Diffusion model assisted channel map rapid construction method and system

Abstract

The invention discloses a diffusion model assisted channel map rapid construction method and system. The method comprises the following steps: firstly, acquiring channel information of a target area, and constructing channel map paired data; secondly, establishing a Markov chain between a fine-grained channel map and a coarse-grained channel map by using channel map residual transfer based on a latent diffusion model architecture, and constructing a channel map residual estimation network embedded with time steps; combining the channel map residual error estimation network with a pre-trained fine-tuned generative adversarial network to construct a latent diffusion model based on channel map residual error transfer; then training the constructed channel map residual error estimation network by adopting a supervised training mode to obtain an optimal parameter; and finally, quickly reconstructing a fine-grained channel map of a target area in the wireless communication system by using the trained model. Compared with a traditional generation diffusion model, the channel map construction method provided by the invention has the advantages that the fine-grained channel map can be constructed with fewer reasoning steps, and meanwhile, higher construction precision is realized.

Description

Technical Field

[0001] This invention belongs to the field of communication technology and relates to a method and system for rapid construction of channel maps assisted by diffusion model. Background Technology

[0002] The overall performance of a massive MIMO (Multiple-Input Multiple-Output) system hinges on the accuracy of the acquired Channel State Information (CSI). However, with the rapid increase in antenna array size, 6G mobile communication networks will face the challenge of estimating channels with considerable dimensionality. An emerging enabling technology—Channel Fingerprint (CF)—shows great potential in solving this problem. As a localized database, the Channel Fingerprint stores wireless channel information related to potential base stations (BS) for any B2X (Browser-to-Everything) link. This database stores key channel knowledge, including channel gain, shadowing fading, and angle of arrival / departure.

[0003] Existing research on channel map construction can be broadly categorized into data-driven interpolation methods and model-driven parameter fitting methods. However, existing research often relies on physical propagation characteristics or model-based assumptions to construct channel maps. Due to limitations in hardware cost and storage overhead, research on reconstructing fine-grained channel maps from coarse-grained channel maps is relatively limited. Furthermore, due to limitations in actual channel measurements and base station storage overhead, currently available channel maps are mostly coarse-grained, resulting in a lack of accurate site-specific channel knowledge. Therefore, rapidly constructing fine-grained channel maps from coarse-grained channel maps to obtain real-time channel information is one of the challenges faced in practical communication scenarios. Summary of the Invention

[0004] Purpose of the invention: To address the shortcomings of existing technologies, the purpose of this invention is to provide a diffusion model-assisted method and system for rapid construction of channel maps, thereby improving the efficiency and accuracy of generating fine-grained channel maps from coarse-grained channel maps.

[0005] Technical Solution: To achieve the above-mentioned objectives, this invention provides a diffusion model-assisted method for rapid construction of channel maps, comprising the following steps:

[0006] Obtain channel information for the target area and construct paired channel map data at different granularities;

[0007] Based on the latent diffusion model architecture, a Markov chain between fine-grained and coarse-grained channel maps is established using channel map residual transfer.

[0008] A channel map residual estimation network embedded with time steps is constructed to predict the channel map residual at a given time step to recover the fine-grained channel map. The inputs are a coarse-grained channel map latent representation and a noisy fine-grained channel map latent representation at a given time step, and the output is the estimated fine-grained channel map latent representation. The channel map residual estimation network includes a first processing stage composed of a convolutional module, a residual module, a window attention module, and a downsampling module, used to progressively extract deep features of the input; a second processing stage composed of a residual module and a window attention module, used for feature transformation of the channel map latent representation; and a third processing stage composed of a convolutional module, a residual module, a window attention module, and an upsampling module, used to progressively reconstruct the latent representation of the fine-grained channel map.

[0009] The channel graph residual estimation network is combined with a pre-trained fine-tuned generative adversarial network to construct a channel graph residual transfer-based latent diffusion model (CF-LRDM). The constructed channel graph residual estimation network is trained in a supervised manner to obtain the optimal parameters.

[0010] We utilize a latent diffusion model based on channel map residual transfer to rapidly reconstruct fine-grained channel maps of target regions in wireless communication systems.

[0011] Preferably, in the latent diffusion model based on channel graph residual transfer, the fine-tuned VQGAN network learns to map channel graph data to latent space encoding-decoding functions, and at the same time designs a corresponding codebook, optimizing both the codebook and encoding / decoding accuracy during the pre-training process.

