Uplink ultra-large-scale intelligent reflection surface beam training method based on deep learning

Abstract

The present invention discloses an uplink ultra-large-scale intelligent reflection surface beam training method based on deep learning. The base station constructs and trains a Resnet estimation model based on a deep residual network. Based on the trained Resnet estimation model, the base station can accurately estimate the optimal codeword in the near-field codebook based on only the received signal of a part of the codeword. Therefore, in the stage when the user sends a pilot, the ultra-large-scale intelligent reflection surface only needs to test part of the codeword, thereby greatly reducing the pilot overhead. Finally, the additional codeword test based on the estimated probability can greatly improve the estimation accuracy of the optimal codeword. The present invention estimates the optimal codeword based on testing only part of the codeword by utilizing a deep residual network, thereby reducing the pilot overhead and improving the estimation accuracy.

Description

Technical Field

[0001] The present invention relates to a low pilot overhead beam training scheme design for ultra-large-scale intelligent reflection surface assisted communication based on deep learning. Background Art

[0002] Smart reflective surface technology is considered to be one of the potential key technologies for the next generation of mobile communications. This technology provides an additional reliable reflection link by deploying a large number of reconfigurable smart reflective elements between users and base stations, greatly improving the spectrum efficiency and throughput of wireless communication systems. In addition, due to the advantages of low cost and low power consumption of smart reflective surfaces, smart reflective surfaces are expected to deploy more reflective elements to compensate for the severe path loss in the reflection link. Therefore, smart reflective surfaces are gradually developing towards ultra-large scale. In wireless communication systems assisted by small-scale smart reflective surfaces, beam training based on predefined codebooks is widely used to optimize the reflection phase coefficient of smart reflective surfaces. Specifically, smart reflective surfaces determine the optimal codeword by testing all codewords in the codebook and construct the phase coefficient with the optimal codeword. However, the introduction of ultra-large-scale smart reflective surfaces will lead to an expansion of the near-field range, which makes it more likely that users or targets are within the range of the near-field. Therefore, in wireless communication systems assisted by ultra-large-scale smart reflective surfaces, the design of predefined codebooks and beam training need to consider the spherical wave channel model of the near field. However, considering the spherical wave channel model of the near field, the number of codewords in the near-field codebook increases significantly compared to the far-field codebook, and the complexity of beam training also increases accordingly. Specifically, the ultra-large-scale smart reflection surface requires more pilot signals to test the codewords, which will lead to a sharp increase in pilot overhead and a sharp decrease in the effective communication time within the coherence time of each channel. On the one hand, the existing beam training schemes are difficult to apply to the near-field wireless propagation environment, and on the other hand, they cannot effectively reduce the pilot overhead of beam training in the near field. Therefore, the design of a beam training scheme based on a predefined codebook for ultra-large-scale source smart reflection surfaces has great research significance, but also has technical challenges. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a low pilot overhead beam training scheme design for ultra-large-scale intelligent reflection surface assisted communication based on deep learning, so as to effectively reduce the pilot overhead in near-field beam training. The base station constructs and trains a Resnet estimation model based on a deep residual network. Based on the trained Resnet estimation model, the base station can accurately estimate the optimal codeword in the near-field codebook based only on the received signal of some codewords. Therefore, in the stage when the user sends the pilot, the ultra-large-scale intelligent reflection surface only needs to test some codewords, thereby greatly reducing the pilot overhead. Finally, the additional codeword test based on the estimated probability can greatly improve the estimation accuracy of the optimal codeword.

[0004] In order to achieve the above object, the present invention adopts the following technical solution:

[0005] A low pilot overhead beam training method for ultra-large-scale intelligent reflective surface assisted communication based on deep learning, comprising the following steps:

[0006] Step S1: The base station constructs and trains a Resnet estimation model based on a deep residual network to obtain model parameters.

[0007] Step S2: The base station is configured with M antennas, the ultra-large-scale intelligent reflection surface is configured with N reflection units with reconfigurable reflection phases, and the user is configured with a single antenna. The user sends a pilot signal to the ultra-large intelligent reflection surface, and the ultra-large intelligent reflection surface selects part of the codewords in the predefined codebook to form a reflection phase coefficient and reflects the pilot signal to the base station. The base station obtains the received signal of the corresponding part of the codeword and forms it into a vector form.

