Digital twin channel modeling method based on generative adversarial network and transfer learning

Abstract

The invention provides a digital twin channel modeling method based on a generative adversarial network and transfer learning, which relates to the technical field of wireless communication, and comprises the following steps: firstly, constructing a semi-physical digital twin channel simulation platform, generating a baseband signal by using the semi-physical digital twin channel simulation platform, and processing the baseband signal; the channel frequency response data is obtained, and the cost of collecting the real channel data is effectively reduced while the precision of the channel data is kept; the channel frequency response data is utilized to train the generative adversarial network, the transfer learning strategy is utilized to accelerate the training process of the generative adversarial network, and parameters obtained by training at the semi-physical digital twin channel simulation platform are transferred to the target task, so that the time for updating each parameter again is shortened, the training efficiency is improved, and the training efficiency of the semi-physical digital twin channel simulation platform is improved. And finally, the trained generative adversarial network is utilized to generate channel data, so that the feature extraction capability of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication, and more specifically, to a digital twin channel modeling method based on generative adversarial network and transfer learning. Background Art

[0002] Traditional channel modeling methods are generally divided into deterministic models and statistical models; deterministic channel models mainly rely on the relative positions between the transmitter and the receiver and the electromagnetic characteristics of scatterers, and describe the channel by accurately calculating the signal propagation path and attenuation characteristics; while statistical channel models adopt signal propagation mechanisms and mathematical statistics methods to describe and predict channel characteristics using statistical data, and this method is applicable to situations where the channel dynamics are strong and it is difficult to accurately model. However, traditional channel modeling methods often require a large amount of time and computing resources, especially in complex environments, where the computational amount and time consumption are even larger. With the development of wireless communication networks towards B5G and 6G, especially in the Internet of Things environment where the application scenarios are becoming more diverse, the complexity and dynamic changes of channels require a more flexible and efficient modeling method. Therefore, traditional channel modeling has become unrealistic under the conditions of limited resources and high real-time requirements, and there is an urgent need to introduce more efficient and intelligent technical means, such as using digital twin technology to achieve accurate modeling and simulation of wireless channels, so as to meet the strict requirements of modern communication systems in terms of performance, timeliness, and computing resources.

[0003] The rapid development of artificial intelligence and deep learning has provided new solutions for efficient and accurate channel modeling. As a deep learning model, generative adversarial network can generate data similar to the real data distribution through the adversarial training of the generator and the discriminator. In recent years, many studies have begun to apply generative adversarial network to generate channel data. However, the channel data used in current studies are mainly divided into two categories: one is the data generated by software simulation, and the other is the data collected from the actual environment. Although the channel data generated by software can be produced in batches, its accuracy is usually low and it is difficult to truly reflect the channel characteristics in the real world. In contrast, although the channel data collected from the real world can more accurately reflect the channel characteristics of the actual environment, its collection process is complex, the time cost is high, and it is limited by measurement devices and environmental factors, and the acquisition of channel data is usually limited.

[0004] The prior art proposes a channel modeling method based on a channel feature generative adversarial network. First, a channel feature generative adversarial network model including a generator, a discriminator and a classifier is constructed; multiple groups of combination vectors are obtained according to the measured channel data; the combination vectors are input into the generator to obtain the generated channel data; the generator, the discriminator and the classifier in the channel feature generative adversarial network model are alternately trained through the measured channel data and the generated channel data; until the channel feature generative adversarial network model converges, a generator that can simulate the channel response of the measured channel data is output. This scheme uses a generative adversarial network to generate channel data, which improves the accuracy of the generated channel data while reducing the cost of generating channel data. However, this scheme first needs to collect the measured channel data from the actual environment, resulting in a high cost of collecting data; and each time training is performed for a new communication scenario, each parameter needs to be updated again, resulting in a low training efficiency of the model; the generator and discriminator network structure design of this scheme is relatively simple, and the feature extraction capability is relatively weak.