[0012] Preferably, the coarse-grained channel map in the latent space is obtained from the Markov chain between the fine-grained and coarse-grained channel maps established using channel map residual transfer. With fine-grained channel maps The residual between them is Design a monotonically increasing condition that satisfies and sequence and define As a forward process The residual transfer coefficient in; For time steps, This is a potential representation of the initial coarse-grained channel map. Represents the forward diffusion process The probability distribution, The mean is variance is The Gaussian distribution, where It is a hyperparameter that controls the noise variance. It is an identity matrix.

[0013] Preferably, the loss function for training the channel graph residual estimation network model is: ,in As a weighting factor, For network models, Represents the learning parameters of the network. This represents the Euclidean norm.

[0014] Preferably, the first processing stage of the channel map residual estimation network includes: stage 0 consisting of convolutional layers; stage 1 consisting of two residual blocks and a window attention module; stage 2 consisting of a downsampling module, two residual blocks, and a window attention module; and stage 3 consisting of a downsampling module, two residual blocks, and a window attention module. The second processing stage is stage 4 consisting of two residual blocks and a window attention module. The third processing stage includes: stage 5 consisting of residual blocks and a window attention module; stage 6 consisting of an upsampling module, residual blocks, and a window attention module; stage 7 consisting of an upsampling module, residual blocks, and a window attention module; and stage 8 consisting of convolutional layers. The data output from stage 8 is pre-trained and fine-tuned before being decoded by a generative adversarial network to obtain the constructed fine-grained channel map.

[0015] This invention provides a diffusion model-assisted rapid channel map construction system for implementing the aforementioned diffusion model-assisted rapid channel map construction method. The system includes: a sample acquisition module for acquiring channel information of a target region and constructing paired channel map data of different granularities; a model construction module for establishing a Markov chain between fine-grained and coarse-grained channel maps based on a latent diffusion model architecture using channel map residual transfer; and a channel map residual estimation network embedded with time steps for predicting the channel map residual at a given time step to recover the fine-grained channel map, with the input being the coarse-grained channel map. The system comprises: a channel map latent representation and a noisy fine-grained channel map latent representation at a given time step; outputting an estimated fine-grained channel map latent representation; constructing a latent diffusion model based on channel map residual transfer by combining a channel map residual estimation network with a pre-trained fine-tuned generative adversarial network; a model training module for training the constructed channel map residual estimation network in a supervised manner to obtain optimal parameters; and a fast reconstruction module for rapidly reconstructing the fine-grained channel map of the target region in the wireless communication system using the latent diffusion model based on channel map residual transfer.

[0016] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the steps of the diffusion model-assisted rapid channel map construction method described in any one of the present invention.

[0017] The present invention also provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the diffusion model-assisted rapid channel map construction method described above.

[0018] The present invention also provides a computer program product, including a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the diffusion model-assisted rapid channel map construction method.

[0019] This invention also provides a wireless communication system, including a base station and multiple user terminals. The base station or user terminals are equipped with: a sample acquisition module for acquiring channel information of a target area and constructing paired channel map data of different granularities; a model construction module for establishing a Markov chain between fine-grained and coarse-grained channel maps using channel map residual transfer based on a latent diffusion model architecture; a channel map residual estimation network embedded in time steps for predicting the channel map residual at a given time step to recover the fine-grained channel map, with inputs being the coarse-grained channel map latent representation and the noisy fine-grained channel map latent representation at a given time step, and outputting the estimated fine-grained channel map latent representation; and a combination of the channel map residual estimation network and a pre-trained fine-tuned generative adversarial network to construct a channel map residual transfer-based system. The latent diffusion model based on channel map residual estimation network includes: a first processing stage composed of a convolutional module, a residual module, a window attention module, and a downsampling module, used to progressively extract deep features of the input; a second processing stage composed of a residual module and a window attention module, used for feature transformation of the latent representation of the channel map; and a third processing stage composed of a convolutional module, a residual module, a window attention module, and an upsampling module, used to progressively reconstruct the latent representation of the fine-grained channel map; a model training module, used to train the constructed channel map residual estimation network in a supervised manner to obtain optimal parameters; and a fast reconstruction module, used to quickly reconstruct the fine-grained channel map of the target region in the wireless communication system using the latent diffusion model based on channel map residual transfer.

[0020] Beneficial Effects: This invention provides a channel map construction scheme based on a latent diffusion model using channel map residual transfer. It establishes a Markov chain between coarse-grained and fine-grained channel maps through residual transfer, designs corresponding transfer schemes, and trains a latent space mapping channel map data to improve model inference efficiency and reduce training complexity. The finally trained network architecture can sample a fine-grained channel map from a given coarse-grained channel map using fewer sampling steps. Compared to traditional generative diffusion models, the proposed fast channel map construction method can construct fine-grained channel maps with fewer inference steps and achieves higher construction accuracy compared to traditional baseline methods. This is more beneficial for rapidly constructing fine-grained channel maps from coarse-grained channel maps to obtain real-time channel information in practical communication scenarios. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of a rapid channel map construction method according to an embodiment of the present invention.

