Wireless communication receiver data enhancement method based on diffusion model

Abstract

The invention discloses a wireless communication receiver data enhancement method based on a diffusion model, and belongs to the technical field of wireless communication. The method comprises the following steps: S1, collecting channel measurement data and corresponding user positions and speeds; s2, training a diffusion model by using channel measurement data under the conditions of user position and speed; s3, generating synthetic channel data by using a diffusion model according to the user position and speed of the scene to be enhanced; s4, combining the channel measurement data with the synthesized channel data to obtain an enhanced channel data set; and S5, generating a receiver training data set based on the enhanced channel data set, and training a receiver neural network model. According to the invention, the scene-adaptive synthetic channel data is generated based on the diffusion model, the demand for channel measurement data is reduced, the adaptive capability of a receiver model to different scenes is improved, and the problems of high data acquisition cost and poor scene adaptability in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to a method for enhancing wireless communication receiver data based on a diffusion model, belonging to the technical field of wireless communication. Background Art

[0002] With the evolution and development of wireless communication systems, the important technology of Multiple-Input Multiple-Output (MIMO) has gradually moved towards deploying ultra-large-scale antenna arrays to achieve higher spectral and energy efficiency, and wider and more flexible network coverage. However, the proliferation of antenna elements poses severe challenges to the design of multi-antenna receivers (such as channel state information acquisition and data detection).

[0003] In recent years, the introduction of Artificial Intelligence (AI) and machine learning has provided new ideas for overcoming the bottleneck in receiver design. However, AI-enabled receiver schemes usually rely on a large amount of training data sets, especially wireless channel data closely related to wireless communication. However, the acquisition of channel data consumes a large amount of time, manpower, and material costs, especially for the high-dimensional channel information corresponding to ultra-large-scale MIMO systems. In addition, the deployment environment of wireless communication systems is often complex, diverse, and dynamically changing. The generalization of receivers based on AI models in different scenarios is a major problem that has not been solved. Especially considering that the amount of training data matching the deployment scenario is often very limited, traditional data enhancement methods based on data transformation are difficult to ensure the diversity and scenario matching of data, while existing data enhancement methods based on generative AI often have problems such as insufficient quality of generated data, high training difficulty, and difficulty in covering various data distributions in different scenarios. Summary of the Invention

[0004] Technical Problem: In view of the above deficiencies in the prior art, the present invention provides a method for enhancing wireless communication receiver data based on a diffusion model, which uses a diffusion model to generate scene-adapted synthetic channel data, reduces the demand for channel measurement data, and improves the adaptability of the receiver model to different scenarios.

[0005] Technical Solution: A method for enhancing wireless communication receiver data based on a diffusion model according to the present invention includes the following steps:

[0006] S1. Collect channel measurement data and the corresponding user location and speed;

[0007] S2. Use the channel measurement data to train a diffusion model with the user location and speed as conditions;

[0008] S3. Generate synthetic channel data using the trained diffusion model according to the user location and speed of the scenario for which the receiver is to be enhanced with data;

[0009] S4. Combine the channel measurement data with the synthesized channel data to obtain an enhanced channel data set;

[0010] S5. Generate a receiver training data set based on the enhanced channel data set and train a receiver neural network model.

[0011] Wherein,

[0012] The diffusion model includes a conditional processing module and a diffusion generation module; wherein, the conditional processing module includes a sine position encoder and a plurality of fully connected neuron layers; the diffusion generation module includes two convolutional neural networks, wherein, the first convolutional neural network is used for processing angle delay domain information, and the second convolutional neural network is used for processing time domain information.

[0013] The step S2 specifically includes:

[0014] S201. Set the total number of diffusion time steps T of the diffusion model, the noise intensity corresponding to different time step numbers, set the training iteration number variable i = 1 and the maximum training iteration number I;

[0015] S202. Randomly extract several pieces of data from the channel measurement data. For each piece of data, randomly select a time step number t (t ∈ {1, 2,..., T}), generate a Gaussian white noise variable with the intensity corresponding to the time step number t, and add it to the piece of channel measurement data to obtain noisy channel data;

[0016] S203. For each piece of noisy channel data, use the conditional processing module of the diffusion model to encode its corresponding time step number, user position and speed into an embedding vector, input the noisy channel data and the embedding vector into the diffusion generation module of the diffusion model, and output a fitted noise variable;

[0017] S204. Calculate the cost function value between the fitted noise variable and the Gaussian white noise variable, calculate the gradient of the cost function value and backpropagate it to optimize the parameters of the diffusion model; set the training iteration number variable i = i + 1;

[0018] S205. Determine whether the training iteration number variable i is greater than the maximum training iteration number I. If not, jump to step S202; if so, end step S2.