[0005] Therefore, how to design a method that does not require direct collection of measured data from the actual environment but can reflect the actual environmental conditions, and then combine it with a generative adversarial network to achieve digital twin wireless channel modeling based on the cooperation of "data-driven + model". While maintaining accuracy, it can generate a large amount of high-quality channel data. This is a difficulty that needs to be solved urgently. Summary of the invention

[0006] In order to solve the problems of high cost of collecting channel data, low training efficiency and weak feature extraction ability in the current channel modeling methods in the communication field, the present invention proposes a digital twin channel modeling method based on generative adversarial networks and transfer learning. By constructing a semi-physical digital twin channel simulation platform, the cost of collecting channel data is reduced, and the transfer learning strategy is used to accelerate the training process of the generative adversarial network, thereby improving the training efficiency and enhancing the feature extraction ability of the model.

[0007] In order to achieve the above technical effects, the technical solution of the present invention is as follows: This application proposes a digital twin channel modeling method based on generative adversarial networks and transfer learning, comprising the following steps: S1. Build a semi-physical digital twin channel simulation platform; S2. Generate baseband signals using a semi-physical digital twin channel simulation platform and process the baseband signals to obtain channel frequency response data; S3. Use the channel frequency response data to train the generative adversarial network, and use the transfer learning strategy to accelerate the training process of the generative adversarial network to obtain a trained generative adversarial network; S4. Generate channel data using the trained generative adversarial network.

[0008] In this technical solution, first, a semi-physical digital twin channel simulation platform is constructed. The semi-physical digital twin channel simulation platform is used to generate baseband signals and process the baseband signals to obtain channel frequency response data. While maintaining the accuracy of the channel data, the cost of collecting real channel data is effectively reduced. The generative adversarial network is trained using the channel frequency response data, and the transfer learning strategy is used to accelerate the training process of the generative adversarial network. The parameters trained from the semi-physical digital twin channel simulation platform are transferred to the target task, reducing the time to update each parameter again and improving the training efficiency. Finally, the trained generative adversarial network is used to generate channel data, enhancing the feature extraction ability of the model.

[0009] Preferably, the semi-physical digital twin channel simulation platform includes: a transmitting end baseband signal processing device, a receiving end baseband signal processing device, a channel simulator, a transmitting end USRP device, and a receiving end USRP device; The transmitting end baseband signal processing device, the transmitting end USRP device, the channel simulator, the receiving end USRP, and the receiving end baseband signal processing device are sequentially connected by connecting lines. The channel simulator twins and simulates the channel effect of the real wireless communication environment on the communication signal.

[0010] Preferably, the process of using the semi-physical digital twin channel simulation platform to generate baseband signals and process the baseband signals to obtain channel frequency response data in step S2 is as follows: Use the transmitting end baseband signal processing device to generate baseband signals and transmit the signals to the transmitting end USRP device; Use the transmitting end USRP device to modulate the frequency of the baseband signal to the carrier frequency to obtain a modulated signal, and transmit the modulated signal to the channel simulator; Use the channel simulator to simulate the actual wireless communication environment and apply the channel effect to the modulated signal to obtain a modulated signal with the channel effect; Use the receiving end USRP device to perform down-conversion, filtering, down-sampling, and analog-to-digital conversion operations on the modulated signal with the channel effect to convert the modulated signal into a baseband signal; Use the receiving end USRP device to process and analyze the baseband signal and extract the frequency response characteristics of the baseband signal to obtain channel frequency response data.

[0011] Preferably, before training the generative adversarial network using the channel frequency response data, it also includes performing time-domain conversion and feature processing on the channel frequency response data to obtain the training data of the generative adversarial network. The process is as follows: Perform an inverse discrete Fourier transform on the channel frequency response data to obtain channel impulse response data; Clip and normalize the channel impulse response data to obtain the training data of the generative adversarial network.

[0012] Preferably, the generative adversarial network includes a generator and a discriminator. The generator is used to generate generated data with the same distribution as the training data and input it into the discriminator. The discriminator is used to determine whether the channel data input to itself is the channel data generated by the generator or the real channel data obtained through actual sampling.