[0022] Figure 2 This is a flowchart illustrating the implementation of a specific scenario in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the principle framework of the CF-LRDM model in an embodiment of the present invention.

[0024] Figure 4 This is a detailed structural diagram of the CF-LRDM model in an embodiment of the present invention.

[0025] Figure 5 This is a comparison diagram of the construction results of the construction method in this embodiment of the invention and the existing method. Detailed Implementation

[0026] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0027] like Figure 1 As shown in the figure, this invention discloses a method for rapid construction of a diffusion model-assisted channel map, which mainly includes:

[0028] S1. Obtain channel information for the target area and construct paired channel map data at different granularities;

[0029] S2. Based on the latent diffusion model architecture, a Markov chain between fine-grained and coarse-grained channel maps is established using channel map residual transfer.

[0030] S3. Construct a channel map residual estimation network embedded with time steps to predict the channel map residual at a given time step to recover the fine-grained channel map. The inputs are the coarse-grained channel map latent representation and the noisy fine-grained channel map latent representation at a given time step, and the output is the estimated fine-grained channel map latent representation. The channel map residual estimation network includes a first processing stage composed of a convolutional module, a residual module, a window attention module, and a downsampling module, used to progressively extract deep features of the input; a second processing stage composed of a residual module and a window attention module, used for feature transformation of the channel map latent representation; and a third processing stage composed of a convolutional module, a residual module, a window attention module, and an upsampling module, used to progressively reconstruct the latent representation of the fine-grained channel map.

[0031] S4. Combine the channel graph residual estimation network with the pre-trained fine-tuned generative adversarial network to construct a latent diffusion model based on channel graph residual transfer. Use supervised training to train the constructed channel graph residual estimation network to obtain the optimal parameters.

[0032] S5. Utilize a latent diffusion model based on channel map residual transfer to quickly reconstruct fine-grained channel maps of target regions in wireless communication systems.

[0033] In this embodiment, a Markov chain is established between coarse-grained and fine-grained channel maps based on the latent diffusion model framework of spectrum residual transfer. The trained network architecture can then sample a fine-grained channel map from a given coarse-grained channel map using only a few steps. Compared to traditional generative diffusion models, this fast channel map construction method achieves higher inference efficiency in constructing fine-grained channel maps and higher construction accuracy compared to traditional baseline methods.

[0034] like Figure 2 As shown, in specific applications, the method for rapidly constructing a diffusion model-assisted channel map can include the following steps: collecting coarse-grained and fine-grained channel map data for a specified environment according to different sampling intervals; dividing the paired channel map data into training and testing sets; training the latent diffusion model based on channel map residual transfer using the training set, continuously optimizing the model parameters, and obtaining the network and its weight parameters that perform best on the testing set; and rapidly constructing the corresponding channel map using the trained model for the input coarse-grained channel map.

[0035] The latent diffusion model based on channel map residual transfer specifically includes, within the latent space: establishing a Markov chain based on the residuals between paired data using a diffusion model framework; constructing a time-step embedded channel map residual estimation network; and reconstructing a fine-grained channel map based on network output sampling. This embodiment establishes a Markov chain between fine-grained and coarse-grained channel maps based on residual transfer and is based on a diffusion model architecture. The difference from traditional diffusion models is that a noise-adding scheme is designed to noisily add noise to the fine-grained channel map during the forward process to obtain a coarse-grained channel map, significantly reducing the number of noise-adding steps. This allows the trained model to obtain relatively ideal fine-grained channel map data from the coarse-grained channel map with only a limited number of sampling steps.

[0036] In step S2, the coarse-grained channel map within the latent space is obtained. With fine-grained channel maps The residual between them is Design a monotonically increasing condition that satisfies and sequence and define As a forward process The residual transfer coefficient in; For time steps, This is a potential representation of the initial coarse-grained channel map. Represents the forward diffusion process The probability distribution, The mean is variance is The Gaussian distribution, where It is a hyperparameter that controls the noise variance. It is an identity matrix.

