Construction method and system for channel fingerprint twinning

Abstract

The invention discloses a channel fingerprint twinning construction method and a channel fingerprint twinning construction system. Coarse-grained channel fingerprints and fine-grained channel fingerprints are regarded as physical objects and digital twinborn objects, and the relation between the physical objects and the digital twinborn objects is constructed by using an image super-resolution technology. According to the method, an evidence lower bound of fine-grained channel fingerprint twinborn likelihood is derived according to a variational reasoning and re-parameterization theory to serve as a target function, coarse-grained channel fingerprints are introduced to serve as side information to design a conditional generation type diffusion model for generating the fine-grained channel fingerprints, and the conditional generation type diffusion model can be deployed in a core computing center of channel fingerprint twinborn. In addition, the invention further introduces a one-time pruning algorithm and a multi-objective knowledge distillation technology to obtain a lightweight condition generation type diffusion model. The channel fingerprint twinning construction method provided by the invention not only ensures the reconstruction precision, but also has relatively strong expandability and generalization ability in a wireless communication scene with channel fingerprints with different fine granularities.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology and relates to a channel fingerprint twin construction method and system based on an image super-resolution method. Background Art

[0002] With the dramatic increase in antenna array size in MIMO-OFDM (Multi-input Multi-output Orthogonal Frequency Division Multiplexing) communication systems, the sharp rise in the number and density of user devices, and the utilization of wider bandwidth, 6G networks will face the challenge of handling ultra-large dimensional MIMO channels. Traditional pilot-based channel training and feedback methods to obtain real-time channel state information (CSI) will result in excessive overhead. In addition, traditional channel modeling methods usually adopt stochastic modeling based on specific assumptions and probability distribution of channel parameters, but these strict assumptions may be difficult to guarantee in highly dynamic and complex wireless propagation environments. It is worth noting that the wireless propagation environment is a key factor affecting channel parameters and the performance of wireless communication systems. Therefore, wireless communication with environment awareness has received increasing attention in academia and industry. Channel Fingerprint (CF), also known as channel knowledge map or channel knowledge base, is an emerging environment-aware wireless communication technology that provides channel knowledge of specific locations related to potential base stations for any base station to user equipment pair. Specifically, fine-grained CF is similar to a channel knowledge base containing all precise locations within the target communication area, marked with the precise locations of transmitters and receivers and their corresponding channel knowledge. This database stores channel-related knowledge at specific locations, such as channel gain, arrival / departure angles, etc., which can meet the challenges of real-time CSI acquisition and provide support for the design of wireless transmission technology.

[0003] Obviously, finer-grained channel fingerprints can help base stations obtain more accurate and location-specific channel information, which can be used to guide wireless transmission design to significantly improve communication gain. However, the degree of sophistication of the channel fingerprints obtained in actual communication scenarios is limited by the number of sensor nodes actually deployed, as well as the measurement and storage costs, resulting in most of the acquired channel fingerprints being coarse-grained. In this case, the performance of the wireless communication system will be degraded because the error in obtaining channel information cannot be ignored. At present, most of the related work focuses on constructing channel fingerprints using the characteristics of the physical propagation environment or the prior assumptions of the physical propagation model, and few studies focus on reconstructing fine-grained channel fingerprints from actually measured coarse-grained channel fingerprints. Therefore, the present invention defines the mapping of coarse-grained channel fingerprints to fine-grained channel fingerprint twins as channel fingerprint twins, and proposes a method for constructing channel fingerprint twins. Summary of the invention

[0004] Purpose of the invention: In view of the shortcomings of the prior art, the present invention discloses a method for constructing a channel fingerprint twin (ChannelFingerprint Twin), which can reconstruct finer-grained channel information from coarse-grained channel information collected from the sensor / channel test side, thereby realizing more accurate environmental perception-enabled wireless communications.

[0005] Technical solution: In order to construct fine-grained channel fingerprints, the present invention proposes a channel fingerprint twin construction solution with the conditional generative diffusion model (CGDM) as the core computing unit, which is used to efficiently construct the connection between coarse-grained channel fingerprints and fine-grained channel fingerprints. The present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a method for constructing a channel fingerprint twin, comprising the following steps:

[0007] Acquire the channel knowledge at some locations within the region of interest, and rearrange the acquired channel knowledge into a coarse-grained channel fingerprint with image structure through spatial discretization and grayscale conversion methods, and then transform the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twin into an image super-resolution problem;

[0008] Build a conditional generative diffusion model CGDM, given a training set containing paired coarse-grained and fine-grained channel fingerprints, train the CGDM model, and solve the optimal parameters of the model;

[0009] The finally trained CGDM model is used as the core computing center of the channel fingerprint twin, and the coarse-grained channel fingerprint is used as side information to reconstruct the corresponding fine-grained channel fingerprint twin to predict the channel knowledge of user devices at potential locations in the area of ​​interest.

[0010] Preferably, the spatial discretization and grayscale conversion is to grid the wireless propagation environment map in the spatial dimension, and use minimum-maximum normalization to convert the channel knowledge into an interval range of 0 to 1, so as to facilitate subsequent neural network processing.

[0011] Preferably, the objective function of training the CGDM model is the negative log-likelihood of the target data distribution conditional on the side information. The evidence lower bound of the objective function is derived using variational inference and reparameterization techniques, and then used as the loss function of the proposed CGDM for training. The loss function is expressed as:

[0012]

[0013] Where T is the number of time steps for noise diffusion, To find the expected operation, ε t is the noise at step t, ε θ (·,·,·) is a denoising neural network, θ is a model parameter, α t =1-β t , β t is the variance of the noise added in the tth step, is a Gaussian noise with a mean vector of 0 and a covariance matrix of the identity matrix I. G0 is the fine-grained channel fingerprint input to the CGDM model. The coarse-grained channel fingerprint G is used as the input of the model as side information. t is the fine-grained channel fingerprint intermediate latent variable between t=1 and t=T, and ||·||2 represents the L2 norm.