[0013] Preferably, the generator includes a fully connected layer, a batch normalization layer, a recombination operation layer, six upsampling modules, and a clipping operation layer connected in sequence. The fully connected layer expands the dimension of the input training data and inputs it into the batch normalization layer. The batch normalization layer normalizes the training data with the expanded dimension and inputs it into the recombination operation layer. The recombination operation layer recombines the normalized data. The upsampling module magnifies and extracts the high-level semantic features of the recombined data, generates channel impulse response data, and inputs it into the clipping operation layer. The clipping operation layer clips the channel impulse response data generated by the upsampling module to obtain channel impulse response data of a fixed size.

[0014] Preferably, the upsampling module includes an upsampling layer, a convolutional layer, a batch normalization layer, and six activation function layers connected in sequence. Among them, the upsampling layer magnifies the high-level semantic features of the data, the convolutional layer extracts the high-level semantic features of the data, the batch normalization layer normalizes the high-level semantic features extracted by the upsampling layer. The activation function layers of the first to fifth upsampling modules use the LeakyReLU function to activate the normalized high-level semantic features into channel impulse response data, and the activation function layer of the sixth upsampling module uses the tanh function to compress the amplitude of the channel impulse response data to the interval of [-1, 1].

[0015] Preferably, the discriminator includes a zero-padding operation layer, seven downsampling modules, a flattening operation layer, and a fully connected layer connected in sequence. The zero-padding operation layer expands the dimension of the input channel data and outputs it to the downsampling module. The downsampling module extracts the core semantic features of the channel data and outputs it to the flattening operation layer. The flattening operation layer flattens the core semantic features of the extracted channel data into the data dimension of the fully connected layer and outputs it to the fully connected layer. The fully connected layer synthesizes the flattened feature vectors and outputs the discrimination result.

[0016] Preferably, in step S3, the generative adversarial network is trained using the channel frequency response data, and the transfer learning strategy is used to accelerate the training process of the generative adversarial network to obtain a trained generative adversarial network. The process is as follows: Randomly initialize the network parameters of the generator and discriminator of the generative adversarial network and ; Use the channel frequency response data to train the generative adversarial network to obtain the network parameters of the preliminarily trained generator and discriminator , ; Take the network parameters of the preliminarily trained generator and discriminator , as the initial values, and use the channel data of the target task to fine-tune the network parameters , to generate network parameters that meet the target task , .

[0017] Preferably, in step S5, the process of generating digital twin channel data using the trained generative adversarial network is as follows: Use the trained generator to generate channel data with the same distribution as the channel data of the target task .

