Method for millimeter wave and terahertz channel modeling based on migration generative adversarial network

Abstract

The application discloses a millimeter wave and terahertz channel modeling method based on a migration generative adversarial network, and the method comprises the following steps: constructing a generative adversarial network model in an offline stage, pre-training the generative adversarial network model by using a simulation data set, performing migration learning and fine-tuning on the pre-trained generative adversarial network model, and obtaining a millimeter wave and terahertz channel of the migration adversarial network model; and in an online stage, a large amount of channel data consistent with measured data can be generated, thereby solving the problem of lack of measured data of the millimeter wave and terahertz, and helping the design of a millimeter wave and terahertz communication system. Compared with a standard channel model in a traditional third-generation mobile communication cooperation plan, the application has good performance in channel modeling, the root mean square error is increased by 9 dB, the structural similarity index measurement is higher, the problem of lack of measured data is solved, and the application is verified in an actual measured channel.

Description

Technical Field

[0001] This invention relates to a technology in the field of wireless communication, specifically a method for modeling millimeter-wave and terahertz channels based on migration generative adversarial networks. Background Technology

[0002] Existing random channel modeling methods suffer from low accuracy under certain assumed distributions and empirical parameters. For example, geometry-based random channel models assume that the scattering locations follow certain statistical distributions, such as the transmitter and receiver being uniformly distributed within a circle. However, the scattering locations are difficult to characterize using statistical distributions, making geometry-based random channel models inaccurate for millimeter-wave and terahertz bands. Furthermore, obtaining extensive channel measurements for millimeter-wave and terahertz channel modeling is both time-consuming and expensive, resulting in a lack of abundant measurement data.

[0003] Existing high-frequency channel modeling techniques based on generative adversarial networks (GANs) generally improve the original channel modeling by using GANs. However, this technique requires a large amount of data for supervised learning and cannot be implemented in scenarios with small sample sizes. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention proposes a millimeter-wave and terahertz channel modeling method based on transfer learning and generative adversarial networks (GANs). By applying transfer learning to GAN-based channel modeling, a large amount of GAN-based millimeter-wave and terahertz channel data can be obtained using only a small amount of millimeter-wave and terahertz channel measurement datasets, which can be used to assist in the design of communication systems. The power delay distribution generated by the GAN shows good agreement with the measurement results. Compared with the standard channel model in the traditional 3G cooperative communication scheme, the GAN achieves superior performance in channel modeling, improving the root mean square error by 9 dB and exhibiting a higher structural similarity index. This solves the problem of scarce measurement data and has been validated in actual channel measurements.

[0005] This invention is achieved through the following technical solution:

[0006] This invention relates to a method for modeling millimeter-wave and terahertz channels based on transfer generative adversarial networks (GANs). The method involves constructing a GAN model offline and pre-training it using a simulated dataset, then performing transfer learning and fine-tuning on the pre-trained GAN model to obtain the millimeter-wave and terahertz channels. In the online phase, it can generate a large amount of channel data that closely matches measurement data, thus solving the problem of insufficient measurement data for millimeter-wave and terahertz communication and contributing to the design of millimeter-wave and terahertz communication systems.

[0007] The simulated dataset was generated from the standard channel model of the 3rd Generation Mobile Communications Cooperative Initiative.

[0008] The pre-training mentioned above refers to training the generative adversarial network using a simulated dataset.

[0009] The aforementioned transfer learning refers to the transfer of knowledge learned from a simulated dataset to a measurement dataset.

[0010] The fine-tuning refers to retraining a generative adversarial model that has already been trained on a simulated dataset using a measurement dataset.

[0011] Technical effect

[0012] This invention utilizes transfer learning to transfer knowledge learned from simulated datasets to measurement datasets. The resulting transfer-based generative adversarial network can generate channel data that highly matches the measurement data, significantly expanding the measurement dataset. Compared to existing technologies, this invention greatly expands the measurement dataset, thus solving the problem of scarce data in millimeter-wave and terahertz systems, and contributing to the design of millimeter-wave and terahertz communication systems. Attached Figure Description

[0013] Figure 1 A schematic diagram of an adversarial network is generated for this invention;

[0014] Figure 2 This is a schematic diagram of transfer learning;

[0015] Figure 3 For channel data measurement scenarios;

[0016] Figure 4 The average power delay distribution;

[0017] Figure 5 A schematic diagram of the similarity measure (SSIM) between the generated power delay distribution and the actual measurement data. Detailed Implementation

[0018] like Figure 1 As shown, this embodiment relates to a modeling method for millimeter-wave and terahertz channels based on transfer generative adversarial networks, including:

[0019] Step 1: Generate a simulation dataset based on the standard channel model of the 3G Mobile Cooperative Scheme. The parameters of this simulation dataset are derived from measurement data and include delay spread, angle domain, and path loss parameters.

