High-frequency channel estimation method based on generative adversarial network

Abstract

The present invention belongs to the technical field of terahertz channel estimation, and specifically is a high-frequency channel estimation method based on a generative adversarial network. The present invention utilizes the spatial spectrum coupling characteristics of leaky wave antennas to make the designed new neural network architecture insensitive and robust to the Doppler effect in terahertz vehicle-to-infrastructure (V2I) networks. The generator generates estimated samples that are as similar as possible to real samples to deceive the discriminator, and the discriminator learns to distinguish between the estimated samples generated by the generator and the real samples. After the training converges, the generator is used as a channel estimator. The channel estimator enables each vehicle to predict the channel power gain based on the received signal power strength (RSS). The present invention has universality, can accurately estimate the channel power gain, and has a faster convergence speed.

Description

Technical Field

[0001] The present invention belongs to the technical field of terahertz channel estimation, and in particular relates to a high-frequency channel estimation method based on a generative adversarial network. Background Art

[0002] Emerging services such as edge computing, immersive communications, and the tactile Internet require high-speed data transmission. To enable these services, the large bandwidth of the millimeter wave (mmW) band is utilized in 5G (currently 24-50GHz). However, as the number of smart devices and autonomous vehicles in the Internet of Things increases, more bandwidth is always needed to strengthen ultra-reliable and low-latency communications (URLLC). Due to the abundance of terahertz (THz) frequency bands, THz communication is seen as a promising 6G technology that can achieve unprecedented high data rates.

[0003] To offset the severe path loss in higher frequencies, larger antenna arrays with narrow beams are used in THz communications compared to mmWave communications. However, existing channel estimation methods exhaustively scan the space to find the best alignment, and their complexity increases with the number of antennas. At the same time, the larger bandwidth of THz communication introduces too many pilots. Therefore, the huge training overhead makes traditional channel estimation at higher frequencies gradually infeasible. The random effects of molecular absorption losses in the THz band prevent empirical models from accurately reflecting actual transmission scenarios, which also poses a huge challenge to THz channel estimation. Therefore, it is necessary to explore a completely new THz channel estimation framework. Machine learning has become a fundamental technique for high-frequency channel estimation because it is able to learn complex mapping functions. Summary of the invention

[0004] In view of this, the purpose of the present invention is to propose a high-frequency channel estimation method based on a generative adversarial network to accurately estimate the high-frequency channel power gain.

[0005] The high-frequency channel estimation method based on the generative adversarial network provided by the present invention is described in detail in Figure 1 , Figure 2As shown. Using the GAN framework, two feedforward networks, the generator and the discriminator, are designed. The output of the generator and the discriminator is guided by splicing the conditional information vector at the input layer, so that the GAN is transformed from unsupervised learning to supervised learning, that is, it becomes a conditional generative network (cGAN); the conditional generative network (cGAN) has a specific structure: a generator and a discriminator; the generator is composed of a three-layer neural network structure consisting of a fully connected layer, a batch normalization layer, and a PReLU activation layer, and a four-layer neural network structure consisting of a fully connected layer, a batch normalization layer, a PReLU activation layer, and a batch normalization layer; the generator structure contains two vector splicing; the discriminator is composed of a fully connected layer and a ReLU activation layer, and a two-layer neural network structure consisting of a fully connected layer and a Sigmoid activation layer; wherein the goal of the generator is to generate estimated samples that are as similar as possible to the real samples, and to make the discriminator provide higher probability values ​​for the estimated samples; while the goal of the discriminator is to give higher probability values ​​to the real samples and lower probability values ​​to the estimated samples generated by the generator; in addition, additional cross-layer splicing of conditional information vectors and noise vectors is added, so that the conditional information vector is spliced ​​twice to compensate for the data information loss of consecutive layers in the neural network, and the proposed GAN model generates samples more efficiently; the noise vector and the conditional information vector are also expanded to increase the richness of information and improve the efficiency of generating accurate samples.

[0006] The high-frequency channel estimation method based on a generative adversarial network provided by the present invention comprises the following specific steps:

[0007] Step 1: The generator first selects a set of frequency points f i And the received signal strength corresponding to the frequency point As a conditional information vector, the conditional information vector is And randomly generate a noise vector z=[z 1 ,z 2 ,…,z n ], firstly, the conditional information vector and the noise vector are vector-expanded, and the vectors are processed in sequence using the fully connected layer, the batch normalization layer, and the PReLU activation layer, thereby expanding their dimensions; the received signal strength is used as the conditional information vector due to the special frequency-angle coupling characteristics of the leaky wave antenna used, and the coupling relationship between frequency and angle is contained in the received signal power intensity;