[0018] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows: The present invention proposes a digital twin channel modeling method based on a generative adversarial network and transfer learning. First, a semi-physical digital twin channel simulation platform is constructed. The semi-physical digital twin channel simulation platform is used to generate baseband signals and process the baseband signals to obtain channel frequency response data, effectively reducing the cost of collecting real channel data while maintaining the accuracy of the channel data; the generative adversarial network is trained using the channel frequency response data, and the transfer learning strategy is used to accelerate the training process of the generative adversarial network. The parameters trained from the semi-physical digital twin channel simulation platform are transferred to the target task, reducing the time for re-updating each parameter and improving the training efficiency. Finally, the trained generative adversarial network is used to generate channel data, enhancing the feature extraction ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Represents the flowchart of the digital twin channel modeling method based on a generative adversarial network and transfer learning proposed in Embodiment 1 of the present invention; Figure 2 Represents the overall composition structure diagram of the semi-physical digital twin channel simulation platform proposed in Embodiment 2 of the present invention; Figure 3 Schematic diagram of the generator proposed in Embodiment 2 of the present invention; Figure 4 Schematic diagram of the upsampling module proposed in Embodiment 2 of the present invention; Figure 5 Schematic diagram of the discriminator proposed in Embodiment 2 of the present invention; Figure 6 Schematic diagram of the downsampling module proposed in Embodiment 2 of the present invention; Figure 7 Shows the comparison of PDP and FCF between the real channel and the generated channel in the A1-LOS scenario after training the generative adversarial network for 40 rounds without introducing the transfer learning strategy, as proposed in Embodiment 2 of the present invention; Figure 8 Shows the comparison of PDP and FCF between the real channel and the generated channel in the C1-LOS scenario after training the generative adversarial network for 40 rounds without introducing the transfer learning strategy, as proposed in Embodiment 2 of the present invention; Figure 9 Shows the comparison of PDP and FCF between the real channel and the generated channel in the EPA scenario after training the generative adversarial network for 40 rounds without introducing the transfer learning strategy, as proposed in Embodiment 2 of the present invention; Figure 10 Shows the comparison of PDP and FCF between the real channel and the generated channel in the C1-LOS scenario after training the generative adversarial network for 10 rounds without introducing the transfer learning strategy, as proposed in Embodiment 2 of the present invention; Figure 11 Shows the comparison of PDP and FCF between the real channel and the generated channel in the EPA scenario after training the generative adversarial network for 10 rounds without introducing the transfer learning strategy, as proposed in Embodiment 2 of the present invention; Figure 12 Shows the comparison of PDP and FCF between the real channel and the generated channel in the C1-LOS scenario after training the generative adversarial network for 10 rounds by introducing the transfer learning strategy; Figure 13 Shows the comparison of PDP and FCF between the real channel and the generated channel in the EPA scenario after training the generative adversarial network for 10 rounds by introducing the transfer learning strategy. Detailed implementation manners

[0020] The drawings are only for illustrative purposes and should not be construed as a limitation of this patent; For a better illustration of this embodiment, some parts of the drawings are omitted, enlarged or reduced, and do not represent the actual size; For those skilled in the art, it is understandable that some well-known content descriptions in the drawings may be omitted.

[0021] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] The description of the positional relationship in the accompanying drawings is only for illustrative purposes and should not be construed as a limitation of this patent; Embodiment 1 This embodiment proposes a digital twin channel modeling method based on a generative adversarial network and transfer learning. The flowchart of this method is shown in Figure 1 This method includes the following steps: S1. Construct a semi-physical digital twin channel simulation platform; S2. Use the semi-physical digital twin channel simulation platform to generate a baseband signal and process the baseband signal to obtain channel frequency response data; S3. Use the channel frequency response data to train the generative adversarial network, and use the transfer learning strategy to accelerate the training process of the generative adversarial network to obtain a trained generative adversarial network; S4. Use the trained generative adversarial network to generate channel data.

[0023] In this embodiment, first, a semi-physical digital twin channel simulation platform is constructed. The semi-physical digital twin channel simulation platform is used to generate a baseband signal and process the baseband signal to obtain channel frequency response data. While maintaining the accuracy of the channel data, the cost of collecting real channel data is effectively reduced; the channel frequency response data is used to train the generative adversarial network, and the transfer learning strategy is used to accelerate the training process of the generative adversarial network. The parameters trained from the semi-physical digital twin channel simulation platform are transferred to the target task, reducing the time for re-updating each parameter and improving the training efficiency. Finally, the trained generative adversarial network is used to generate channel data, enhancing the feature extraction ability of the model.

[0024] Embodiment 2 In this embodiment, the semi-physical digital twin channel simulation platform includes: a transmitting end baseband signal processing device, a receiving end baseband signal processing device, a channel simulator, a transmitting end USRP device, and a receiving end USRP device; The transmitting end baseband signal processing device, the transmitting end USRP device, the channel simulator, the receiving end USRP device, and the receiving end baseband signal processing device are sequentially connected by connecting lines; the overall composition structure diagram of the semi-physical digital twin channel simulation platform is shown in Figure 2 as shown; The channel simulator twins and simulates the channel effect of the communication signal on the real wireless communication environment.

