Space-time-frequency multi-dimensional channel prediction method based on digital expression

Abstract

The invention provides a space-time-frequency multi-dimensional channel prediction method based on a digital expression, and belongs to the technical field of mobile communication. The method comprises the following steps: receiving original channel state information data of a sensor terminal; filling missing values of the original channel state information data through affinity propagation clustering to obtain local actual measurement channel state information data; obtaining global approximate channel state information data through historical channel state information data by using a channel generation model based on a digital table; fusing the global approximate channel state information data and the local actual measurement channel state information data to obtain fused channel state information data; and obtaining a global channel state information prediction value based on the fused channel state information data by using a channel prediction network. According to the method disclosed by the invention, efficient and accurate prediction in a complex channel environment can be realized by utilizing local channel data in a rapid time-varying communication scene in a complex shielding environment.

Description

Technical Field

[0001] The present invention relates to the field of mobile communication technology, and in particular to a time-space-frequency multi-dimensional channel prediction method based on digital cousins. Background Art

[0002] Future mobile communication systems will develop towards wider frequency bands, larger antenna arrays, higher mobile speeds and more user devices to support instantaneous, ultra-high-speed wireless connections. Channel prediction, as one of the means of obtaining channel information, is a key technology to support low-latency and high-precision communications. In complex shielded environments such as underground tunnels and dense urban buildings, signal propagation is affected by factors such as obstacle obstruction, multipath effects and Doppler frequency shift, resulting in significant differences in channel propagation characteristics at different locations. Especially in strong shielding areas, it is difficult for sensor terminals to obtain effective channel measurement data and can only collect local channel information at limited locations, which makes the measured data unable to reflect the dynamic changes and propagation characteristics of the global channel. In addition, wireless channels have obvious fast time-varying and non-stationary characteristics, and channel state information (CSI) shows highly dynamic changes in time, frequency and space domains.

[0003] At present, the methods of channel prediction are mainly divided into statistical modeling-based methods and data-driven methods. The statistical model-based methods usually use the time correlation of the channel historical data to describe the channel characteristics through parameterized models, and use time series methods to predict the future channel state. The data-driven channel prediction method regards the channel state as a mapping process from one parameter subspace to another parameter subspace, and learns the mapping relationship of the subspace through a deep learning model. This type of scheme usually uses Bayesian linear estimators or sensor data as raw inputs, selects and learns channel features by constructing deep neural networks, and finally predicts the channel state. However, in complex channels with fast time variations and non-stationary, the existing channel prediction schemes based on statistical models are difficult to adapt to the dynamic changes of the channel. In particular, in time-frequency selective fading channels, channel state information usually presents highly nonlinear characteristics, and autoregressive models based on linear assumptions (such as Kalman filtering) cannot accurately capture these changes. Therefore, this type of method is only applicable to channels with slow parameter changes and low fading rates. In addition, in complex propagation environments such as underground tunnels and dense urban buildings, signal propagation characteristics vary significantly depending on the location. Since it is difficult for sensor terminals to effectively measure channel data in severely shielded areas, the actual available data samples are very limited. Existing data-driven methods are mostly verified in ideal simulation environments or complete data sets, and have low adaptability to real complex scenarios under scarce samples, resulting in poor generalization of the model in actual channel environments. At the same time, most current solutions lack the mining of multi-domain correlations of wireless channels. In complex time-varying channels, the channel state information changes rapidly over time, resulting in low correlation in the time domain. It is difficult to achieve accurate prediction by simply relying on the correlation of the channel in the time domain.