[0019] The step S3 specifically includes:

[0020] S301. Collect several user position and speed labels from the data enhancement scenarios to be applied to the receiver. The number of labels is denoted as N. Use the conditional processing module of the trained diffusion model to encode each position and speed label into an embedding vector, and input the embedding vector into the diffusion generation module of the trained diffusion model;

[0021] S302. Randomly generate N Gaussian noise samples from the standard normal distribution, and use the diffusion generation module of the trained diffusion model to process the Gaussian noise samples T times to obtain N synthetic channel data, where T is the total number of diffusion time steps. The label embedding vectors input in the T - time processing are obtained from step S301, and the time - step sequence numbers corresponding to the input time - step embedding vectors decrease from T to 1.

[0022] Step S5 generates a pilot signal and a data signal. Based on the enhanced channel data set obtained in step S4, a received signal is acquired. The pilot signal, the data signal, the enhanced channel data set, and the received signal together constitute a training data set for training a receiver neural network, where the pilot signal and the received signal serve as input features of the receiver neural network, and the enhanced channel data set and the data signal serve as labels.

[0023] Advantageous effects: Compared with the prior art, the beneficial effects of the technical solution provided by the present invention are as follows:

[0024] The present invention uses a limited amount of channel measurement data to train a diffusion model. With the powerful distribution coverage ability and high - quality sample generation ability of the diffusion model, synthetic channel data adapted to the scenario is generated according to the user location and speed labels, expanding the limited amount of channel measurement data and reducing the demand for channel measurement data. Compared with the existing data enhancement technologies, the distribution of the samples generated by the present invention is closer to the real - channel data distribution of different scenarios, thereby improving the adaptation ability of the receiver model to different scenarios. Description of the Drawings

[0025] Figure 1 It is a flowchart of a wireless communication receiver data enhancement method based on a diffusion model according to an embodiment of the present invention;

[0026] Figure 2 It is a network structure diagram of a diffusion model according to an embodiment of the present invention;

[0027] Figure 3 It is a comparison diagram of the bit - error rate performance between the method of the present invention and a comparative method. Detailed Embodiments

[0028] The present invention will be further described below with reference to the drawings and embodiments.

[0029] I. System Model Adopted in This Embodiment

[0030] Consider the uplink transmission of a large - scale MIMO system. The user acts as a transmitter equipped with N t antennas, and the base station acts as a receiver equipped with N rRoot antenna. Orthogonal Frequency Division Multiplexing (OFDM) transmission is adopted, and the number of subcarriers is N f . Both the transmitter and the receiver use uniform linear arrays. The channel response matrix on the k-th subcarrier can be expressed as:

[0031]

[0032] where L is the number of propagation paths, i is the path sequence number, γ i,k is the complex gain, which is related to the channel bandwidth and subcarrier frequency, a r (·) and a t (·) are the array response vectors at the receiving end and the transmitting end respectively, (·) H denotes the conjugate transpose of the matrix, and are the azimuth of arrival and the azimuth of departure respectively. The variables L, γ i,k , and are all conditional random variables of the user position coordinate vector x, such that the distribution of the channel response matrix H k is determined by the user position coordinate vector x

[0033] Consider the spatio-frequency domain channel matrix f with dimensions M×N where M = N r N t , the channel vector h k (k = 1, 2,..., N f ) is obtained by vectorizing the channel response matrix H k . Perform a two-dimensional discrete Fourier transform on the spatio-frequency domain channel matrix H′, and retain the first D non-zero columns in the time-delay domain to obtain an angular time-delay domain channel matrix with dimensions M×D. Further, considering the change of the channel in the time domain, the degree of change is determined by the user speed v. The number of transmission symbols in one transmission time slot is denoted as τ, then the channel data in one transmission time slot can be expressed as a channel tensor H with dimensions M×D×τ, and its distribution is jointly determined by the user position coordinate vector x and the user speed v