[0025] In this embodiment, the process of using the semi-physical digital twin channel simulation platform to generate a baseband signal and process the baseband signal to obtain channel frequency response data in step S2 is: The baseband signal is generated by the baseband signal processing device at the transmitting end and transmitted to the USRP device at the transmitting end; The USRP device at the transmitting end modulates the frequency of the baseband signal to the carrier frequency to obtain a modulated signal, and transmits the modulated signal to the channel emulator; The channel emulator simulates the actual wireless communication environment, applies channel effects to the modulated signal, and obtains a modulated signal with channel effects; The USRP device at the receiving end performs down-conversion, filtering, down-sampling, and analog-to-digital conversion operations on the modulated signal with channel effects, and converts the modulated signal into a baseband signal; The USRP device at the receiving end processes and analyzes the baseband signal and extracts the frequency response characteristics of the baseband signal to obtain channel frequency response data.

[0026] In actual implementation, the USRPs at the transmitting and receiving ends are controlled by a self-developed B5G link-level simulator based on MATLAB software. Among them, the baseband signal processing device at the transmitting end generates an IQ baseband signal after operations such as channel coding, symbol mapping, pilot insertion, and multi-carrier modulation through MATLAB software, and transmits the baseband signal to the USRP at the transmitting end through a network port. In the USRP at the transmitting end, the baseband signal undergoes digital-to-analog conversion, up-sampling, filtering, and up-conversion to the carrier frequency, and then is sent into the Spirent professional channel emulator through a feeder connection. The channel emulator applies various channel effects to the transmitted signal, such as path loss, large-scale fading, small-scale fading, multipath and Doppler effects, noise, etc., to simulate the actual wireless communication environment.

[0027] In this embodiment, the channel scenarios of the channel emulator are respectively set to the A1-LOS (line-of-sight in indoor office environment), B1-NLOS (urban small cell), B2-NLOS (severe urban small cell), C1-LOS (suburban macro cell), C2-NLOS (urban macro cell), C3-NLOS (severe urban macro cell) channel scenarios in WINNER II, the EPA (pedestrian mobility scenario), EVA (vehicle mobility scenario), HT (hilly terrain scenario) channel scenarios in 3GPP (3rd Generation Partnership Project) LTE (Long Term Evolution), and the SUI-2 and SUI-6 channel scenarios of Stanford University.

[0028] The signal after passing through the channel simulator continues to be transmitted to the receiving end USRP device through the feeder. The USRP device at the receiving end performs operations such as down-conversion, filtering, down-sampling, and analog-to-digital conversion on the high-frequency signal, and converts it into a baseband signal. Subsequently, the baseband IQ signal is transmitted over the network cable to the receiving end baseband signal processing device for multi-carrier demodulation, channel estimation and equalization, symbol demapping, channel decoding, and further processing and analysis of data such as bit error rate and throughput.

[0029] The frequency response characteristics of the channel extracted by the signal estimation module of the receiving end baseband signal processing device, including information such as the channel gain, phase change, and path loss suffered by the signal when propagating in the channel, can provide an important basis for wireless channel modeling, optimization of signal processing algorithms, and performance evaluation of communication systems.

[0030] Specifically, in the transmitting end baseband signal processing device and the receiving end baseband signal processing device, the number of sub-carriers is set to 512, the transmission time interval is set to 100, and the number of orthogonal frequency division multiplexing symbols is set to 14.

[0031] In this embodiment, before training the generative adversarial network using the channel frequency response data, it also includes performing time-domain conversion and feature processing on the channel frequency response data to obtain the training data of the generative adversarial network. The process is as follows: Perform an inverse discrete Fourier transform on the channel frequency response data to obtain the channel impulse response data; Crop and normalize the channel impulse response data to obtain the training data of the generative adversarial network; Specifically, the channel frequency response data can be obtained by analyzing the transmitted signal and the received signal through MATLAB in the transmitting end baseband signal processing device and the receiving end baseband signal processing device. The dimension of the channel frequency response data is 1400×512. The first dimension is the number of samples, and the second dimension is the number of sub-carriers. Then, perform an inverse discrete Fourier transform on the channel frequency response data to obtain the channel impulse response data. Considering that most of the tap energy of the channel impulse response is concentrated in the first few or dozens of taps, the channel impulse response data is cropped to 200 taps, and then the dimension is transformed to obtain the data dimension suitable for the generative adversarial network. Finally, the amplitude of all data is normalized between -1 and 1 to ensure the stability of network training.

