Channel knowledge map generation method based on de-noising diffusion probability model

Abstract

The invention discloses a channel knowledge map generation method based on a denoising diffusion probability model, and the method comprises the steps: generating high-precision channel data through the mapping relation between environment coordinate features and CSI data through DDPM. Compared with the existing method, the method has the advantages that the measurement overhead is reduced, the accuracy and adaptability of data generation are improved, and an efficient and low-cost solution is provided for generation of the channel knowledge map in a wireless communication system.

Description

Technical Field

[0001] The present invention relates to a technology in the field of wireless communication, specifically a method for generating a channel knowledge map based on a denoising diffusion probability model. Background Art

[0002] A channel knowledge map is a key tool in a wireless communication system, used to establish a channel model and analyze the channel characteristics in the communication system to improve the robustness of the communication system and the reliability of data transmission. However, since the channel characteristics are affected by factors such as the environment and frequency, they need to be continuously updated to adapt to the transmission conditions of different scenarios, and high-precision depends on a large amount of measurement data and complex computational processing. Existing methods require high-cost measurement equipment and computational resources and are difficult to promote on a large scale. Therefore, there are many challenges in constructing a high-precision channel knowledge map. Summary of the Invention

[0003] Aiming at the deficiencies of the existing channel knowledge map construction technology, such as low accuracy of the generated results and high computational complexity, the present invention proposes a method for generating a channel knowledge map based on a denoising diffusion probability model. Through the mapping relationship between environmental coordinate features and CSI data, high-precision channel data is generated using DDPM. Compared with the existing methods, the present invention reduces the measurement overhead, improves the accuracy and adaptability of data generation, and provides an efficient and low-cost solution for generating a channel knowledge map in a wireless communication system.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to a method for generating a channel knowledge map based on a denoising diffusion probability model, including:

[0006] Step 1: Obtain the CSI data of the communication system through channel measurement in reality or simulation based on a ray tracing algorithm, and perform preprocessing;

[0007] Step 2: Collect the point cloud data containing the corresponding coordinate information of the propagation environment corresponding to the CSI data in Step 1, and perform preprocessing;

[0008] Step 3: While extracting CSI features from the preprocessed CSI data obtained in Step 1 as the feature labels of the point cloud data using a Vision Transformer (ViT), extract the point cloud data features from the preprocessed point cloud data obtained in Step 2 as the CSI feature labels using a PointNet network; jointly input the obtained CSI features and point cloud data features into a Contrastive Pre-training of Channel State Information and Point Cloud Data (CP-CLIP) model for contrastive learning training to optimize the mapping relationship between CSI and environmental features;

[0009] Step 4: Convert the CSI data obtained in Step 1 from the original space to the latent space through a Variational Autoencoder (VAE) and train it based on a Diffusion Probability Model (DDPM) network;

[0010] Step 5: Through the DDPM obtained in Step 4, perform online inference based on the coordinates of the input target location and the point cloud data. After generating the CSI data of the target location, perform clustering processing and then construct a channel knowledge map.

[0011] The present invention relates to a system for implementing the above method, including: a CP-CLIP unit, a VAE unit, and a DDPM unit, where: the CP-CLIP unit processes the collected CSI information and the point cloud data of the corresponding propagation environment and inputs them into the CP-CLIP model to obtain the mapping relationship between CSI and point cloud data features; the VAE unit converts the CSI data to the latent space through a VAE based on the original CSI data to obtain the latent representation of CSI; the DDPM unit generates the CSI of the target location based on the CSI in the latent space through a DDPM based on the environmental coordinate features. Technical Effects

[0012] The present invention combines environmental information with deep learning technology in a wireless communication system by jointly learning the features of CSI data and point cloud environmental information through a CP-CLIP model, mapping the CSI data to the latent space through a VAE, and generating CSI based on environmental coordinate features through a DDPM to construct a channel knowledge map. Compared with the prior art, the present invention generates a more accurate channel knowledge map. Brief Description of the Drawings

[0013] Figure 1 It is a flowchart of the present invention;

[0014] Figure 2 It is a schematic diagram of the CP-CLIP model structure and training process;

[0015] Figure 3 It is the training flow chart of DDPM under the influence of the cross-attention mechanism;

[0016] Figure 4 It is the performance result graph based on measured data. Specific implementation manners

[0017] As Figure 1 shown, this embodiment relates to a method for generating a channel knowledge map based on a denoising diffusion probability model, including:

[0018] Step 1: Obtain CSI data of a communication system through channel measurement in reality or simulation based on a ray tracing algorithm, and perform preprocessing.