[0034] II. Specific steps of this embodiment

[0035] As Figure 1 shown, the embodiment of the present invention provides a flowchart of a method for enhancing wireless communication receiver data based on a diffusion model, including the following steps:

[0036] S1. Collect channel measurement data and the corresponding user positions and speeds;

[0037] S2. Using the channel measurement data, train a diffusion model based on the user location and speed as conditions;

[0038] S3. According to the user location and speed of the data enhancement scenario to be applied to the receiver, use the trained diffusion model to generate synthetic channel data;

[0039] S4. Combine the channel measurement data and the synthetic channel data to obtain an enhanced channel dataset;

[0040] S5. Generate a receiver training dataset based on the enhanced channel dataset and train a receiver neural network model.

[0041] Further, in this embodiment, each piece of channel measurement data collected in step S1 is recorded in the form of a channel tensor with dimensions M×D×τ, and each piece of user location and speed data is recorded in the form of p = [x T , v] T , where p represents the user location and speed label, and (·) T represents the transpose of a matrix.

[0042] The diffusion model is one of the state-of-the-art generative AI models. By establishing a Markov chain of diffusion steps, gradually introducing random noise into the data, and learning to reverse this forward process, it can generate high-quality data samples from the noise. By introducing conditional information in training, the model can be guided to generate data with different specified distributions, thereby enhancing the adaptability for downstream tasks.

[0043] Further, the diffusion model includes a conditional processing module and a diffusion generation module; among them, the conditional processing module includes a sine position encoder and multiple fully connected neuron layers; the diffusion generation module includes two convolutional neural networks, where the first convolutional neural network is used for angle delay domain information processing, and the second convolutional neural network is used for time domain information processing.

[0044] Further, Figure 2The network structure diagram of the diffusion model adopted in this example is shown. The two convolutional neural networks of the diffusion generation module adopt the U-Net structure, which includes several residual convolutional network units, several upsampling units, and several downsampling units; the input of the diffusion generation module is a real-valued first channel tensor with a dimension of [B, 2, M, D, τ], where B is the size of a small batch; after dimension reconstruction and dimension permutation, the first channel tensor obtains a second channel tensor with a dimension of [B×τ, 2, M, D], and after being processed by the first convolutional neural network, a third channel tensor with a dimension of [B×τ, 2, M, D] is obtained; after dimension reconstruction and dimension permutation, the third channel tensor obtains a fourth channel tensor with a dimension of [B, 2, M×D, τ], and after being processed by the second convolutional neural network, a fifth channel tensor with a dimension of [B, 2, M×D, τ] is obtained; after dimension reconstruction, the fifth channel tensor obtains an output tensor with a dimension of [B, 2, M, D, τ].

[0045] Further, step S2 specifically includes:

[0046] S201. Set the total number of diffusion time steps T of the diffusion model and the noise intensity corresponding to different time step numbers, set the training iteration number variable i = 1 and the maximum training iteration number I;

[0047] S202. Randomly extract several pieces of data from the channel measurement data. Further, in this embodiment, the number of data entries extracted is B; for each piece of data, randomly select a time step number t (t ∈ {1, 2,..., T}), generate a Gaussian white noise variable with the intensity corresponding to the time step number t, and add it to this piece of channel measurement data to obtain the noisy channel data; further, in this embodiment, the dimension of the Gaussian white noise variable is [B, 2, M, D, τ], and the noisy channel data is a real-valued channel tensor with a dimension of [B, 2, M, D, τ];

[0048] S203. For each piece of noisy channel data, use the conditional processing module of the diffusion model to encode its corresponding time step number, user location, and speed into an embedding vector, input the noisy channel data and the embedding vector into the diffusion generation module of the diffusion model, and output a fitted noise variable; further, the fitted noise variable is an output tensor with a dimension of [B, 2, M, D, τ].

[0049] S204. Calculate the cost function value between the fitted noise variable and the Gaussian white noise variable, calculate the gradient of this cost function value and backpropagate it to optimize the parameters of the diffusion model; set the training iteration number variable i = i + 1;

[0050] S205. Determine whether the training iteration number variable i is greater than the maximum training iteration number I. If not, jump to step S202; if so, end step S2.