[0014] Preferably, the channel fingerprint twin construction method further includes:

[0015] A one-time pruning method is used to remove redundant CGDM model parameters and obtain a lightweight conditional generative diffusion model LiCGDM.

[0016] The multi-objective knowledge distillation technology is used to fine-tune the parameters of the pruned LiCGDM model to solve the problem of model performance degradation caused by one-time pruning.

[0017] The finally trained LiCGDM model is used as the core computing center of the channel fingerprint twin.

[0018] As a preferred method, a one-time pruning method is used to obtain a lightweight conditional generative diffusion model (LiCGDM) to further reduce memory consumption and computational delay pruning. The specific pruning objective function is expressed as:

[0019]

[0020] Among them, ε tea is the output of the CGDM model before pruning, For different network layers, is the set of all prunable network layers, Params(p i ) indicates the pth i The number of parameters of the layer, represents the amount of parameters that need to be pruned, ε(p i ) is the pth pruning i The output of the model after the layer, Indicates an undefined variable.

[0021] As a preferred method, multi-objective knowledge distillation technology is used to further adjust the weights of the pruned model to restore its performance. A task objective and two knowledge distillation objectives are combined as the loss function to achieve overall retraining of the model. The specific distillation loss function is expressed as:

[0022]

[0023] in, ε t is the noise at step t, ε S (·) and ε T (·) are the outputs of the denoising neural network in LiCGDM (student model) and CGDM (teacher model), respectively. and are the feature maps output by the i-th network layer of CGDM and LiCGDM, respectively, O With λ F are the distillation loss weights respectively.

[0024] Preferably, the denoising neural network framework of the CGDM model is implemented based on the U-Net framework, which is mainly composed of three stages: downsampling, intermediate layer and upsampling. Its main modules include a time embedding module based on sine-cosine position encoding, a residual module and a self-attention module.

[0025] As a preference, the three stages of downsampling, intermediate layer and upsampling in CGDM include multiple integrated residual and self-attention modules; define X, Γ tare the feature map corresponding to the channel fingerprint and the time embedding vector corresponding to time t respectively. The output X′ of the residual module is expressed as:

[0026]

[0027] Where, Ω = g(g(X)+f Ful (f Swi (f Ful (Γ t )))) is the intermediate quantity in the feature extraction process, g(X) = f Conv,3 (f Dro (f Swi (f Gn (X)))), f Conv,3 (·), f Conv,1 (·) are 3×3 and 1×1 convolution operations, respectively, f Ful (·) and f Swi (·) represent the linear layer and activation function layer respectively, f Dro (·) and f Gn (·) are the dropout layer and group normalization layer functions, respectively, c in and c out Represents the number of feature map channels of input and output respectively.

[0028] In a second aspect, the present invention provides a wireless communication channel fingerprint twin system, comprising:

[0029] The coarse-grained channel fingerprint acquisition module is used to obtain channel knowledge at some locations in the region of interest; through spatial discretization and grayscale conversion methods, the acquired channel knowledge is rearranged into a coarse-grained channel fingerprint with an image structure, and then the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twins is converted into an image super-resolution problem;

[0030] The channel fingerprint twin module is used to reconstruct the fine-grained channel fingerprint twin in the target area through the trained conditional generative diffusion model CGDM or the lightweight model LiCGDM which is based on the one-time pruning and multi-objective knowledge distillation method; the CGDM model is trained based on a training set containing paired coarse-grained and fine-grained channel fingerprints to obtain the optimal parameters of the model; the trained CGDM model or LiCGDM model serves as the core computing center of the channel fingerprint twin, and uses the coarse-grained channel fingerprint as side information to reconstruct the corresponding fine-grained channel fingerprint twin, and predicts the channel knowledge of user devices at potential locations in the area of ​​interest.

[0031] In a third aspect, the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of a method for constructing a channel fingerprint twin.

[0032] Beneficial effects: Compared with the prior art, the present invention transforms the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twin into an image super-resolution problem, designs a conditional generative diffusion model, uses coarse-grained channel fingerprint as side information, guides the direction of iterative denoising of the model, and then completes the construction of fine-grained channel fingerprint twin. At the same time, in order to reduce the model calculation amount and memory usage, the present invention also introduces a one-time pruning and multi-objective knowledge distillation method, which can accurately obtain the channel information of user equipment in ultra-high precision position in the target area with low computational complexity and storage complexity while ensuring the reconstruction accuracy. The obtained fine-grained channel fingerprint can further assist in the acquisition of real-time channel state information for large-scale wireless communications, and be used to guide wireless transmission design, thereby greatly improving the overall transmission efficiency of the system. In addition, the method proposed in the present invention has a strong generalization ability and is suitable for channel fingerprint reconstruction under different fine-grained requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of channel fingerprint twin construction according to an embodiment of the present invention;

[0034] Figure 2 A flowchart of channel fingerprint twin construction according to another embodiment of the present invention;

[0035] Figure 3 Schematic diagram of the deployment method of base station antennas in an embodiment of the present invention;

[0036] Figure 4 It is a framework diagram of the model adopted in the channel fingerprint twin core computing hub in an embodiment of the present invention;

[0037] Figure 5 This is a flow chart of an algorithm for solving the one-time pruning problem in an embodiment of the present invention;

[0038] Figure 6 Schematic diagram of the effect of the number of basic channels of the input feature map in the latent space on the performance effect of the implementation of this example in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0040] like Figure 1As shown, an embodiment of the present invention discloses a method for constructing a channel fingerprint twin. First, the channel knowledge at a partial position within the region of interest is obtained, and the acquired channel knowledge is rearranged into a coarse-grained channel fingerprint with an image structure through spatial discretization and grayscale conversion methods, and then the mapping task of the coarse-grained channel fingerprint to the fine-grained channel fingerprint twin is converted into an image super-resolution problem; then, a conditional generative diffusion model CGDM is constructed, and a training set containing paired coarse-grained and fine-grained channel fingerprints is given to train the CGDM model to solve the optimal parameters of the model; finally, the finally trained CGDM model is used as the core computing center of the channel fingerprint twin, and the coarse-grained channel fingerprint is used as side information to reconstruct the corresponding fine-grained channel fingerprint twin, so as to predict the channel knowledge of the user device at the potential position in the region of interest.