[0004] Therefore, in order to achieve efficient and accurate prediction of channel status using scarce channel data in complex shielding environments, a time-space-frequency multidimensional channel prediction method based on digital cousins ​​is urgently needed. Summary of the invention

[0005] In view of this, the present invention provides a method for predicting multi-dimensional channels based on digital cousins ​​in space-time and frequency. First, the distortion of measurement data caused by channel noise is eliminated by affinity propagation (AP) clustering; a lightweight digital cousin channel is constructed by using a cross-dimensional generative adversarial network (Cross-Dimensional GAN, CrossDGAN) to generate global channel data with similar space-time and frequency characteristics to the real measurement data; a graph convolutional neural network is used for feature propagation, and the optimal fusion is achieved based on a progressive spiral data fusion strategy to obtain globally accurate channel information; a convolutional long short-term memory network is combined to realize the prediction of channel state information of the joint space-time and frequency channel characteristics. This method solves the problem that the local measured channel data in a complex shielding environment is difficult to accurately reflect the global channel state, and the time domain correlation of the time-varying non-stationary channel is low, resulting in insufficient channel state information prediction accuracy.

[0006] To this end, the present invention provides the following technical solutions: A time-space-frequency multi-dimensional channel prediction method based on digital cousins, comprising: Receive original channel state information data from the sensor terminal; Filling missing values ​​of original channel state information data by affinity propagation clustering to obtain local measured channel state information data; Using a digital cousin-based channel generation model to obtain global approximate channel state information data through historical channel state information data; fusing global approximate channel state information data and local measured channel state information data to obtain fused channel state information data; A channel prediction network is used to obtain a global channel state information prediction value based on the fused channel state information data.

[0007] Further, the filling missing values ​​of the original channel state information data by affinity propagation clustering includes: Normalizing each attribute value of the original channel state information data; The Euclidean distance between data targets is calculated based on the normalized attribute values, and the similarity matrix between data targets is constructed; Dividing the original channel state information data set into a complete data subset and an incomplete data subset; Calculate density indicators of data targets based on complete data subsets and incomplete data subsets; The data target with the highest density is taken as the cluster center, and the relevant radius of the cluster neighborhood is obtained based on the similarity matrix between the data targets, and affinity propagation clustering is performed; When the maximum number of iterations is reached or when the cluster center does not change, the similarity between the incomplete data subset and the complete data subset in the same cluster is calculated; Get the entropy value of the complete data subset; The entropy value of the complete data subset and the similarity between the incomplete data subset and the complete data subset in the same cluster are combined to obtain the attribute value of the incomplete data subset, thereby completing the orderly filling of the missing values ​​of the original channel state information data.

[0008] Further, the method of obtaining global approximate channel state information data by using the digital cousin-based channel generation model through historical channel state information data includes: The local historical channel state information data is mapped into a real potential sequence through a segmented embedding layer, and a second-order attention layer based on an attention mechanism is introduced to capture the cross-dimensional dependencies between the time domain, frequency domain, and spatial domain; A random noise sequence is introduced into the simulation environment to simulate random factors, and a synthetic potential sequence is generated based on the random noise sequence through a sequence generator; reconstructing a synthetic time-frequency spatial channel coefficient matrix based on the synthetic latent sequence; Denormalizing the composite channel coefficient matrix to obtain global approximate channel state information data; The weights of the embedding network, restoration network, sequence generator, and sequence discriminator are updated according to the reconstruction loss, unsupervised loss, and supervised loss.

[0009] Furthermore, the sequence generator and the sequence discriminator are subjected to adversarial training to optimize the quality of the synthesized potential sequence, including: The sequence generator aims to maximize the score given by the sequence discriminator and updates the parameters by minimizing its loss function; The sequence discriminator aims to maximize the score of the true latent sequence while minimizing the score of the synthetic latent sequence, and updates its parameters by minimizing the loss function.

[0010] Furthermore, the channel prediction network includes: a convolutional long short-term memory network.

[0011] Further, fusing the global approximate channel state information data and the local measured channel state information data through a graph convolutional neural network includes: Using sensor terminals as nodes, construct an initial graph of local measured channel state information data and an initial graph of global approximate channel state information data; Obtaining a number of measured node embedding matrices and approximate node embedding matrices based on an initial graph of local measured channel state information data and an initial graph of global approximate channel state information data through a data fusion model; Adaptive complementary weighted fusion method is used to obtain the comprehensive node matrix based on the measured node embedding matrix and the approximate node embedding matrix; Concatenate all comprehensive node embedding matrices to obtain the global node matrix; The decoder is used to obtain the fused channel state information data feature map based on the global node matrix.