[0032] In this embodiment, the generative adversarial network includes a generator and a discriminator. Among them, the generator is used to generate generated data with the same distribution as the training data and input it into the discriminator; the discriminator is used to discriminate whether the channel data input to itself is the channel data generated by the generator or the real channel data obtained through actual sampling.

[0033] In this embodiment, the generator includes a fully connected layer, a batch normalization layer, a recombination operation layer, six upsampling modules, and a cropping operation layer connected in sequence; the structural schematic diagram of the generator is as shown in Figure 3 shown; The fully connected layer expands the dimension of the input training data and inputs it into the batch normalization layer; The batch normalization layer normalizes the training data after dimension expansion and inputs it into the recombination operation layer; The recombination operation layer recombines the normalized data; The upsampling module magnifies and extracts the high-level semantic features of the recombined data, generates channel impulse response data and inputs it into the cropping operation layer; The cropping operation layer crops the channel impulse response data generated by the upsampling module to obtain channel impulse response data of a fixed size.

[0034] In this embodiment, the upsampling module includes an upsampling layer, a convolutional layer, a batch normalization layer, and six activation function layers connected in sequence; the structural schematic diagram of the upsampling module is as shown in Figure 4 shown; among them, the upsampling layer magnifies the high-level semantic features of the data, the convolutional layer extracts the high-level semantic features of the data, the batch normalization layer normalizes the high-level semantic features extracted by the upsampling layer, and the activation function layers of the first to fifth upsampling modules use the LeakyReLU function to activate the normalized high-level semantic features into channel impulse response data, and the activation function layer of the sixth upsampling module uses the tanh function to compress the amplitude of the channel impulse response data to the interval of [-1, 1].

