Large-scale MIMO system channel map construction method based on hidden space diffusion model

Abstract

The invention provides a large-scale MIMO (Multiple Input Multiple Output) system channel atlas construction method based on a hidden space diffusion model (LDM). According to the method, the spatial consistency of a large-scale MIMO system channel is utilized, and LDM learning is used for counting conditional probability distribution of channel state information (sCSI) about the position of a user terminal, so that a channel map is constructed. An improved variational auto-encoder (VAE) is used to learn the low-dimensional representation of high-dimensional sCSI, and at the same time, an LDM is improved to improve the generation precision of the low-dimensional representation of the sCSI, so that a channel map containing the position and the low-dimensional representation data of the sCSI is constructed. The storage and transmission cost of the channel map is greatly reduced compared with the direct use of the high-dimensional sCSI, the low-dimensional representation of the sCSI in the channel map can be accurately restored to the high-dimensional sCSI by using the VAE, the generation process does not depend on pilot signals, and the pilot overhead is greatly saved while the sCSI is accurately generated.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications and relates to a method for constructing a channel spectrum of a large-scale MIMO system based on a latent space diffusion model (LDM). Background Art

[0002] Massive Multiple-Input Multiple-Output (Massive MIMO) technology is a core enabling technology for fifth-generation (5G) mobile communications and will continue to evolve and play a key 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 receive processing. However, the continuous growth 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 usually use 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; and as the number of base station antenna units, the number of parallel data streams, and the system bandwidth further increase significantly, the required pilot overhead will become extremely large. Therefore, obtaining channel information increasingly relies on statistical channel state information (sCSI). In recent years, using channel maps to obtain sCSI has become an important technical approach. The channel map can be defined as a database of channel statistics (or channel fingerprints) based on location index within the base station coverage area, or a generative intelligent module. Through the location information (or its equivalent data representation) and mobility information provided by the terminal, the base station can obtain the required channel statistical information.

[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 extract fingerprints through uplink channel sounding. Online interpolation involves generating the channel fingerprint required for the target location using an interpolation algorithm or a generative intelligent module based on the channel fingerprints at the existing grid points. However, existing technologies directly store high-dimensional sCSI when constructing channel maps, which results in significant storage and transmission overhead. Furthermore, the accuracy of sCSI generation needs to be improved. Summary of the Invention

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

[0006] Technical solution: In order to achieve the above-mentioned purpose, the present invention provides the following technical solution:

[0007] A method for constructing a channel map of a massive MIMO system based on LDM, comprising the following steps:

[0008] Based on the massive MIMO system configuration, 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. 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 reference point positions in the target area and their corresponding sCSI are collected to train the conditional LDM. In the online generation stage, the user position is input to generate the corresponding sCSI.

[0010] During the offline training phase, the training process of the conditional LDM includes: first, training a variational autoencoder (VAE) specific to sCSI to learn a low-dimensional representation of the high-dimensional sCSI in the latent space; wherein, when training the VAE, a modified loss function is used, a hyperparameter is introduced as the coefficient of the reconstruction loss term, and the hyperparameter is increased during training; then, a conditional diffusion model is trained in the latent space to transform the standard Gaussian distribution into a low-dimensional latent space conditional distribution with respect to position;

[0011] During 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. The sampling variance term in the LDM generation process is discarded to improve the accuracy of the generation results. A low-dimensional representation of the sCSI under specific user location conditions is generated, thereby constructing a channel map containing the low-dimensional representations of location and 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, arrival angle and arrival delay of each path of the channel between the base station and the user, and they are closely related to the user position.