[0012] Furthermore, the data fusion model:

[0013]

[0014] in, The graph normalized Laplacian matrix representing the local measured channel state information data, represents the measured node embedding matrix, , The graph minimum-maximum normalized node feature matrix representing the local measured channel state information data, represents a multi-layer perceptron; A graph normalized Laplacian matrix representing the global approximate channel state information data; represents the approximate node embedding matrix, , A graph min-max normalized node feature matrix representing the global approximate channel state information data; is the sum function; Indicates the number of iterations.

[0015] Furthermore, the method of using the adaptive complementary weighted fusion method to obtain a comprehensive node matrix based on the measured node embedding matrix and the approximate node embedding matrix includes: With the minimum total mean square error as the constraint, the corresponding optimal weighting factors are assigned to the measured node embedding matrix and the approximate node embedding matrix.

[0016] Further, the using the channel prediction network to obtain a global channel state information prediction value based on the fused channel state information data includes: A multi-dimensional convolution kernel is used to extract the spatiotemporal features of the fused channel state information data feature map to obtain output features of a preset length; A layer of convolutional long short-term memory network is used as a predictor to update the output features; The convolutional neural network is used to perform feature recovery based on the updated output features to obtain the global channel state information prediction value.

[0017] Beneficial effects of the present invention: The present invention proposes a time-space-frequency multi-dimensional channel prediction method based on digital cousins, which can use local channel data to achieve efficient and accurate prediction in complex channel environments in fast time-varying communication scenarios under complex shielding environments. The method of the present invention includes three parts: CSI compensation based on AP clustering, cousin channel construction based on CrossDGAN, and channel prediction by fusion of virtual and real data.

[0018] The CSI compensation method based on AP clustering of the present invention divides the data into different subsets through normalization processing, which can reduce the imbalance of the data and improve the quality and filling efficiency of the data. By updating the attraction and attribution matrices, the normal transmission of messages is ensured to achieve neighbor propagation. The similarity between data is clustered, and the CSI missing values ​​are filled according to the information entropy, which solves the distortion problem of the measured channel data and ensures that high-precision channel data can still be obtained in complex shielding environments.

[0019] The cousin channel construction method based on CrossDGAN of the present invention can effectively capture the multi-dimensional correlation of the channel in the time domain, frequency domain and space domain by learning the high-dimensional characteristics of the wireless channel from historical real data. This method constructs a digital cousin channel, which can generate channel simulation data with similar distribution characteristics to the real environment at low cost and on a large scale, reducing the overhead of frequent collection of measured data. It can also cover locations that are difficult for sensors to measure, obtain richer channel samples, supplement the deficiencies of measured data, and has good adaptability when facing fast time-varying non-stationary channels.