[0019] The specific communication system is an OFDM communication system using N t transmitting antennas and N r receiving antennas, which is simulated in a scattering environment composed of multiple clusters. The frequency response of this system is: where: each cluster |ξ| in the scattering environment has P sub-paths, i and j are time indices and sub-carrier indices, Δt and Δf are the symbol duration and sub-carrier interval respectively, and τ n,p respectively represent the complex amplitude, departure elevation angle, departure azimuth angle, arrival azimuth angle, Doppler frequency shift and propagation delay of the p-th path in the n-th cluster, is the imaginary unit, and are the response vectors of the transmitting array and the receiving array.

[0020] The data preprocessing refers to: after transforming the CSI data into the spatio-frequency domain, normalizing its real part value and imaginary part value to the interval [-1, 1] respectively to improve the stability of data processing and model training.

[0021] Step 2: Collect point cloud data including coordinate information of the propagation environment corresponding to the CSI data in Step 1, and perform preprocessing.

[0022] The point cloud data includes: the point cloud of the building environment and the point cloud representing the coordinates of the transceiver. Specifically: regarding the coordinates of the receiving end and the transmitting end as a sphere, the dimension of the point cloud is n×m×6, where: n is the number of digits of the corresponding coordinate points, m is the environment and the point cloud of the coordinate points, and the first 3 dimensions of the dimension 6 are the environmental coordinate information [X, Y, Z], and the last 3 dimensions are the environmental color information [R, G, B]. The color information is used to distinguish different types of environmental scatterers.

[0023] The preprocessing described above includes: first, performing normalization to ensure the consistency of data input, and then using downsampling and resampling methods on the environmental point cloud to reduce the computational amount while maintaining the integrity of the environmental structure information and reducing the dependence of the model on specific samples.

[0024] Step 3: While using a Vision Transformer (ViT) to extract CSI features from the preprocessed CSI data obtained in Step 1 as the feature labels of the point cloud data, use a PointNet network to extract point cloud data features from the preprocessed point cloud data obtained in Step 2 as the CSI feature labels; input the obtained CSI features and point cloud data features into the CP-CLIP model for contrastive learning training to optimize the mapping relationship between CSI and environmental features.

[0025] As Figure 2 shown, the CP-CLIP model described above includes: a point cloud encoder and a CSI encoder, where: the point cloud encoder uses a PointNet model to obtain point cloud data features, and the CSI encoder uses a ViT to obtain CSI features.

[0026] The contrastive learning training mentioned above means that for a training batch containing N×N paired samples, each CSI feature will be paired with the corresponding point cloud feature after being mapped to the same feature space E b and form an N×N similarity matrix based on the cosine similarity between the CSI feature and the point cloud feature. In this matrix, the diagonal elements correspond to N positive samples, that is, the real CSI-point cloud feature pairs, while the non-diagonal elements are N 2 -N negative samples. Then, calculate the cross-entropy loss in the row and column directions and take the average of the two as the final loss function. By optimizing this loss function, the mapping relationship between the CSI feature and the point cloud feature can be made more accurate, thereby improving the modeling ability of the CP-CLIP model for channel characteristics.

[0027] The cosine similarity mentioned above where: the features extracted by the CSI encoder and the point cloud encoder are C n1 and T n2 respectively, and n1, n2 ∈ [1, N], that is, after L2 normalization processing.

