Method for constructing channel map of large-scale MIMO system based on latent space diffusion model

Abstract

The application proposes a large-scale MIMO system channel graph construction method based on a latent space diffusion model (LDM). The method uses the spatial consistency of the large-scale MIMO system channel to learn the conditional probability distribution of the statistical channel state information (sCSI) about the user terminal position to construct the channel graph. An improved variational autoencoder (VAE) is used to learn the low-dimensional representation of high-dimensional sCSI, and the LDM is improved to improve the generation accuracy of the low-dimensional representation of sCSI, and then a channel graph containing position and sCSI low-dimensional representation data is constructed. The storage and transmission cost of the channel graph is greatly reduced compared with directly using high-dimensional sCSI, and the sCSI low-dimensional representation in the channel graph can be accurately restored to high-dimensional sCSI using VAE, and the generation process does not depend on the pilot signal, which greatly saves the pilot overhead while accurately generating sCSI.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology and relates to a method for constructing channel maps for large-scale MIMO systems based on the latent space diffusion model (LDM). Background Technology

[0002] Massive MIMO technology is a core enabling technology for fifth-generation (5G) mobile communication and will continue to evolve and play a crucial role in sixth-generation (6G) mobile communication systems. In terrestrial massive MIMO systems, base stations rely on instantaneous channel state information (iCSI) to perform user scheduling and time-frequency resource allocation, downlink precoding transmission, and uplink reception processing. However, the continuous increase in base station antenna size, the number of parallel data streams, and system bandwidth has significantly exacerbated the challenges of channel acquisition.

[0003] Existing systems typically employ orthogonal pilot sequences for channel estimation and prediction. However, in scenarios with a large number of active users, traditional orthogonal pilot channel detection methods are difficult to implement; furthermore, with the significant increase in the number of base station antenna elements, parallel data streams, and system bandwidth, the required pilot overhead becomes extremely large. Therefore, obtaining channel information increasingly relies on statistical channel state information (sCSI). In recent years, obtaining sCSI using channel maps has become an important technical approach. A channel map can be defined as a database of channel statistics (or channel fingerprints) based on location indexes within the base station coverage area, or a generative intelligent module. Using location information (or its equivalent data representation) and mobility information provided by the terminal, the base station can obtain the required channel statistics.

[0004] Channel maps can be constructed using methods such as offline measurement, online measurement, and online interpolation. Offline measurement utilizes dedicated channel measurement equipment to perform channel measurements and fingerprint extraction at pre-planned grid points within the base station coverage area. Online measurement utilizes online terminal equipment to achieve fingerprint extraction through uplink channel probing. Online interpolation refers to generating the channel fingerprint required for the target location based on the channel fingerprints of existing grid points using interpolation algorithms or generative intelligent modules. However, existing technologies directly store high-dimensional sCSIs when constructing channel maps, which leads to significant storage and transmission overhead. Furthermore, the accuracy of sCSI generation needs improvement. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method for constructing channel maps for large-scale MIMO systems based on LDM, which does not rely on pilot signals, can generate sCSI with higher accuracy for the target area, and can greatly save the storage and transmission costs of channel map data.

[0006] Technical Solution: To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for constructing channel maps for a large-scale MIMO system based on LDM includes the following steps:

[0008] Based on the configuration of a large-scale MIMO system, the space-frequency domain channel model is expressed as the product of the spatial domain sampling beam matrix, the dual-beam domain channel, and the time domain sampling beam matrix, and a dual-beam domain channel model is established; the beam domain channel energy matrix is ​​extracted as the user's sCSI;

[0009] The construction of the channel map is divided into offline training and online generation stages. In the offline training stage, the location of reference points in the target area and their corresponding sCSIs are collected to train the conditional LDM. In the online generation stage, the user location is input to generate the corresponding sCSI.

[0010] In the offline training phase, the training process of the conditional LDM includes: First, training a variational autoencoder (VAE) specific to sCSI to learn the low-dimensional representation of high-dimensional sCSI in the latent space; wherein, when training the VAE, an improved loss function is used, a hyperparameter is introduced as the coefficient of the reconstruction loss term, and the hyperparameter is increased during the training process; then, training the conditional diffusion model in the latent space to realize the transformation from the standard Gaussian distribution to the low-dimensional latent space conditional distribution with respect to location;

[0011] In the online generation phase, the geographic coordinates of the user terminal are obtained, the conditional LDM diffusion model is used, and the generation process is improved by discarding the sampling variance term in the LDM generation process to improve the accuracy of the generation results. A low-dimensional representation of sCSI under specific user location conditions is generated, thereby constructing a channel map containing the location and the low-dimensional representation of sCSI. When using the channel map, the low-dimensional representation is restored to sCSI through VAE.

[0012] Furthermore, the dual-beam domain energy matrix reflects the distribution of channel energy at different distinguishable angles and delays. That is, each element in the matrix is ​​related to the power, angle of arrival, and delay of arrival of each path of the channel between the base station and the user, and they are closely related to the user's location.

[0013] Furthermore, during the offline training phase, the locations of all reference points in the target area are collected, and the dual-beam domain channel energy matrix at the corresponding locations is extracted to form a dataset as training data. Using the training data, an LDM is trained, and the sCSI at a given location can be obtained. During the online generation phase, using the trained LDM and given the user location, the sCSI between the base station and the user and its low-dimensional representation can be obtained. Thus, the channel map can be constructed using online interpolation.