[0008] Step 2: Perform the first vector concatenation of the conditional information vector and the noise vector after dimension expansion; then process them through a three-layer neural network structure including a fully connected layer, a batch normalization layer, and a PReLU activation layer; then use five identical four-layer neural network structures (each four-layer neural network structure includes a fully connected layer, a batch normalization layer, a PReLU activation layer, and a batch normalization layer) for processing; then process them through two identical three-layer neural networks (each three-layer neural network structure includes a fully connected layer, a batch normalization layer, and a PReLU activation layer); the batch normalization layer is used many times in the processing of the three major network architectures mentioned above, and its purpose is to convert the data in the network to a state with a mean of zero and a variance of 1, so that the distribution of each layer of data is the same, thereby accelerating the training convergence of the neural network; using the PReLU activation layer for processing can retain more feature information in the shallow layer of the neural network and have a higher feature recognition rate in the deep layer; as the depth of the neural network increases, the activation function becomes more nonlinear, which leads to stronger fitting ability; the main function of the fully connected layer is to expand the vector dimension to increase the richness of information;

[0009] Step 3: The vector output in step 2 is concatenated with the conditional information vector after dimension expansion for the second time, so that the conditional information vector is concatenated twice to compensate for the data information loss of consecutive layers in the neural network and enable the proposed GAN model to generate samples more efficiently; then the vector dimension is expanded through three layers of fully connected layers to increase the richness of information and adjust the output vector dimension to be consistent with the estimated vector dimension, and the output is the estimated channel power gain vector in Representative frequency f i The corresponding estimated channel power gain value;

[0010] Step 4: The discriminator converts the true channel power gain vector and the estimated channel power gain vector They are concatenated with the conditional information vector as the input of the discriminator, and then processed by two identical two-layer neural network structures (consisting of a fully connected layer and a ReLU layer), and then processed by a fully connected layer and a Sigmoid activation layer respectively, and the output result is a probability value.

[0011] The ReLU layer is used multiple times in the processing of the two network modules mentioned above, in order to reduce the interdependence of parameters and alleviate the occurrence of overfitting problems. The Sigmoid activation layer is used for processing to map the output between 0 and 1. The main function of the fully connected layer is to expand the vector dimension to increase the richness of information and adjust the output vector dimension to be consistent with the output vector dimension.

[0012] Compared with the existing methods, the high-frequency channel estimation method based on GAN proposed in this invention has the following characteristics and advantages:

[0013] (1) By leveraging the spatial-spectral coupling characteristics of leaky-wave antennas, a novel neural network architecture is designed to be insensitive and robust to the Doppler effect in terahertz vehicle-to-infrastructure (V2I) networks.

[0014] (2) The cross-layer concatenation of additional conditional information vectors and noise vectors is considered, so that the conditional information vector is concatenated twice. This can compensate for the data information loss of consecutive layers in the neural network and enable the proposed GAN model to generate samples more efficiently;

[0015] (3) The noise vector and conditional information vector are expanded to increase the richness of information and improve the efficiency of generating accurate samples. It can accurately estimate the channel power gain and has a faster convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The framework diagram of the high-frequency channel estimation method based on the generative adversarial network of the present invention.

[0017] Figure 2 It is a diagram of the network structure in the high-frequency channel estimation method based on the generative adversarial network of the present invention.

[0018] Figure 3 The figure illustrates an application scenario of the Internet of Vehicles of the high-frequency channel estimation method based on the generative adversarial network of the present invention.

[0019] Figure 4 It is a diagram illustrating the channel power gain estimation effect of the high-frequency channel estimation method based on the generative adversarial network of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described below with reference to the embodiments and drawings, so as to more clearly understand the essence of the present invention, as well as the effectiveness and superiority of the present invention. However, it should be understood that the specific embodiments described herein are not intended to limit the present invention.

[0021] like Figure 2As shown, the conditional generative network (cGAN) includes: a generator and a discriminator; the generator is mainly composed of a three-layer neural network structure including a fully connected layer, a batch normalization layer and a PReLU activation layer, and a four-layer neural network structure including a fully connected layer, a batch normalization layer, a PReLU activation layer and a batch normalization layer; the generator structure contains two vector splicing. The discriminator is mainly composed of a two-layer neural network structure including a fully connected layer and a ReLU activation layer, and a fully connected layer and a Sigmoid activation layer. Among them, the goal of the generator is to generate estimated samples that are as similar as possible to the real samples, and to make the discriminator provide a higher probability value for the estimated samples; and the goal of the discriminator is to give a higher probability value to the real samples and a lower probability value to the estimated samples generated by the generator; wherein, the cross-layer splicing of the additional conditional information vector and the noise vector is added, so that the conditional information vector is spliced ​​twice to compensate for the data information loss of the continuous layers in the neural network, and the proposed GAN model generates samples more efficiently; the noise vector and the conditional information vector are also expanded to increase the richness of information and improve the efficiency of generating accurate samples.

