Method for constructing channel fingerprint twin, and system

Abstract

Disclosed in the present invention are a method for constructing a channel fingerprint twin, and a system. In the present invention, a coarse-grained channel fingerprint and a fine-grained channel fingerprint are respectively regarded as a physical object and a digital twin object, and image super-resolution technology is used to construct a relationship between the coarse-grained channel fingerprint and the fine-grained channel fingerprint. In the present invention, on the basis of variational inference and a re-parameterization theory, an evidence lower bound of a fine-grained channel fingerprint twin likelihood is derived to serve as an objective function, and the coarse-grained channel fingerprint is introduced as side information to design a conditional generative diffusion model for generating the fine-grained channel fingerprint, wherein the conditional generative diffusion model can be deployed in a core computing center of a channel fingerprint twin. In addition, in the present invention, a one-shot pruning algorithm and multi-objective knowledge distillation technology are further introduced to acquire a lightweight conditional generative diffusion model. The method for constructing a channel fingerprint twin provided in the present invention not only ensures the reconstruction accuracy, but also has relatively strong scalability and generalization capability in wireless communication scenarios with different fine-grained channel fingerprints.

Description

A method and system for constructing channel fingerprint twins Technical Field

[0001] This invention belongs to the field of communication technology and relates to a method and system for constructing channel fingerprint twins based on image super-resolution methods. Background Technology

[0002] With the dramatic increase in antenna array size, the surge in the number and density of user equipment (MAUs), and the utilization of wider bandwidth in MIMO-OFDM communication systems, 6G networks will face the challenge of handling ultra-high-dimensional MIMO channels. Traditional pilot-based channel training and feedback methods for obtaining real-time channel state information (CSI) will result in excessive overhead. Furthermore, traditional channel modeling methods typically employ stochastic modeling based on specific assumptions and probability distributions of channel parameters; however, these rigorous assumptions may be difficult to guarantee in highly dynamic and complex wireless propagation environments. It is noteworthy that the wireless propagation environment is a key factor affecting channel parameters and the performance of wireless communication systems; therefore, environment-aware wireless communication is receiving increasing attention in academia and industry. Channel fingerprinting (CF), also known as a channel knowledge map or channel knowledge base, is an emerging environment-aware wireless communication technology that provides any base station-to-user equipment pair with channel knowledge specific to the potential base station location. Specifically, fine-grained channel knowledge (CF) is similar to a channel knowledge base containing all precise locations within the target communication area, labeled with the precise locations of transmitters and receivers and their corresponding channel information. This database stores location-specific channel-related knowledge, such as channel gain and arrival / departure angles, addressing the challenges of real-time CSI acquisition and supporting the design of wireless transmission technologies.

[0003] It is evident that finer-grained channel fingerprints can help base stations obtain more accurate, location-specific channel information, guiding wireless transmission design to significantly improve communication gain. However, in real-world communication scenarios, the fineness of the acquired channel fingerprints is limited by the number of deployed sensor nodes and measurement and storage costs, resulting in most acquired channel fingerprints being coarse-grained. In this case, the inability to ignore errors in acquiring channel information leads to a degradation in the performance of the wireless communication system. Current work largely focuses on constructing channel fingerprints using physical propagation environment characteristics or prior assumptions of physical propagation models, with few studies focusing on reconstructing fine-grained channel fingerprints from actual measured coarse-grained channel fingerprints. Therefore, this invention defines the mapping from 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: To address the shortcomings of existing technologies, this invention discloses a method for constructing a Channel Fingerprint Twin, which can reconstruct finer-grained channel information from coarse-grained channel information collected from the sensor / channel test side, thereby achieving more accurate environmentally-aware wireless communication.

[0005] Technical Solution: To construct fine-grained channel fingerprints, this invention proposes a channel fingerprint twin construction scheme with a Conditional Generative Diffusion Model (CGDM) as the core computational unit. This scheme efficiently constructs the connection between coarse-grained and fine-grained channel fingerprints. The invention provides the following technical solution:

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

[0007] We acquire channel knowledge at certain locations within a region of interest, and then rearrange the acquired channel knowledge into a coarse-grained channel fingerprint with an image structure through spatial discretization and grayscale conversion methods. This transforms the task of mapping coarse-grained channel fingerprints to fine-grained channel fingerprint twins into an image super-resolution problem.

[0008] A conditional generative diffusion model (CGDM) is constructed. Given a training set containing paired coarse-grained and fine-grained channel fingerprints, the CGDM model is trained, and the optimal parameters of the model are solved.

[0009] The final trained CGDM model is used as the core computational hub for channel fingerprint twinning. Coarse-grained channel fingerprints are used as side information to reconstruct the corresponding fine-grained channel fingerprint twinning, predicting the channel knowledge of user equipment at potential locations within the region of interest.

[0010] Preferably, the spatial discretization and grayscale conversion involves gridding the wireless propagation environment map in the spatial dimension and using min-max normalization to convert the channel knowledge into a range of 0 to 1, which facilitates subsequent neural network processing.

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

[0012] Where T is the number of time steps for noise propagation. To achieve the desired operation, ε t Let ε be the noise at step t. θ (·,·,·) represents the denoising neural network, and θ represents the model parameters. α t =1-β t ,β t The variance of adding noise at step t, The noise follows a Gaussian pattern with a mean vector of 0 and a covariance matrix of identity matrix I. G0 is a fine-grained channel fingerprint of the CGDM model input. G serves as a coarse-grained channel fingerprint as input to the model, acting as side information. t Let ||·||2 be the intermediate latent variable of the fine-grained channel fingerprint between t=1 and t=T, and let ||·||2 denote the L2 norm.