[0014] Furthermore, VAE is used to learn the low-dimensional representation of the high-dimensional dual-beam domain channel energy matrix and restore the low-dimensional representation to the high-dimensional dual-beam domain channel energy matrix data space; the diffusion model is used to learn the latent space distribution of the low-dimensional representation. By applying the conditional diffusion model, a low-dimensional representation of the dual-beam domain channel energy matrix at a given location is generated, and VAE is used to restore the low-dimensional representation to the high-dimensional dual-beam domain channel energy matrix data space.

[0015] Furthermore, the compression process of the high-dimensional dual-beam domain channel energy matrix is ​​as follows:

[0016] Step a) Initialize encoder parameters ψ and decoder parameters φ;

[0017] Step b) Sample M samples from the training set containing N sCSIs;

[0018] Step c) Perform steps d) through g) for each sample until all M samples have undergone the above steps;

[0019] Step d) Randomly sample from standard Gaussian noise;

[0020] Step e) Calculate the mean and variance terms of the latent space vector of the encoder output of the VAE;

[0021] Step f) Calculate the low-dimensional representation of sCSI in the latent space based on the mean and variance terms and the sampled Gaussian noise;

[0022] Step g) Input the low-dimensional representation of sCSI into the VAE decoder to calculate sCSI;

[0023] Step h) Calculate the improved VAE loss function, where the reconstruction loss term of the loss function increases with the number of training cycles;

[0024] Step i) Calculate and update the gradients of the encoder and decoder parameters in the VAE;

[0025] Step j) Calculate and update the encoder parameters ψ and decoder parameters φ using the updated gradient;

[0026] Step k) If the network parameters converge, output the encoder parameters ψ and decoder parameters φ; otherwise, return to step b).

[0027] Furthermore, the U-Net network and VAE in the trained LDM are used to construct the channel map and recover the high-dimensional sCSI data; the specific process is as follows:

[0028] Step a) Generate a blank channel map for a specific location region and load the existing locations and corresponding low-dimensional representations of sCSIs in the dataset into the channel map;

[0029] Step b) Load the decoder parameters φ and diffusion model parameters θ of the trained VAE;

[0030] Step c) Given a location vector to generate sCSI, randomly sample a matrix that conforms to the latent space dimension from a standard Gaussian distribution;

[0031] Step d) Based on the trained diffusion model, use the improved generation process to generate a low-dimensional representation of the sCSI at the current position;

[0032] Step e) Repeat steps c) and d) until the low-dimensional representations of sCSI at all locations in the channel map are obtained, and then interpolate them into the channel map to establish a complete channel map;

[0033] Step f) Use the VAE decoder to recover the low-dimensional representation of sCSI in the channel map into a high-dimensional sCSI.

[0034] Furthermore, when generating the low-dimensional representation of sCSI, skip sampling is used to accelerate the generation process, while discarding the sampling variance term in the accelerated generation process to improve the accuracy of the generation results.

[0035] Furthermore, the VAE encoder includes two 2D convolutional layers, one Flatten layer, and four fully connected layers, while the VAE decoder includes three fully connected layers, one Reshape layer, two 2D deconvolutional layers, and two 2D convolutional layers.

[0036] A computer system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the steps of the LDM-based large-scale MIMO system channel map construction method.

[0037] A computer program product includes a computer program that, when executed by a processor, implements the steps of the method for constructing a channel map for a large-scale MIMO system based on LDM.

[0038] Beneficial Effects: The proposed method for constructing channel maps for large-scale MIMO systems based on LDM in this invention uses an improved VAE to learn the low-dimensional representation of high-dimensional sCSI, while simultaneously improving LDM to enhance the generation accuracy of the low-dimensional sCSI representation, thereby constructing a channel map containing location and low-dimensional sCSI representation data. The storage and transmission costs of this channel map are significantly reduced compared to directly using high-dimensional sCSI. Furthermore, using VAE, the low-dimensional sCSI representation in the channel map can be accurately reconstructed into high-dimensional sCSI, and the generation process does not rely on pilot signals, greatly saving pilot overhead while accurately generating sCSI. This invention also proposes an improved sampling method to accelerate the generation process, significantly speeding up the sCSI generation process. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the channel map construction process for a large-scale MIMO system based on LDM, as described in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of a conditional LDM.

[0041] Figure 3 This is a comparison chart of the normalized mean square error (NMSE) performance of the present invention with other methods in an embodiment of the present invention.

[0042] Figure 4 This is a comparison chart of the generation time of the accelerated sampling method and the non-accelerated sampling method in this invention. Detailed Implementation

[0043] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0044] like Figure 1 As shown in the figure, the method for constructing a channel map of a large-scale MIMO system based on LDM disclosed in this invention includes the following steps:

[0045] Step 1: Based on the configuration of the large-scale MIMO system, the space-frequency domain channel model is expressed as the product of the spatial domain sampling beam matrix, the dual-beam domain channel, and the time domain sampling beam matrix to establish the dual-beam domain channel model; the beam domain channel energy matrix is ​​extracted as the user's sCSI.

[0046] Step 2: The construction of the channel map is divided into offline training and online generation stages. In the offline training stage, the location of reference points in the target area and their corresponding sCSIs are collected to train the conditional LDM. In the online generation stage, the user location is input to generate the corresponding sCSIs and the channel map is constructed accordingly.