[0028] The loss function mentioned above where:

[0029] Step 4: As Figure 3As shown, after the CSI data is transformed from the original space to the latent space through a pre-trained variational autoencoder (VAE) including an encoder and a generator, a compressed representation is achieved, which is used to train a diffusion probabilistic model (DDPM) network based on the U-Net architecture to reduce the generation complexity. Specifically, it includes:

[0030] 4.1 Construct the forward diffusion process: Starting from the input sample x0 ∼ q(x) extracted from a specific data distribution, Gaussian noise is incrementally added to the sample over T time steps. Specifically: Among them: x1, x2, …, x T is the sample sequence, q represents the probability distribution in the diffusion process, represents the Gaussian distribution, I is the identity matrix, is the noise schedule parameter sequence, satisfying 0 < β1 < β2 < … < β T < 1, and the cumulative parameter and α t = 1 - β t , and a noise sample sequence is generated by recursively adding Gaussian noise:

[0031] 4.2 Construct the DDPM ∈ θ (x t , t) to learn the probability model p θ (x t-1 |x t ) parameterized by θ. By sampling from to generate data samples from q(x) for parameterizing the reverse distribution, specifically: Among them: the mean vector ∈ θ is the predicted noise obtained by the DDPM ∈ θ (x t , t) according to the noise sample x t and the time step t. The covariance matrix

[0032] The described DDPM ∈ θ (x t , t) adopts the loss function to minimize the difference between the target mean vector and the neural network predicted value μ θ (x t , t) for training. Specifically: Among them: t ∼ Unif[T] means selecting a time step t from 1 to T, represents the square of the L2 norm, ∈0 represents the initial noise, ∈t Denote the true noise at time step \(t\).

[0033] 4.3 Map the channel \(x\) in the empty frequency domain to a latent \(z = \epsilon(x)\) through the encoder \(\epsilon\) of the pre-trained variational autoencoder, and then reconstruct the original channel from this latent representation through the decoder \(D\) of the variational autoencoder where: N s is the number of subcarriers, and \(N\) t is the number of transmit antennas. and are the downsampled numbers of the number of subcarriers in the latent space respectively.

[0034] 4.4 Combine the reconstructed original channel obtained in step 4.3 and the downsampling factor to form an effective low-dimensional latent space, where: This latent space is more suitable for probability-based generative models because it focuses more on the key feature information of the data and is trained in a computationally more efficient low-dimensional space.

[0035] 4.5 Construct a DDPM based on the U-Net architecture as Figure 3 shown and train it. Specifically, the loss function where: \(z\) t represents the noise latent variable at time step \(t\), and \(\tau\) is the intermediate representation θ (c).

[0036] The described DDPM is implemented by adopting, but not limited to, the techniques recorded by Rombach et al. in "High-Resolution Image Synthesis with Latent Diffusion Models" (CVPR 2022, Stable Diffusion).

[0037] Step Five: Through the DDPM obtained in Step Four, perform online inference based on the coordinates of the input target position and the point cloud data to generate the CSI data of the target position, thereby constructing a channel knowledge map.

[0038] Preferably, in order to eliminate the unnecessary variability of the model, cluster the CSI data output by the DDPM obtained in step four according to the occurrence frequency of the main path position as Figure 4 shown, and after identifying and selecting the most frequent clustering result, average the clustering result in the angle-delay domain as the final output. Specifically: where: \(i\) d is the index set matching the main path position, and \(l\) e is the number of samples matching the main path position. is the finally averaged CSI data, is the M generated by DDPM c CSI data, N s is the number of subcarriers, N t is the number of transmit antennas, and j is a variable used to traverse the data index.

[0039] After specific actual experiments, 1,606 groups of CSI data containing line-of-sight and non-line-of-sight scenarios were actually collected in an empty classroom on the Taicang Campus of Northwestern Polytechnical University, and the point cloud data of this environment was also collected. 1,580 randomly selected groups of CSI and environmental coordinate feature pairs were used for network training, and the remaining 26 groups of data were used as the test set to verify the model generation effect. After summarizing the results through the clustering algorithm, the cosine similarity in the angle domain is 0.8201 and the cosine similarity in the time-delay domain is 0.765411 in the line-of-sight scenario; the cosine similarity in the angle domain is 0.6407 and the cosine similarity in the time-delay domain is 0.6008 in the non-line-of-sight scenario. The results of the probability density function (CDF) graph of the cosine similarity are shown in the figure. The calculation result of the cosine similarity represents the comparison between the result generated by the model and the actual result, and the closer it is to 1, the better the performance.