[0013] Preferably, the method for constructing the channel fingerprint twin also includes:

[0014] A one-time pruning method was used to remove redundant CGDM model parameters, resulting in a lightweight conditional generative diffusion model, LiCGDM.

[0015] We employ multi-objective knowledge distillation techniques to fine-tune the parameters of the pruned LiCGDM model in order to address the performance degradation caused by one-time pruning.

[0016] The final trained LiCGDM model is used as the core computational hub for channel fingerprint twinning.

[0017] As a preferred approach, a one-time pruning method is employed to obtain a lightweight conditional generative diffusion model (LiCGDM), further reducing memory consumption and computational latency. The specific pruning objective function is expressed as follows:

[0018] Where, ε tea This is the output of the CGDM model before pruning. For different network layers, For the set of all pruningable network layers, Params(p i ) represents the p-th i The number of parameters of the layer This represents the parameter quantity that needs to be pruned, ε(p) i ) is the pth pruning i The output of the model after the layer, Represents an uncertain variable.

[0019] As a preferred approach, multi-objective knowledge distillation is employed to further adjust the weights of the pruned model to restore its performance. A task objective and two knowledge distillation objectives are combined as loss functions to achieve overall model retraining. The specific distillation loss function is expressed as follows:

[0020] in, ε t Let ε be the noise at step t. S (·) and ε T (·) represent the outputs of the denoising neural network in LiCGDM (student model) and CGDM (teacher model), respectively. and These are the feature maps output by the i-th layer of the CGDM and LiCGDM networks, respectively, and λ is the value of λ. O With λ F These are the weights for the distillation loss values, respectively.

[0021] As a preferred approach, 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 layers, and upsampling. Its main modules include a temporal embedding module based on sine and cosine position encoding, a residual module, and a self-attention module.

[0022] Preferably, the downsampling, intermediate layer, and upsampling stages in CGDM include multiple integrated residual and self-attention modules; define X, Γ t Let X' be the feature map corresponding to the channel fingerprint and the temporal embedding vector at time t, respectively. The output X′ of the residual module is expressed as:

[0023] Where Ω=g(g(X)+f Ful (f Swi (f Ful (Γ t )))) is an 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 (·) represent convolution operations for 3×3 and 1×1, respectively, f Ful (·) and f Swi (·) represent the linear layer and the activation function layer, respectively, f Dro (·) and f Gn (·) represent the dropout layer and group normalization layer functions, respectively. in and c out These represent the number of feature map channels for the input and output, respectively.

[0024] In a second aspect, the present invention provides a wireless communication channel fingerprint twinning system, comprising:

[0025] 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, thereby transforming the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twin into an image super-resolution problem.

[0026] The channel fingerprint twin module is used to reconstruct fine-grained channel fingerprint twins within a target area using a trained Conditional Generative Diffusion Model (CGDM) or a lightweight model (LiCGDM) derived using a one-time pruning and multi-objective knowledge distillation method. The CGDM model is trained 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 computational hub for channel fingerprint twinning, using coarse-grained channel fingerprints as side information to reconstruct the corresponding fine-grained channel fingerprint twins and predict the channel knowledge of user equipment at potential locations within the region of interest.

[0027] Thirdly, the present invention provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method for constructing a channel fingerprint twin.

[0028] Beneficial Effects: Compared with existing technologies, this invention transforms the mapping task from coarse-grained channel fingerprints to fine-grained channel fingerprint twins into an image super-resolution problem. A conditional generative diffusion model is designed, using coarse-grained channel fingerprints as side information to guide the iterative denoising direction of the model, thereby completing the construction of fine-grained channel fingerprint twins. Simultaneously, to reduce model computation and memory usage, this invention also introduces one-time pruning and multi-objective knowledge distillation methods, which can accurately obtain the channel information of ultra-high-precision location user equipment within the target area with lower computational and storage complexity while ensuring reconstruction accuracy. The obtained fine-grained channel fingerprints can further assist in the acquisition of real-time channel state information for large-scale wireless communication, guiding wireless transmission design and significantly improving the overall transmission efficiency of the system. Furthermore, the method proposed in this invention has strong generalization ability and is applicable to channel fingerprint reconstruction under different fine-grained requirements. Attached Figure Description

[0029] Figure 1 is a flowchart of channel fingerprint twin construction according to an embodiment of the present invention;

[0030] Figure 2 is a flowchart of channel fingerprint twin construction according to another embodiment of the present invention;

[0031] Figure 3 is a schematic diagram of the base station antenna deployment method in an embodiment of the present invention;

[0032] Figure 4 is a framework diagram of the model used in the core computing hub of channel fingerprint twin in this embodiment of the invention;

[0033] Figure 5 is a flowchart of the algorithm for solving the one-time pruning problem in an embodiment of the present invention;

[0034] Figure 6 is a schematic diagram illustrating the impact of the number of basic channels in the latent space of the input feature map on the performance of this embodiment of the invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0036] As shown in Figure 1, this embodiment of the invention discloses a method for constructing channel fingerprint twins. First, channel knowledge at some locations within the region of interest is acquired. Then, through spatial discretization and grayscale conversion, the acquired channel knowledge is rearranged into a coarse-grained channel fingerprint with an image structure. This transforms the mapping task from coarse-grained channel fingerprints to fine-grained channel fingerprint twins into an image super-resolution problem. Next, a conditional generative diffusion model (CGDM) is constructed. Given a training set containing paired coarse-grained and fine-grained channel fingerprints, the CGDM model is trained, and the optimal parameters of the model are solved. Finally, the finally trained CGDM model is used as the core computational hub for channel fingerprint twins. The coarse-grained channel fingerprints are used as side information to reconstruct the corresponding fine-grained channel fingerprint twins, predicting the channel knowledge of user equipment at potential locations within the region of interest.