[0013] Furthermore, during the offline training phase, the positions of all reference points in the target area are collected and the dual-beam domain channel energy matrices at the corresponding positions are extracted to form a data set as training data. By using the training data to train an LDM, the sCSI at a given position can be obtained. During the online generation phase, the trained LDM is used and given the user position, the sCSI between the base station and the user and its low-dimensional representation can be obtained. As a result, 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, and by applying the conditional diffusion model, a low-dimensional representation of the dual-beam domain channel energy matrix at a given position 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) sampling M samples from a training set containing N sCSIs;

[0018] Step c) performing steps d) to g) for each sample until all M samples have performed the above steps;

[0019] Step d) randomly sampling from standard Gaussian noise;

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

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

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

[0023] Step h) calculating an improved VAE loss function, wherein the reconstruction loss term of the loss function increases as the training cycle increases;

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

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

[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 trained U-Net network and VAE in the 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 area and load the existing locations and corresponding low-dimensional representations of sCSI in the dataset into the channel map;

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

[0030] Step c) Given a position vector to generate sCSI, a matrix that matches the dimension of the latent space is randomly sampled from a standard Gaussian distribution;

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

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

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

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

[0035] Furthermore, the VAE encoder includes two 2D convolutional layers, one Flatten layer and four fully connected layers, and 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 the computer program is executed by the processor, the steps of the method for constructing a channel map of a large-scale MIMO system based on LDM are implemented.

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

[0038] Beneficial effects: The LDM-based large-scale MIMO system channel map construction method proposed in the present invention uses an improved VAE to learn the low-dimensional representation of high-dimensional sCSI, and at the same time improves the LDM to improve the generation accuracy of the low-dimensional representation of sCSI, thereby constructing a channel map containing position and sCSI low-dimensional representation data. The storage and transmission costs of the channel map are greatly reduced compared to directly using high-dimensional sCSI, and the use of VAE can accurately restore the low-dimensional representation of sCSI in the channel map to high-dimensional sCSI. The generation process does not rely on pilot signals, and while accurately generating sCSI, it greatly saves pilot overhead. The present invention also proposes an accelerated generation method with improved sampling, which greatly speeds up the generation process of sCSI. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The present invention provides a flowchart for constructing a channel map for a massive MIMO system based on LDM in an embodiment of the present invention.

[0040] Figure 2 Schematic diagram of conditional LDM.

[0041] Figure 3 3 is a comparison chart of the normalized mean square error (NMSE) performance of the present invention and other methods in the embodiments 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 the present invention. DETAILED DESCRIPTION

[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 used to limit the scope of the present invention.

[0044] like Figure 1 As shown, the method for constructing a channel map of a massive MIMO system based on LDM disclosed in an embodiment of the present invention includes the following steps:

[0045] Step 1: Based on the massive MIMO system configuration, 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. A dual-beam domain channel model is established. The beam domain channel energy matrix is ​​extracted as the user's sCSI.

[0046] Step 2: The channel map construction is divided into offline training and online generation phases. In the offline training phase, the reference point locations in the target area and their corresponding sCSI are collected to train the conditional LDM. In the online generation phase, the user locations are input to generate the corresponding sCSI, and the channel map is constructed accordingly.

[0047] Step 3, offline training, involves two steps in the training of the conditional LDM. First, a variational autoencoder (VAE) specific to sCSI is trained to learn a low-dimensional representation of the high-dimensional sCSI in the latent space, i.e., to perform data compression. The loss function is modified during VAE training by introducing a hyperparameter as the coefficient of the reconstruction loss term, which is increased during training. Then, a conditional diffusion model is trained in the latent space to transform the standard Gaussian distribution into a low-dimensional latent space conditional distribution with respect to position. This involves using standard Gaussian noise data to generate a low-dimensional representation of the sCSI at the corresponding position.

[0048] In step 4, during the online generation phase, the base station obtains the geographic coordinates of the user terminal, utilizes the conditional LDM diffusion model, and improves the generation process. This model discards the sampling variance term in the LDM generation process to improve the accuracy of the generated results. This generates a low-dimensional representation of the sCSI under specific user location conditions, thereby constructing a channel map containing the location and low-dimensional representations of the sCSI. When using the channel map, the low-dimensional representation is restored to the sCSI through VAE.

[0049] Specifically, the spatial beam sampling matrix and the time beam sampling matrix in step 1 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; each column of the time beam sampling matrix corresponds to a time beam.

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

[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 represented as the product of the space-domain sampling beam matrix, the dual-beam domain channel and the time-domain sampling beam matrix.