[0047] Step 3, offline training phase: The training process of conditional LDM is divided into two steps: First, a variational autoencoder (VAE) specific to sCSI is trained to learn the low-dimensional representation of high-dimensional sCSI in the latent space, i.e., data compression is performed; wherein, the loss function is improved when training the VAE, a hyperparameter is introduced as the coefficient of the reconstruction loss term, and the hyperparameter is increased during the training process; then, the conditional diffusion model is trained in the latent space to realize the transformation from the standard Gaussian distribution to the low-dimensional latent space conditional distribution with respect to location, i.e., using a standard Gaussian noise data to generate a low-dimensional representation of sCSI at the corresponding location;

[0048] Step 4, the online generation stage, the base station can obtain the geographic location coordinates of the user terminal, use the conditional LDM diffusion model, and improve the generation process by discarding the sampling variance term in the LDM generation process to improve the accuracy of the generation results, and generate a low-dimensional representation of sCSI under specific user location conditions, thereby constructing a channel map containing the location and the low-dimensional representation of sCSI; when using the channel map, the low-dimensional representation is restored to sCSI through VAE.

[0049] Specifically, in step 1, the spatial beam sampling matrix and the temporal beam sampling matrix are matrices formed by splicing the sampling rudder vectors corresponding to a set of receiving angle direction cosines and time delay sampling points selected by the base station; each column of the spatial beam sampling matrix corresponds to a spatial beam; and each column of the temporal beam sampling matrix corresponds to a temporal beam.

[0050] For example, the sampling range for the direction cosine of the receiving angle is -1 to 1, and the sampling range for the delay is 0 to the maximum delay spread; the sampling method is uniform sampling; the number of sampling points for the direction cosine of the receiving angle and the delay is greater than or equal to the number of user-end receiving antennas and the number of equivalent delay spread points; the number of equivalent delay spread points is obtained by multiplying the ratio of the number of effective subcarriers to the total number of subcarriers by the cyclic prefix length and rounding up.

[0051] The space-frequency domain channel model is sampled in the angle-delay domain to establish a dual-beam domain channel in the space-frequency domain. The space-frequency domain channel model is characterized as the product of the spatial domain sampled beam matrix, the dual-beam domain channel, and the time domain sampled beam matrix.

[0052] In this embodiment, the energy matrix of the dual-beam domain channel is extracted as sCSI to construct a knowledge graph. The dual-beam domain energy matrix reflects the distribution of channel energy at different distinguishable angles and delays. That is, each element in the matrix is ​​related to the power, angle of arrival, and delay of arrival of each path of the channel between the base station and the user, and they are closely related to the user's location. Due to the large number of angle samples and delay samples and the small number of channel paths, the dual-beam domain energy matrix exhibits sparsity; and the non-zero elements in the energy matrix are usually clustered, with each cluster reflecting a scatterer in the physical environment.

[0053] In step 2, during the offline training phase, the locations of all reference points in the target area are collected, and the dual-beam domain channel energy matrix at the corresponding locations is extracted to form a dataset as training data. Using the training data, an LDM is trained, and the sCSI at a given location can be obtained. In the online generation phase, using the trained LDM and given the user location, the sCSI between the base station and the user can be obtained. Thus, the channel map can be constructed using online interpolation.

[0054] Step 3, LDM training, consists of two steps: training a VAE and training a diffusion model. The VAE can learn the low-dimensional representation of the high-dimensional dual-beam domain channel energy matrix and can restore the low-dimensional representation to the high-dimensional dual-beam domain channel energy matrix. The diffusion model can learn the latent space distribution of the low-dimensional representation. By applying the conditional diffusion model, a low-dimensional representation of the sCSI at a given location can be generated, and the VAE can be used to restore the low-dimensional representation to high-dimensional data, i.e., the dual-beam domain channel energy matrix.

[0055] In this embodiment, the VAE learns the low-dimensional latent space distribution of the dual-beam domain channel energy matrix and converts each sCSI matrix in the dataset into a low-dimensional representation in its latent space. The dimension of the latent space is much lower than the dimension of the data space, which is a data compression process. By applying the converted low-dimensional representation, the computational complexity of the training process can be greatly reduced. The VAE contains an encoder and a decoder, where the encoder can convert the sCSI matrix into its low-dimensional representation, and the decoder can restore the low-dimensional representation to a high-dimensional sCSI matrix.

[0056] The diffusion model in this embodiment is divided into a forward process and a backward process. The forward process can also be called the noise-adding process or the diffusion process, while the backward process is called the denoising process. The forward process progressively adds Gaussian noise multiple times to the low-dimensional representation, eventually transforming it into standard Gaussian noise. Based on the "variance scheduling value," the variance of the noise added in each step of the forward process increases. The forward process is modeled as a Markov chain, where the data after each noise addition is only related to the data after the previous noise addition. The backward process first samples a Gaussian noise, then progressively removes the noise added in the forward process to recover the data before noise addition. A U-Net network is used to learn the noise added at each step of the forward process to recover the undisturbed data.

[0057] In step 4, the U-Net network in the trained LDM and the VAE are used for online sCSI generation of users and to construct a channel map. In online mode, the base station extracts the user's location, samples to obtain a standard Gaussian noise matrix, obtains the accumulated noise through the trained U-Net, and gradually removes the noise to finally obtain un-noised low-dimensional data. The low-dimensional data is input into the VAE decoder to obtain a dual-beam domain energy matrix. Based on this, the channel map of the target area can be constructed using the online interpolation method.