[0037] This invention transforms the mapping task from coarse-grained channel fingerprints (considered as physical objects) to fine-grained channel fingerprints (considered as digital twin objects) into an image super-resolution problem. A conditional generative diffusion model is designed, introducing coarse-grained channel fingerprints as side information to guide the model in a directional, iterative construction of fine-grained channel fingerprints. In some preferred embodiments, one-time pruning and multi-objective knowledge distillation methods are introduced to reduce computational and memory consumption, thereby reducing computational and storage resource consumption. As shown in Figure 2, the specific steps include: (1) providing actually deployed sensors or channel test vehicles to collect channel knowledge at some locations within the target area. (2) rearranging the measured channel knowledge within the target area into coarse-grained channel fingerprints with a 2D image structure using spatial discretization and grayscale conversion, thus transforming the mapping task from coarse-grained channel fingerprints to fine-grained channel fingerprint twins into an image super-resolution problem. (3) constructing a conditional generative diffusion model (CGDM) with a Markov chain structure, training the CGDM model with a training set containing paired coarse-grained and fine-grained channel fingerprints, and solving for the optimal parameters of the model. Specifically, in step (3), variational inference and reparameterization techniques are used to derive the lower bound of the negative log-likelihood of the target data distribution with respect to conditional information, which serves as the loss function for model training. To reduce memory consumption and latency, a one-time pruning method is used to remove redundant model parameters, resulting in a lightweight model, LiCGDM, which is beneficial for practical deployment. Furthermore, multi-objective knowledge distillation is introduced to fine-tune the parameters of the pruned lightweight model obtained from the previous training, addressing the performance degradation caused by one-time pruning. In this embodiment, the dataset is divided into a training set (5000 samples), a validation set (1000 samples), and a test set (1000 samples) in a 5:1:1 ratio. The model is trained using the Adam optimization algorithm and unsupervised training. The batch size is set to 16, the learning rate to 0.00005, and this training strategy is used for over 500,000 iterations. Starting from the 5000th iteration, an exponential moving average algorithm is introduced with a decay factor of 0.9999. Additionally, the model employs a dropout strategy with a dropout rate of 0.1. During training, the model's weight parameters are adjusted using the validation set, and the final performance of the model is tested using the test set. (4) The trained model or the pruned lightweight model is used as the core computing hub for channel fingerprint twinning and is used for the efficient construction of fine-grained channel fingerprint twinning in the target area in the field of wireless communication.

[0038] The specific steps of the embodiments of the present invention will be described below with reference to specific scenarios:

[0039] I. System Model

[0040] Consider a large-scale multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDMOFDM) communication scenario within a square communication area. The location of the user equipment is defined as follows: As shown in Figure 3, the base station (BS) is equipped with an N-type... r =N r,v ×N r,h A uniform linear array (ULA) consisting of N antenna elements, with the antennas arranged at half-wavelength intervals, is used to serve M single-antenna users within a cell, where N r,v and N r,h These represent the number of antennas in the vertical column and horizontal row, respectively. Assume that in the considered wireless propagation environment, from BS to the antenna located at x... m User equipment exists There are N physical propagation paths. The system employs N... c OFDM modulation of subcarriers, with a spacing of Δ between subcarriers.f Typically, in N c The center of each subcarrier is selected as N. k One active subcarrier is used for signal transmission, while the remaining subcarriers serve as a guard band, and the set of active subcarriers is defined as follows:

[0041] Between the base station and the user equipment on active subcarriers The space-frequency domain channel response is modeled as follows:

[0042] in, and They respectively represent the locations located at x m The complex gain and propagation delay of the l-th path of the user equipment, Represents the Kronecker product. Normalized horizontal angle. With elevation angle Related to, normalized vertical angle Then through the relational formula With elevation angle Correspondingly, azimuth angle Defined as Guiding vector and It can be represented as:

[0043] in, and Typically, the baseband signal received at the terminal is described as: y(x m ) = h H (x m )s+z(x m ), (5)

[0044] Where s is the power P X The transmitted signal symbol, z(x) m ) represents the additive noise of the single-sided power spectral density N0.

[0045] The channel in Equation (1) typically includes large-scale fading (path loss and shadowing) and small-scale fading. The former usually varies over distances measured in meters, while the latter fluctuates over distances measured in wavelengths. Considering that large-scale MIMO-OFDM systems operate at wavelengths on the millimeter or centimeter scale, estimating this small-scale fading component would require knowing the terminal's location with millimeter-level accuracy. This is impractical for current technologies, such as Global Navigation Satellite Systems. Therefore, it is generally assumed that the effects of small-scale fading have been averaged out, which is quite consistent with the actual situation in scenarios where channel measurements are performed while the test vehicle is moving.