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

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

[0054] The training of the LDM in step 3 is divided into two steps: training a VAE and training a diffusion model. The VAE can learn a 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. The diffusion model can learn the latent space distribution of the low-dimensional representation. By applying the conditional diffusion model, it can generate a low-dimensional representation of the sCSI at a given position and use the VAE to restore this low-dimensional representation to high-dimensional data, namely the dual-beam domain channel energy matrix.

[0055] In this embodiment, 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 that 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. VAE includes an encoder and a decoder, where the encoder can convert the sCSI matrix into its low-dimensional representation, while the decoder restores 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 reverse process. The forward process can also be called a noise addition process or a diffusion process, and the reverse process is also called a denoising process. The forward process gradually adds multiple Gaussian noises to the low-dimensional representation, eventually turning it into standard Gaussian noise; according to the "variance scheduling value", the variance of the noise added each time in the forward process becomes larger and larger; the forward process is modeled as a Markov chain, and the data after each noise addition is only related to the data after the previous noise addition. The reverse process first samples a Gaussian noise, and then gradually removes the noise added in the forward process to restore the data before the noise addition; the noise added in each step of the forward process is learned by using the U-Net network to restore the unnoised data.

[0057] In step 4, the trained U-Net network and VAE in the LDM are used for online sCSI generation of users and constructing a channel map. In the online mode, the base station extracts the user position and samples it to obtain a standard Gaussian noise matrix. The accumulated noise is obtained through the trained U-Net and gradually removed to obtain low-dimensional data without noise. The low-dimensional data is input into the VAE decoder to obtain a dual-beam domain energy matrix. In this way, the channel map of the target area can be constructed using the online interpolation method.

[0058] The following describes in detail the method for constructing a channel map for a massive MIMO system based on LDM disclosed in an embodiment of the present invention in conjunction with a specific system model.

[0059] 1. Massive MIMO System Channel Model

[0060] Consider a single-cell massive MIMO communication system, which consists of a r The system uses time division duplex (TDD) mode and orthogonal frequency division multiplexing (OFDM) for transmission. c subcarriers, of which N v subcarriers are used for pilot signal transmission. The system sampling interval and cyclic prefix (CP) length are T s and N g . Note f c and are the carrier frequency and subcarrier spacing respectively. Assume that the channel is quasi-static on an OFDM symbol and the base station antenna spacing is half a wavelength. represents the spatial frequency domain channel matrix between the base station and user u, G u It can be specifically expressed as

[0061]

[0062] Among them, P u is the number of paths between the base station and user u, βu,p is the complex channel gain of the p-th path between the base station and user u, u,p and τ u,p are the direction cosines of the arrival angle and path delay of the pth path between the base station and user u at the base station side, and the steering vector v(Ξ u,p ) and u(τ u,p ) are respectively expressed as

[0063]

[0064] The present invention defines the space-frequency domain channel coefficient as

[0065]

[0066] The space-frequency domain channel matrix G can be u Written as

[0067]

[0068] in and are sets of numerical ranges of Ξ and τ respectively, and are defined as the union of the following disjoint sets.

[0069]

[0070] where N a =F a N r and means round up, and F a and F d is the refinement factor. a and N d When large enough, it will satisfy The rudder vector v(Ξ u,p ) and u(τ u,p ) can be approximated as the sampling rudder vector v(Ξ i ) and u(τ j ),in Therefore, the space-frequency channel matrix G u It can be approximated as

[0071]

[0072] in

[0073]

[0074] The present invention defines the spatial beam sampling matrix and the time beam sampling matrix Therefore, the space-frequency channel matrix G uThe dual beam domain channel matrix express

[0075] G u =VH u U T

[0076] in Based on this, the present invention gives the channel energy matrix expression of the dual beam domain:

[0077]

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

[0079] 2. High-dimensional sCSI compression process

[0080] Since N a and N d Large enough, the dual beam domain channel matrix Ω u The dimension is also relatively large. Training the diffusion model directly in the data dimension requires a large amount of video 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, maps the dataset into the latent space, and then trains the diffusion model.