[0041] The embodiment of the present invention converts the mapping task of coarse-grained channel fingerprint (regarded as a physical object) to fine-grained channel fingerprint (regarded as a digital twin object) into an image super-resolution problem, designs a conditional generative diffusion model, introduces coarse-grained channel fingerprint as side information, and guides the model to perform directional iterative fine-grained channel fingerprint construction. In some more preferred embodiments, in order to reduce the amount of model calculation and memory usage, a one-time pruning and multi-objective knowledge distillation method is introduced to reduce the consumption of computing and storage resources. Figure 2 As shown, the specific steps include: (1) providing actually deployed sensors or channel test vehicles to collect channel knowledge at some locations in the target area. (2) through spatial discretization method and grayscale conversion, the channel knowledge measured in the target area is rearranged into a coarse-grained channel fingerprint with a 2D image structure, and then the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twin is converted into an image super-resolution problem. (3) Build a conditional generative diffusion model CGDM with a Markov chain structure, given a training set containing paired coarse-grained and fine-grained channel fingerprints, train the CGDM model, and solve the optimal parameters of the model. Specifically, in step (3), the variational inference technique and the re-parameterization technique are used to derive the evidence lower bound of the negative log-likelihood of the target data distribution conditional on the side information as the loss function for model training, that is, In order to reduce a large amount of memory consumption and delay, a one-time pruning method is used to remove redundant model parameters to obtain a lightweight model LiCGDM, which is beneficial to the actual deployment of the model. And further introduce multi-objective knowledge distillation technology to fine-tune the lightweight model parameters after pruning obtained by the above training to solve the problem of model performance degradation caused by one-time pruning. In this embodiment, the data set is divided into a training set (5000), a validation set (1000) and a test set (1000) in a ratio of 5:1:1. The above model is trained using the Adam optimization algorithm and unsupervised training. The batchsize set during the training process is 16, and the learning rate is set to 0.00005. This training strategy is used for more than 500,000 iterations, and the exponential moving average algorithm is introduced at the 5000th iteration. The attenuation factor is set to 0.9999. In addition, the model adopts a dropout strategy and the discard rate is set to 0.1. During the training process, the weight parameters of the model are adjusted with the validation set, and the final performance of the model is tested with the test set. (4) The above-mentioned trained model or the pruned lightweight model is used as the core computing center of the channel fingerprint twin, and is used for the efficient construction of fine-grained channel fingerprint twins in the target area in the field of wireless communications.

[0042] The specific steps of the embodiment of the present invention are described below in conjunction with specific scenarios:

[0043] 1. System Model

[0044] Consider a large-scale multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDMOFDM) communication scenario in a square communication area. The location of the user device is defined as like Figure 3 As shown, the base station (BS) is equipped with a r =N r,v ×N r,h A uniform linear array (ULA) consisting of antenna units is arranged at half-wavelength intervals to serve M single-antenna users in a cell, where N r,v and N r,h denotes the number of antennas in the vertical column and horizontal row respectively. Assume that in the wireless propagation environment considered, the distance from the BS to the antenna at x m User devices exist The system uses N physical propagation paths. c OFDM modulation of subcarriers, the spacing between subcarriers is Δf . Usually, in N c The center of the subcarrier is selected as N k The active subcarriers are used for signal transmission, and the remaining subcarriers are used as guard bands, and the active subcarrier set is defined as

[0045] The spatial-frequency domain channel response between the base station and the user equipment on the active subcarrier K is modeled as:

[0046]

[0047] in, and They respectively represent the positions corresponding to m The complex gain and propagation delay of the lth path of the user equipment, represents the Kronecker product. Normalized horizontal angle With elevation Related, normalized vertical angle Then through the relation With elevation Correspondingly, the azimuth Defined as Guide vector and It can be expressed as:

[0048]

[0049] in, and Typically, the baseband signal received at the terminal is described as:

[0050] y(x m )=h H (x m )s+z(x m ), (5)

[0051] Where s is the power P X The transmitted signal symbol, z(x m ) is the additive noise with single-sided power spectral density N0.

[0052] The channel in equation (1) typically contains large-scale fading (path loss and shadowing) and small-scale fading, the former of which typically varies over distances in meters, while the latter fluctuates over distances in the order of wavelengths. Considering that massive MIMO-OFDM systems operate at wavelengths in the millimeter or centimeter scale, estimating this small-scale fading component would require knowing the location of the terminal with millimeter-level accuracy. This is impractical for current technologies, such as global navigation satellite systems. Therefore, it is usually assumed that the effects of small-scale fading have been averaged out, which is quite realistic in scenarios where channel measurements are made while the test vehicle is moving.

[0053] 2. Channel Gain Model

[0054] The channel gain attenuation observed at the terminal may be dominated by various environmental factors, such as propagation losses from different paths, reflections from buildings, and diffraction. Therefore, among these effects, the relatively slowly varying parts contribute to the channel gain function, which is then expressed as Indicates the user's location The large-scale signal attenuation measured on the surface is , where e represents the propagation environment. In addition, the small-scale effect is usually modeled as a complex Gaussian random variable with unit variance Therefore, the received baseband signal in formula (5) can be expressed in a more general form as:

[0055]

[0056] Among them, the average received energy of each symbol can be expressed as:

[0057]

[0058] Where B is the signal bandwidth. m The channel gain of the user equipment at , in dB, is defined as:

[0059] G(e,x m )=(P Y ) dB -(P X ) dB , (8) which reflects the mThe received power change of the user equipment at the propagation environment e. Then, the set of channel gains at potential locations is called a channel fingerprint (CF). It is worth noting that if the channel gain at any location in the area of ​​interest can be accurately obtained, an embodiment of the present invention can construct a fine-grained CF, which is particularly important for the design of transceivers in large-scale MIMO systems. It should be noted that in this embodiment, channel gain is used as an example of channel fingerprint. It can be understood that the channel fingerprint can be other channel knowledge such as path loss and channel delay.