[0046] II. Channel Gain Model

[0047] The channel gain attenuation observed at the terminal may be governed by various environmental factors, such as propagation loss from different paths, reflections from buildings, and diffraction. Therefore, among these effects, the relatively slowly varying components collectively contribute to the channel gain function, which is subsequently expressed as... Indicates the user's location The large-scale signal attenuation is measured above, where e represents the propagation environment. Furthermore, small-scale effects are typically modeled as a complex Gaussian random variable with unit variance. Therefore, the baseband signal received in formula (5) can be expressed in a more general form as:

[0048] The average received energy per symbol can be expressed as:

[0049] Where B is the signal bandwidth. Located at x m The channel gain of the user equipment at the location, in dB, is defined as: G(e,x) m )=(P Y ) dB -(P X ) dB (8)

[0050] It reflects the location of x m The received power variation of the user equipment at location e in the propagation environment is then determined. The set of channel gains at potential locations is then termed the channel fingerprint (CF). It is worth noting that if the channel gain at any location within the region of interest can be accurately obtained, this embodiment of the invention can construct a fine-grained CF, which is particularly important for transceiver design in large-scale MIMO systems. It should be noted that this embodiment uses channel gain as an example of a channel fingerprint; it is understood that a channel fingerprint can be path loss, channel delay, or other channel knowledge.

[0051] III. Problem Statement

[0052] Since the communication area under consideration is a square area Embodiments of the present invention can perform spatial discretization along the X and Y axes. Specifically, considering the size of the region of interest is... A resolution factor σ is defined such 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 represented as: Λ i,j :=[iΔ x ,jΔ y ] T (9)

[0053] Given a resolution factor σ, the CF corresponding to the potential user equipment locations within the target geometric region can be rearranged into a two-dimensional tensor, denoted as: [G] i,j =G(e,Λ) i,j (10)

[0054] Note that when the target area size is When the resolution factor is σ, the number of interpolation points required is As the resolution factor σ increases, the complexity of traditional interpolation algorithms increases. This increase poses a significant challenge to building fine-grained channel knowledge (CF) in real-world scenarios. On the other hand, increasing the density of test nodes can improve the spatial resolution of the CF, which can be achieved by significantly increasing the number of sensor nodes or using test vehicles to collect channel knowledge over smaller geographical intervals. However, neither of these approaches may be impractical due to high hardware and labor costs. A coarse-grained factor σ is defined. LR and a fine-grained factor σ HR , σ HR Usually σ LR Powers of 2, such as 2×, 4×, 8×. For a resolution factor of σ... LR The CF is defined as G LR Target area Discretized into units of size σ LR ×σ LR The two-dimensional tensor, G HR It is also defined in a similar way. CF twins focus on tasks derived from a given coarse-grained CFG. LR Reconstructing fine-grained CFG HR This is especially true in scenarios where measurement costs, privacy, or security are a constraint.

[0055] In typical image super-resolution (ISR) tasks in computer vision, the goal is to reconstruct a high-resolution (HR) image from a given low-resolution (LR) corresponding image, enhancing the image's details and overall quality. It can be seen that the task of this invention's embodiments is consistent with the ISR task; therefore, this invention's embodiments analyze the fine-grained CF reconstruction problem from the perspective of ISR. Specifically, this invention's embodiments will... LRThe elements of the matrix are treated as pixels, and coarse-grained CFG is used. LR Treat it as an LR image, and use fine-grained CFG HR This is considered as an HR image. The goal is then to learn a specific mapping from a given LR CFG. LR Effective reconstruction of HR CFG HR ,Right now:

[0056] Where Θ is a mapping The parameter set, U is the number of training samples. However, this task presents a classic and challenging inverse problem involving the efficient reconstruction of fine details from a given LR CF. Given that the conditional distribution of the HR CF output given the LR CF input usually does not follow a simple parameter distribution, most ISR regression methods based on feedforward neural networks typically perform poorly at high amplification factors and fail to accurately recover details. Deep generative models have achieved success in learning complex empirical distributions of target data. Specifically, if the implicit prior information of the HR CF distribution can be learned, such as the gradient of the data log density, it is possible to transform to the target CF distribution through an iterative sampling step from the standard normal distribution, similar to Langevin dynamics. Therefore, the mapping relationship in Equation (11) can be solved by optimizing the following objective:

[0057] in, The gradient of the logarithmic density of the HR CF, also known as the Stein fraction, P Θ The learning density is represented by . Note that the noise learned by traditional GDM has been shown to be equivalent to the Stein score, which makes it possible to generate HR CF samples consistent with the target data distribution. To accurately reconstruct an HR CF from a given LR CF, the task is ill-conditioned, meaning that the reconstruction process may produce multiple possible solutions for the HR CF. In other words, different HR CFs may map to the same LR CF. Therefore, unlike traditional GDM, it starts with a pure Gaussian noise tensor, and introducing additional source signals as side information (also guiding conditions) is crucial for achieving the optimal solution. To this end, the method of this embodiment combines the learned prior information and introduces LR CFG. LR As an additional source signal to guide the iterative optimization process.

[0058] IV. Fine-grained Channel Fingerprint Twin Construction Method Based on CGDM

[0059] 1. Overall Design of CGDM

[0060] To simplify the notation, G LR and G HR Use respectively And G represents. Given a paired LR CF input and HR CF output dataset, it is defined as Represents an unknown distribution The samples are extracted from the data. These datasets are typically collected from sensor nodes and test vehicles, and have different resolution factors depending on the specific scenario, such as σ. LR and σ HR In the task of this embodiment of the invention, the focus is on learning through a targeted iterative refinement process guided by source information. The approximate value of the parameter makes this possible. The mapping to G becomes possible. Given the powerful implicit prior learning capability of GDM, this embodiment of the invention designs a CGDM to facilitate the generation of G.