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

[0082] p φ (Ω)≈p * (Ω)

[0083] According to the generation process of the dual beam domain statistical channel energy matrix given above, N reference points are selected in the target area and the corresponding sCSI is generated. The sCSI is used as the data set for training the variational autoencoder. Therefore, the optimal parameter φ is solved by maximizing the likelihood logarithm estimation method. * ,Right now

[0084]

[0085] However, since the posterior distribution p φ (Z|Ω) is difficult to solve, which makes the above equation impossible to solve directly. Therefore, a cognitive model (also called encoder) q is introduced ψ (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] Among them 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 the two distributions, which is non-negative 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) of 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) 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 The KL divergence between (Z|Ω).

[0088] Next, we will derive the expression of the variational lower bound to maximize it. The variational lower bound mentioned above can be organized as follows

[0089]

[0090] in p(Z) is the latent space distribution, usually a standard Gaussian distribution, that is, As can be seen from the above formula, the variational lower bound consists of the expectation term and the KL divergence term, where the expectation term represents the reconstruction loss, that is, a sample Ω (n) After encoding by the variational autoencoder, the latent space variable Z is obtained (n) Then the reconstructed sample is obtained through the decoder When , 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-set distribution p(Z). Assume that the approximate posterior distribution follows a Gaussian distribution, that is, So 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 expectation term, that is, the reconstruction loss term, assume in is the output of the decoder, 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, the maximum expectation term can be approximated as shown in the following formula

[0095]

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

[0097] From the above expression of the expected term, it can be seen that maximizing the expected term is to minimize the error between the sample and the reconstructed sample, so this term is also called the reconstruction error term. Since channel information is crucial to the performance of the wireless communication system, in order to obtain more accurate sCSI, the present invention weights the reconstruction error term and proposes an improved loss function as follows:

[0098]

[0099] Among them, λ is a hyperparameter introduced to increase the weight of the reconstruction loss term. During the training process, the value of this hyperparameter will increase with the increase of the training cycle.

[0100] VAE through the encoder ε ψThe high-dimensional sCSI is converted into the mean and variance of the low-dimensional latent space distribution, and the low-dimensional latent space matrix is ​​generated through the reparameterization technique and passed through the decoder Recover high-dimensional sCSI as shown below

[0101] (μ,logΣ)=ε ψ (Ω)

[0102] Z=μ+exp(0.5×logΣ)∈

[0103]

[0104] The reason for taking the logarithm of the variance above is to avoid the negative values ​​of the neural network output not corresponding to 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) In the training set Sample M samples in the middle;

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

[0108] Step d) randomly sampling 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 gradient

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

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

[0117] 3. sCSI low-dimensional distribution learning process

[0118] The diffusion model consists of two parts: a forward process and a reverse process. The forward process adds a series of Gaussian noises with different variances to the training data in T steps to transform the training data into data that conforms to the standard Gaussian noise distribution. The reverse process randomly samples a noise data from the standard Gaussian distribution and restores the original data by gradually removing the noise added in the forward process. The present invention involves generating sCSI based on location, so it is necessary to input location information into the LDM as a condition. The schematic diagram of the conditional LDM proposed in the present invention is shown in Figure 2. Figure 2 As shown. Figure 2 As shown in the figure, the left half uses the trained VAE to transform the data Ω in the data space and the data Z in the latent space. The right half is the forward and reverse process of the row diffusion model in the latent space. The encoder of the variational autoencoder is used to transform the sCSI matrix in the dataset into Data converted to latent space Combined with the corresponding positions to form the training data set of the diffusion model

[0119] The forward process of the conditional diffusion model is expressed as the following conditional probability

[0120]

[0121] Where Z0 is sampled from the unnoised data of the dataset,

[0122] is a predefined constant that increases monotonically with t. When T is large enough, the unnoised data Z0 is added with Gaussian noise T times and eventually becomes the Gaussian noise data Z T According to the above formula, the reparameterization method can be applied to obtain the noised data at any time step t

[0123]

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

[0125]

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

[0127] According to the Markov hypothesis, the reverse process of the conditional diffusion model is defined as follows

[0128]