[0022] Channel parameter setting: The ITU gas attenuation calculation model is used to generate actual channel power gain data from 100 GHz to 400 GHz. Figure 3 A scenario where a roadside unit (RSU) serves multiple vehicles is shown. Based on the high temporal resolution of terahertz communication, it is reasonable to assume that the spatial state of each vehicle is roughly unchanged when performing channel estimation at very small time intervals.

[0023] Model parameter setting: Assume that the length of the noise vector z is 4 and the length of the estimated channel power gain vector is 5, so the length of the conditional information vector y is 10. For the generator, the lengths of the noise vector and the conditional information vector are first extended to 500 and 1000, respectively. After the first concatenation, the vector length is kept at 1500 through a three-layer structure and five identical four-layer structures, and then the vector length is extended to 10000 and reduced to 1000 before the second concatenation. Finally, through three fully connected layers, the vector length is reduced to 1500, 1000, and 5, respectively. For the discriminator, the length of the concatenated vector is first extended to 64, and then the length is reduced to 32, 1 in the last two fully connected layers. In addition, for each training, the training data input to the model in batches is 128. The generator G and the discriminator D are trained 200 times in each iteration, and the learning rate of the generator and the discriminator is the same, which is 0.003.

[0024] Effect analysis: Figure 4The average normalized mean square error values ​​of the traditional cGAN and the proposed cGAN as the training process progresses are shown. It can be seen that the proposed method has a faster convergence speed, which is due to the expansion of the noise vector z and the conditional information vector y. The vector expansion increases the richness of information and enables the conditional information vector y to more efficiently guide the model to generate accurate prediction data. At the same time, the proposed cGAN has a lower average normalized mean square error value, which indicates that it can more accurately estimate the actual channel power gain vector h, which is the result of the invention's clever double splicing of the conditional vector y.

Claims

1. A high-frequency channel estimation method based on generative adversarial networks, It is characterized in that Using the GAN framework, two feedforward networks, a generator and a discriminator, are designed. The output of the generator and the discriminator is guided by splicing the conditional information vector at the input layer, so that the GAN is transformed from unsupervised learning to supervised learning, that is, a conditional generative network (cGAN); the specific structure of the conditional generative network (cGAN) includes: a generator and a discriminator; the generator is composed of a three-layer neural network structure consisting of a fully connected layer, a batch normalization layer and a PReLU activation layer, and a four-layer neural network structure consisting of a fully connected layer, a batch normalization layer, a PReLU activation layer and a batch normalization layer; the generator structure contains two vector splicing; the discriminator is composed of a two-layer neural network structure consisting of a fully connected layer and a ReLU activation layer, and a fully connected layer and a Sigmoid activation layer; The goal of the generator is to generate estimated samples that are as similar as possible to the real samples, and to make the discriminator provide higher probability values ​​for the estimated samples; while the goal of the discriminator is to give higher probability values ​​to the real samples and lower probability values ​​to the estimated samples generated by the generator; wherein, additional conditional information vectors and cross-layer concatenation of noise vectors are added, so that the conditional information vectors are concatenated twice to compensate for the data information loss of consecutive layers in the neural network and enable the proposed GAN model to generate samples more efficiently; the noise vectors and conditional information vectors are also expanded to increase the richness of information and improve the efficiency of generating accurate samples; The specific steps are: Step 1: The generator first selects a set of frequency points f i And the received signal strength corresponding to the frequency point As a conditional information vector, the conditional information vector is And randomly generate a noise vector z=[z 1 ,z 2 , … ,z n ], first perform vector expansion on the conditional information vector and the noise vector, and use the fully connected layer, batch normalization layer, and PReLU activation layer to process the vectors in sequence, thereby expanding their dimensions; Step 2: Perform the first vector concatenation of the dimensionally expanded conditional information vector and the noise vector; then process them through a three-layer neural network structure including a fully connected layer, a batch normalization layer, and a PReLU activation layer, and then use five identical four-layer neural network structures, each of which includes a fully connected layer, a batch normalization layer, a PReLU activation layer, and a batch normalization layer, and then process them through two identical three-layer neural networks; each three-layer neural network structure includes a fully connected layer, a batch normalization layer, and a PReLU activation layer; Step 3: Concatenate the vector output in step 2 with the dimensionally expanded conditional information vector for the second time, and then pass it through three fully connected layers to output the estimated channel power gain vector in Representative frequency f i The corresponding estimated channel power gain value; Step 4: The discriminator converts the true channel power gain vector and the estimated channel power gain vector They are concatenated with the conditional information vector as the input of the discriminator, and then pass through two identical double-layer neural network structures, the fully connected layer and the ReLU layer, and then pass through the fully connected layer and the Sigmoid activation layer respectively, and the output result is a probability value.

Images