[0008] Step S3: The base station inputs the obtained received signal vector into the Resnet estimation model, and outputs the estimated probability of each codeword in the codebook. The greater the estimated probability, the greater the possibility that the codeword is the optimal codeword.

[0009] Step S4: The base station arranges all codewords in descending order based on the estimated probability, and feeds back the codeword numbers with the highest estimated probabilities to the ultra-large-scale intelligent reflecting surface through a separate control link. The ultra-large-scale intelligent reflecting surface performs additional tests on the codewords transmitted by the base station, and the base station obtains the final optimal codeword by measuring the signal energy.

[0010] Step S5: The base station feeds back the calculated optimal codeword sequence number to the ultra-large-scale intelligent reflective surface through a separate control link. The ultra-large-scale intelligent reflective surface constructs a phase coefficient with the optimal codeword to perform uplink signal transmission.

[0011] The step S1 comprises the following steps:

[0012] Step S101: The base station constructs a Resnet estimation model. The Resnet estimation model includes an input module, a convolution module and an output module. The input module consists of a convolution layer and a maximum pooling layer. The convolution module consists of 8 basic modules connected end to end, one of which consists of two convolution layers, and each module adopts an identity mapping structure, that is, the input data is added to the output data of the last convolution layer and used as the final output of the module. The output module consists of an average pooling layer and a fully connected layer.

[0013] Step S102: The base station randomly initializes all model parameters of the Resnet estimation model. The input is set to the received signal obtained by the base station, and the output is set to the estimated probability of all codewords in the codebook.

[0014] Step S103: The base station randomly generates cascade channels in the ultra-large-scale intelligent reflective surface-assisted communication system, and uses an exhaustive search method to determine the optimal codeword under each cascade channel, and uses it as a label for the cascade channel.

[0015] Step S104: The base station uses the generated cascade channels and corresponding labels as training samples, and uses the Adama optimizer and the learning rate decay strategy to train the Resnet estimation model. During the training process, the model parameters are continuously updated to minimize the loss function. The loss function is expressed as follows:

[0016]

[0017] Where S represents the number of codewords in the original codebook W; represents the probability of the sth codeword estimated by the Resnet estimation model; p s represents the true estimated probability of each codeword. When the sth codeword is the optimal codeword, p s =1, otherwise p s =0.

[0018] When the loss function converges, save the model parameters.

[0019] The step S2 comprises the following steps:

[0020] Step S201: design a predefined near-field codebook W suitable for a very large-scale smart reflective surface according to an array steering vector of a cascaded channel in a near-field environment, wherein each codeword in the codebook is a set of reflection phase coefficients of the very large-scale smart reflective surface.

[0021] Step S202: Based on the codebook obtained in step S201, select codewords at intervals D to form a sub-codebook W. sub , the number of codewords in the sub-codebook is correspondingly 1 / D of the number of codewords in the original codebook.

[0022] Step S203: The user transmits a pilot signal to the ultra-large-scale intelligent reflection surface. The ultra-large-scale intelligent reflection surface receives the pilot signal from the subcodebook W sub A code word is selected to form a phase coefficient and reflect the pilot signal to the base station, and the base station obtains the corresponding received signal.

[0023] Step S204: Repeat step S203 until the subcodebook W sub All code words in the sub-codebook are tested. The base station obtains the corresponding received signals of all code words in the sub-codebook and forms them into vector form.

[0024] The step S4 comprises the following steps:

[0025] Step S401: The base station sorts all codewords in the source codebook W in descending order according to the estimated probability of each codeword obtained in step S3, and takes the first K codewords to form a codebook F that requires additional testing.

[0026] Step S402: The base station feeds back information of the additional test codebook F to the ultra-large-scale intelligent reflective surface through a separate control link.

[0027] Step S403: The user transmits a pilot signal to the ultra-large-scale intelligent reflecting surface again. The ultra-large-scale intelligent reflecting surface selects a codeword from the codebook F to form a phase coefficient and reflects the pilot signal to the base station. The base station obtains a corresponding received signal.

[0028] Step S404: The base station repeats step S403 until all code words in the codebook F are tested. The base station obtains corresponding received signals of the K code words.

[0029] Step S405: The base station calculates the modulus values ​​of the K received signals, and the codeword corresponding to the received signal with the largest modulus value is the optimal codeword.