[0060] 3. Problem Statement

[0061] Since the communication area considered is a square area The embodiment of the present invention can perform spatial discretization along the X-axis and the Y-axis. Specifically, considering the size of the region of interest is A resolution factor σ is defined so that the minimum spacing unit in the spatial discretization process is and Each spatial grid defines Λ i,j ,in The (i,j)th spatial grid can be expressed as:

[0062]

[0063] Given a resolution factor σ, the CF corresponding to the potential user device location within the target geometric area can be rearranged into a two-dimensional tensor and recorded as:

[0064] [G] i,j =G(e,Λ i,j ). (10)

[0065] Note that when the target area size is When the resolution factor is σ, the number of points required for interpolation is As the resolution factor σ increases, the complexity of the traditional interpolation algorithm increases. This poses a significant challenge to building fine-grained CF in practical scenarios. On the other hand, increasing the density of test nodes can improve the spatial resolution of CF, which can be achieved by significantly increasing the number of sensor nodes or using test vehicles to collect channel knowledge at smaller geographical intervals. However, both approaches may be impractical due to high hardware and labor costs. Define a coarse-grained factor σ LR and a fine-grained factor σ HR , σ HR Usually σ LR 2 times the power, for example 2 =, 4 ×, 8 ×. For the resolution factor σ LR The CF is defined as GLR , target area is discretized into pieces of size σ LR ×σ LR A two-dimensional tensor, G HR is also defined in a similar way. The task of CF twins focuses on LR Reconstructing fine-grained CFG HR , especially in scenarios constrained by measurement cost, privacy, or security.

[0066] In the typical image super-resolution (ISR) task in the field of computer vision, the goal is to reconstruct a high-resolution (HR) image from a given low-resolution (LR) corresponding image to enhance the details and overall quality of the image. It can be seen that the task of the embodiment of the present invention is consistent with the ISR task, so the embodiment of the present invention analyzes the fine-grained CF reconstruction problem from the perspective of ISR. Specifically, the embodiment of the present invention converts G LR The elements of the matrix are regarded as pixels, and the coarse-grained CFG LR As an LR image, the fine-grained CFG HR is regarded as a HR image. Then, the goal is to learn a specific mapping relationship from a given LR CFG LR Efficiently reconstruct HR CFG HR ,Right now:

[0067]

[0068] Where Θ is a mapping , where U is the parameter set of the target image, and U is the number of training samples. However, this task poses a classic and challenging inverse problem involving the efficient reconstruction of fine details from a given LR CF. Considering that the conditional distribution of the HR CF output given the LR CF input usually does not follow a simple parametric distribution, most ISR regression methods based on feedforward neural networks usually perform poorly at high magnification factors and cannot accurately restore details. Deep generative models have achieved success in learning the complex empirical distribution of target data. Specifically, if implicit prior information of the HR CF distribution can be learned, such as the gradient of the logarithmic density of the data, it can be converted to the target CF distribution through iterative sampling steps from the standard normal distribution, similar to Langevin dynamics. Therefore, the mapping relationship in formula (11) can be solved by optimizing the following objective:

[0069]

[0070] in, represents the gradient of the logarithmic density of HR CF, also known as the Stein score, P Θ Represents the learned density. Note that the noise learned by the traditional GDM has been shown to be equivalent to the Stein score, which makes it possible to generate HR CF samples that are consistent with the target data distribution. In order to accurately reconstruct an HR CF from a given LR CF, the task is pathological, which means that the reconstruction process may produce multiple possible solutions for the HR CF. In other words, different HR CFs may be mapped to the same LR CF. Therefore, unlike the traditional GDM, which starts from a pure Gaussian noise tensor, the introduction of additional source signals as side information (also a guiding condition) is crucial to achieving the optimal solution. To this end, the method of the embodiment of the present invention combines the learned prior information and introduces the LR CFG LR As an additional source signal to guide the iterative optimization process.

[0071] 4. Fine-grained channel fingerprint twin construction method based on CGDM

[0072] 1. Overall design of CGDM

[0073] To simplify the notation, G LR and G HR Use And G represents. Given a paired LR CF input and HR CF output dataset, it is defined as Represents an unknown distribution These datasets are usually collected from sensor nodes and test vehicles, and have different resolution factors depending on the specific scenario, for example, σ LR and σ HR In the task of the embodiment of the present invention, the focus is on learning through a directed iterative refinement process guided by source information The approximate values ​​of the parameters make The mapping to G becomes possible. In view of the powerful implicit prior learning ability of GDM, a CGDM is designed in the embodiment of the present invention to promote the generation of G.

[0074] Specifically, CGDM can generate a target HR CF defined as G0 through T refinement time steps. CGDM is derived from a pure Gaussian noise At the beginning, according to the source signal and the prior information learned during training, that is, the conditional distribution This initial input is iteratively refined, where 0 and I represent the mean vector to be 0 and the covariance matrix to be the identity matrix. As each time step t evolves, it produces a series of functions defined as {G T-1 ,G T-2 ,...,G0} output CF, and finally get the target Specifically, the distribution of intermediate CFs in the iterative refinement chain is controlled by a forward diffusion process that gradually adds noise to the output CF via a fixed Markov chain, defined as q(G t |G t-1 ). The model of the embodiment of the present invention attempts to The inverse Markov chain conditioned by α iteratively recovers the signal from the noise, thereby reversing the Gaussian diffusion process. To achieve this, the embodiment of the present invention optimizes the denoising neural network ε by using the following objective function formula (35): θ (·) to learn the backward chain. The trained CGDM takes the LRCF and the noisy image as input to estimate the noise, and after T refinement time steps, generates the target HR CF.