[0058] The following section provides a detailed explanation of the channel map construction method for large-scale MIMO systems based on LDM disclosed in this invention, using a specific system model as an example.

[0059] I. Channel Model for Large-Scale MIMO Systems

[0060] Consider a single-cell massive MIMO communication system, including a system equipped with N r The system consists of a base station with one uniform linear array antenna (ULA) and U single-antenna users. It employs Time Division Duplex (TDD) mode and Orthogonal Frequency Division Multiplexing (OFDM) for transmission. Uplink transmission has N... c N subcarriers, of which N are allocated v One subcarrier is used for pilot signal transmission. The system sampling interval and cyclic prefix (CP) length are T and T, respectively. s and N g Let f be an example. c and These represent the carrier frequency and subcarrier spacing, respectively. Assume the channel is quasi-static over one OFDM symbol and the base station antenna spacing is half a wavelength. G represents the space-frequency domain channel matrix between the base station and user u. u It can be specifically expressed as

[0061]

[0062] Among them, P u β is the number of paths between the base station and user u.u,p It is the complex channel gain of the p-th path between the base station and user u, Ξ u,p and τ u,p These are the direction cosine of the arrival angle and the path delay of the p-th path between the base station and user u at the base station side, respectively, and the rudder vector v(Ξ) u,p ) and u(τ u,p ) are respectively represented as

[0063]

[0064] This invention defines the spatial frequency domain channel coefficient as:

[0065]

[0066] The spatial frequency domain channel matrix G can be used u Written as

[0067]

[0068] in and Let be the sets of numerical ranges of Ξ and τ, respectively, and define them as the union of the following disjoint sets.

[0069]

[0070] Where N a =F a N r and This indicates rounding up, and F a and F d This is the refinement factor. When N a and N d When it is large enough, it will satisfy rudder vector v(Ξ) u,p ) and u(τ u,p It can be approximated as the sampling rudder vector v(Ξ) i ) and u(τ j ),in Therefore, the spatial frequency domain channel matrix G u It can be approximated as

[0071]

[0072] in

[0073]

[0074] This invention defines a space beam sampling matrix. and time beam sampling matrix Therefore, the spatial frequency domain channel matrix G uIt can be derived from the dual beam domain channel matrix express

[0075] G u =VH u U T

[0076] in Based on this, the present invention provides a channel energy matrix expression for the dual-beam domain.

[0077]

[0078] Since most of the channel energy is concentrated in a few distinguishable spatial directions and path delays, the dual-beam domain channel energy matrix has values ​​in only a finite number of beams, and its non-zero elements are usually clustered together, making it a sparse matrix. Ω u Each cluster corresponds to a scatterer in the physical environment and reflects the angle of arrival, delay, and signal energy on each path from the base station to the user u.

[0079] II. High-Dimensional sCSI Compression Process

[0080] Due to N a and N d Large enough, dual-beam domain channel matrix Ω u The dimensions are also relatively large, and training the diffusion model directly on the data dimension requires significant GPU memory and computational complexity. Therefore, the LDM in this invention first trains a VAE specific to the dual-beam domain channel energy matrix to find a low-dimensional latent space, and then maps the dataset to the latent space to train the diffusion model.

[0081] Visual Encoding Algorithms (VAEs) are powerful generative models that combine deep learning and probabilistic graphical models. They can not only learn an effective low-dimensional representation of data (the latent space), but also generate new data samples from this latent space. The core idea of ​​VAEs is a probabilistic generative model, where we assume that the statistical channel energy matrix Ω is composed of some unobserved, low-dimensional latent space variables. Generated through a complex process, where h×w< <N a ×N d We use a parameterized selection model p φ (Ω) is used to approximate the true distribution of the channel energy matrix p * (Ω), that is

[0082] p φ (Ω)≈p * (Ω)

[0083] Based on the generation process of the dual-beam domain statistical channel energy matrix given above, N reference points are selected in the target area and corresponding sCSIs are generated. The sCSIs are then used as the dataset for training the variational autoencoder. Therefore, the optimal parameter φ is solved by maximizing the likelihood logarithm estimate. * ,Right now

[0084]

[0085] However, due to the posterior distribution p φ (Z|Ω) is difficult to solve, making the above equation unsolvable directly. Therefore, a cognitive model (also called an encoder) q is introduced. ψ Using (Z|Ω) to approximate the true posterior distribution, the log-likelihood function of the sampling distribution can be expressed as follows:

[0086] logp φ (Ω)=D KL (q ψ (Z|Ω)||p φ (Z|Ω))+L(φ,ψ;Ω)

[0087] Where D KL (q ψ (Z|Ω)||p φ (Z|Ω))+L(φ,ψ;Ω)=∫q ψ (Z|Ω)log(q ψ (Z|Ω) / p φ (Z|Ω))dZ represents the Kullback-Leibler (KL) divergence between two distributions. It is a non-negative value, and the smaller the value, the closer the two distributions are. L(φ,ψ;Ω)=∫q ψ (Z|Ω)log(p φ (Z,Ω) / q ψ (Z|Ω))dZ is the variational lower bound (ELBO) for the objective. By adjusting the network parameters φ and ψ to maximize the variational lower bound L(φ,ψ;Ω), the following two objectives will be optimized simultaneously: (1) to approximately maximize the log-likelihood function logp φ (Ω); (2) Minimize the cognitive model q ψ (Z|Ω) and the true posterior distribution p φ KL divergence between (Z|Ω).