[0061] Specifically, CGDM can generate a target HR CF defined as G0 through T refinement time steps. CGDM starts from a source consisting of pure Gaussian noise. Initially, based on the source signal and prior information learned during training, i.e., the conditional distribution... The initial input is iteratively refined, where 0 and I represent a mean vector of 0 and a covariance matrix of identity, respectively. As it evolves at each time step t, it generates a series of values ​​defined as {G}. T-1 G T-2 The output CF of ,...,G0} ultimately yields the target. Specifically, the distribution of intermediate CFs in the iterative refinement chain is controlled by a forward diffusion process. This process gradually adds noise to the output CF through a fixed Markov chain, defined as q(G t |G t-1 The model in this embodiment of the invention attempts to use the source... A conditional inverse Markov chain iteratively recovers the signal from noise, thereby reversing the Gaussian diffusion process. To achieve this, embodiments of the invention optimize the denoising neural network ε using the following objective function formula (35). θ (·) is used to learn the inverse chain. The trained CGDM estimates the noise using the LR CF and the noisy image as input, and generates the target HR CF after T refinement time steps.

[0062] 2. The positive diffusion process that begins with HR CF

[0063] Consider the HR CF samples extracted from the distribution of interest, defined as G0~q(G). GDM uses a fixed diffusion process q(G). 1:TTraining is performed using |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 a variance table. Confirmed. Specifically, the forward diffusion process is defined as:

[0064] Note that the forward diffusion subprocess, also known as the encoder, does not need to be learned at each time step t. Instead, it is a fixed and predefined linear Gaussian model, which can be represented as...

[0065] Where ε is Gaussian noise, its distribution is as follows: Define α t =1-β t and At the transfer density q(G) t |G t-1 Under the linear Gaussian assumption, and combining formulas (14) and (15), this embodiment of the invention can utilize reparameterization techniques to sample G in a closed-form manner at any given time step t. t :

[0066] Typically, variance variation is set as β1 < β2 < ... < β T When β T When the value is infinitely close to 1, for any initial state G0, G T It converges to a standard Gaussian distribution, i.e.

[0067] 3. The inverse process of LR CF under condition

[0068] For traditional GDM, the inverse process can be viewed as a decoding process, where at each time step t, G... t Denoising and restored to G t-1 The transition probability at each step is represented as p(G). t-1 |G t Based on the Markov density transformation transfer property, the joint distribution of the reverse process is expressed as:

[0069] However, using Bayes' theorem to derive p(G) t-1 |G t The expression for reveals that its denominator contains an integral, thus lacking a closed-form solution. Therefore, a denoising neural network with parameter θ is needed to approximate these conditional probabilities in order to perform the reverse process. Since the reverse process is also a Markov chain, it can be represented as:

[0070] In the task of this embodiment of the invention, unlike the traditional GDM, the denoising model ε θ (·) with LR CF The lateral information of the form is used as a condition to guide it from the Gaussian distribution G. T The noise is gradually reduced and HR CFG0 is generated. Therefore, equations (19) and (20) need to be rewritten as follows:

[0071] 4. CGDM Optimization Based on Evidence Lower Bound

[0072] To make the network model ε θ (·) can effectively approximate the inverse process, requiring optimization of the model parameters θ with a specific objective. Mathematically, the latent variable G 1:T and conditions The observed sample G0 can be represented by the conditional joint distribution. This is represented by the likelihood-based approach used in generative modeling to maximize the conditional joint probability distribution of all observed samples. To optimize the model. However, embodiments of the present invention can only access the observed sample G0, and cannot access the latent variable G. 1:T Therefore, embodiments of the present invention need to maximize the conditional marginal distribution. Right now

[0073] Within the framework of variational reasoning, the condition is... The likelihood of the observed sample G0, i.e., the evidence, allows the embodiments of the present invention to derive the Evidence Lower Bound (ELBO) as a surrogate objective function for optimizing the CGDM:

[0074] in, This comes from Jensen's inequality. To further derive ELBO, equation (24) can be rewritten as follows:

[0075] Among them, D KL (·) represents the KL divergence (KL, Kullback-Leibler). Then, the parameters θ of the CGDM can be learned by optimizing the negative ELBO:

[0076] in, The objective function for training CGDM.

[0077] for The embodiments of the present invention note It is a constant and can be excluded from optimization, while Monte Carlo estimation can be used for approximation and optimization. As can be seen from the above analysis, the training objective of CGDM is mainly composed of... This can be determined. It can be observed that the training objective of CGDM is to approximate the transfer density as closely as possible to the reverse process, thereby... and The KL divergence between them is minimized. Therefore, embodiments of the present invention require obtaining q(G) t_1 |G t The mean of G0) and variance The explicit expression. Specifically, q(G) t-1 |G t G0) can be represented as

[0078] Combining formulas (16) and (17), (27) can be further rewritten as:

[0079] Based on (28), q(G) t-1 |G t The mean and variance of G0 can be explicitly expressed as:

[0080] Typically, variance Set as a constant Therefore, to ensure that the denoised transfer density is close to the true denoised transfer density, the embodiments 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 embodiments of the present invention only need to make the CGDM predict...