[0075] 2. Forward diffusion process starting from HR CF

[0076] Consider HR CF samples drawn from a distribution of interest, defined as G0~q(G). GDM uses a fixed diffusion process q(G 1:T |G0), a process involving relatively high-dimensional latent variables. This process defines a forward diffusion mechanism, a deterministic Markov chain, where Gaussian noise is gradually introduced into the samples over T time steps. The noise level at each step is represented by the variance Specifically, the forward diffusion process is defined as:

[0077]

[0078] Note that the forward diffusion subprocess, also called the encoder, does not need to be learned at each time step t. Instead, it is a fixed and predefined linear Gaussian model that can be expressed as

[0079]

[0080] Among them, ε is Gaussian noise, and its distribution is Define α t =1-β t and At the transfer density q(G t |G t-1 ), combined with formulas (14) and (15), the embodiment of the present invention can use the reparameterization technique to sample G in a closed form at any given time step t. t :

[0081]

[0082] Usually, the variance change is generally set to β1<β2<...<β T When β TWhen infinitely close to 1, for any initial state G0, G T Converges to a standard Gaussian distribution, that is,

[0083] 3. Reverse process of conditional LR CF

[0084] For the traditional GDM, the inverse process can be regarded as a decoding process. At each time step t, G t is denoised and restored to G t-1 , the transition probability of each step is expressed as p(G t-1 |G t ). Based on the Markov density transition property, the joint distribution of the reverse process is expressed as:

[0085]

[0086] However, using Bayes' theorem to derive p(G t-1 |G t ), which reveals that its denominator contains an integral, so there is no closed form solution. Therefore, it is necessary to use a denoising neural network with parameter θ to approximate these conditional probabilities in order to perform the reverse process. Since the reverse process is also a Markov chain, it can be expressed as:

[0087]

[0088] In the task of the embodiment of the present invention, unlike the traditional GDM, the denoising model ε θ (·) LR CF The side information in the form of is conditioned to guide it from the Gaussian distribution G T Step by step, denoise and generate HR CFG0. Therefore, formulas (19) and (20) need to be rewritten as:

[0089]

[0090] 4. CGDM Optimization Based on Evidence Lower Bound

[0091] In order to make the network model ε θ (·) can effectively approximate the reverse process, which requires the optimization of the model parameters θ for a specific goal. Mathematically, the latent variable G 1:T and conditions in The observed sample G0 can be distributed conditionally In generative modeling, a common likelihood-based approach is to maximize the conditional joint probability distribution of all observed samples. However, the embodiment of the present invention can only access the observed sample G0, but cannot access the latent variable G 1:TTherefore, the embodiment of the present invention needs to maximize the conditional marginal distribution Right now

[0092]

[0093] In the framework of variational inference, conditional on The likelihood of the observed sample G0, i.e., the evidence, enables the embodiment of the present invention to derive the evidence lower bound (ELBO, Evidence Lower Bound) as a proxy objective function for optimizing CGDM:

[0094]

[0095] in, From Jensen inequality. To further derive ELBO, formula (24) can be rewritten as:

[0096]

[0097] Among them, D KL (·) is the KL divergence (KL, Kullback-Leibler). Then, the parameters θ of CGDM can be learned by optimizing the negative ELBO:

[0098]

[0099] in, is the objective function for training CGDM.

[0100] for The present invention embodiment notices is a constant and can be excluded from the optimization, while It can be approximated and optimized by Monte Carlo estimation. From the above analysis, it can be seen that the training objective of CGDM is mainly composed of It can be found that the training goal of CGDM is to be as close as possible to the transfer density in the reverse process, so that and The KL divergence between is the smallest. Therefore, the embodiment of the present invention needs to obtain q(G t-1 |G t ,G0) and variance Specifically, q(G t-1 |G t ,G0) can be expressed as

[0101]

[0102] Combining formulas (16) and (17), (27) can be further rewritten as:

[0103]

[0104] Based on (28), q(G t-1 |G t ,G0) can be explicitly expressed as:

[0105]

[0106] Usually, the variance Set to a constant Therefore, in order to ensure that the denoised transfer density is close to the true denoised transfer density, the embodiment of the present invention can simplify the optimization of the KL divergence term, thereby minimizing the difference between the means of the two distributions. In this case, the embodiment of the present invention only needs to make the CGDM predict

[0107]

[0108] The KL divergence is defined as:

[0109]

[0110] Where d represents the dimension of x. Substituting equations (29), (31) and (33) into equation (25) get:

[0111]

[0112] Substituting formula (25) and (34) into formula (26), the objective function of CGDM is It can be further simplified to:

[0113]

[0114] Based on the trained CGDM, given any noise-contaminated image G t , the trained CGDM can utilize side information To predict the noise ε t , and then the target image is obtained by converting formula (16) The approximate value of is:

[0115]

[0116] By reparameterization, formula (36) represents the result of iterative refinement, and each iterative step in the CGDM proposed in the embodiment of the present invention is given by the following formula:

[0117]

[0118] in, The noise estimation step in equation (37) is analogous to the Langevin dynamics step in score-based generative models, which is equivalent to estimating the first-order derivative of the log-likelihood of the observed sample, also known as the gradient or Stein score.

[0119] 5. CGDM’s network architecture

[0120] The backbone network used by the conditional denoising neural network in CGDM is a variant of the Unet network, which consists of three main stages: downsampling (Down), middle layer (Mid), and upsampling (Up). The key components of each stage include the time embedding block, Res + Block, self-attention block, downsampling block and upsampling block. The architecture of the model is as follows Figure 4 As shown, the specific module contents will be introduced one by one.