[0040] Compared with the prior art, the present invention greatly improves the accuracy in generating CSI at the target location and thus constructing the channel knowledge map, providing a framework with great application potential for the generation of the channel knowledge map.

[0041] The above specific implementation can be locally adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation, and all implementation solutions within its scope are subject to the constraints of the present invention.

Claims

1. A method for generating a channel knowledge map based on a denoising diffusion probability model, characterized in that Including: Step 1: Obtain the CSI data of the communication system through actual channel measurement or simulation based on the ray tracing algorithm, and perform preprocessing; Step 2: Collect the point cloud data containing the corresponding coordinate information of the propagation environment corresponding to the CSI data in Step 1, and perform preprocessing; Step 3: While using the Vision Transformer to extract CSI features from the preprocessed CSI data obtained in Step 1 as the feature labels of the point cloud data, use the PointNet network to extract the point cloud data features from the preprocessed point cloud data obtained in Step 2 as the CSI feature labels; Input the obtained CSI features and point cloud data features into the contrastive learning pre-training model based on the comparison of channel state information and point cloud data for contrastive learning training to optimize the mapping relationship between CSI and environmental features; Step 4: Convert the CSI data obtained in Step 1 from the original space to the latent space through a variational autoencoder, and train based on the diffusion probability model network; Step 5: Through the DDPM obtained in Step 4, perform online inference based on the coordinates of the input target position and the point cloud data, generate the CSI data of the target position, and then construct a channel knowledge map through clustering processing.

2. The method for generating a channel knowledge map based on a denoising diffusion probability model according to claim 1, wherein The communication system described in Step 1 uses N t transmitting antennas and N r receiving antennas in an OFDM communication system, and is simulated in a scattering environment composed of multiple clusters. The frequency response of the system is: where each cluster |ξ| in the scattering environment has P sub-paths, i and j are the time index and sub-carrier index respectively, Δt and Δf are the symbol duration and sub-carrier interval respectively, a n,p , f n,p}3]]and τ n,p represent the complex amplitude, departure elevation angle, departure azimuth angle, arrival azimuth angle, Doppler frequency shift and propagation delay of the p-th path in the n-th cluster respectively, is the imaginary unit, and are the response vectors of the transmitting array and the receiving array respectively.

3. The method for generating a channel knowledge map based on a denoising diffusion probability model according to claim 1, wherein The point cloud data described in Step 2 includes: the point cloud of the building environment and the point cloud representing the coordinates of the transceiver. Specifically: regarding the coordinates of the receiver and the transmitter as a sphere, the dimension of the point cloud is n×m×6, where: n is the number of digits of the corresponding coordinate points, m is the environmental and coordinate point cloud, and the first 3 dimensions of the dimension 6 are the environmental coordinate information [X, Y, Z], and the last 3 dimensions are the environmental color information [R, G, B]. The color information is used to distinguish different types of environmental scatterers.

4. The method for generating a channel knowledge map based on a denoising diffusion probability model according to claim 1, characterized in that, The CP-CLIP model described includes: a point cloud encoder and a CSI encoder, where: the point cloud encoder uses the PointNet model to obtain the point cloud data features, and the CSI encoder uses ViT to obtain the CSI features; The described contrastive learning training means that for a training batch containing N×N paired samples, each CSI feature will be paired with the corresponding point cloud feature after being mapped to the same feature space E b and then paired. Based on the cosine similarity between the CSI feature and the point cloud feature, an N×N similarity matrix is formed. In this matrix, the diagonal elements correspond to N positive samples, that is, the true CSI-point cloud feature pairs, and the non-diagonal elements are N 2 -N negative samples. Then, the cross-entropy loss is calculated in the row and column directions, and the mean of the two is taken as the final loss function. By optimizing this loss function, the mapping relationship between the CSI feature and the point cloud feature is made more accurate, thereby improving the CP-CLIP model's ability to model channel characteristics; The described loss function Where:

5. The method for generating a channel knowledge map based on a denoising diffusion probability model according to claim 1, characterized in that, In Step 4, after converting the CSI data from the original space to the latent space through a pre-trained variational autoencoder (VAE) including an encoder and a generator, compression representation is achieved, which is used to train the diffusion probability model (DDPM) network based on the U-Net architecture to reduce the generation complexity.