[0129] Therefore, if we know the single-step transition probability distribution function p(Z t-1 |Z t ), we can get the data distribution of noise. According to this Bayesian formula, we can get

[0130]

[0131] Although p(Z t-1 ) and p(Z t ) are unknown, but continuing the derivation of the forward process can get p(Z t-1 |Z0) and p(Z t |Z0), so if we know the initial distribution p(Z0), we have

[0132]

[0133] Among them, C(Z t , Z0) is the same as Z t-1 It can be seen from the above formula that p(Z t-1 |Z t ,Z0) also obeys Gaussian distribution, and its mean and variance can be calculated as follows

[0134]

[0135] From this we can get

[0136]

[0137] The above formula is the true conditional distribution under the known initial distribution p(Z0). Our goal is to make the conditional distribution p learned by the diffusion model θ (Z t-1 |Z t ,Z0) is as close to the real conditional distribution as possible. From the above formula, we can see that the variance is a fixed amount, so the diffusion model only needs to learn the mean of the real distribution. According to the forward process, Z0 can be replaced by Z t and ∈ t Indicates, as follows

[0138]

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

[0140]

[0141] After simplification, we have μ(Z0,Z t ) becomes μ(Z0,∈ t ), where ∈ t is the aforementioned cumulative noise, so we only need to let the diffusion model learn this cumulative noise, that is, the diffusion model loss function is defined as follows

[0142]

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

[0144] Step a) Initialize network parameters θ;

[0145] Step b) In the training set Sample M samples in the middle;

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

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

[0148] Step e) At time step t m Randomly sample from standard Gaussian noise

[0149] Step f) Calculate the loss function

[0150]

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

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

[0153] In step i), if the network parameters converge, the model parameters θ are output; otherwise, the process returns to step b).

[0154] 4. sCSI Generation Process Based on LDM and Channel Map Construction Process

[0155] After training the LDM, we can generate the position-based sCSI and construct the channel map. First, according to the above formula, we know the cumulative noise ∈ θ (Z t ,p,t) and the noise matrix Z t In the case of , the noise matrix of step t-1 is obtained according to the following improved single-step sampling formula

[0156]

[0157] It's important to note that this invention targets sCSI generation, requiring accuracy rather than sample diversity. Therefore, this invention improves the single-step sampling process. The above equation represents this improved single-step sampling process, which discards the noise variance term used in the traditional diffusion model sampling process and retains only the mean term to improve the accuracy of sCSI data generation.

[0158] According to the trained diffusion model parameters θ, a matrix Z is sampled from the standard Gaussian distribution T , Z can be obtained by stepwise sampling T-1 ,...,Z1,Z0, and finally get the noise-free data Z=Z0. Then, use the decoder of the trained VAE variational autoencoder Convert 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 sCSI in the dataset into the channel map;

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

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

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

[0163]

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

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

[0166] Since the traditional diffusion model needs to perform a single-step sampling process when generating samples, the generation speed is slow. Based on this, the present invention proposes an accelerated generation method with improved sampling. This method accelerates the generation process by skipping sampling and discards the sampling variance term in the accelerated generation process to improve the accuracy of the generated results. The single-step sampling formula of the accelerated generation process after improved sampling is as follows

[0167]

[0168] in

[0169]

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

[0171] V. Implementation Effect

[0172] To help those skilled in the art better understand the present invention, the following compares the performance of the LDM-based massive MIMO system channel map construction method of this embodiment with other sCSI generation methods under specific system configurations.

[0173] To create a simulation environment that replicates real-world characteristics, the present invention uses the mainstream QuaDRiGa channel model to generate simulation scenarios, taking into account 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 a wavelength. The target area for channel mapping is a square area centered at (200,0) meters with a side length of 101 meters. To obtain sCSI, the target area is divided into N = 101 × 101 square areas at intervals of 1 meter. Fifty single-antenna user terminals are randomly placed in each square area. The sCSI for each square area is generated based on the 50 user terminals in each area. The center position of this small area and its sCSI are used as the dataset for training the LDM, which is divided into a training set and a test set in a 9:1 ratio.