[0030] The beneficial effects of the present invention are:

[0031] Compared with the existing beam training scheme, the present invention takes into account the codebook and channel model in the near-field environment, and adopts the Resnet estimation model based on the deep residual network to replace the traditional exhaustive test scheme, which only needs to test part of the codewords, effectively reducing the pilot overhead of beam training in the near-field environment. In addition, the present invention performs additional codeword tests based on the Resnet estimation model, which greatly improves the estimation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Design a flow chart for a deep learning based low pilot overhead beam training scheme for very large scale smart reflective surface assisted communications;

[0033] Figure 2 A schematic diagram of ultra-large-scale intelligent reflective surface assisted communication;

[0034] Figure 3 This is a network architecture diagram of the Resnet estimation model of the present invention;

[0035] Figure 4 The spectrum efficiency effect diagram obtained by the embodiment of the present invention and the exhaustive scanning solution is shown. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Example 1

[0038] See also Figure 1 , Figure 2 and Figure 3 This example provides a low pilot overhead beam training solution design for ultra-large-scale intelligent reflective surface assisted communication based on deep learning. By building and training the Resnet estimation model, the base station can accurately estimate the optimal codeword in the near-field codebook based on the received signal of only part of the codeword, effectively reducing the pilot overhead required for beam training in the ultra-large-scale intelligent reflective surface assisted wireless communication system. At the same time, additional codeword testing improves the accuracy of beam training. The specific steps include:

[0039] Step 1. In a wireless communication system assisted by a very large-scale intelligent reflective surface, the direct path between the user and the base station is blocked by obstacles, and only the cascade path from the user to the base station through the very large-scale intelligent reflective surface is considered. The base station is equipped with M = 512 antennas, the very large-scale intelligent reflective surface is configured with N = 512 reconfigurable phase reflection units, and the user is configured with a single antenna. In addition, there is a separate control link between the base station and the very large-scale intelligent reflective surface. First, the base station needs to build and train a Resnet estimation model based on a deep residual network to obtain the model parameters.

[0040] a1) The base station is constructed as follows Figure 3 The Resnet estimation model is designed based on the deep residual network structure shown in the figure, wherein the Resnet estimation model includes an input module, a convolution module and an output module. The input module consists of a convolution layer and a maximum pooling layer, wherein the convolution kernel size of the convolution layer is 2*2, and the number of output channels is 64. The received signal vector will be split into a real part and an imaginary part before input, and the two parts will be converted into real matrix forms as feature maps of two channels and input into the input module. The convolution module is composed of 8 basic modules connected end to end, one of which is composed of two convolution layers, and each module adopts an identity mapping structure, that is, the input data is added to the output data of the last convolution layer and used as the final output of the module ( Figure 2 The convolution kernel size of the convolution layer in the basic module is 3*3, and the number of output channels is Figure 2As shown, it increases from 64 to 512. In addition, each convolutional layer of the basic module is attached with a Relu activation layer and a Batchnorm layer. The output module consists of an average pooling layer and a fully connected layer, where the activation function of the fully connected layer adopts the Softmax activation function, and the number of neurons in the fully connected layer is equal to the number of codewords in the source codebook W.

[0041] a2) The base station randomly initializes all model parameters of the Resnet estimation model. The input is set to the received signal obtained by the base station, and the output is set to the estimated probability of all codewords in the codebook.

[0042] a3) The base station randomly generates 20,000 cascade channel training samples based on the near-field spherical wave channel model, of which 18,000 samples are used to train the neural network model and 2,000 samples are used to test the estimation performance of the neural network model. For each cascade channel in the sample, an exhaustive search method is used to determine the optimal codeword under the cascade channel and use it as the label of the training sample.

[0043] a4) The base station uses the generated cascade channels and corresponding labels as training samples, and uses the Adama optimizer and learning rate decay strategy to train the Resnet estimation model. During the training process, the model parameters are continuously updated to minimize the loss function. The loss function is expressed as follows:

[0044]

[0045] Where S represents the number of codewords in the original codebook W; represents the probability of the sth codeword estimated by the Resnet estimation model; p s represents the true estimated probability of each codeword. When the sth codeword is the optimal codeword, p s =1, otherwise p s =0.

[0046] During the training process, the initial learning rate is set to 0.005, and when the loss function does not decrease for 5 consecutive training rounds, the learning rate will decay to half of the original value. Each training round traverses 2000 samples until all training samples are traversed.

[0047] When the loss function converges, save the model parameters.

[0048] Finally, the 2000 samples used for testing can be used to test the estimation performance of the final Resnet estimation model to ensure the generalization performance of the final model parameters.