6. The method for generating a channel knowledge map based on a denoising diffusion probability model according to claim 1 or 5, characterized in that Step 4 specifically includes: 4.1 Constructing the forward diffusion process: Starting from the input sample x0 ∼ q(x) extracted from a specific data distribution, Gaussian noise is incrementally added to the sample over T time steps, specifically as follows: where: x1, x2, …, x T is the sample sequence, q represents the probability distribution in the diffusion process, represents the Gaussian distribution, I is the identity matrix, is the sequence of noise scheduling parameters, satisfying 0 < β1 < β2 < … < β T < 1, and the cumulative parameter and α t = 1 - β t , and a sequence of noise samples is generated by recursively adding Gaussian noise: 4.2 Construct DDPM∈ Based on U-Net Architecture θ (x t , t) learns the probability model p θ (x t-1 ∣x t ) parameterized by θ, and parameterizes the reverse distribution by sampling from q(x) to generate data samples, specifically: where: the mean vector ∈ θ is the predicted noise obtained by DDPM∈ θ (x t , t) according to the noise sample x t and the time step t, and the covariance matrix 4.3 The encoder ε of the pre-trained variational autoencoder maps the channel x in the empty frequency domain to a latent variable z = ε(x), and then the decoder D of the variational autoencoder reconstructs the original channel from this latent representation. Where: N s is the number of subcarriers, N t is the number of transmit antennas, and are the downsampled numbers of the subcarriers in the latent space respectively; 4.4 Combine the reconstructed original channel and the downsampling factor obtained in step 4.3 to form an effective low-dimensional latent space, where: This latent space is more suitable for probability-based generative models because it focuses more on the key feature information of the data and is trained in a computationally more efficient low-dimensional space; 4.5 Construct the DDPM based on the U-Net architecture and train it, specifically: the loss function where: z t represents the noise latent variable at time step t, and the middle represents τ θ (c).

7. The method for generating a channel knowledge map based on a denoising diffusion probability model according to claim 6, characterized in that, The described DDPM belongs to θ (x t , t) uses a loss function to minimize the difference between the target mean vector and the neural network predicted value μ θ (x t , t), specifically: where: t ∼ Unif[T] means selecting a time step t from 1 to T, represents the square of the L2 norm, ∈0 represents the initial noise, ∈ t represents the true noise at time step t.

8. The method for generating a channel knowledge map based on a denoising diffusion probability model according to claim 1, wherein For the CSI data output by the obtained DDPM, cluster them according to the occurrence frequency of the main path position. After identifying and selecting the most frequent clustering result, average the clustering result in the angle-delay domain as the final output, specifically: where: i d is the index set matching the main path position, l e is the number of samples matching the main path position, is the CSI data after the final averaging, is the M c CSI data generated by DDPM, N s is the number of subcarriers, N t is the number of transmit antennas, and j is a variable used to traverse the data index.

9. A denoising diffusion probability model-based channel knowledge map generation system for implementing the method according to any one of claims 1-8, comprising: A CP-CLIP unit, a VAE unit, and a DDPM unit, where: the CP-CLIP unit, according to the collected CSI information and the point cloud data of the corresponding propagation environment, inputs through data processing to the CP-CLIP model to obtain the relationship between the CSI and the point cloud data feature mapping. The VAE unit, according to the original CSI data, converts the CSI data to the latent space through the VAE to obtain the latent representation of the CSI. The DDPM unit, according to the CSI in the latent space, generates the CSI of the target position through the DDPM based on the environmental coordinate features.

Images