[0088] Next, we will derive the expression for the variational lower bound to maximize it. The variational lower bound described above can be rearranged as follows:

[0089]

[0090] in p(Z) is the latent space distribution, usually a standard Gaussian distribution, i.e. As can be seen from the above equation, the variational lower bound consists of an expectation term and a KL divergence term, where the expectation term represents the reconstruction loss, i.e., the loss of a single sample Ω. (n) The latent space variable Z is obtained after encoding by the variational autoencoder. (n) Then the reconstructed sample is obtained through the decoder. At that time, the difference between the reconstructed sample and the original sample; the KL divergence term represents the distribution q. ψ The difference between (Z|Ω) and the pre-defined distribution p(Z). Assume the approximate posterior distribution follows a Gaussian distribution. Therefore, for a single sample Ω (n) The KL divergence term in its variational lower bound can be calculated as:

[0091]

[0092] Where J = h × w, for all samples have

[0093]

[0094] For the expected term, i.e. the reconstruction loss term, we assume... in For the decoder output, c0I represents the distribution with a constant covariance matrix. Using the Markov Chain Monte Carlo (MCMC) method, for each latent space variable Z... (n) Taking R sampling points, maximizing the expected term can be approximated by the following formula.

[0095]

[0096] Among them, || || F This represents the Frobenius norm of the matrix.

[0097] As can be seen from the above expression for the expectation term, maximizing the expectation term means minimizing the error between the sample and the reconstructed sample; therefore, this term is also called the reconstruction error term. Since channel information is crucial to the performance of wireless communication systems, in order to obtain more accurate sCSI, this invention weights the reconstruction error term and proposes an improved loss function as follows:

[0098]

[0099] Here, λ is a hyperparameter introduced to increase the weight of the reconstruction loss term, and its value will increase with the increase of the training cycle during the training process.

[0100] VAE via encoder ε ψThe high-dimensional sCSI is transformed into the mean and variance of a low-dimensional latent space distribution, and a low-dimensional latent space matrix is ​​generated through reparameterization. This matrix is ​​then decoded. The high-dimensional sCSI is recovered as shown in the following formula.

[0101] (μ,logΣ)=ε ψ (Ω)

[0102] Z = μ + exp(0.5 × logΣ) ∈

[0103]

[0104] The logarithm of the variance is used above to avoid a mismatch between the negative values ​​of the neural network output and the variance. The high-dimensional sCSI compression process is as follows:

[0105] Step a) Initialize encoder parameters ψ and decoder parameters φ;

[0106] Step b) on the training set M samples were sampled.

[0107] Step c) For each sample Ω (m) For m∈{1,2,...,M}, execute steps d) to g) until m = M;

[0108] Step d) Randomly sample a noise matrix ∈ from standard Gaussian noise;

[0109] Step e) Calculate (μ) (m) ,logΣ (m) )=ε ψ (Ω (m) );

[0110] Step f) Calculate Z (m) =μ (m) +exp(0.5×logΣ (m) )∈;

[0111] Step g) Calculation

[0112] Step h) Calculate the improved loss function

[0113]

[0114] Step i) Calculate and update the neural network gradients

[0115] Step j) Calculate and update the encoder parameters ψ and decoder parameters φ using the updated gradient;

[0116] Step k) If the network parameters converge, output the encoder parameters ψ and decoder parameters φ; otherwise, return to step b).

[0117] III. sCSI Low-Dimensional Distribution Learning Process

[0118] The diffusion model comprises two parts: a forward process and a backward process. The forward process transforms the training data into data conforming to a standard Gaussian noise distribution by adding a series of Gaussian noises with different variances in T steps. The backward process randomly samples a noisy data point from the standard Gaussian distribution and recovers the original data by removing the noise added in the forward process step by step. This invention relates to location-based sCSI generation, therefore, location information needs to be input as a condition into the LDM. A schematic diagram of the conditional LDM proposed in this invention is shown below. Figure 2 As shown. According to Figure 2 As shown, the left half uses the trained VAE to transform the data Ω in the data space to the data Z in the latent space. The right half shows the forward and backward processes of the row diffusion model in the latent space. The encoder of the variational autoencoder transforms the sCSI matrix in the dataset... Data transformed into the latent space The training dataset for the diffusion model is formed by combining the corresponding positions.

[0119] The forward process of the conditional diffusion model is represented by the following conditional probabilities.

[0120]

[0121] Z0 is sampled from the noisy data in the dataset.

[0122] Let be a predefined constant that monotonically increases with t. When T is sufficiently large, the unnoised data Z0 is transformed into Gaussian noise data Z after T times of adding Gaussian noise. T Based on the above formula, by applying the reparameterization method, we can obtain the noisy data at any time step t.

[0123]

[0124] in α t =1-β t And ∈ t For standard Gaussian noise. In the above equation, ∈ t Characterized by Z0 to Z t The cumulative value of the added noise t times. It can also be seen from the above formula that when T is large enough, we have... therefore

[0125]

[0126] Once we know the cumulative noise added at each time step t, we can reverse the forward process step by step to reconstruct the data before the noise was added from a standard Gaussian noise.