[0020] The channel prediction method of the present invention, which combines virtual and real data, can extract the instantaneous propagation characteristics of the channel by modeling the graph structure of the CSI data at multiple node positions. The progressive spiral fusion strategy effectively fuses the measured channel data with the approximate channel data, and can obtain global accurate channel information, ensuring the spatiotemporal consistency of fast time-varying channel data. The convolutional long short-term memory network with excellent spatiotemporal data processing capabilities is used to realize the joint prediction of channel state information, providing efficient and accurate channel prediction solutions for various complex wireless scenarios and multiple communication systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0022] Figure 1 This is a flow chart of a channel prediction method according to an embodiment of the present invention; Figure 2 This is a flow chart of a method for compensating missing values ​​of CSI data in an embodiment of the present invention; Figure 3 This is a diagram of the digital cousin channel model architecture in an embodiment of the present invention; Figure 4 A flow chart of simulation channel data generation in an embodiment of the present invention; Figure 5 Schematic diagram of a virtual-real data fusion model in an embodiment of the present invention; Figure 6 This is a flowchart of the progressive spiral data fusion in an embodiment of the present invention. Figure 7 This is a diagram of a channel prediction network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] The present invention provides a time-space-frequency multi-dimensional channel prediction method based on digital cousins. Through the CSI compensation strategy based on AP clustering, it ensures to obtain more accurate channel measurement data and solves the problems of amplitude attenuation and phase distortion caused by environmental noise; by constructing a lightweight digital cousin channel with similar distribution characteristics and multi-domain correlation to the real channel, it realizes the filling and expansion of global channel state information, solves the problem that the propagation characteristics of different positions are quite different due to the complex channel environment, and the measured data is limited by the acquisition space, and only effective channel measurement data can be obtained in the area with less shielding, resulting in the limited channel data samples cannot reflect the global characteristics of the channel; by introducing the local detail features of the measured data, it is fused with the synthetic data of the cousin channel, so as to accurately capture the real channel state at the current moment. The prediction model extracts the time-space characteristics of the fused channel data sequence, makes full use of the correlation of the channel in the time, space and frequency domains, so as to realize the accurate prediction of the fast time-varying channel state; further solves the problem that the digital cousin channel can reflect the overall statistical characteristics and long-term change trend of the global based on the complex dynamic characteristics of the fast time-varying channel, but cannot accurately simulate the instantaneous state of the actual propagation environment at a certain moment.

[0026] Combination Figure 1 As shown, the present invention includes three parts: CSI compensation based on AP clustering, channel modeling based on digital cousins, and a channel prediction method of virtual and real data fusion, specifically: Consider a wireless communication system in a complex environment, where the base station deploys K uniform rectangular antenna arrays and there are M single-antenna sensor terminals. The spatial channel vector between the sensor terminal and the base station It is expressed as:

[0027] in, is the number of paths for signal multipath transmission, , and Respectively represent The gain, azimuth and elevation of each transmission path. is the antenna array response matrix, whose values ​​vary with the array geometry.

[0028] Channel state information is the superposition of signals from multiple different transmission paths. After defining the channel response of a single sensor terminal device, the spatial channel matrix of all sensor terminal devices is It is expressed as:

[0029] The sensor measures the channel information based on the pilot signal. Each sensor terminal periodically sends a Q-dimensional pilot signal to the base station. , so the signal matrix received by the terminal is for:

[0030] in, is the pilot matrix, Represents the noise matrix, whose elements follow Gaussian distribution , is a sign function.

[0031] S1. Compensation for missing values ​​of channel state information based on AP clustering method; Combination Figure 2 , for this embodiment, the method of orderly filling of missing values ​​based on AP clustering: in order to improve data quality, the data is divided into different subsets, and the data is normalized according to the standard deviation idea, and the numerical value is controlled within a fixed interval to reduce the imbalance of the data. The attraction degree and the belonging degree update matrix are constructed separately to ensure the normal transmission of messages to achieve the purpose of neighbor propagation. The similarity between different data is used as the basis for clustering, the radius parameter of the clustering neighborhood is obtained, the cluster center is determined by the data point density index, and the adjacent data are clustered together. Using the concept of entropy value, the weighting coefficient is calculated according to the data similarity, the attribute value of the missing data is determined, and the orderly filling of the missing values ​​is realized. The specific steps are as follows: Step 1-1: For channel data set , where any data target All by Attributes The feature vector Indicates. For data targets Normalize the attribute values ​​within:

[0032] in, and Respectively represent data targets The mean and standard deviation of all vector attributes in , Represents the channel vector In Properties The value on .

[0033] Step 1-2: Calculate the Euclidean distance between each data target based on the attribute value and establish a similarity matrix between targets; Step 1-3: Update the attraction and belongingness matrix: During the execution of the AP clustering algorithm, the attraction matrix R must be continuously updated. Assign to the corresponding matrix row, The data of the degree of attribution matrix and the degree of membership matrix corresponding to the row will be put into the same class for calculation, and then the first Column attraction matrix .