[0037] In this embodiment, the model as a whole learns the mapping relationship between fine-grained and coarse-grained channel spectra within a latent space, thereby reducing the complexity of model training. For the channel spectra data used for training, an additional network for latent space mapping is trained, reducing the dimensionality of the data within the latent space. For the channel spectra data, a fine-tuned VQGAN network can be used to learn the data mapping to the latent space encoding-decoding function, while a corresponding codebook is designed, and the codebook and encoding / decoding accuracy are optimized simultaneously during training.

[0038] In step S4, the loss function of the entire residual estimation network is designed to be the mean square error between the fine-grained channel map constructed by the time-step embedded residual estimation network and the true channel map, which is... ,in As a weighting factor, For network models, Represents the learning parameters of the network. This represents the Euclidean norm.

[0039] For example, in a specific large-scale MIMO-OFDM scenario, it is possible to generate Channel information in the communication area is used to obtain paired coarse-grained and fine-grained channel spectrum data using different sampling intervals. The dataset was divided into training and test sets in a 4:1 ratio. The Adam optimization algorithm and supervised training were used to train the model parameters to minimize the loss function and obtain the optimal residual estimation model parameters. The batch size was set to 64, and the learning rate was set to... This training strategy is used to iterate 500k times. Finally, the trained model is used for fast reconstruction of the channel map of the target region in a wireless communication system. Based on the input coarse-grained channel map, samples are taken from it. The fine-grained channel map is obtained step by step.

[0040] In this embodiment, in step S3, a feasible implementation of the channel map residual estimation network is as follows: the first processing stage includes: stage 0 consisting of convolutional layers, stage 1 consisting of two residual blocks and a window attention module, stage 2 consisting of a downsampling module, two residual blocks and a window attention module, and stage 3 consisting of a downsampling module, two residual blocks and a window attention module; the second processing stage is stage 4 consisting of two residual blocks and a window attention module; the third processing stage includes: stage 5 consisting of residual blocks and a window attention module, stage 6 consisting of an upsampling module, residual blocks and a window attention module, stage 7 consisting of an upsampling module, residual blocks and a window attention module, and stage 8 consisting of convolutional layers; the data output from stage 8 is pre-trained and fine-tuned before being decoded by a generative adversarial network to obtain the constructed fine-grained channel map.

[0041] To enable those skilled in the art to better understand the present invention, the specific steps of the embodiments of the present invention will be described below with reference to a specific scenario, taking channel power prediction as an example.

[0042] I. System Model

[0043] Consider a large-scale MIMO-OFDM communication scenario, where a base station communicates with multiple stations distributed across a geographical area. Each user equipment communicates using a single antenna. The base station is equipped with a uniform planar array containing... One antenna element, , These represent the number of antennas in the vertical and horizontal directions, respectively. For locations... The user equipment, whose radio channel contains Different physical paths. Let... The total number of OFDM subcarriers, of which One is the active subcarrier. Located at position The space-frequency channel response model between the user equipment and the base station on the active subcarrier is as follows:

[0044] (1)

[0045] in Indicates the Kronecker product. Indicates complex gain. , and They respectively represent the locations UE at the location The path has a propagation delay Normalized horizontal angle and vertical angle The guide vector on.

[0046] Considering that the channel power received by the user equipment will vary depending on the propagation environment (e.g., propagation loss and building reflection) attenuate. Based on formula (1), we can define the location attenuation. The channel power of the user equipment at that location is

[0047] (2)

[0048] in This represents a value in dB. This represents the expected operation. In this embodiment, the set of channel power values ​​at potential UE locations is defined as the unstructured CF. Furthermore, to obtain the structured CF, we divide the communication region into grids and collect the channel power values ​​within each spatial grid. Given a size of... The square communication area, we along shaft and The axis discretizes this region. This is done by introducing a resolution factor. With minimum grid spacing unit and The communication area is discretized. Each spatial grid... The two-dimensional coordinates can be represented as ,in and In this way, we rearrange these channel power values ​​along the spatial dimension into a two-dimensional tensor:

[0049] (3)

[0050] Therefore, the dimension of the obtained structured CF is With resolution factor With the increase of , the granularity of CF is improved.

[0051] II. Problem Statement

[0052] set up and These represent coarse-grained CF and fine-grained CF The resolution scale, of which Usually Integer multiples of. The core of this problem lies in utilizing the provided coarse-grained CF. To reconstruct fine-grained CF .