[0121] 1) In order to encode the time parameters [1,...,t,...,T] in the diffusion process, a time embedding coding method based on sinusoidal position coding is used, and the resulting time embedding Γ t Effectively captures the temporal characteristics. Definition Represents the jth component of the time embedding vector corresponding to time t:

[0122]

[0123] Where j = 0, 1, 2, ..., c time / 2-1, c time is the dimension of the time embedding vector. Then the embedding vector of time constant t can be expressed as:

[0124]

[0125] in, Using formula (39), for any given time Γ t+Δt The embedding vector of can be obtained by the following linear transformation:

[0126]

[0127] M Δt is a linear transformation matrix, defined as:

[0128]

[0129] in, In order to enable the model to capture more complex temporal features, the temporal embedding is further enhanced by two fully connected layers and one activation layer, denoted as Γ′ t =f Ful (f Swi (fFul (Γ t ))). f Ful (·)f Swi (·) denote the fully connected layer and the swish activation function layer, respectively.

[0130] 2)Res + Module: Design a deeper feature extraction network, aiming to learn higher-level semantic information from CGM. Nevertheless, deep neural networks often experience model degradation due to challenges such as gradient vanishing and exploding. To this end, residual connections are introduced as a way to alleviate these problems. Specifically, define Γ t are feature maps and time embedding vector inputs, Res + The output of the module can be expressed as:

[0131]

[0132] Among them, f Conv,1 (·) represents a 1×1 convolution operation. Defined as:

[0133] Ω=g(g(X)+f Ful (f Swi (f Ful (Γ t )))), (43)

[0134] Where g(X) = f Conv,3 (f Dro (f Swi (f Gn (X)))), f Conv,3 (·) is a 3×3 convolution operation. f Dro (·), f Gn (·) are the dropout layer and group normalization layer, respectively.

[0135] 3) Self-attention module: In order to capture global structural information, a self-attention mechanism is introduced to improve the interaction between global and local features. Specifically, a normalized attention matrix is ​​introduced to represent different degrees of attention to the input, and larger weights are assigned to more important input components. The input is weighted according to the attention weights indicated in the attention matrix to generate the final output. Given an input d m represents the number of image blocks, d n Represents the feature dimension of each block, for z j Perform three different linear transformations:

[0136] k j =z j Wk ,j=1,...,d n ,(44)

[0137] q j =z j W q ,j=1,...,d n ,(45)

[0138] v j =z j W v ,j=1,...,d n ,(46)Among them, and are the key, query, and value vectors, respectively. and They represent their respective trainable transformation matrices. Specifically, the weight distribution function is composed of k j and q j′ OK. A higher correlation This means that the jth input block z j The features of have higher importance for the j′th output block. Usually, this correlation can be obtained based on the input Z and the matrix W k and W q For the sake of clarity, the matrix form of formulas (44)-(46) is expressed as K=ZW k , Q=ZW q , V=ZW v ,in and Using K and Q, we can get the attention matrix

[0139]

[0140] Among them, Softmax(Z)=exp(z j ) / ∑exp(z j ), is a scaling factor. Finally, the output Z′ of the self-attention block can be expressed as:

[0141]

[0142] in,

[0143] Note that the three stages in the proposed conditional denoising neural network all contain N RA ≥2 Res + And the corresponding self-attention module, such as Figure 4 shown.

[0144] 5. LiCGDM based on one-shot pruning and multi-objective knowledge distillation

[0145] 1. Model lightweighting method based on one-time pruning

[0146] Considering that CGDM sacrifices a lot of memory consumption and latency to obtain excellent performance, its deployment on personal computers and even mobile devices is highly restricted. To this end, the embodiment of the present invention utilizes the additivity of the network layer and designs a one-time pruning method to obtain LiCGDM.

[0147] Given a specific pruning ratio, the goal of layer pruning is to remove a subset of prunable layers from the network structure of CGDM while minimizing the degradation of model performance. The embodiment of the present invention minimizes the mean square error (MSE) loss between the final output of the original network CGDM (called the "teacher") and the final output of the pruned network LiCGDM (called the "student") as the pruning goal. Specifically, the pruning objective function is expressed as:

[0148]

[0149] Among them, ε tea is the output of the CGDM model before pruning, is the set of all prunable network layers, Indicates the i-th prunable network layer, Params(p i ) indicates the pth i The number of parameters of the layer, Indicates pruning The network output after the layer, represents the amount of parameters that need to be pruned, ε(p i ) is the pth pruning i The model output after the layer, is an uncertain variable. However, solving the above equation is an NP-hard problem, and the upper bound of the objective function needs to be obtained by using the triangular inequality:

[0150]

[0151] in, Therefore, the optimization problem of formula (49) is transformed into the problem of minimizing the upper bound:

[0152]

[0153] Although equation (51) is still an NP-hard problem, by utilizing the additivity of the network layer, the output distortion caused by multiple perturbations can be approximated as the sum of the distortions caused by each individual perturbation, that is, the additivity can be expressed as:

[0154]

[0155] Substituting equation (52) into equation (51), we can obtain a new approximate proxy objective:

[0156]

[0157] From the above formula, we can see that the pruning standard is completely determined by OK, so we only need to calculate Optimize and solve. It is worth noting that the time complexity of calculating this objective function for each network layer is Therefore, this problem can be converted into a 0-1 knapsack problem, and the classic dynamic programming algorithm can be used to solve the 0-1 knapsack problem. The specific algorithm flow is as follows Figure 5 shown.