[0081] The KL divergence is defined as:

[0082] Where d represents the dimension of x. Substituting formulas (29), (31), and (33) into formula (25) get:

[0083] Substituting formulas (25) and (34) into formula (26), the objective function of CGDM is obtained. It can be further simplified to:

[0084] Based on the trained CGDM, given any noise-contaminated image G t A well-trained CGDM can utilize side information. To predict noise ε tThe target image is then obtained through the conversion formula (16). The approximate value, that is:

[0085] By reparameterizing, formula (36) represents the result of iterative refinement. Each iteration step in the CGDM proposed in this embodiment of the invention is given by the following formula:

[0086] in, The noise estimation step in Equation (37) is similar to the Langevin dynamics step in fraction-based generative models, which is equivalent to estimating the first derivative of the log-likelihood of the observed sample, also known as the gradient or Stein score.

[0087] 5. CGDM Network Architecture

[0088] The conditional denoising neural network in CGDM uses a variant of the Unet network as its backbone, which consists of three main stages: downsampling, mid-levels, and upsampling. Key components of each stage include temporal embedding blocks, Res... + The model consists of blocks, self-attention blocks, downsampling blocks, and upsampling blocks. The model architecture is shown in Figure 4, and the specific modules will be introduced one by one.

[0089] 1) To encode the time parameters [1,...,t,...,T] during the diffusion process, a time embedding coding method based on sinusoidal position coding is used, and the resulting time embedding Γ t It effectively captures temporal characteristics. Definition The j-th component of the time embedding vector corresponding to time t:

[0090] Where j = 0, 1, 2, ..., c time / 2-1, c time Let be the dimension of the time embedding vector. Then the embedding vector of the time constant t can be expressed as:

[0091] in, Using formula (39), for any given time Γ t+Δt The embedding vector can be obtained through the following linear transformation:

[0092] M Δt It is a linear transformation matrix, defined as:

[0093] in, To enable the model to capture more complex temporal features, the temporal embedding is further enhanced through two fully connected layers and an activation layer, denoted as Γ′. t =f Ful (f Swi (f Ful (Γ t ))). f Ful (·)f Swi (·) represent fully connected layers and swish activation function layers, respectively.

[0094] 2) Res + Module: Designing deeper feature extraction networks to learn higher-level semantic information from CGM. However, deep neural networks often suffer from model degradation due to challenges such as vanishing and exploding gradients. To address this, residual connections are introduced as a way to mitigate these problems. Specifically, defining... Γ t The inputs are the feature map and the temporal embedding vector, respectively, and Res + The module's output can be represented as:

[0095] Among them, f Conv,1 (·) represents a 1×1 convolution operation. Defined as: Ω=g(g(X)+f Ful (f Swi (f Ful (Γ t (43)

[0096] Where g(X)=f Conv,3 (f Dro (f Swi (f Gn (X)))),f Conv,3 (·) represents a 3×3 convolution operation. f Dro (·), f Gn (·) represent the dropout layer and the group normalization layer, respectively.

[0097] 3) Self-Attention Module: 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 levels of attention to the input, with larger weights 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 d represents the number of image patches. n This represents the feature dimension of each block, for z j Perform three different linear transformations: kj =z j W k ,j=1,...,d n , (44) q j =z j W q ,j=1,...,d n (45) v j =z j W v ,j=1,...,d n (46)

[0098] in, and These are the key, query, and value vectors, respectively. and These represent their respective trainable transformation matrices. Specifically, the weight assignment function is determined by k. j and q j′ Confirmed. A higher correlation. This means the j-th input block z j The characteristics of [the input] are of higher importance for the j′-th output block. Typically, this correlation can be determined based on the input Z and the matrix W. k and W q Adaptive adjustments are made. For 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 obtain the attention matrix.

[0099] Where Softmax(Z) = exp(z) j ) / ∑exp(z j ), It is a scaling factor. Finally, the output Z′ of the self-attention block can be expressed as:

[0100] in,

[0101] Note that N is included in all three stages of the proposed conditional denoising neural network. RA ≥2 Res + And the corresponding self-attention module, as shown in Figure 4.

[0102] V. LiCGDM Based on One-Time Pruning and Multi-Objective Knowledge Distillation

[0103] 1. A lightweight model-based method using one-time pruning

[0104] Given that CGDM sacrifices significant memory consumption and latency to achieve excellent performance, its deployment on personal computers and even mobile devices is highly limited. Therefore, embodiments of this invention utilize the additivity of network layers to design a one-time pruning method to obtain LiCGDM.

[0105] Given a specific pruning ratio, the goal of layer pruning is to remove a subset of pruneable layers from the CGDM network structure while minimizing the degradation in model performance. In this embodiment of the invention, minimizing the mean-square error (MSE) loss between the final output of the original CGDM network (referred to as the "teacher") and the final output of the pruned LiCGDM network (referred to as the "student") is used as the pruning objective. Specifically, the pruning objective function is expressed as:

[0106] Where, ε tea This is the output of the CGDM model before pruning. For the set of all pruningable network layers, Represents the i-th pruning network layer, Params(p i ) represents the p-th i The number of parameters of the layer Indicates pruning The network output after the layer, This represents the parameter quantity that needs to be pruned, ε(p) i ) is the pth pruning i The model output after the layer, Let be an uncertain variable. However, solving the above equation is an NP-hard problem, requiring the use of triangle inequalities to transform and obtain the upper bound of the objective function:

[0107] in,

[0108] Therefore, the optimization problem of formula (49) is transformed into minimizing the upper bound:

[0109] Although equation (51) is still an NP-hard problem, by utilizing the additivity of network layers, the output distortion caused by multiple perturbations can be approximated by the sum of distortions caused by each individual perturbation. That is, this additivity can be expressed as:

[0110] Substituting equation (52) into equation (51), we obtain a new approximate proxy target:

[0111] As can be seen from the above formula, the pruning standard is entirely determined by... Therefore, only calculation is required. Optimization is then performed. It's worth noting that the time complexity of this objective function for each network layer is O(log n). Therefore, this problem can be transformed into a 0-1 knapsack problem, which can be solved using the classic dynamic programming algorithm. The specific algorithm flow is shown in Figure 5.