[0020] The simulated dataset generated in this embodiment contains 10,000 channel data points, each channel data point having a 401-dimensional power delay distribution, corresponding to the power within a 400ns reception time.

[0021] Step 2: Train using the simulated dataset generated in Step 1, as follows Figure 1 The generative adversarial network shown.

[0022] The generative adversarial network includes a generator G and a discriminator D, wherein the generator G maps random noise to a simulated power delay distribution based on 100-dimensional random noise; and the discriminator D determines whether the input data is real based on the real power delay distribution or the simulated power delay distribution from the generator.

[0023] The generator G and discriminator D are both composed of five fully connected layers, with the number of neurons being 128, 128, 128, 128, 401 and 512, 256, 128, 64, 1, respectively.

[0024] The training process involves alternating training sessions for the generator and discriminator for a total of 10,000 rounds, with the loss function being... Where: E is the expected data, x is the input real data, and G(z) is the generated fake data. Let G(z) be the point obtained by randomly linearly sampling between x and G(z), i.e. λ is an adjustable parameter that controls the gradient loss.

[0025] Step 3, as follows Figure 2 As shown, the trained generative adversarial network is retrained on the measurement data to obtain the transfer generative adversarial network T-GAN, which generates channel data that matches the measurement data in the online stage.

[0026] Through specific practical experiments, at a communication frequency of 0.3THz, in an indoor setting in an office corridor, 21 millimeter-wave and terahertz channel measurement points were distributed within a communication distance of 31m. Using these 21 channel measurement points, a migration generative adversarial network can be used to model the channel, and a large amount of channel power delay distribution data that matches the measurement data can be generated.

[0027] Compared with existing technologies, this method can perform channel modeling in millimeter-wave and terahertz bands even when measurement data is scarce. In an indoor scenario of an office corridor at a communication frequency of 0.3 THz, 21 millimeter-wave and terahertz channel measurement points are distributed within a 31m communication distance. Using these 21 measurement points, a transfer generative adversarial network (GAN) can be used to model the channel. Experimental results show that the power delay distribution generated by the GAN agrees well with the measurement results. Compared with the standard channel model in the traditional 3G cooperative scheme, the GAN achieves good performance in channel modeling. Figure 4 and Figure 5 As shown, its root mean square error was improved by 9 dB, and its structural similarity index was also higher.

[0028] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for modeling millimeter-wave and terahertz channels based on transfer generative adversarial networks, characterized in that, By constructing a generative adversarial network model offline and pre-training it using a simulated dataset, and then performing transfer learning and fine-tuning on the pre-trained adversarial network model, millimeter-wave and terahertz channels of the transfer adversarial network model are obtained. During the online phase, a large amount of channel data that matches the measurement data is generated, specifically including: Step 1: Generate a simulation dataset based on the standard channel model of the 3G Mobile Cooperative Initiative. The parameters of this simulation dataset are derived from measurement data and include delay spread, angle domain, and path loss parameters. The simulated dataset contains 10,000 channel data points, each channel data point having a 401-dimensional power delay distribution, corresponding to the power within a 400ns reception time. Step 2: Train a generative adversarial network using the simulated dataset generated in Step 1; The generative adversarial network includes a generator G and a discriminator D, wherein the generator G maps random noise to a simulated power delay distribution based on 100-dimensional random noise; the discriminator D determines whether the input data is real based on the real power delay distribution or the simulated power delay distribution from the generator. The generator G and discriminator D are both composed of five fully connected layers, with the number of neurons being 128, 128, 128, 128, 401 and 512, 256, 128, 64, 1, respectively. The training process involves alternating training sessions for the generator and discriminator for a total of 10,000 rounds, with the loss function being... Where: E is the expected data, x is the input real data, and G(z) is the generated fake data. Let G(z) be the point obtained by randomly linearly sampling between x and G(z), i.e. = x+ , These are adjustable parameters for controlling the gradient loss; Step 3: Retrain the trained generative adversarial network on the measurement data to obtain the transfer generative adversarial network T-GAN, thereby generating channel data that matches the measurement data in the online stage.

Images