[0127] Based on the Markov assumption, the reverse process of the conditional diffusion model is defined as follows:

[0128]

[0129] Therefore, if the single-step transition probability distribution function p(Z) is known... t-1 |Z t This allows us to obtain the data distribution with added noise. Based on this Bayes' theorem, we can obtain...

[0130]

[0131] Despite p(Z) t-1 ) and p(Z t All of these are unknown, but by continuing the derivation of the forward process, we can obtain p(Z). t-1 |Z0) and p(Z t |Z0), therefore, if the initial distribution p(Z0) is known, then we have

[0132]

[0133] Where C(Z) t Z0) is related to Z t-1 An irrelevant constant. From the above equation, we can see that p(Z) t-1 |Z t The distribution (Z0) also follows a Gaussian distribution, and its mean and variance can be calculated as follows:

[0134]

[0135] From this, we can obtain

[0136]

[0137] The above equation represents the true conditional distribution given the initial distribution p(Z0). Our goal is to enable the diffusion model to learn the conditional distribution p. θ (Z t-1 |Z t The diffusion model aims to approximate the true conditional distribution as closely as possible to Z0. As seen in the equation above, the variance is a fixed quantity; therefore, the diffusion model only needs to learn the mean of the true distribution. Based on the forward process, Z0 can be represented by Z... t and ∈ t It is indicated as follows:

[0138]

[0139] Therefore, the mean of the true distribution can be expressed as follows:

[0140]

[0141] After simplification, we will use μ(Z0,Z) t ) becomes μ(Z0,∈ t ), where ∈ t Since the aforementioned accumulated noise is the primary noise, we only need to train the diffusion model to learn this accumulated noise; that is, the loss function of the diffusion model is defined as follows:

[0142]

[0143] The training process of LDM is given below:

[0144] Step a) Initialize network parameters θ;

[0145] Step b) on the training set M samples were sampled.

[0146] Step c) For each sample p (m) ,Ω (m ), m∈{1,2,...,M}, execute steps d) to e) until m=M;

[0147] Step d) Randomly sample a time step from {1,2,...,T}, denoted as t. m ;

[0148] Step e) at time step t m Random sampling from standard Gaussian noise

[0149] Step f) Calculate the loss function

[0150]

[0151] Step g) Calculate and update the neural network gradients

[0152] Step h) Calculate and update the model parameters θ using the updated gradient;

[0153] Step i) If the network parameters converge, output the model parameters θ; otherwise, return to step b).

[0154] IV. LDM-based sCSI generation process and channel map construction process

[0155] After training the LDM, location-based sCSI can be generated and a channel map can be constructed. First, according to the aforementioned formula, given the accumulated noise ∈ [the previous t steps]... θ (Z t (p,t) and noise matrix Z t In this case, the noise matrix at step t-1 is obtained according to the improved single-step sampling formula below.

[0156]

[0157] It is important to note that this invention is designed for generating sCSI data, thus prioritizing accuracy over sample diversity. Therefore, this invention improves the single-step sampling process during generation. The above equation is the improved single-step sampling process expression, which discards the noise variance term from the traditional diffusion model sampling process, retaining only the mean calculation term to improve the accuracy of the generated sCSI data.

[0158] Based on the trained diffusion model parameters θ, a matrix Z is sampled from a standard Gaussian distribution. T Z can then be obtained through stepwise sampling. T-1 ..., Z1, Z0, finally obtaining noise-free data Z = Z0. Then, the decoder of the trained VAE variational autoencoder is used. Transforming noise-free latent space data into high-dimensional statistical channel information. The specific process is as follows:

[0159] Step a) Generate a blank channel map for a specific location area and load the existing locations and corresponding sCSIs from the dataset into the channel map;

[0160] Step b) Load the decoder parameters φ and diffusion model parameters θ of the trained variational autoencoder;

[0161] Step c) Given a location vector p to generate sCSI, randomly sample a matrix Z from a standard Gaussian distribution that conforms to the latent space dimension. T ;

[0162] Step d) For t = T, ..., 2, 1, calculate Z according to the following formula. T-1 ...,Z1,Z0, thus obtaining the low-dimensional latent space representation Z = Z0 of sCSI.

[0163]

[0164] Step e) Repeat steps c) and d) until the low-dimensional representations of sCSI at all locations in the channel map are obtained, and then interpolate them into the channel map to establish a complete channel map;

[0165] Step f) Use the VAE decoder to recover the low-dimensional representation of sCSI in the channel map into a high-dimensional sCSI.

[0166] Traditional diffusion models require a single-step sampling process to generate samples, resulting in slow generation speed. Therefore, this invention proposes an improved sampling method for accelerated generation. This method accelerates the generation process by skipping steps in the sampling process and discards the sampling variance term in the accelerated generation process to improve the accuracy of the generated results. The single-step sampling formula for the accelerated generation process after improvement is as follows:

[0167]

[0168] in

[0169]

[0170] Δt>0 represents the skip sampling interval. To accelerate the number of sampling steps in the sampling process, the accelerated sampling process is performed according to a sequence. To execute and construct the channel map using the accelerated generation method with improved sampling, simply execute step d) of the original channel map construction process according to the new sampling sequence and single-step sampling formula.