[0034] Similarly, calculate the first Column belonging Through the attraction matrix Column Value , cluster the matrix according to the key value Assign to the corresponding column to obtain the corresponding attribution matrix .

[0035] Step 1-4: Assemble the data Partition into subsets of complete data and incomplete data subsets , and obtain the relevant radius parameters of the cluster neighborhood and :

[0036] Step 1-5: Calculate the density index of the data points and use the data with the highest density as the cluster center:

[0037] in, Indicates The cluster center selected.

[0038] Step 1-6: By continuously modifying the density index, the data with the highest density value is used as the next cluster center; modifying the density index includes:

[0039] Through steps 1-3 to 1-6, the attraction and belonging between data are continuously updated to transmit messages and realize neighbor propagation. After selecting the cluster center, the data clustering process is completed according to the selected cluster center. The weighted value of the corresponding attribute of the missing data in the cluster cluster is used as the predicted value. By introducing the concept of entropy value and combining the similarity between data, the weighted coefficient is calculated to obtain an accurate weighted coefficient.

[0040] Step 1-7: Obtain missing data using similarity calculation method and other complete data in the same cluster The similarity is expressed as .

[0041]

[0042] in, , Indicates channel data and In Properties The probability of having the same value.

[0043] Step 1-8: Similarity for missing data After unitization, we get the normalized similarity:

[0044] Step 1-9: Calculate the entropy of all complete data:

[0045] in, Indicates The channel data The entropy value of an attribute.

[0046] Step 1-10: Calculate attribute values ​​for missing data:

[0047] Through step 1, the attribute values ​​of the missing data are obtained, and then the missing values ​​are filled in an orderly manner.

[0048] S2, construct a channel model based on digital cousins; In order to obtain accurate and effective channel information, the real physical world is digitized to construct a digital cousin wireless channel with similar characteristics to the real channel, and the channel state information of the global position is obtained. Since digital twins need to synchronize data in real time, there are problems such as high cost and difficulty in expansion. As a new virtualization method, the goal of digital cousin is to create a virtual version that is geometrically and semantically similar to the object, without being restricted to precise reproduction. Therefore, building a digital cousin channel has stronger advantages in cost, flexibility, and cross-domain generalization.

[0049] In this embodiment, the target problem is firstly modeled, and then the high-dimensional feature space of the wireless channel is learned from the historical real data collected by the sensor terminal to generate synthetic channel data with similar distribution characteristics and time domain, frequency domain and spatial domain correlation as the real channel data.

[0050] The channel data of length T is discretized into multiple time stamps, and the channel between each sensor terminal and the base station is modeled in each time stamp. , will The actual channel frequency response between the sensor terminals and the base station is arranged into a vector , then the time series channel data set collected by the sensor terminal can be expressed as:

[0051] The goal is to use the dataset To estimate a joint distribution closest to the time-frequency-space wireless channel matrix The probability density function of This goal is usually difficult to estimate mathematically through high-dimensional probability density functions. Therefore, in this embodiment, autoregressive decomposition is used to decompose the feature space learning problem into two sub-problems: local distribution learning and step-by-step conditional distribution learning. Synthetic sequence The channel coefficient matrix in Should be consistent with the real sequence The channel coefficient matrix in has the same distribution, ignoring the time index In the case of , the local distribution learning problem is expressed as:

[0052] in, is the Kullback-Leibler divergence, which measures the gap between two distributions. and They are and Obey the distribution. and It is a channel coefficient matrix containing frequency domain and spatial domain information, so local distribution learning can ensure that the synthetic data set has a channel coefficient distribution and frequency domain and spatial domain correlation similar to the real data set, but ignores the time domain correlation of the wireless channel. Therefore, the step-by-step conditional distribution learning problem is expressed as:

[0053] in, is the stepwise conditional distribution of the synthetic sequence, and is the stepwise conditional distribution of the real sequence. By learning the time domain correlation through the stepwise conditional distribution, the synthetic channel dataset can have the same time domain correlation as the real channel dataset. By solving the local distribution problem and the stepwise conditional distribution problem, the feature space of the high-dimensional wireless channel is learned so that sampling can be performed in it, thereby constructing a statistical digital cousin channel with similar distribution characteristics and time, frequency and spatial domain correlation as the real channel.

