematic diagram of the channel estimation method proposed by the present invention; Figure 3 The NMSE performance and signal-to-noise ratio diagram of the channel estimation method proposed in the present invention; Figure 4 Comparison of NMSE performance under pilot contamination for the channel estimation method proposed in the present invention; Figure 5 The NMSE performance comparison of the channel estimation method proposed in the present invention under insufficient pilot conditions is shown. DETAILED DESCRIPTION

[0044] The embodiments of the present invention are further described below in conjunction with the accompanying drawings:

[0045] A channel estimation method for a MIMO communication system assisted by an intelligent reflective surface, the specific steps are as follows:

[0046] Step 1: In order to reduce the complexity, deep learning technology is introduced into the channel estimation problem of the MIMO communication system assisted by the intelligent reflector to reduce the frequency of channel information update. At the same time, the conditional generative adversarial network (Conditional GAN) is introduced to address the problem of insufficient channel estimation accuracy of the intelligent reflector assisted communication system and the problem of loss function design in the traditional deep learning method.

[0047] Step 2: Design a new objective function for the auxiliary training generator, and concatenate the objective function with the objective function of the original generator G to form a new objective function, thereby ensuring the appropriate optimization direction of the conditional generative adversarial network.

[0048] Step 3: By using the received signal as a conditional input to generate a conditional adversarial network, the information loss during the neural network training process is reduced to effectively improve the channel estimation accuracy and robustness performance of the smart reflector-assisted communication. The two neural networks, the generator and the discriminator, are trained adversarially.

[0049] The intelligent reflector-assisted massive MIMO communication system described in step 1 is:

[0050] In a massive MIMO communication system assisted by a smart reflector, a transmitting base station with M antennas serves k users through an IRS composed of N reflective elements. The base station uses a baseband precoder Send K data symbols Then the transmitted downlink signal can be expressed as in p k represents the transmission power allocated to the kth user. Therefore, the signal received by the kth user can be expressed as

[0051]

[0052] The equivalent baseband channels from the base station to the kth user, from the base station to the smart reflection surface, and from the smart reflection surface to the kth user are represented by h D，k , G and h R，k express, represents the diagonal reflection coefficient matrix, β n ∈[0, 1] and θ n ∈[0, 2π) represent the reflection amplitude and phase shift of the nth reflection element respectively.

[0053] A mutually orthogonal pilot signal is transmitted within each channel coherence time. And P ≥ M. The cascade channel matrix between the kth user and the base station can be expressed as where Θ k =diag{h R，k}. Then, there is H k ψ=GΨh R，k , where Ψ = diag{ψ}. Therefore, the signal received by the kth user can be expressed as

[0054]

[0055] in, Indicates receiving signal, represents the pilot signal matrix, To obey Gaussian distribution The noise vector of the kth user.

[0056] When P = M pilot signals are sent, only one element of the smart reflector is turned on at a time through the backhaul link, and the reflection beamforming vector of the nth frame can be expressed as ψ n =[0,…,0,ψ n , 0, …, n] T Then, the received signal of the kth user can be expressed as

[0057]

[0058] in, and represents a 1×P row vector, h k，n is the cascade channel matrix H k The nth column, that is, h k，n =H k ψ n .

[0059] Therefore, the estimated concatenated channel matrix for the Kth user is Can be constructed as The received signal matrix Y of the kth user kIt can be expressed as

[0060] The new objective function for training the generator described in step 2 is as follows:

[0061] Step 2.1: According to the smart reflector-assisted MIMO communication system described in step 1, the Saleh-Valenzuela channel model is used to generate h R，k 、h D，k and G channels and obtain the cascade channel H k At the same time, the corresponding received signal y is generated k of the sample.

[0062] Step 2.2: Based on the channel samples obtained in step 2.1, construct the input-output pair {Y k , H k}, and the received signal matrix and cascade channels

[0063] Step 2.3, based on the received signal as the conditional information of conditional GAN, design the target loss function for training the generator and discriminator in conditional GAN, using the following formula:

[0064]

[0065]

[0066] Among them, γ G is the parameter set of the generator G, γ D is the parameter set of the discriminator D, z represents the prior distribution p z The noise vector, p data represents the target distribution containing the true channel samples, and G(z) represents the generator G.

[0067]

[0068]

[0069] Here, the outputs of the generator G and the discriminator D become G(z, Y k ) and D(H k , Y k). During the training process, the generator G learns to generate more realistic channel samples G(z) to disrupt the correct judgment of the discriminator D, while the discriminator D aims to distinguish the real channel from the channel samples generated by the generator G. At the same time, the parameter sets of the generator G and the discriminator D are updated by the loss functions of formulas (6) and (7), respectively. When the parameters of one of the two neural networks are updated, the parameters of the other network are fixed.

[0070] In order to ensure the appropriate optimization direction of conditional GAN, an additional loss function is added to the original target loss function of the generator G. Right now:

[0071]

[0072] Among them, η represents the weight of the loss function. Therefore, the new target loss function for training the generator G can be expressed as:

[0073]

[0074] Therefore, updating the parameters of the generator G is completed by the objective loss function of formula (9), thereby ensuring the appropriate optimization direction of the conditional GAN.