[0051] Further, step S3 specifically includes:

[0052] S301: Collect a number of user position and velocity tags from the data enhancement scenario to be applied to the receiver. The number of tags is denoted as N. Use the conditional processing module of the trained diffusion model to encode each position and velocity tag into an embedding vector, and input the embedding vector into the diffusion generation module of the trained diffusion model;

[0053] S302: Randomly generate N Gaussian noise samples from the standard normal distribution. Use the diffusion generation module of the trained diffusion model to process the Gaussian noise samples T times to obtain N synthetic channel data, where T is the total number of diffusion time steps. The input tag embedding vectors in the T - time processing are obtained from step S301, and the time step sequence numbers corresponding to the input time - step embedding vectors decrease from T to 1.

[0054] Further, in this embodiment, step S302 is carried out in small batches. The dimensions of the Gaussian noise samples processed in each small batch and the generated synthetic channel data are both [B, 2, M, D, τ]. The diffusion model learns the channel distribution of a given scenario (characterized by user position and velocity) during training. During deployment, it can generate scenario - adapted channel data according to the user position and velocity tags of the current scenario, so as to perform targeted data enhancement on the receiver neural network model, significantly improve the performance of the receiver in the deployment scenario, relieve the burden of the AI model relying on a large amount of channel measurement data, and overcome the problem of insufficient channel measurement data for adapting to different scenarios.

[0055] Further, step S5 generates a transmit pilot signal and a data signal. Based on the enhanced channel data set obtained in step S4, a received signal is acquired. The pilot signal, data signal, enhanced channel data set, and received signal together constitute the training data set for training the receiver neural network, where the pilot signal and the received signal are used as the input features of the receiver neural network, and the enhanced channel data set and the data signal are used as labels.

[0056] III. Implementation Effects

[0057] To enable those skilled in the art to better understand the solution of the present invention, the following presents the performance comparison between the wireless communication receiver data enhancement method based on the diffusion model and the comparative method in this embodiment.

[0058] The considered scenario uses the 3GPP TR 38.901 channel model, the uplink carrier center frequency is 2.655 GHz, the channel bandwidth is 10 MHz, the total number of sub - carriers is N f = 512, the number of OFDM symbols in one transmission time slot is τ = 14, the delay spread length is D = 32, and the base station is equipped with Nr = 32 receiving antennas, with each user equipped with N t = 1 transmitting antenna, the user's moving speed is fixed at 72 km / h, the base station is located at the origin of coordinates, and users are evenly distributed in 5 regions as shown in Table 1. For specific configurations, please refer to the literature X. Li, J. Guo, C.-K. Wen, S. Jin, S. Han, and X. Wang, “Multi-task learning-based CSI feedback design in multiple scenarios,” IEEE Transactions on Communications, vol. 71, no. 12, pp. 7039–7055, Dec. 2023.

[0059] Table 1 Parameters of User Distribution Regions

[0060] Area Central position (m) Channel scenario Number of scatterer clusters 1 Commercial area (50，0) 3GPP-38.901.UMi-NLOS 40 2 Residential area (-100，-50) 3GPP-38.901.UMi-NLOS 40 3 Park (10，-70) 3GPP-38.901.UMi-LOS 5 4 Parking lot (90，-160) 3GPP-38.901.UMi-LOS 5 5 Warehouse (0，170) 3GPP-38.901.UMi-NLOS 10

[0061] The transmission adopts the Superimposed Pilots (SIP) format to improve spectral efficiency. The number of effective subcarriers is 72, the pilot-to-data power ratio is 0.3:0.7, the pilot symbols are modulated using Quadrature Phase Shift Keying (QPSK), the data symbols are modulated using Quadrature Amplitude Modulation (QAM), and the modulation order is 16. For a detailed introduction to SIP transmission and the neural network receiver model used, please refer to the literature H. Xiao et al, “Interference cancellation based neural receiver for superimposed pilot in multi-layer transmission,” China Commun., vol. 22, no. 1, pp. 75–88, Jan. 2025.