[0054] In order to realize the joint mining of temporal dependency, spatial dependency, and dependency between environment and channel, the CrossDGAN joint model with the ability to mine cross-dimensional dependencies is used to learn the high-dimensional features of wireless channels from the collected historical real data. The model mainly consists of a dimensional segmentation embedding layer, a second-order attention layer, a recovery network layer, a sequence generator, and a sequence discriminator. The overall architecture of the model is as follows: Figure 3 shown.

[0055] The digital cousin channel constructed in this embodiment is not limited to a single channel feature dimension, but a multi-dimensional potential space is established to express the spatiotemporal dependency and multi-level features of the channel data. Therefore, the dimensional segmented embedding network is used for representation learning to reduce the dimension of the time-frequency-space channel coefficient matrix. The segmented embedding layer divides each dimension of the three-dimensional input time series into The slices are then embedded and positionally encoded:

[0056] in, represents the learnable projection matrix, is the dimension after embedding, yes The dimension fragments, For location The embedding network provides a mapping from feature space to latent space, i.e. The resulting sequence generator is able to capture the underlying properties of channel coefficient distribution and multi-domain correlation by learning low-dimensional features.

[0057] In order to better explore the cross-dimensional dependencies between time, space, and frequency, a second-order attention layer based on the attention mechanism is introduced to enhance the cross-dimensional modeling capability. The second-order attention layer is divided into a cross-time layer and a cross-dimensional layer. The cross-time layer uses a multi-head self-attention mechanism for different time series:

[0058] in, is the input two-dimensional vector array, represents the number of segments of the time series, Representation layer normalization operation, Represents a multi-layer feed-forward network. , express and Output. The dependencies between data segments are captured and then fed into the cross-dimensional layer. In the cross-dimensional layer, the same learnable routing vector is used for the time series segments in each dimension. ;in To obtain the number of learnable vectors for the router, the routing vector first collects information from all dimensions and distributes the aggregated information to different dimensions. By capturing the cross-dimensional dependencies between the time domain, frequency domain, and spatial domain, the second-order attention layer can help the generator generate more consistent and realistic channel data from the latent space, so that the synthesized channel data has similar distribution characteristics.

[0059] Therefore, the restoration network is used to provide an inverse mapping from the latent space to the feature space. The intermediate vector output by the restoration network is and the recovered channel coefficient matrix Respectively expressed as:

[0060]

[0061] The restoration network accurately reconstructs the time-frequency spatial channel coefficient matrix according to the potential sequence output by the embedding network, so the reconstruction loss is defined as:

[0062] in, represents the batch size, Indicates Samples in Channel matrix sequence under time slots; It represents the restored channel matrix sequence of the corresponding sample. The digital cousin channel emphasizes the pseudo-randomness of the virtual channel. Even in different environments, the channel characteristics can still generate a reasonable distribution similar to the real channel. This pseudo-randomness requires that the generated channel data is not restricted to specific physical parameters, but is randomly generated according to certain generation strategies and environmental conditions. Therefore, the present invention introduces random noise sequences to simulate random and unpredictable factors in the environment, thereby providing random perturbations for the generation model. The sequence generator is used to convert the random noise sequence Mapped to a latent sequence, the score given by the sequence discriminator is maximized. The vector in the synthesized latent sequence is represented as:

[0063] The sequence discriminator is used to map the input latent sequence into a classification vector, maximizing the score of the true latent sequence while minimizing the score of the synthetic latent sequence. The intermediate vector output by the sequence discriminator is The scalars in the and classification vectors are represented as:

[0064]

[0065] in, Represents the classification of the input real data Or input synthetic data classification .