[0112] 2. A Network Model Enhancement Method Based on Multi-Objective Knowledge Distillation

[0113] Removing certain network layers from the teacher network typically leads to a performance degradation in the model. Therefore, this embodiment of the invention employs a multi-objective knowledge distillation technique to further adjust the weights of the pruned model to restore performance. Combining one task objective and two knowledge distillation objectives as the loss function, the model is retrained as a whole. The loss function is expressed as follows:

[0114] in, ε S (·) and ε T (·) represent the outputs of the denoising neural network in LiCGDM and the denoising neural network in CGDM, respectively. and These are the feature maps output by the i-th layer of the CGDM and LiCGDM networks, respectively, and λ is the value of λ. O With λ F These are the weights for the distillation loss values, respectively. It's worth noting that, by default, λ0 and λ... F It was set to 1 for training.

[0115] VI. Implementation Results

[0116] To enable those skilled in the art to better understand the present invention, the following provides a comparison of the prediction performance results of the channel fingerprint prediction method and existing methods under several specific system configurations in the embodiments of the present invention.

[0117] The system configuration for ultra-large-scale MIMO-OFDM is given below: the number of base station antennas is N. r,v ×N r,h=8×8, target area size A = 128m×128m, center frequency is 2.4GHz, subcarrier spacing is 15kHz, base station height is set to 25m, user height is set to 1.5m. This embodiment of the invention uses the QuaDRiGa generator (version 2.6.1) to simulate a real wireless communication system, while considering the typical 5G NR urban microcell scenario "3GPP 38.901 UMa NLOS" as the specific communication scenario for 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, β1 = 10 -4 ,β T =0.01, Batch size is 16, training iterations are 500,000, and an exponential moving average algorithm is introduced starting from the 5000th iteration, with a decay factor set to 0.9999, and coarse-grained and fine-grained factors σ. LR With σ HR The values ​​are 32 and 128 respectively.

[0118] Figure 6 illustrates the impact of the number of basic channels in the latent space of the input feature map on the performance of the proposed model. It can be observed that when the number of basic channels c1 in the latent space of the input feature map increases from 8 channels to 128 channels, the proposed model achieves better performance in terms of convergence speed and training stability. This is attributed to the fact that each feature channel can capture different types of information; increasing the number of feature channels enhances the model's expressive power and aggregates simple, low-level features with abstract, high-level features, thereby improving the model's training efficiency.

[0119] Table 1. Quantitative comparison of performance on fine-grained channel fingerprint twin reconstruction task.

[0120] Table 1 presents a quantitative comparison of the reconstruction performance of CGDM and LiCGDM in the embodiments of the present invention with existing methods. The methods compared are: SRGAN and SRGAN-MSE methods proposed in "Photo-realistic single image super-resolution using a generative adversarial network," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017, pp. 4681–4690; C-VAE method in "Auto-encoding variational Bayes," in Proc. Int. Conf. Learn. Represent. (ICLR), Banff, AB, Canada, Apr. 2014, pp. 1–14; and DRRN method in "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, SRGAN-MSE, SRGAN, C-VAE, and DRRN methods of this invention are applied to a ×4 super-resolution channel fingerprint dataset (32 2 →128 2 The performance of CGDM was compared using metrics including Normalized Mean Squared Error (NMSE), Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structure Similarity Index Measure (SSIM). The model parameters for CGDM were set as follows: c1 = 64, N... RA =2, c1:c2:c3:c4:c5 = 1:2:4:8:16, and the pruning ratio used in LiCGDM is 35%. It can be observed that compared with SRGAN, CGDM reduces NMSE and MSE by 65.236 × 10⁻⁶. -5The PSNR and SSIM values ​​are improved by 11.747 and 0.006 respectively, and the reconstruction accuracy is also significantly superior to the other comparison methods. In addition, LiCGDM shows a slight decrease in performance metrics compared to CGDM, but the decrease is still within an acceptable range, which confirms that the proposed lightweight method effectively compresses CGDM so that it can be quickly deployed on personal computers and even mobile devices.

[0121] Based on the same inventive concept, this invention discloses a wireless communication channel fingerprint twinning system, comprising: a coarse-grained channel fingerprint acquisition module, used to acquire channel knowledge at partial locations within a 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, thereby transforming the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twinning into an image super-resolution problem; and a channel fingerprint twinning module, used to reconstruct fine-grained channel fingerprint twinning within a target region using a trained Conditional Generative Diffusion Model (CGDM) or a lightweight model LiCGDM obtained through one-time pruning and multi-objective knowledge distillation; 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 serves as the core computational hub for channel fingerprint twinning, using the coarse-grained channel fingerprint as side information to reconstruct the corresponding fine-grained channel fingerprint twinning, predicting the channel knowledge of user equipment at potential locations within the region of interest. Specific implementation details are detailed in the foregoing embodiments and will not be repeated here.