[0075] The step 3 comprises the following steps:

[0076] Step 3.1. Obtain the constructed received signal and cascade channel samples from step 2.1, and feed the real and imaginary parts of each received signal matrix and cascade channel into the conditional generative adversarial network conditional GAN.

[0077] The real and imaginary parts of the received signal matrix of the kth user are expressed as Re{Y k} and Im{Y k}, the real and imaginary parts of the concatenated channel of the kth user are represented as Re{H k} and Im{H k}.

[0078] Step 3.2: Use the objective loss function designed in step 2.3 to update the parameters of the generator G and the discriminator D during the training process. When training one network, the parameters of the other network must be fixed. The training algorithm of the conditional GAN ​​is shown in the following table:

[0079]

[0080] Step 3.3: Use the cascade channel and the received signal matrix to train the conditional GAN ​​adversarially. After the conditional GAN ​​is trained, the new cascade channel and the received signal matrix are fed into the network, so that the trained conditional GAN ​​outputs the estimated cascade channel. The network structure of the conditional GAN ​​is as follows:

[0081] Conditional GAN ​​consists of a generator G and a discriminator D. The {2nd, 3rd, 4th} layers of the generator G are all convolutional layers with 2×2 filters. These three layers are used to change the dimension of the input to keep consistent with Hk. The {5th, 6th, 7th, 8th} layers of the generator G are 4 identical encoding blocks, each of which contains an LReLU activation function, a convolutional layer with a 3×3 filter, and a batch normalization layer. Adding batch normalization to the convolutional layer can avoid gradient explosion and overfitting. The {9th, 10th, 11th} layers of the generator G are 3 identical decoding blocks, each of which contains a ReLU activation function, a batch normalization layer, and a convolutional layer. The first layer of the generator G is the input layer, and its last layer is the hyperbolic tangent (Tanh) activation function, which can scale the output value of the generator to [-1, 1] to avoid mode collapse. The {1, 2, 3} layers of the discriminator D are all 3 encoding blocks. The 4th layer of the discriminator D is a convolutional layer with a 3×3 filter and zero padding, followed by an LReLU activation function and a batch normalization layer. The last layer of the discriminator D is a convolutional layer with a Sigmoid activation function, which averages all responses of the receptive field and outputs the average value between [0, 1]. In addition, the U-Net structure and the patching structure are used in the generator and the discriminator respectively to improve the generator's ability to generate more realistic data and recognize detailed information of the input.

[0082] The normalized mean square error (NMSE) is used to measure the channel estimation performance. NMSE can be expressed as:

[0083]

[0084] Among them, H and represent the true concatenated channel and the estimated concatenated channel respectively.

[0085] In the experiment, in order to verify the performance of the proposed channel estimation method (Conditional GAN), we set M = P = N = 32, ∈ 0 = ∈ 1 = 0, K = 4. In the environment we modeled, we have L R =L G =LD = 10 paths, and the direction of each user is uniformly selected from the interval [-π, π]. In order to train and test the proposed conditional generative adversarial network, 40,000 channel samples of size 32×32 are generated and divided into training set, validation set and test set according to the ratio of 60%, 20% and 20% respectively. The RMSProp optimizer is used to train the two networks, the generator and the discriminator, and the learning rates used to train the generator and the discriminator are set to 0.0002 and 0.00002 respectively.

[0086] Depend on Figure 3 It can be seen that the NMSE performance of the conditional GAN ​​proposed in this paper is better than that of existing deep learning methods such as ChannelNet and MLP and the traditional channel estimation method LS. At the same time, as the SNR value increases, the gap between conditional GAN ​​and existing DL methods such as ChannelNet becomes larger and larger. Figure 4 The NMSE performance comparison of different methods under pilot contamination is shown. Figure 4 It can be seen that when the pilot signal is contaminated, the NMSE performance of the traditional LS method drops sharply, while the NMSE performance of the deep learning-based method is not greatly affected, showing strong robustness. At the same time, conditional GAN ​​still achieves better performance than existing DL methods. Figure 5 In , we also tested the NMSE performance of conditional GAN ​​under insufficient pilots. It can be seen from the figure that when the pilot is insufficient (P<M), the NMSE performance of conditional GAN ​​is not greatly affected, and conditional GAN ​​can still achieve satisfactory performance.

[0087] It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. All components not specified in this embodiment can be implemented using existing technologies.