[0066] The sequence generator and the sequence discriminator run an adversarial process in the latent space to play a minimax two-player game to achieve Nash equilibrium. During the game, the generator continuously adjusts its generation strategy and eventually generates high-quality channel samples to ensure that the generated channel data has pseudo-randomness, so that the sequence generator can generate high-quality time series channel samples. , , and They represent the weights of the embedding network, restoration network, sequence generator, and sequence discriminator respectively.

[0067] S3, the process of generating digital cousin channel data using the digital cousin-based channel model is as follows Figure 4 As shown, the specific steps are as follows: Step 3-1: Initialize batch size , training hyperparameters, number of trainings, and embedding networks , restore the network layer , Sequence Generator and sequence discriminator ; Step 3-2: For the real channel dataset Perform minimum-maximum normalization; Step 3-3: Extraction training samples and map them to the true potential sequence:

[0068] Step 3-4: Reconstruct the true time-frequency-space channel coefficient matrix based on the true potential sequence:

[0069] Step 3-5: Perform Gaussian distribution sampling on the reconstructed sequence to simulate the random disturbance of the environment and obtain the sampling sequence , and then generate the synthetic latent sequence:

[0070] Step 3-6: Generate classification vectors for real latent sequences and synthetic latent sequences:

[0071]

[0072] Step 3-7: Update the weights of the embedding network, restoration network, sequence generator, and sequence discriminator according to the reconstruction loss, unsupervised loss, and supervised loss; Step 3-8: Reconstruct the synthesized time-frequency-space channel coefficient matrix based on the synthesized latent sequence:

[0073] Step 3-9: Denormalize the synthesized channel coefficient matrix.

[0074] By inputting the measured channel data ,The digital cousin channel model can synthesize approximate channel data covering the entire domain at low cost and large scale.

[0075] S4, predicting channel state information by using a channel prediction method combining virtual and real data; The measured data provides the real channel state information at a limited location, and the synthetic data provides the approximate channel information covering the whole world. The measured channel data and the synthetic channel data are fully integrated through the graph convolutional neural network to extract the spatial correlation relationship of the channel and obtain the global real state information of the current channel. The input is the time series of the measured channel and the synthetic channel. The graph convolutional network is used to receive the spatial information of the topology, and then the fused space-frequency channel matrix sequence is input into the convolutional long short-term memory network to extract the spatiotemporal features, and the dynamic information changes between each unit are obtained, so as to realize the joint prediction of the channel state.

[0076] 1) Fusion of virtual and real associations based on graph convolutional networks; First, the spatial correlation between nodes is modeled using the graph structure, and then feature propagation is performed through graph convolution operations. Then, a progressive spiral virtual-real channel data fusion strategy is adopted to gradually optimize the representation of the node embedding matrix in multiple iterations, and adaptive weighting is used to achieve the optimal fusion of virtual and real data in the final embedding matrix. Finally, the decoder calculates the existence probability of the edges between nodes, thereby constructing a globally accurate channel spatial correlation model. The model architecture of virtual-real data fusion is shown in Figure 1. Figure 5 shown.

[0077] Step 4-1: Initial graph generation; Map the measured sensor terminal nodes at a certain moment into a weighted undirected graph , where the graph consists of two elements, Represents a collection of nodes; Represents the spatial correlation between nodes. The graph can be represented by the adjacency matrix Indicates that the correlation between nodes is calculated using the Pearson correlation coefficient. Similarly, the adjacency matrix of cousin channel data is constructed , which contains the global approximate channel information.

[0078] Step 4-2: Normalize the node feature matrix using min-max and ; Step 4-3: Use a multi-layer perceptron to transform the feature matrix:

[0079]

[0080] Step 4-4: Calculate the normalized graph Laplacian matrix:

[0081]

[0082] Step 4-5: Use the progressive spiral virtual-real data fusion strategy: Feature fusion is performed within the network layer and is carried out in a progressive spiral manner to ensure the complete integration of virtual and real data. The fusion process is as follows Figure 6 As shown. The data fusion equation at the iteration is:

[0083]