[0122] Based on the same inventive concept, this invention discloses a computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of a channel fingerprint twinning construction method. The program / instruction code for implementing the method of this invention can be written in any combination of one or more programming languages. This program / instruction code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program / instruction code causes the steps of the method of this invention to be implemented. The program / instruction code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

[0123] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways without departing from the spirit and scope of this application. The current embodiments are merely exemplary examples and should not be considered limiting, nor should the specific content given limit the purpose of this application. For example, some features may be omitted or not implemented.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing a channel fingerprint twin, characterized in that, The method includes the following steps: We acquire channel knowledge at certain locations within a region of interest, and then rearrange the acquired channel knowledge into a coarse-grained channel fingerprint with an image structure through spatial discretization and grayscale conversion methods. This transforms the task of mapping coarse-grained channel fingerprints to fine-grained channel fingerprint twins into an image super-resolution problem. A conditional generative diffusion model (CGDM) is constructed. Given a training set containing paired coarse-grained and fine-grained channel fingerprints, the CGDM model is trained, and the optimal parameters of the model are solved. The final trained CGDM model is used as the core computational hub for channel fingerprint twinning. Coarse-grained channel fingerprints are used as side information to reconstruct the corresponding fine-grained channel fingerprint twinning, predicting the channel knowledge of user equipment at potential locations within the region of interest.

2. The method for constructing a channel fingerprint twin according to claim 1, characterized in that, The spatial discretization and grayscale conversion involve gridding the wireless propagation environment map in the spatial dimension and using min-max normalization to convert channel knowledge into a range of 0 to 1, which facilitates subsequent neural network processing.

3. The method for constructing a channel fingerprint twin according to claim 1, characterized in that, The objective function for training the CGDM model is the negative log-likelihood of the target data distribution conditionally dependent on side information. The lower bound of evidence for the objective function is derived using variational inference and reparameter techniques, and then used as the loss function for training the proposed CGDM. This loss function is expressed as: Where T is the number of time steps for noise propagation. To achieve the desired operation, ε t Let ε be the noise at step t. θ (·,·,·) represents the denoising neural network, and θ represents the model parameters. β t The variance of adding noise at step t, The noise follows a Gaussian pattern with a mean vector of 0 and a covariance matrix of identity matrix I. G0 is a fine-grained channel fingerprint of the CGDM model input. G serves as a coarse-grained channel fingerprint as input to the model, acting as side information. t Let ||·||2 be the intermediate latent variable of the fine-grained channel fingerprint between t=1 and t=T, and let ||·||2 denote the L2 norm.

4. The method for constructing a channel fingerprint twin according to claim 1, characterized in that, Also includes: A one-time pruning method was used to remove redundant CGDM model parameters, resulting in a lightweight conditional generative diffusion model, LiCGDM. We employ multi-objective knowledge distillation techniques to fine-tune the parameters of the pruned LiCGDM model in order to address the performance degradation caused by one-time pruning. The final trained LiCGDM model is used as the core computational hub for channel fingerprint twinning.

5. The method for constructing a channel fingerprint twin according to claim 4, characterized in that, A one-time pruning method is used to obtain a lightweight conditional generative diffusion model (LiCGDM). The specific pruning objective function is expressed as follows: Where, ε tea This is the output of the CGDM model before pruning. For different network layers, For the set of all pruningable network layers, Params(p i ) represents the p-th i The number of parameters of the layer This represents the parameter quantity that needs to be pruned, ε(p) i ) is the pth pruning i The output of the model after the layer, Represents an uncertain variable.

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

7. The 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. It mainly consists of three stages: downsampling, intermediate layers, and upsampling. Its main modules include a temporal embedding module based on sine and cosine position encoding, a residual module, and a self-attention module.

8. The method for constructing a channel fingerprint twin according to claim 1, characterized in that, The downsampling, intermediate layer, and upsampling stages in CGDM contain multiple integrated residual and self-attention modules; define X, Γ t Let X' be the feature map corresponding to the channel fingerprint and the temporal embedding vector at 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 an 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 (·) represent convolution operations for 3×3 and 1×1, respectively, f Ful (·) and f Swi (·) represent the linear layer and the activation function layer, respectively, f Dro (·) and f Gn (·) represent the dropout layer and group normalization layer functions, respectively. in and c out These represent the number of feature map channels for the input and output, respectively.

9. A wireless communication channel fingerprint twin system, characterized in that, include: A coarse-grained channel fingerprinting module is used to obtain channel knowledge at certain locations within a region of interest. By using spatial discretization and grayscale conversion methods, the acquired channel knowledge is rearranged into a coarse-grained channel fingerprint with an image structure, thereby transforming the mapping task from coarse-grained channel fingerprint to fine-grained channel fingerprint twin into an image super-resolution problem. The channel fingerprint twin module is used to reconstruct fine-grained channel fingerprint twins within a target area using a trained Conditional Generative Diffusion Model (CGDM) or a lightweight model (LiCGDM) derived using a one-time pruning and multi-objective knowledge distillation method. The CGDM model is trained 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 computational hub for channel fingerprint twinning, using coarse-grained channel fingerprints as side information to reconstruct the corresponding fine-grained channel fingerprint twins and predict the channel knowledge of user equipment at potential locations within the region of interest.

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

Images