[0174] In order to verify the generation effect, the generated statistical channel information can be used The normalized mean square error (NMSE) between the actual statistical channel information Ω in the test set is used to evaluate the NMSE. The NMSE defined in dB is as follows

[0175]

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

[0177] During the verification process of the present invention, the VAE encoder includes two two-dimensional convolutional layers with 64 and 128 convolution kernels, a Flatten layer, and four fully connected layers with 512, 128, 16, and 64 units, respectively, with a latent space dimension of 8×8. The decoder includes three fully connected layers with 128, 512, and 65536 units, a Reshape layer with a target shape of (32, 32, 64), two two-dimensional deconvolution layers with 128 and 64 convolution kernels, respectively, and two two-dimensional convolution layers with 16 and 1 convolution kernels, respectively. To improve the reconstruction effect of the VAE, the fully connected layers added to the VAE encoder and decoder of the present invention play a crucial role. This paper 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 cycle was set to 10,000, and the initial value of the hyperparameter λ was set to 1. It was increased to 1000 and 10^6 in the 3000th and 6000th training cycles, respectively, to increase the proportion of the reconstruction loss term in the loss function. Using NMSE as the verification metric for codec 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 paper has excellent data compression and recovery effects.

[0178] To verify the effectiveness of the proposed LDM-based sCSI generation, we compared it with traditional linear interpolation and a conditional VAE. For the linear interpolation method, we used the training set positions and the sCSI as indexes to interpolate the sCSI at the test set positions. For the conditional VAE method, we set the latent space dimension to 16 and denoted this method as VAE-16. Figure 3 The changes in the normalized mean square error (NMSE) of the proposed method and two comparison schemes over 100,000 training cycles are shown: the proposed LDM, the improved sampling accelerated LDM method, the conditional VAE-16, and the linear interpolation method. The comparison sampling interval is Δt = 10 and the accelerated sampling LDM method is denoted as ALDM-10. LDM, ALDM-10, and VAE-16 all show a continuous NMSE downward trend, while the interpolation method remains unchanged due to its non-parametric characteristics. It should be noted that the strategy of saving parameters when training neural networks in this invention is to save network parameters only when the loss function decreases, so Figure 3 The results shown in the figure do not fail to reach all training cycles. As can be seen from the figure, LDM achieves NMSE gains of 7.08dB and 2.19dB compared to interpolation and VAE-16, respectively. The ALDM-10 method suffers a performance loss of approximately 1dB compared to the LDM method, but still achieves a certain performance improvement over interpolation and VAE-16. Figure 4 This paper compares the generation speed of the LDM method and the improved sampling accelerated generation method (denoted as ALDM). The time in the figure is 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 greatly accelerates the generation process and the construction of the channel map. Moreover, when the sampling interval is Δt = 100, its NMSE performance still reaches -37.4318dB, which still has a performance gain of approximately 4dB compared to the interpolation method.

[0179] An embodiment of the present 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, the steps of the method for constructing a channel map of a large-scale MIMO system based on LDM are implemented.

[0180] An embodiment of the present invention further discloses a computer program product, including a computer program, which, 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 for implementing the inventive method can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that the program code, when executed by the processor or controller, causes the steps of the inventive method to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as an independent software package and partially on a remote machine, or completely on a remote machine or server. The present invention is not described in detail herein, and all of these are known techniques to those skilled in the art.

[0182] In the embodiments provided herein, it should be understood that the disclosed methods may be implemented in other ways without departing from the spirit and scope of the present application. The present embodiments are merely illustrative examples and should not be construed as limiting. The specific details provided herein should not limit the purpose of the present application. For example, some features may be omitted or not implemented.