[0053] With coarse-grained CF For input, fine-grained CF For the output, the key is to estimate the residual between the two to obtain the intrinsic mapping relationship, which is expressed as:

[0054] (4)

[0055] in It is a mapping function The set of learnable parameters This represents the size of the training dataset. This is typically not a deterministic inverse problem because the distribution of high-resolution outputs based on low-resolution inputs rarely follows a simple parameter distribution, leading to poor performance of many feedforward neural network-based super-resolution regression methods. Generative diffusion models have proven effective in capturing complex empirical data distributions across domains. However, traditional generative diffusion models suffer from ill-posedness when generating fine-grained target CFs and are limited in inference efficiency due to the need for hundreds or even thousands of sampling steps. Therefore, this embodiment establishes a Markov chain between the high-resolution CF and the low-resolution CF via residual offset, rather than diffusing to a pure Gaussian noise tensor as in traditional diffusion models. Thus, equation (4) can be further expressed as...

[0056] (5)

[0057] in Indicates according to Find the expected value of the probability distribution. The gradient represents the logarithmic density of high-resolution CF. This represents the learned density. In summary, our method establishes a Markov chain between the high-resolution and low-resolution CFs via residual transfer, meaning the endpoint of the Markov process is not pure Gaussian noise, but rather the low-resolution CF data. In this way, the target high-resolution CF can be iteratively sampled from a given low-resolution CF in significantly fewer steps.

[0058] III. HR CF Reconstruction Method Based on CF-LRDM

[0059] Based on the aforementioned system model and problem statement, in step S2, within the Markov chain established using channel graph residual transfer based on the latent diffusion model architecture, the residual transfer diffusion process using HR CF is described as follows:

[0060] make and These represent a pair of fine-grained and coarse-grained CF data, respectively. The core of this method is to use a data set of length... The Markov process is used to progressively shift the residuals, thereby constructing a transformation from high-resolution CF to low-resolution CF. Specifically, high-resolution CF... With low resolution CF The residual between them is denoted as Its forward conversion process can be described as follows:

[0061] (6)

[0062] in For time steps, Derived from satisfaction and monotonically increasing sequence , and when hour ,and .also, It is a hyperparameter that controls the noise variance. It is an identity matrix. In this way, a forward diffusion process can be established. Based on... and The relationship, for a given pair of CF data We can derive the arbitrary time step Marginal distribution:

[0063] (7)

[0064] The reverse diffusion process starting from LR CF is as follows:

[0065] The inverse process aims to obtain information through the posterior distribution. Perform backsampling as follows:

[0066] (8)

[0067] Wherein, according to formula (7), in Under the conditions, Approximately ;and Indicates that it has learnable parameters from arrive Inverse conversion kernel:

[0068] (9)

[0069] Considering the variance parameter and and It is irrelevant, therefore we will and The settings are as follows:

[0070] (10)

[0071] in Indicates prediction The embedded time-step channel graph residual estimation network model has the following parameters: .

[0072] like Figure 3 As shown, in step S4, the channel graph residual estimation network is combined with the pre-trained fine-tuned VQGAN to construct a latent diffusion model based on channel graph residual transfer. Specifically, the pre-trained fine-tuned VQGAN is used to transfer the fine-grained CF... and coarse-grained CF Mapped into the latent space and Then, within the latent space, a fine-grained CF is established based on the channel spectrum residual transfer. and coarse-grained CF The Markov chain between the two is used to train a channel graph residual estimation network embedded at time steps to recover the latent representation of fine-grained CF, and finally the fine-grained CF is obtained by decoding.

[0073] To accurately approximate the reverse process, the parameters need to be optimized by minimizing the lower bound of negative evidence. ,Right now

[0074] (11)

[0075] in This represents the Kullback-Leibler divergence. Specifically, through this formula, we can see the Kullback-Leibler divergence over the entire... Within a step, the derived distribution is minimized. With estimated distribution Obtained by the divergence between them Combining formulas (6) and (7), we can derive the target distribution in formula (11). as follows:

[0076] (12)

[0077] According to formulas (9), (12) and (10), the objective function (11) can be simplified to

[0078] (13)

[0079] in This embodiment references the design of LDM and uses a fine-tuned version of VQGAN in the CF scenario. By learning the discrete representation of CF data, the data can be mapped into the latent space, and then the above model can be trained and used in the latent space.