[0049] Step 2: The user sends a pilot signal to the ultra-large intelligent reflective surface, which selects some codewords in the predefined codebook to form a reflection phase coefficient and reflects the pilot signal to the base station. The base station obtains the corresponding received signal of the partial codeword and forms it into a vector form.

[0050] b1) According to the array steering vector of the cascade channel in the near-field environment, a predefined near-field codebook W suitable for the ultra-large-scale smart reflective surface is designed, wherein each codeword in the codebook is a set of phase coefficients of the ultra-large-scale smart reflective surface. Finally, the number of codewords in the near-field codebook W is 6000.

[0051] b2) Based on the codebook obtained in b1), select codewords at intervals D = 20 to form a sub-codebook W sub , the number of codewords in the sub-codebook is correspondingly 1 / 20 of the number of codewords in the original codebook.

[0052] b3) The user transmits a pilot signal to the ultra-large-scale intelligent reflection surface. The ultra-large-scale intelligent reflection surface is based on the subcodebook W sub A code word is selected to form a phase coefficient and reflect the pilot signal to the base station, and the base station obtains the corresponding received signal.

[0053] b4) Repeat step b3) until the subcodebook W sub All code words in the sub-codebook are tested. The base station obtains the corresponding received signals of all code words in the sub-codebook and forms them into vector form.

[0054] Step 3: The base station inputs the received signal vector into the Resnet estimation model, and outputs the estimated probability of each codeword in the codebook. The larger the estimated probability, the greater the possibility that the codeword is the optimal codeword.

[0055] Step 4: The base station arranges all codewords in descending order based on the estimated probability, and feeds back the codeword numbers with the highest estimated probabilities to the ultra-large-scale intelligent reflecting surface through a separate control link. The ultra-large-scale intelligent reflecting surface performs additional tests on the codewords transmitted by the base station, and the base station obtains the final optimal codeword by measuring the energy of the received signal.

[0056] c1) The base station sorts all codewords in the source codebook W in descending order according to the estimated probability of each codeword obtained in step S3, and takes the first K codewords to form a codebook F that requires additional testing.

[0057] c2) The base station feeds back the information of the additional test codebook F to the ultra-large-scale intelligent reflective surface through a separate control link.

[0058] c3) The user transmits a pilot signal to the ultra-large-scale intelligent reflecting surface again. The ultra-large-scale intelligent reflecting surface selects a codeword from the codebook F to form a phase coefficient and reflects the pilot signal to the base station. The base station obtains a corresponding received signal.

[0059] c4) The base station repeats step S403 until all code words in the codebook F are tested. The base station obtains the corresponding received signals of the K code words.

[0060] c5) The base station calculates the modulus values ​​of the K received signals, and the codeword corresponding to the received signal with the largest modulus value is the optimal codeword.

[0061] Step 5: The base station feeds back the calculated optimal codeword sequence number to the ultra-large-scale intelligent reflective surface through a separate control link. The ultra-large-scale intelligent reflective surface constructs a phase coefficient with the optimal codeword to perform uplink signal transmission.

[0062] The results of the above specific embodiments are as follows Figure 4 As shown, the result graphs show the achievable rate performance of the proposed scheme when K=1, K=5, and K=10, and the performance curve of the exhaustive scanning scheme is also added for comparison. The effect graph of this embodiment shows that the present invention can approach and achieve the achievable rate performance of the exhaustive scanning scheme in uplink data transmission, but the pilot overhead is reduced to nearly 1 / 20 of the exhaustive scanning scheme.