[0158] 2. Network model enhancement method based on multi-objective knowledge distillation

[0159] Usually, deleting some network layers from the teacher network will lead to a decrease in model performance. Therefore, the embodiment of the present invention adopts a multi-objective knowledge distillation technology to further adjust the weights of the pruned model to achieve performance recovery. A task goal and two knowledge distillation goals are combined as a loss function to achieve overall retraining of the model. The loss function is expressed as:

[0160]

[0161] in, ε S (·) and ε T (·) are the outputs of the denoising neural network in LiCGDM and CGDM, respectively. and are the feature maps output by the i-th network layer of CGDM and LiCGDM, respectively, O With λ F are the distillation loss weights respectively. It is worth noting that by default, λ0 and λ F is set to 1 for training.

[0162] VI. Implementation Effect

[0163] In order to enable those skilled in the art to better understand the solution of the present invention, the following provides a comparison of the prediction performance results of the channel fingerprint prediction method in the embodiments of the present invention and the existing methods under several specific system configurations.

[0164] First, the system configuration of ultra-large-scale MIMO-OFDM is given as follows: The number of base station antennas is N r,v ×Nr,h =8×8, target area size A=128 meters×128 meters, center frequency is 2.4GHz, subcarrier spacing is 15kHz, base station height is set to 25 meters, and user height is set to 1.5 meters. The embodiment of the present invention uses the QuaDRiGa generator (version 2.6.1) to simulate the real wireless communication system, and considers the 5G NR typical urban micro-cell scenario "3GPP 38.901UMa NLOS" as the specific communication scenario of the simulation experiment. Secondly, the hyperparameter configuration of the model is given as follows: N RA =2, dropout probability d r =0.1, learning rate α = 0.00005, diffusion steps T = 1000, 61 = 10 -4 , 6 T =0.01, the batch size is 16, the number of training iterations is 500000, the exponential moving average algorithm is introduced at the 5000th iteration, the decay factor is set to 0.9999, the coarse-grained factor and the fine-grained factor σ LR With σ HR They are 32 and 128 respectively.

[0165] Figure 6 The influence of the number of basic channels of the input feature map in the implicit space on the performance of the proposed model is shown. It can be found that when the number of basic channels c1 of the input feature map in the implicit space increases from 8 channels to 128 channels, the proposed model achieves better performance in terms of convergence speed and training stability. The analysis believes that this is because each feature channel can capture different types of information. The increase in feature channels can improve the expressiveness of the model and aggregate the underlying simple features and high-level abstract features, thereby improving the training efficiency of the model.

[0166] Table 1 Quantitative comparison of performance on fine-grained channel fingerprint twin reconstruction task

[0167]

[0168] Table 1 gives a quantitative comparison of the reconstruction performance of CGDM and LiCGDM in the embodiments of the present invention with the existing methods. The compared methods are the SRGAN and SRGAN-MSE methods proposed in the literature: "Photo-realistic single image super-resolution using agenerative adversarial network," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017, pp. 4681–4690, the C-VAE method in the literature "Auto-encoding variational Bayes," in Proc. Int. Conf. Learn. Represent. (ICLR), Banff, AB, Canada, Apr. 2014, pp. 1-14, and the DRRN method in the literature "Image super-resolution via deep recursive residual network," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017, pp. 3147–3155. The CGDM, LiCGDM and SRGAN-MSE, SRGAN, C-VAE and DRRN methods of the present invention are compared on the ×4 super-resolution channel fingerprint dataset (32 2 →128 2 ) are used to compare the performance of the two models, and the performance indicators involved include normalized mean square error (NMSE), mean square error (MSE), peak signal-to-noise ratio (PSNR) and structure similarity index measure (SSIM). The model parameters involved in CGDM are set as: c1=64, N RA =2, c1:c2:c3:c4:c5=1:2:4:8:16, and the pruning ratio used by LiCGDM is 35%. It can be found that compared with SRGAN, the NMSE and MSE of CGDM are reduced by 65.236×10 -5and 9.542, PSNR and SSIM are improved by 11.747 and 0.006 respectively, and the reconstruction accuracy is significantly superior to the other comparison methods. In addition, compared with CGDM, LiCGDM has a slight decrease in performance indicators, but the decrease is still within an acceptable range, which confirms that the proposed lightweight method effectively compresses CGDM to enable it to be quickly deployed on personal computers and even mobile devices.

[0169] Based on the same inventive concept, a wireless communication channel fingerprint twin system disclosed in an embodiment of the present invention includes: a coarse-grained channel fingerprint acquisition module, which is used to obtain channel knowledge at some locations in the area of ​​interest; through spatial discretization and grayscale conversion methods, the acquired channel knowledge is rearranged into a coarse-grained channel fingerprint with an image structure, and then the mapping task from the coarse-grained channel fingerprint to the fine-grained channel fingerprint twin is converted into an image super-resolution problem; a channel fingerprint twin module, which is used to reconstruct the fine-grained channel fingerprint twin in the target area through a trained conditional generative diffusion model CGDM or a model LiCGDM that is lightweighted by a one-time pruning and multi-objective knowledge distillation method; wherein the CGDM model is trained based on a training set containing paired coarse-grained and fine-grained channel fingerprints to obtain the optimal parameters of the model; the trained CGDM model or LiCGDM model is used as the core computing center of the channel fingerprint twin, and the coarse-grained channel fingerprint is used as side information to reconstruct the corresponding fine-grained channel fingerprint twin, and predict the channel knowledge of the user equipment at the potential location in the area of ​​interest. For specific implementation details, please refer to the aforementioned embodiment, which will not be repeated here.

[0170] Based on the same inventive concept, a computer program product disclosed in an embodiment of the present invention includes a computer program / instruction, which implements the steps of a channel fingerprint twin construction method when executed by a processor. The program / instruction code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program / instruction 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 / instruction code implements the steps of the method of the present invention when executed by a processor or controller. The program / instruction 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 entirely on a remote machine or server. What is not described in detail in the present invention is a well-known technology for those skilled in the art.

[0171] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways without exceeding the spirit and scope of this application. The current embodiment is only an illustrative example and should not be used as a limitation. The specific content given should not limit the purpose of this application. For example, some features can be ignored or not executed.