[0080] In general, such as Figure 4 As shown, this embodiment first fine-tunes the VQGAN network based on the constructed channel map pairwise data, learns the discrete representation of the data, and then maps the data into the latent space, i.e., the dimension of... Data mapping as The latent representation; in the latent space, based on the residuals of the mapped fine-grained and coarse-grained channel maps, a Markov chain from the fine-grained channel map to the coarse-grained channel map is constructed using residual transfer. Then, a residual estimation network embedded in the time step is constructed to predict the residual at a given time step to recover the fine-grained channel map. The input is the coarse-grained channel map and the noisy fine-grained channel map at a given time step (dimension: ...). The output is an estimated fine-grained map of the channel (with dimensions in the latent space). Specifically, the network consists of three main stages: a first processing stage, a second processing stage, and a third processing stage. The first processing stage includes the following steps: Stage 0, composed of convolutional layers (changing the input data dimension from...). Become In this embodiment Stage 1 (output data dimensionality preserved) consists of two residual blocks and a window attention module. Stage 2, consisting of a downsampling module, two residual blocks, and a window attention module, comprises output data dimensions from... Become Stage 3, consisting of a downsampling module, two residual blocks, and a window attention module, comprises output data dimensions from... Become The second processing stage is stage 4, which consists of two residual blocks and a window attention module (output data dimensionality preservation). The third processing stage includes the following stages: Stage 5 (output data dimensionality preservation) consists of residual blocks and window attention modules. Stage 6, consisting of an upsampling module, residual blocks, and a window attention module (output data dimensions are...), consists of... Become Stage 7, consisting of an upsampling module, residual blocks, and a window attention module (output data dimensions are...), consists of... Become Stage 8 (output data dimension) consists of convolutional layers. The final stage 8 output data, after being fine-tuned by the VQGAN, yields the constructed fine-grained channel map (data dimension: ).

[0081] IV. Implementation Results

[0082] To enable those skilled in the art to better understand the present invention, the following provides a comparison of the estimation performance results of the CF fast reconstruction method in this embodiment with existing methods under several specific system configurations.

[0083] To comprehensively evaluate the effectiveness of the proposed method, we employ four metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Mean Squared Error (MSE), and Normalized Mean Squared Error (NMSE). Subsequently, we... Experiments were conducted on the CF reconstruction task, and its performance was compared with several baseline methods.

[0084] First, Table 1 shows the proposed CF-LRDM and baseline methods in terms of performance. Quantitative results on the CF reconstruction task.

[0085] Table 1. Comparison of performance metrics between CF-LRDM and baseline methods

[0086]

[0087] The baseline methods compared to these baseline methods include bilinear interpolation, nearest neighbor interpolation, the SRGAN method proposed in the paper "Photo-realistic singleimage super-resolution using a generative adversarial network," and the DRRN method proposed in the paper "Image super-resolution via deep recursive residual network." Specifically, compared to the baseline methods, the proposed method achieves competitive reconstruction performance, with NMSE, MSE, and PSNR values ​​of [values ​​to be filled in]. The SSIM scores were 1.570 and 46.329. Since the SSIM metric primarily reflects the perceptual characteristics of the human visual system, it is not a critical consideration in our CF application scenario. Nevertheless, the proposed method still achieved quite good SSIM results, slightly outperforming the bilinear interpolation method. For example... Figure 4 As shown, for qualitative comparison, we visualized random samples of CF reconstruction results using the proposed CF-LRDM and baseline methods. It can be seen that the proposed method effectively refines the CF, closely approximating the true CF with minimal error.

[0088] Next, Table 2 shows the CF-LRDM in this embodiment at different diffusion steps. The results under the configuration include four metrics and the inference time for generating fine-grained CF.

[0089] Table 2 Performance and inference time of CF-LRDM with different diffusion steps

[0090]

[0091] It's worth noting that as the number of diffusion steps increases, the inference time increases, while the four metrics only show slight fluctuations. This indicates that the model's performance stabilizes early on with fewer diffusion steps, and increasing the number of steps...

[0092] This did not result in a significant improvement. Therefore, the proposed model can operate efficiently with relatively few diffusion steps, thereby generating competitive reconstruction results in a shorter inference time. This characteristic highlights its applicability in practical scenarios where a balance between computational efficiency and output quality is required.

[0093] Based on the same inventive concept, this invention discloses a diffusion model-assisted rapid channel map construction system for implementing the diffusion model-assisted rapid channel map construction method of any of the foregoing embodiments. The system includes: a sample acquisition module for acquiring channel information of a target region and constructing paired channel map data of different granularities; a model construction module for establishing a Markov chain between fine-grained and coarse-grained channel maps using channel map residual transfer based on a latent diffusion model architecture; and a channel map residual estimation network embedded with time steps for predicting the channel map residual at a given time step to recover the fine-grained channel map. The system takes a channel map as input, a coarse-grained channel map latent representation, and a noisy fine-grained channel map latent representation at a given time step, and outputs an estimated fine-grained channel map latent representation. It also constructs a latent diffusion model based on channel map residual transfer by combining a channel map residual estimation network with a pre-trained, finely tuned generative adversarial network. A model training module trains the constructed channel map residual estimation network using supervised training to obtain optimal parameters. Finally, a fast reconstruction module rapidly reconstructs the fine-grained channel map of a target region in a wireless communication system using the latent diffusion model based on channel map residual transfer.