[0171] V. Implementation Results

[0172] To enable those skilled in the art to better understand the present invention, the following embodiment of the method for constructing a channel map of a large-scale MIMO system based on LDM under a specific system configuration is compared with the performance results of other sCSI generation methods.

[0173] To construct a simulation environment capable of reproducing real-world characteristics, this invention employs the mainstream QuaDRiGa channel model to generate simulation scenarios, including the "3GPP 38.901UMa NLOS" scenario. In the simulation environment, a ground base station equipped with a uniform linear array antenna is located at coordinates (0,0) meters. 128 antennas are arranged along the Y direction at half-wavelength intervals. The target area for constructing the channel map is a square region centered at (200,0) meters with sides of 101 meters. To obtain the sCSI, the target area is divided into N = 101 × 101 small square regions at 1-meter intervals. Fifty single-antenna user terminals are randomly placed within each small square region. The sCSI of each small square region is generated based on these 50 user terminals. The center location of this small region and its sCSI are used as the dataset for training the LDM, and the dataset is divided into training and test sets at a 9:1 ratio.

[0174] To verify the generation effect, the generated statistical channel information can be used. The normalized mean square error (NMSE) between the channel information and the true statistical channel information Ω in the test set is used for evaluation. The NMSE, defined in dB, is as follows:

[0175]

[0176] Where L is the number of samples in the test set, Ω (l) This is the l-th sCSI matrix in the test set.

[0177] In the verification process of this invention, the VAE encoder includes two 2D convolutional layers with 64 and 128 kernels respectively, a Flatten layer, and four fully connected layers with 512, 128, 16, and 64 units respectively, wherein the latent space dimension is 8×8; the decoder includes three fully connected layers with 128, 512, and 65536 units respectively, a Reshape layer with a target shape of (32, 32, 64), two 2D deconvolutional layers with 128 and 64 kernels respectively, and two 2D convolutional layers with 16 and 1 kernel respectively. To improve the reconstruction effect of the VAE, the fully connected layers added to the VAE encoder and decoder in this invention play a crucial role. This invention uses an improved loss function, which introduces a hyperparameter λ to control the weight of the reconstruction loss term. During the experiment, the maximum training epochs were set to 10000, and the initial value of the hyperparameter λ was 1. It was then increased to 1000 and 10^6 in the 3000th and 6000th training epochs, respectively, to increase the proportion of the reconstruction loss term in the loss function. Using NMSE as the verification standard for encoding and decoding performance, performance of -57.95dB and -57.81dB was achieved on the training and test sets, respectively, demonstrating that the sCSI-specific VAE in this invention has excellent data compression and recovery effects.

[0178] To verify the effectiveness of the proposed LDM-based sCSI generation method, the traditional linear interpolation method and the conditional VAE method were compared. For the linear interpolation method, the positions in the training set and the sCSIs were interpolated using the positions as indices to obtain the sCSIs at the positions in the test set; for the conditional VAE method, the latent space dimension was set to 16, and this method was denoted as VAE-16. Figure 3 The changes in normalized mean squared error (NMSE) of the proposed method and two comparative schemes over 100,000 training cycles are shown: the proposed LDM, the improved sampling acceleration LDM method, conditional VAE-16, and the linear interpolation method. The accelerated sampling LDM method with a sampling interval of Δt = 10 is compared and denoted as ALDM-10. LDM, ALDM-10, and VAE-16 all show a continuous decreasing trend in NMSE, while the interpolation method remains unchanged due to its non-parametric nature. It should be noted that the strategy of this invention in saving parameters during neural network training is to save network parameters only when the loss function decreases. Figure 3 The results shown do not cover all training cycles. As can be seen from the figure, LDM achieves an NMSE gain of 7.08 dB compared to interpolation and 2.19 dB compared to VAE-16. The ALDM-10 method suffers approximately a 1 dB performance loss compared to LDM, but still offers some performance improvement compared to interpolation and VAE-16. Figure 4 This paper compares the generation speed of the LDM method and the improved sampling accelerated generation method (ALDM). The time in the figure represents the average generation time per sample after generating 1020 samples. As can be seen from the figure, the ALDM method is approximately 15 times faster than the LDM method, which significantly accelerates the generation process and the construction of the channel map. Furthermore, at a sampling interval of Δt = 100, its NMSE performance still reaches -37.4318 dB, representing a performance gain of approximately 4 dB compared to the interpolation method.

[0179] This invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the method for constructing a channel map of a large-scale MIMO system based on LDM.

[0180] This invention also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method for constructing a channel map of a large-scale MIMO system based on LDM.

[0181] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

[0182] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways without departing from the spirit and scope of this application. The current embodiments are merely exemplary examples and should not be considered limiting, nor should the specific content given limit the purpose of this application. For example, some features may be omitted or not implemented.