[0062] Figure 3Shows the comparison of the bit error rate performance of the receiver trained by the present invention and the comparative method. Among them, the diffusion model used in the present invention is trained based on 500 channel measurement data and the corresponding user location and speed labels obtained from 5 regions in Table 1. Then, 4500 synthetic channel data are generated according to the user location and speed labels additionally obtained from the 5 regions, and merged with the original 500 channel measurement data to obtain an enhanced channel data set and construct a receiver training data set to train the receiver neural network model. The performance has an improvement of more than 2 dB compared with the receiver neural network model directly trained using the original 500 channel measurement data, and approaches the ideal scheme trained using 5000 channel measurement data, avoiding the overhead of a large number of channel measurements. In addition, a data augmentation scheme based on noisy data transformation is also compared. This scheme expands the channel data set by adding Gaussian white noise (controlling the signal-to-noise ratio to 15 dB) to the original 500 channel measurement data. From Figure 3 It can be seen that the present invention can also achieve a performance gain of more than 1.5 dB compared with the noise addition scheme.

[0063] The above-described is only a preferred embodiment of the present invention illustrated in conjunction with the accompanying drawings, and cannot be used to limit the scope of rights covered by the present invention. It should be understood that any equivalent changes made without departing from the spirit of the present invention fall within the scope of protection covered by the claims of the present invention.

Claims

1. A wireless communication receiver data enhancement method based on a diffusion model, characterized in that: The method comprises the following steps: S1. Collect channel measurement data and corresponding user position and speed; S2, using channel measurement data to train a diffusion model based on user position and velocity; S3, generating synthetic channel data using the trained diffusion model according to the user position and speed of the scenario in which data enhancement is to be performed on the receiver; S4, merging the channel measurement data with the synthetic channel data to obtain an enhanced channel data set; S5. Generate a receiver training data set based on the enhanced channel data set to train a receiver neural network model.

2. The method for data enhancement of a wireless communication receiver based on a diffusion model according to claim 1, characterized in that: The diffusion model includes a conditional processing module and a diffusion generation module; wherein the conditional processing module includes a sinusoidal position encoder and a plurality of fully connected neuron layers; the diffusion generation module includes two convolutional neural networks, wherein the first convolutional neural network is used for angle delay domain information processing, and the second convolutional neural network is used for time domain information processing.

3. The method for data enhancement of a wireless communication receiver based on a diffusion model according to claim 1, characterized in that: The step S2 specifically includes: S201, setting the total diffusion time steps T of the diffusion model, the noise intensity corresponding to different time step numbers, setting the training iteration number variable i=1 and the maximum training iteration number I; S202, randomly extracting a number of data from the channel measurement data, and for each data, randomly selecting a time step number t, generating a Gaussian white noise variable with an intensity corresponding to the time step number t, and adding it to the channel measurement data to obtain noisy channel data; S203, for each strip of noisy channel data, use the conditional processing module of the diffusion model to encode the corresponding time step number, user position and speed into an embedded vector, input the noisy channel data and the embedded vector into the diffusion generation module of the diffusion model, and output the fitted noise variable; S204, calculating the cost function value between the fitting noise variable and the Gaussian white noise variable, calculating the gradient of the cost function value and back-propagating it, optimizing the parameters of the diffusion model; setting the training iteration number variable i=i+1; S205, determine whether the training iteration number variable i is greater than the maximum training iteration number I, if not, jump to step S202; if yes, end step S2.

4. The method for data enhancement of a wireless communication receiver based on a diffusion model according to claim 1, characterized in that: The step S3 specifically includes: S301, collecting a number of user positions and speed labels from a scenario where data enhancement is to be performed on a receiver, where the number of labels is recorded as N, using the conditional processing module of the trained diffusion model to encode each position and speed label into an embedding vector, and inputting the embedding vector into the diffusion generation module of the trained diffusion model; S302, randomly generate N Gaussian noise samples from a standard normal distribution, and use the diffusion generation module of the trained diffusion model to process the Gaussian noise samples T times to obtain N synthetic channel data, where T is the total number of diffusion time steps, the label embedding vector input in the T times of processing is obtained by step S301, and the time step sequence number corresponding to the input time step embedding vector decreases from T to 1.

5. The method for data enhancement of a wireless communication receiver based on a diffusion model according to claim 1, characterized in that: The step S5 generates a transmission pilot signal and a data signal, and obtains a received signal based on the enhanced channel data set obtained in the step S4. The pilot signal, the data signal, the enhanced channel data set and the received signal together constitute a training data set for training the receiver neural network, wherein the pilot signal and the received signal serve as input features of the receiver neural network, and the enhanced channel data set and the data signal serve as labels.

Images