[0172] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for constructing a channel fingerprint twin, characterized in that: The method comprises the following steps: Acquire the channel knowledge at some locations within the region of interest, and rearrange the acquired channel knowledge into a coarse-grained channel fingerprint with image structure through spatial discretization and grayscale conversion methods, and then transform the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twin into an image super-resolution problem; Build a conditional generative diffusion model CGDM, given a training set containing paired coarse-grained and fine-grained channel fingerprints, train the CGDM model, and solve the optimal parameters of the model; The finally trained CGDM model is used as the core computing center of the channel fingerprint twin, and the coarse-grained channel fingerprint is used as side information to reconstruct the corresponding fine-grained channel fingerprint twin to predict the channel knowledge of user devices at potential locations in the area of ​​interest.

2. A method for constructing a channel fingerprint twin according to claim 1, characterized in that: The spatial discretization and grayscale conversion is to grid the wireless propagation environment map in the spatial dimension, and use minimum-maximum normalization to convert the channel knowledge into an interval range of 0 to 1, so as to facilitate the subsequent neural network processing.

3. A method for constructing a channel fingerprint twin according to claim 1, characterized in that: The objective function of training the CGDM model is the negative log-likelihood of the target data distribution conditional on the side information. The evidence lower bound of the objective function is derived using variational inference and reparameterization techniques, and then used as the loss function of the proposed CGDM for training. The loss function is expressed as: Where T is the number of time steps for noise diffusion, To find the expected operation, ε t is the noise at step t, ε θ (·,·,·) is a denoising neural network, θ is a model parameter, β t is the variance of the noise added in the tth step, is a Gaussian noise with a mean vector of 0 and a covariance matrix of the identity matrix I. G0 is the fine-grained channel fingerprint input to the CGDM model. The coarse-grained channel fingerprint G is used as the input of the model as side information. t is the fine-grained channel fingerprint intermediate latent variable between t=1 and t=T, and ||·||2 represents the L2 norm.

4. A method for constructing a channel fingerprint twin according to claim 1, characterized in that: Also includes: A one-time pruning method is used to remove redundant CGDM model parameters and obtain a lightweight conditional generative diffusion model LiCGDM. The multi-objective knowledge distillation technology is used to fine-tune the parameters of the pruned LiCGDM model to solve the problem of model performance degradation caused by one-time pruning. The finally trained LiCGDM model is used as the core computing center of the channel fingerprint twin.

5. A method for constructing a channel fingerprint twin according to claim 4, characterized in that: A one-time pruning method is used to obtain the lightweight conditional generative diffusion model LiCGDM. The specific pruning objective function is expressed as: Among them, ε tea is the output of the CGDM model before pruning, For different network layers, is the set of all prunable network layers, Params)p i ) indicates the pth i The number of parameters of the layer, represents the amount of parameters that need to be pruned, ε(p i ) is the pth pruning i The output of the model after the layer, Indicates an undefined variable.

6. A method for constructing a channel fingerprint twin according to claim 4, characterized in that: The multi-objective knowledge distillation technology is used to further readjust the weights of the pruned model LiCGDM to achieve performance recovery. Specifically, a task objective and two knowledge distillation objectives are combined as the loss function to achieve overall retraining of the model. The loss function is expressed as: in, ε t is the noise at step t, ε S (·) and ε T (·) are the outputs of the denoising neural network in LiCGDM and CGDM, respectively. and are the feature maps output by the i-th network layer of CGDM and LiCGDM, respectively, O With λ F are the distillation loss weights respectively.

7. A method for constructing a channel fingerprint twin according to claim 1, characterized in that: The denoising neural network framework of the CGDM model is implemented based on the U-Net framework, which mainly consists of three stages: downsampling, intermediate layer and upsampling. Its main modules include the temporal embedding module based on sine-cosine position encoding, the residual module and the self-attention module.

8. A method for constructing a channel fingerprint twin according to claim 1, characterized in that: The three stages of downsampling, intermediate layer and upsampling in CGDM include multiple integrated residual and self-attention modules; define X, Γ t are the feature map corresponding to the channel fingerprint and the time embedding vector corresponding to time t respectively. The output X′ of the residual module is expressed as: Where, Ω = g(g(X)+f Ful (f Swi (f Ful (Γ t ))) is the intermediate quantity in the feature extraction process, g(X) = f Conv,3 (f Dro (f Swi (f Gn (X)))), f Conv,3 (·), f Conv,1 (·) are 3×3 and 1×1 convolution operations, respectively, f Ful (·) and f Swi (·) represent the linear layer and activation function layer respectively, f Dro (·) and f Gn (·) are the dropout layer and group normalization layer functions, respectively, c in and c out Represents the number of feature map channels of input and output respectively.

9. A wireless communication channel fingerprint twin system, characterized in that: include: A coarse-grained channel fingerprint acquisition module is used to obtain channel knowledge at some locations in the area of ​​interest; Through spatial discretization and grayscale conversion methods, the acquired channel knowledge is rearranged into coarse-grained channel fingerprints with image structure, and then the mapping task from coarse-grained channel fingerprints to fine-grained channel fingerprint twins is converted into an image super-resolution problem; The channel fingerprint twin module is used to reconstruct the fine-grained channel fingerprint twin in the target area through the trained conditional generative diffusion model CGDM or the lightweight model LiCGDM which is based on the one-time pruning and multi-objective knowledge distillation method; the CGDM model is trained based on a training set containing paired coarse-grained and fine-grained channel fingerprints to obtain the optimal parameters of the model; the trained CGDM model or LiCGDM model serves as the core computing center of the channel fingerprint twin, and uses the coarse-grained channel fingerprint as side information to reconstruct the corresponding fine-grained channel fingerprint twin, and predicts the channel knowledge of user devices at potential locations in the area of ​​interest.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of a method for constructing a channel fingerprint twin according to any one of claims 1-8 are implemented.

Images