[0094] An embodiment of the present invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the diffusion model-assisted rapid channel map construction method of any of the foregoing embodiments.

[0095] In a specific implementation, the device includes a processor, a communication bus, a memory, and a communication interface. The processor can be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. The communication bus may include a path for transmitting information between the aforementioned components. The communication interface, using any transceiver-like device, is used for communicating with other devices or communication networks. The memory can be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), read-only optical disc (CD-ROM) or other optical disc storage, disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory can exist independently and be connected to the processor via a bus. The memory can also be integrated with the processor.

[0096] The memory stores application code that executes the present invention and is controlled by the processor. The processor executes the application code stored in the memory to implement the method provided in the above embodiments. The processor may include one or more CPUs, or multiple processors, each of which may be a single-core processor or a multi-core processor. Here, "processor" may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0097] This invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the diffusion model-assisted rapid channel map construction method of any of the foregoing embodiments.

[0098] The invention also discloses a computer program product, including a computer program stored in a computer-readable storage medium, wherein the computer program, when executed by a processor, implements the steps of the diffusion model-assisted rapid channel map construction method of any of the foregoing embodiments.

[0099] Based on the same inventive concept, this invention also provides a wireless communication system, including a base station and multiple user terminals. The base station or user terminals are equipped with: a sample acquisition module for acquiring channel information of a target area and constructing paired channel map data of different granularities; a model construction module for establishing a Markov chain between fine-grained and coarse-grained channel maps using channel map residual transfer based on a latent diffusion model architecture; a channel map residual estimation network embedded in a time step for predicting the channel map residual at a given time step to recover the fine-grained channel map, with inputs being the coarse-grained channel map latent representation and the noisy fine-grained channel map latent representation at a given time step, and outputting the estimated fine-grained channel map latent representation; and combining the channel map residual estimation network with a pre-trained fine-tuned generative adversarial network to construct a channel map-based... The latent diffusion model for channel map residual transfer is described. The channel map residual estimation network includes a first processing stage composed of a convolutional module, a residual module, a window attention module, and a downsampling module, used to progressively extract deep features of the input; a second processing stage composed of a residual module and a window attention module, used for feature transformation of the latent representation of the channel map; and a third processing stage composed of a convolutional module, a residual module, a window attention module, and an upsampling module, used to progressively reconstruct the latent representation of the fine-grained channel map; a model training module, used to train the constructed channel map residual estimation network using supervised training to obtain optimal parameters; and a fast reconstruction module, used to rapidly reconstruct the fine-grained channel map of the target region in the wireless communication system using the latent diffusion model based on channel map residual transfer. Specific implementations of each module are described in the aforementioned method embodiments and will not be repeated here.

[0100] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0101] It should also be noted that relational terms such as "first" and "second" in this specification are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] 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.

[0103] 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 rapid construction of channel maps assisted by a diffusion model, characterized in that, Includes the following steps: Obtain channel information for the target area and construct paired channel map data at different granularities; Based on the latent diffusion model architecture, a Markov chain between fine-grained and coarse-grained channel maps is established using channel map residual transfer. A channel map residual estimation network embedded with time steps is constructed to predict the channel map residual at a given time step to recover the fine-grained channel map. The inputs are a coarse-grained channel map latent representation and a noisy fine-grained channel map latent representation at a given time step, and the output is the estimated fine-grained channel map latent representation. The channel map residual estimation network includes a first processing stage composed of a convolutional module, a residual module, a window attention module, and a downsampling module, used to progressively extract deep features of the input; a second processing stage composed of a residual module and a window attention module, used for feature transformation of the channel map latent representation; and a third processing stage composed of a convolutional module, a residual module, a window attention module, and an upsampling module, used to progressively reconstruct the latent representation of the fine-grained channel map. A latent diffusion model based on channel graph residual transfer is constructed by combining a channel graph residual estimation network with a pre-trained fine-tuned generative adversarial network. The constructed channel graph residual estimation network is trained in a supervised manner to obtain the optimal parameters. We utilize a latent diffusion model based on channel map residual transfer to rapidly reconstruct fine-grained channel maps of target regions in wireless communication systems.

2. The method for rapid construction of channel maps assisted by a diffusion model according to claim 1, characterized in that, In the latent diffusion model based on channel graph residual transfer, the fine-tuned VQGAN network learns to map channel graph data to latent space encoding-decoding functions, and designs corresponding codebooks. During pre-training, the codebook and encoding / decoding accuracy are optimized simultaneously.