[0035] In this embodiment, the discriminator includes a zero-padding operation layer, seven downsampling modules, a flattening operation layer, and a fully connected layer connected in sequence; the structural schematic diagram of the discriminator is as shown in Figure 5 shown; The zero-padding operation layer expands the dimension of the input channel data and outputs it to the downsampling module; The downsampling module extracts the core semantic features of the channel data and outputs them to the flattening operation layer; the structural schematic diagram of the downsampling module is as shown in Figure 6 shown; The flattening operation layer flattens the core semantic features of the extracted channel data into the data dimension of the fully connected layer and outputs it to the fully connected layer; The fully connected layer synthesizes the flattened feature vectors and outputs the discrimination result; Specifically, the downsampling module consists of a convolutional layer and an activation function layer, and in order to reduce the model calculation parameters, 25% of the neuron parameters are kept unchanged each time the parameters are updated; Specifically, the fully connected layer outputs all the information as a value of 0 or 1 to discriminate the channel data; Specifically, the generative adversarial network is trained using channel frequency response data without introducing a transfer learning strategy. After 40 rounds of training, the differences in the power delay profile (PDP) and frequency correlation function (FCF) between the real channel data (Real channel) and the data generated by the generative adversarial network (Fake channel) in different channel scenarios are compared respectively. Among them, Figure 7 represents the comparison of the PDP and FCF between the real channel and the generated channel in the A1-LOS scenario after training the generative adversarial network for 40 rounds without introducing a transfer learning strategy; Figure 8 represents the comparison of the PDP and FCF between the real channel and the generated channel in the C1-LOS scenario after training the generative adversarial network for 40 rounds without introducing a transfer learning strategy; Figure 9 represents the comparison of the PDP and FCF between the real channel and the generated channel in the EPA scenario after training the generative adversarial network for 40 rounds without introducing a transfer learning strategy; In this embodiment, in step S3, the generative adversarial network is trained using channel frequency response data, and a transfer learning strategy is used to accelerate the training process of the generative adversarial network to obtain a trained generative adversarial network. The process is as follows: Randomly initialize the network parameters of the generator and discriminator of the generative adversarial network and ; Use the channel frequency response data to train the generative adversarial network to obtain the network parameters of the preliminarily trained generator and discriminator , ; Take the network parameters of the preliminarily trained generator and discriminator , as the initial values, and use the channel data of the target task to fine-tune the network parameters , to generate network parameters that meet the target task , ; Specifically, Figure 10 represents the comparison of the PDP and FCF between the real channel and the generated channel in the C1-LOS scenario after training the generative adversarial network for 10 rounds without introducing a transfer learning strategy proposed in Embodiment 2 of the present invention; Figure 11 represents the comparison of the PDP and FCF between the real channel and the generated channel in the EPA scenario after training the generative adversarial network for 10 rounds without introducing a transfer learning strategy proposed in Embodiment 2 of the present invention; Figure 12 represents the comparison of the PDP and FCF between the real channel and the generated channel in the C1-LOS scenario after training the generative adversarial network for 10 rounds with the introduction of a transfer learning strategy; Figure 13It shows the comparison of PDP and FCF between the real channel and the generated channel in the EPA scenario after introducing the transfer learning strategy to train the generative adversarial network for 10 rounds. It can be analyzed by comparison that the introduction of transfer learning accelerates the convergence of the model. Specifically, to quantitatively demonstrate the effect of transfer learning, Table 1 shows the mean square error between the PDP and FCF of the real channel and the digital twin channel in 12 channel scenarios, where TL represents transfer learning. It can be seen that the introduction of transfer learning can significantly improve the time-frequency domain accuracy of digital twin channel modeling and is expected to effectively reduce the online training complexity and sample requirements in practical system applications.

[0036] In this embodiment, the process of using the trained generative adversarial network to generate digital twin channel data in step S4 is as follows: Use the trained generator to generate channel data with the same distribution as the channel data of the target task. .

[0037] Obviously, the above embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A digital twin channel modeling method based on generative adversarial network and transfer learning, characterized in that, Including the following steps: S1. Construct a semi-physical digital twin channel simulation platform; S2. Use the semi-physical digital twin channel simulation platform to generate a baseband signal and process the baseband signal to obtain channel frequency response data; S3. Use the channel frequency response data to train a generative adversarial network, and use a transfer learning strategy to accelerate the training process of the generative adversarial network to obtain a trained generative adversarial network; S4. Use the trained generative adversarial network to generate channel data.

2. The digital twin channel modeling method based on generative adversarial network and transfer learning according to claim 1, wherein The semi-physical digital twin channel simulation platform includes: a transmitter baseband signal processing device, a receiver baseband signal processing device, a channel simulator, a transmitter USRP device, and a receiver USRP device; the transmitter baseband signal processing device, the transmitter USRP device, the channel simulator, the receiver USRP device, and the receiver baseband signal processing device are sequentially connected by connecting lines; the channel simulator twins and simulates the channel effect of the real wireless communication environment on the communication signal.

3. The digital twin channel modeling method based on generative adversarial network and transfer learning according to claim 2, characterized in that, The process of using the semi-physical digital twin channel simulation platform to generate a baseband signal and process the baseband signal to obtain channel frequency response data in step S2 is as follows: Use the transmitter baseband signal processing device to generate a baseband signal and transmit the signal to the transmitter USRP device; Use the transmitter USRP device to modulate the frequency of the baseband signal to the carrier frequency to obtain a modulated signal, and transmit the modulated signal to the channel simulator; Use the channel simulator to simulate the actual wireless communication environment and apply a channel effect to the modulated signal to obtain a modulated signal with a channel effect; Use the receiver USRP device to perform down-conversion, filtering, down-sampling, and analog-to-digital conversion operations on the modulated signal with a channel effect to convert the modulated signal into a baseband signal; Use the receiver USRP device to process and analyze the baseband signal and extract the frequency response characteristics of the baseband signal to obtain channel frequency response data.