Claims

1. A channel estimation method for a MIMO communication system assisted by an intelligent reflector, characterized in that: The specific steps are as follows: Step 1: Generate the equivalent baseband channels h from the base station to the kth user, from the base station to the smart reflection surface, and from the smart reflection surface to the kth user respectively. D，k , G and h R，k , and obtain the cascade channel H k samples, and at the same time, generate the corresponding received signal y k Samples of Step 2: Construct the input-output pair {Y k , H k }, receiving signal matrix Cascade Channel M represents the number of antennas in the base station, and N represents the number of reflective elements; Step 3: Based on the received signal as the conditional information of conditional GAN, design the target loss function for training the generator G and discriminator D in conditionalGAN; Step 4: The real and imaginary parts of each received signal matrix and cascade channel are fed into the conditional GAN. The real and imaginary parts of the received signal matrix of the kth user are represented as Re{Y k } and Im{Y k }, the real and imaginary parts of the concatenated channel of the kth user are represented as Re{H k } and Im{H k }; Step 5: Use the target loss function to update the parameters of the generator G and the discriminator D during the training process. When training one network, the parameters of the other network must be fixed. Step 6: Use the cascade channel and the received signal matrix to train the conditional GAN ​​adversarially. After the conditional GAN ​​is trained, the new cascade channel and the received signal matrix are fed into the conditional GAN, so that the trained conditional GAN ​​outputs the estimated cascade channel The intelligent reflector-assisted MIMO communication system comprises: a base station uses a baseband precoder Send K data symbols The transmitted downlink signal is expressed as in p k represents the transmission power allocated to the kth user, so the signal received by the kth user is expressed as in, represents the diagonal reflection coefficient matrix, β n ∈[0, 1] and θ n ∈[0, 2π) represent the reflection amplitude and phase shift of the nth reflection element respectively; A mutually orthogonal pilot signal is transmitted within each channel coherence time. and P≤M; the cascade channel matrix between the kth user and the base station can be expressed as where Θ k =diag{h R，k }, H k ψ=GΨh R，k , where Ψ = diag{ψ}, so the signal received by the kth user is expressed as in, Indicates receiving signal, represents the pilot signal matrix, To obey Gaussian distribution The noise vector of the kth user; When P = M pilot signals are sent, only one element of the smart reflector is turned on at a time through the backhaul link, and the reflection beamforming vector of the nth frame can be expressed as ψ n =[0,…,0,ψ n , 0, …, 0] T , then the received signal of the kth user is expressed as in, and represents a 1×P row vector, h k，n is the cascade channel matrix H k The nth column, that is, h k，n =H k ψ n ; Therefore, the estimated concatenated channel matrix for the kth user is Can be constructed as The received signal matrix Y of the kth user k It can be expressed as The objective loss function for training the generator G is expressed as: The new objective loss function for training the discriminator D is expressed as: Among them, G is the parameter set of the generator G, Υ D is the parameter set of the discriminator D, z represents the prior distribution p z The noise vector, p data represents the target distribution containing real channel samples, G(z) represents the generator G; the outputs of the generator G and the discriminator D become G(z, Y k ) and D(H k , Y k );η represents the weight of the loss function; during the training process, the generator G learns to generate more realistic channel samples G(z) to disrupt the correct judgment of the discriminator D, while the discriminator D aims to distinguish the real channel from the channel samples generated by the generator G; at the same time, the parameter sets of the generator G and the discriminator D are updated by the loss functions of formulas (4) and (5) respectively. When the parameters of one of the two neural networks are updated, the parameters of the other network are fixed.

2. The channel estimation method for a MIMO communication system assisted by an intelligent reflector according to claim 1, characterized in that: The training steps of the conditional GAN ​​in step 5 include: Step 5.1: Update the generator G; Step 5.2: Sample the cascaded channel samples of the small batch; Step 5.3: Obtain a small batch of received signal matrices as conditional information; Step 5.4: Sample the noise vector of the mini-batch; Step 5.5: Boost the stochastic gradient of the objective loss function to update the parameters of the generator G; Step 5.6: Update the discriminator D; Step 5.7: Sample the cascaded channel samples of the small batch; Step 5.8: Obtain a small batch of received signal matrices as conditional information; Step 5.9: Sample the noise vector of the mini-batch; Step 5.10: Reduce the stochastic gradient of the objective loss function to update the parameters of the discriminator D.

3. The channel estimation method for a MIMO communication system assisted by an intelligent reflective surface according to claim 1, characterized in that: The network structure of the conditional GAN ​​is as follows: Conditional GAN ​​consists of a generator G and a discriminator D, where the {2nd, 3rd, 4th} layers of the generator G are all convolutional layers with 2×2 filters. These three layers are used to change the dimension of the input to match H. k To keep it consistent, the {5, 6, 7, 8}th layers of generator G are 4 identical encoding blocks, each of which contains an LReLU activation function, a convolutional layer with a 3×3 filter, and a batch normalization layer; the {9, 10, 11}th layers of generator G are 3 identical decoding blocks, each of which contains a ReLU activation function, a batch normalization layer, and a convolutional layer; the first layer of generator G is the input layer, and the last layer is the hyperbolic tangent activation function, which can scale the output value of the generator to [-1, 1] to avoid mode collapse The {1, 2, 3} layers of the discriminator D are all 3 encoding blocks. The 4th layer of the discriminator D is a convolutional layer with a 3×3 filter and zero padding, followed by an LReLU activation function and a batch normalization layer. The last layer of the discriminator D is a convolutional layer with a Sigmoid activation function, which averages all responses of the receptive field and outputs the average value between [0, 1]. In addition, the U-Net structure and the patching structure are adopted in the generator and the discriminator respectively to improve the generator's ability to generate more realistic data and recognize detailed information of the input.

Images