[0183] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for constructing channel maps for a large-scale MIMO system based on a latent space diffusion model, characterized in that, Includes the following steps: Based on the configuration of a large-scale MIMO system, the space-frequency domain channel model is expressed as the product of the spatial domain sampling beam matrix, the dual-beam domain channel, and the time domain sampling beam matrix, and a dual-beam domain channel model is established; the beam domain channel energy matrix is ​​extracted as the user's statistical channel state information (sCSI). The construction of the channel map is divided into offline training and online generation stages. In the offline training stage, the location of reference points in the target area and their corresponding sCSIs are collected to train the conditional latent space diffusion model (LDM). In the online generation stage, the user location is input to generate the corresponding sCSI. During the offline training phase, the training process of the conditional LDM includes: First, training a variational autoencoder (VAE) specific to sCSI to learn the low-dimensional representation of high-dimensional sCSI in the latent space; wherein, when training the VAE, an improved loss function is used, a hyperparameter is introduced as the coefficient of the reconstruction loss term, and the hyperparameter is increased during the training process; then, training the conditional diffusion model in the latent space to realize the transformation from the standard Gaussian distribution to the low-dimensional latent space conditional distribution with respect to location; In the online generation phase, the geographic coordinates of the user terminal are obtained, the conditional LDM diffusion model is used, and the generation process is improved by discarding the sampling variance term in the LDM generation process to improve the accuracy of the generation results. A low-dimensional representation of sCSI under specific user location conditions is generated, thereby constructing a channel map containing the location and low-dimensional representation of sCSI. When using the channel map, the low-dimensional representation is restored to sCSI through VAE. The process of constructing the channel map and recovering the high-dimensional sCSI includes: Step a) Generate a blank channel map for a specific location region and load the existing locations and corresponding low-dimensional representations of sCSIs in the dataset into the channel map; Step b) Load the decoder parameters and diffusion model parameters of the trained VAE; Step c) Given a location vector to generate sCSI, randomly sample a matrix that conforms to the latent space dimension from a standard Gaussian distribution; Step d) Based on the trained diffusion model, use the improved generation process to generate a low-dimensional representation of the sCSI at the current position; Step e) Repeat steps c) and d) until the low-dimensional representations of sCSI at all locations in the channel map are obtained, and then interpolate them into the channel map to establish a complete channel map; Step f) Use the VAE decoder to recover the low-dimensional representation of sCSI in the channel map into a high-dimensional sCSI.

2. The method for constructing channel maps of large-scale MIMO systems based on the latent space diffusion model according to claim 1, characterized in that: The dual-beam domain energy matrix reflects the distribution of channel energy at different distinguishable angles and delays. That is, each element in the matrix is ​​related to the power, angle of arrival, and delay of arrival of each path of the channel between the base station and the user, and they are closely related to the user's location.

3. The method for constructing channel maps of large-scale MIMO systems based on the latent space diffusion model according to claim 1, characterized in that: During the offline training phase, the locations of all reference points in the target area are collected and the dual-beam domain channel energy matrix at the corresponding locations is extracted to form a dataset as training data. By training an LDM using training data, the sCSI at a given location can be obtained; in the online generation stage, by using the trained LDM and given the user's location, the sCSI between the base station and the user and its low-dimensional representation can be obtained. Therefore, it is possible to construct channel maps using online interpolation.

4. The method for constructing channel maps of large-scale MIMO systems based on the latent space diffusion model according to claim 1, characterized in that: The VAE is used to learn the low-dimensional representation of the high-dimensional dual-beam domain channel energy matrix and restore the low-dimensional representation to the high-dimensional dual-beam domain channel energy matrix data space. The diffusion model is used to learn the latent space distribution of the low-dimensional representation. By applying the conditional diffusion model, a low-dimensional representation of the dual-beam domain channel energy matrix at a given location is generated, and the VAE is used to restore the low-dimensional representation to the high-dimensional dual-beam domain channel energy matrix data space.

5. The method for constructing channel maps of large-scale MIMO systems based on the latent space diffusion model according to claim 4, characterized in that: The compression process of the high-dimensional dual-beam domain channel energy matrix is ​​as follows: Step a) Initialize encoder parameters and decoder parameters ; Step b) in containing Sampling from the training set of each sCSI One sample; Step c) Perform steps d) through g) for each sample until all All samples underwent the above steps; Step d) Randomly sample from standard Gaussian noise; Step e) Calculate the mean and variance terms of the latent space vector of the encoder output of the VAE; Step f) Calculate the low-dimensional representation of sCSI in the latent space based on the mean and variance terms and the sampled Gaussian noise; Step g) Input the low-dimensional representation of sCSI into the VAE decoder to calculate sCSI; Step h) Calculate the improved VAE loss function, where the reconstruction loss term of the loss function increases with the number of training cycles; Step i) Calculate and update the gradients of the encoder and decoder parameters in the VAE; Step j) Calculate and update the encoder parameters using the updated gradients. and decoder parameters ; Step k) If the network parameters converge, output the encoder parameters. and decoder parameters Otherwise, return to step b).

6. The method for constructing channel maps of large-scale MIMO systems based on the latent space diffusion model according to claim 1, characterized in that: When generating the low-dimensional representation of sCSI, skip sampling is used to accelerate the generation process, while the sampling variance term in the accelerated generation process is discarded to improve the accuracy of the generated results.

7. The method for constructing channel maps of large-scale MIMO systems based on the latent space diffusion model according to claim 1, characterized in that: The VAE encoder consists of two 2D convolutional layers, one Flatten layer, and four fully connected layers, while the VAE decoder consists of three fully connected layers, one Reshape layer, two 2D deconvolutional layers, and two 2D convolutional layers.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for constructing a channel map of a large-scale MIMO system based on a latent space diffusion model according to any one of claims 1-7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for constructing a channel map of a large-scale MIMO system based on a latent space diffusion model according to any one of claims 1-7.

Images