4. A digital twin channel modeling method based on generative adversarial network and transfer learning according to claim 3, characterized in that, Before using the channel frequency response data to train the generative adversarial network, it also includes performing time-domain conversion and feature processing on the channel frequency response data to obtain training data for the generative adversarial network. The process is as follows: Perform an inverse discrete Fourier transform on the channel frequency response data to obtain channel impulse response data; Crop and normalize the channel impulse response data to obtain training data for the generative adversarial network.

5. A digital twin channel modeling method based on a generative adversarial network and transfer learning according to claim 4, characterized in that, The generative adversarial network includes a generator and a discriminator. Among them, the generator is used to generate generated data with the same distribution as the training data and input it into the discriminator; the discriminator is used to discriminate whether the data input to itself is channel data generated by the generator or real channel data obtained through actual sampling.

6. A digital twin channel modeling method based on a generative adversarial network and transfer learning according to claim 5, characterized in that, The generator includes a fully connected layer, a batch normalization layer, a recombination operation layer, six upsampling modules, and a cropping operation layer connected in sequence; The fully connected layer expands the dimension of the input training data and inputs it into the batch normalization layer; The batch normalization layer normalizes the training data with the expanded dimension and inputs it into the recombination operation layer; The recombination operation layer recombines the normalized data; The upsampling module amplifies and extracts the high-level semantic features of the recombined data to generate channel impulse response data and inputs it into the cropping operation layer; The cropping operation layer crops the channel impulse response data generated by the upsampling module to obtain channel impulse response data of a fixed size.

7. A digital twin channel modeling method based on generative adversarial network and transfer learning according to claim 6, characterized in that, The upsampling module includes an upsampling layer, a convolutional layer, a batch normalization layer, and six activation function layers connected in sequence; among them, the upsampling layer amplifies the high-level semantic features of the data, the convolutional layer extracts the high-level semantic features of the data, the batch normalization layer normalizes the high-level semantic features extracted by the upsampling layer, and the activation function layers of the first to fifth upsampling modules use the LeakyReLU function to activate the normalized high-level semantic features into channel impulse response data. The activation function layer of the sixth upsampling module uses the tanh function to compress the amplitude of the channel impulse response data to the interval of [-1, 1].

8. A digital twin channel modeling method based on a generative adversarial network and transfer learning according to claim 7, characterized in that The discriminator includes a zero-padding operation layer, seven downsampling modules, a flattening operation layer, and a fully connected layer connected in sequence; The zero-padding operation layer expands the dimension of the input channel data and outputs it to the downsampling module; The downsampling module extracts the core semantic features of the channel data and outputs them to the flattening operation layer; The flattening operation layer flattens the core semantic features of the extracted channel data into the data dimension of the fully connected layer and outputs it to the fully connected layer; The fully connected layer synthesizes the flattened feature vectors and outputs the discrimination result.

9. A digital twin channel modeling method based on a generative adversarial network and transfer learning according to claim 8, characterized in that, In step S3, the channel frequency response data is used to train the generative adversarial network, and the transfer learning strategy is used to accelerate the training process of the generative adversarial network to obtain a trained generative adversarial network. The process is as follows: Randomly initialize the network parameters of the generator and discriminator of the generative adversarial network and ; Utilize channel frequency response data Train a generative adversarial network to obtain the network parameters of the preliminarily trained generator and discriminator , ; The network parameters of the initially trained generator and discriminator , are used as initial values, and the channel data of the target task is used to fine-tune the network parameters , to generate network parameters that meet the target task . .

10. A digital twin channel modeling method based on a generative adversarial network and transfer learning according to claim 9, characterized in that In step S4, the trained generative adversarial network is used to generate digital twin channel data. The process is as follows: Generate channel data with the same distribution as the channel data of the target task using the trained generator .

Images