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

Claims

1. A method for constructing a channel map for a massive MIMO system based on a latent diffusion model (LDM), characterized in that: The following steps are involved: Based on the massive MIMO system configuration, 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. 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 reference point positions in the target area and their corresponding sCSI are collected to train the conditional LDM. In the online generation stage, the user position 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 a low-dimensional representation of the high-dimensional sCSI in the latent space; wherein, when training the VAE, a modified loss function is used, a hyperparameter is introduced as the coefficient of the reconstruction loss term, and the hyperparameter is increased during training; then, a conditional diffusion model is trained in the latent space to transform the standard Gaussian distribution into a low-dimensional latent space conditional distribution with respect to position; During 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. The sampling variance term in the LDM generation process is discarded to improve the accuracy of the generation results. A low-dimensional representation of the sCSI under specific user location conditions is generated, thereby constructing a channel map containing the low-dimensional representations of location and sCSI. When using the channel map, the low-dimensional representation is restored to sCSI through VAE.

2. The method for constructing a channel map for a massive MIMO system based on LDM according to claim 1, wherein: 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, arrival angle and arrival delay of each path of the channel between the base station and the user, and they are closely related to the user location.

3. The method for constructing a channel map for a massive MIMO system based on LDM according to claim 1, wherein: During the offline training phase, the positions of all reference points in the target area are collected and the dual beam domain channel energy matrices at the corresponding positions are extracted to form a data set as training data; By training an LDM using training data, the sCSI at a given location can be obtained. In the online generation phase, using the trained LDM and a given user location, the sCSI between the base station and the user and its low-dimensional representation can be obtained. From this, the channel map can be constructed using online interpolation.

4. The method for constructing a channel map for a massive MIMO system based on LDM according to claim 1, wherein: 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 position is generated, and 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 a channel map for a massive MIMO system based on LDM according to claim 4, wherein: The compression process of the high-dimensional dual-beam domain channel energy matrix is ​​as follows: Step a) Initialize encoder parameters ψ and decoder parameters φ; Step b) sampling M samples from a training set containing N sCSIs; Step c) performing steps d) to g) for each sample until all M samples have performed the above steps; Step d) randomly sampling from standard Gaussian noise; Step e) calculating the mean and variance of the latent space vector output by the VAE encoder; Step f) calculating a low-dimensional representation of the sCSI in the latent space based on the mean and variance terms and the sampled Gaussian noise; Step g) inputting the low-dimensional representation of sCSI into the VAE decoder to calculate sCSI; Step h) calculating an improved VAE loss function, wherein the reconstruction loss term of the loss function increases as the training cycle increases; Step i) Calculate and update the encoder and decoder parameter gradients in the VAE; Step j) using the updated gradient to calculate and update the encoder parameters ψ and decoder parameters φ; Step k) If the network parameters converge, output the encoder parameters ψ and decoder parameters φ; otherwise, return to step b).

6. The method for constructing a channel map for a massive MIMO system based on LDM according to claim 1, wherein: The trained U-Net network and VAE in LDM are used to construct the channel map and recover the high-dimensional sCSI data. The specific process is as follows: Step a) Generate a blank channel map for a specific location area and load the existing locations and corresponding low-dimensional representations of sCSI in the dataset into the channel map; Step b) Load the trained VAE decoder parameters φ and diffusion model parameters θ; Step c) Given a position vector to generate sCSI, a matrix that matches the dimension of the latent space is randomly sampled from a standard Gaussian distribution; Step d) Based on the trained diffusion model, a low-dimensional representation of the sCSI at the current position is generated using the improved generation process; Step e) Repeat steps c) and d) until low-dimensional representations of the sCSI at all locations in the channel map are obtained, and interpolate them into the channel map to establish a complete channel map; Step f) Use the VAE decoder to restore the low-dimensional representation of sCSI in the channel map to high-dimensional sCSI.

7. The method for constructing a channel map for a massive MIMO system based on LDM according to claim 6, wherein: When generating the sCSI low-dimensional representation, 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 generated results.

8. The method for constructing a channel map for a massive MIMO system based on LDM according to claim 1, wherein: The encoder of VAE consists of two 2D convolutional layers, one Flatten layer and four fully connected layers, and the decoder of VAE consists of three fully connected layers, one Reshape layer, two 2D deconvolutional layers and two 2D convolutional layers.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the method for constructing a channel map of a massive MIMO system based on LDM are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method for constructing a channel map of a massive MIMO system based on LDM are implemented.

Images