Claims

1. A deep learning-based uplink ultra-large-scale intelligent reflection surface beam training method, characterized in that: The steps include: Step S1: The base station constructs and trains a Resnet estimation model based on a deep residual network to obtain model parameters; Step S2: The base station is configured with M antennas, the ultra-large-scale intelligent reflection surface is configured with N reflection units with reconfigurable reflection phases, and the user is configured with a single antenna; The user sends a pilot signal to the ultra-large intelligent reflecting surface, and the ultra-large intelligent reflecting surface selects part of the codewords in the predefined codebook to form a reflection phase coefficient and reflects the pilot signal to the base station; the base station obtains the corresponding partial codeword receiving signal and forms it into a vector form; Step S3: The base station inputs the received signal vector into the Resnet estimation model, and outputs the estimated probability of each codeword in the codebook. The greater the estimated probability, the greater the possibility that the codeword is the optimal codeword. Step S4: The base station arranges all codewords in descending order based on the estimated probability, and feeds back the codeword sequence number with the highest estimated probability to the ultra-large-scale intelligent reflecting surface through a separate control link. The ultra-large-scale intelligent reflecting surface performs additional tests on the codewords transmitted by the base station, and the base station obtains the final optimal codeword by measuring the signal energy. Step S5: The base station feeds back the calculated optimal codeword sequence number to the ultra-large-scale intelligent reflective surface through a separate control link. The ultra-large-scale intelligent reflective surface constructs a phase coefficient with the optimal codeword to perform uplink signal transmission. In step S2, the ultra-large-scale intelligent reflection surface selects part of the codewords in the predefined codebook to form the reflection phase coefficient and reflects the pilot signal, specifically comprising: Step S201: designing a predefined near-field codebook W of a very large-scale smart reflective surface according to an array steering vector of a cascaded channel in a near-field environment, wherein each codeword in the codebook is a set of reflection phase coefficients of the very large-scale smart reflective surface; Step S202: Based on the codebook obtained in step S201, select codewords at intervals D to form a sub-codebook W. sub , the number of codewords in the sub-codebook is correspondingly 1 / D of the number of codewords in the original codebook; Step S203: The user transmits a pilot signal to the ultra-large-scale intelligent reflection surface; the ultra-large-scale intelligent reflection surface uses the subcodebook W sub A codeword is selected to form a phase coefficient and reflect the pilot signal to the base station, and the base station obtains the corresponding received signal; Step S204: Repeat step S203 until the subcodebook W sub All code words of the sub-codebook are tested; the base station obtains the corresponding received signals of all code words in the sub-codebook and forms them into a vector form; In step S4, the ultra-large-scale intelligent reflective surface performs additional testing on the codeword transmitted by the base station and obtains the final optimal codeword at the base station. The specific steps are: Step S401: The base station sorts all codewords in the source codebook W in descending order according to the estimated probability of each codeword obtained in step S3, and takes the first K codewords to form a codebook F that needs additional testing; Step S402: The base station feeds back information of the additional test codebook F to the ultra-large-scale intelligent reflective surface through a separate control link; Step S403: The user transmits a pilot signal to the ultra-large-scale intelligent reflection surface again. The ultra-large-scale intelligent reflection surface selects a codeword from the codebook F to form a phase coefficient and reflects the pilot signal to the base station. The base station obtains a corresponding received signal. Step S404: The base station repeats step S403 until all code words in the codebook F are tested, and the base station obtains corresponding received signals of K code words; Step S405: The base station calculates the modulus values ​​of the K received signals, and the codeword corresponding to the received signal with the largest modulus value is the optimal codeword.

2. According to the deep learning-based uplink ultra-large-scale intelligent reflection surface beam training method of claim 1, it is characterized in that: In step S1, a Resnet estimation model based on a deep residual network is constructed and trained, and the specific steps are as follows: Step S101: The base station constructs a Resnet estimation model; Step S102: The base station randomly initializes all model parameters of the Resnet estimation model; the input is set to the received signal obtained by the base station, and the output is set to the estimated probability of all codewords in the codebook; Step S103: The base station randomly generates cascade channels in the ultra-large-scale intelligent reflective surface-assisted communication system, and uses an exhaustive search method to determine the optimal codeword under each cascade channel, and uses the optimal codeword as a label of the cascade channel; Step S104: The base station uses the generated cascade channels and corresponding labels as training samples, and uses the Adama optimizer and the learning rate decay strategy to train the Resnet estimation model. During the training process, the model parameters are continuously updated to minimize the loss function. The loss function is expressed as follows: Where S represents the number of codewords in the original codebook W; represents the probability of the sth codeword estimated by the Resnet estimation model; p s represents the true estimated probability of each codeword. When the sth codeword is the optimal codeword, p s =1, otherwise p s =0; When the loss function converges, save the model parameters.

3. The uplink ultra-large-scale intelligent reflection surface beam training method based on deep learning according to claim 2 is characterized in that: The Resnet estimation model includes an input module, a convolution module and an output module; the input module consists of a convolution layer and a maximum pooling layer; the convolution module consists of 8 basic modules connected end to end, one of which consists of two convolution layers, and each module adopts an identity mapping structure, that is, the input data is added to the output data of the last convolution layer and used as the final output of the module; the output module consists of an average pooling layer and a fully connected layer.

Images