3. The method for rapid construction of channel maps assisted by a diffusion model according to claim 1, characterized in that, In the Markov chain between fine-grained and coarse-grained channel maps established using channel map residual transfer, the coarse-grained channel map is located in the latent space. With fine-grained channel maps The residual between them is Design a monotonically increasing condition that satisfies and sequence and define As a forward process The residual transfer coefficient in; For time steps, This is a potential representation of the initial coarse-grained channel map. Represents the forward diffusion process The probability distribution, The mean is variance is The Gaussian distribution, where It is a hyperparameter that controls the noise variance. It is an identity matrix.

4. The method for rapid construction of channel maps assisted by a diffusion model according to claim 3, characterized in that, The loss function for training the channel map residual estimation network model is: ,in As a weighting factor, For network models, Represents the learning parameters of the network. This represents the Euclidean norm.

5. The method for rapid construction of channel maps assisted by a diffusion model according to claim 1, characterized in that, The first processing stage of the channel map residual estimation network includes: stage 0 consisting of convolutional layers; stage 1 consisting of two residual blocks and a window attention module; stage 2 consisting of a downsampling module, two residual blocks, and a window attention module; and stage 3 consisting of a downsampling module, two residual blocks, and a window attention module. The second processing stage is stage 4 consisting of two residual blocks and a window attention module. The third processing stage includes: stage 5 consisting of residual blocks and a window attention module; stage 6 consisting of an upsampling module, residual blocks, and a window attention module; stage 7 consisting of an upsampling module, residual blocks, and a window attention module; and stage 8 consisting of convolutional layers. The data output from stage 8 is pre-trained and fine-tuned before being decoded by a generative adversarial network to obtain the constructed fine-grained channel map.

6. A diffusion model-assisted rapid channel map construction system, used to implement the diffusion model-assisted rapid channel map construction method according to any one of claims 1-5, characterized in that, include: The sample acquisition module is used to acquire channel information of the target area and construct paired channel map data at different granularities. The model building module is used to establish a Markov chain between fine-grained and coarse-grained channel maps based on the latent diffusion model architecture using channel map residual transfer. Construct a channel map residual estimation network embedded with time steps to predict the channel map residual at a given time step in order to recover the fine-grained channel map. The input is the coarse-grained channel map latent representation and the noisy fine-grained channel map latent representation at a given time step. The output is the estimated fine-grained channel map latent representation. Furthermore, a latent diffusion model based on channel graph residual transfer is constructed by combining the channel graph residual estimation network with a pre-trained fine-tuned generative adversarial network. The model training module is used to train the constructed channel map residual estimation network in a supervised manner to obtain the optimal parameters; In addition, a fast reconstruction module is provided for rapidly reconstructing fine-grained channel maps of target regions in a wireless communication system using a latent diffusion model based on channel map residual transfer.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the diffusion model-assisted rapid channel map construction method according to any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the diffusion model-assisted rapid channel map construction method according to any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the diffusion model-assisted rapid channel map construction method according to any one of claims 1-5.

10. A wireless communication system, comprising a base station and multiple user terminals, characterized in that, The base station or user terminal is equipped with: The sample acquisition module is used to acquire channel information of the target area and construct paired channel map data at different granularities. The model building module is used to establish a Markov chain between fine-grained and coarse-grained channel maps based on the latent diffusion model architecture using channel map residual transfer. Construct a channel map residual estimation network embedded with time steps to predict the channel map residual at a given time step in order to recover the fine-grained channel map. The input is the coarse-grained channel map latent representation and the noisy fine-grained channel map latent representation at a given time step. The output is the estimated fine-grained channel map latent representation. Furthermore, a latent diffusion model based on channel graph residual transfer is constructed by combining a channel graph residual estimation network with a pre-trained fine-tuned generative adversarial network. The channel graph residual estimation network includes a first processing stage composed of a convolutional module, a residual module, a window attention module, and a downsampling module, used to progressively extract deep features of the input; a second processing stage composed of a residual module and a window attention module, used for feature transformation of the latent representation of the channel graph; and a third processing stage composed of a convolutional module, a residual module, a window attention module, and an upsampling module, used to progressively reconstruct the latent representation of the fine-grained channel graph. The model training module is used to train the constructed channel map residual estimation network in a supervised manner to obtain the optimal parameters; In addition, a fast reconstruction module is provided for rapidly reconstructing fine-grained channel maps of target regions in a wireless communication system using a latent diffusion model based on channel map residual transfer.

Images