[0084] in, is a function used to mix virtual and real channel data after each neighborhood aggregation. Considering the balance between computational complexity and model performance, the summation is selected as the fusion function. After fusion, the embedding matrices of the sensor nodes will be updated with each other through progressive iteration. In order to better utilize the embedding matrix of each node, the adaptive complementary weighted fusion method is used to perform complementary correlation fusion of virtual and real data. Under the premise of minimizing the total mean square error, the corresponding optimal weighting factor is assigned to the virtual and real channel data:

[0085] in, and are the weight coefficients of the real and cousin node embedding matrices, respectively, which are used to measure the contribution of the corresponding data features, and the sum of the two weight coefficients is 1. It is through The comprehensive node embedding matrix obtained after iterations reflects the The fusion result of skipping neighborhood. After iterations, multiple comprehensive node embedding matrices can be obtained In order to make full use of the information of each hop neighborhood of the central node, the results of each layer are concatenated to obtain a single embedding matrix, which contains more useful information than the result of using only the last layer. The concatenated result is as follows:

[0086] Step 4-6: Generate edge embedding vector; After obtaining a pair of target nodes and their embedding vectors and Afterwards, an edge generation strategy is constructed to determine the existence of edges between target node pairs:

[0087] in, is a function used to obtain the edge vector between the target node pair. is the edge embedding vector between target node pairs.

[0088] Step 4-7: Use the decoder to calculate the existence probability of the corresponding edge:

[0089] in, is a parameter matrix, is a bias term, using sigmoid as the activation function. The global channel state information is reconstructed by calculating the existence probability of the edge of the node adjacency matrix, thereby extracting the spatial correlation of the channel.

[0090] Step 4-8: Based on probability and the true label Calculate the binary cross entropy loss:

[0091] 2) Joint channel prediction based on convolutional long short-term memory network; A three-dimensional convolutional neural network is used to extract channel matrix features with frequency domain and spatial domain dimensions through convolution operations. The time domain correlation of channel data is utilized, and a convolutional long short-term memory network improved by the convolution kernel is used for prediction. Finally, the predicted features are input into the convolutional neural network to adjust the form of the output data, thereby completing the prediction.

[0092] The channel prediction network used in this embodiment includes: an input layer, a convolutional layer, a convolutional long short-term memory network layer and an output layer, and the structure is as follows: Figure 7 As shown. After graph convolutional network fusion, the obtained data can be expressed as , the input sequence length is T, and the sliding value is taken according to the past historical information in the training set, then Indicates The input sequence of training samples is Indicates the fused time domain sampling points.

[0093] Step 5-1: Feature extraction; The multi-dimensional convolution kernel of the three-dimensional convolutional neural network is used to extract the temporal and spatial features of the input to obtain the output features with a sequence length of T. ,in Represents input After the three-dimensional convolutional neural network The characteristics of the corresponding element.

[0094] Step 5-2: Use a layer of convolutional long short-term memory network as a predictor to transform the output of the three-dimensional convolutional neural network Expressed as .

[0095] In the forward propagation process, the output of the last moment is taken as the output of this layer, that is, the output at t=T is , the output of the convolutional long short-term memory network layer is expressed as In the convolutional long short-term memory network layer, the gate structure and unit state are updated to achieve the purpose of deep feature memory. The output is the convolutional long short-term memory network for the first The feature prediction is performed at the time domain sampling point at time T+1 of each sample.

[0096] Step 5-3: Feature recovery; Since the final output needs to meet the dimensional requirements of the channel matrix, feature recovery is performed through a convolutional neural network to obtain an output that meets the dimensional requirements of the channel matrix.

[0097] Step 5-4: Design loss function; Parameter regularization is introduced to solve the overfitting problem and weight the trainable parameters to the loss function:

[0098] in, represents the regularization coefficient, Represents a trainable parameter in the model.

[0099] The present invention fuses virtual and real channel data and fully exploits the multi-domain correlation of channels in time, space and frequency dimensions, thereby using local channel data samples to achieve accurate prediction of fast time-varying and non-stationary channels in complex shielding environments: Aiming at the problems of sensor measurement data error and state information loss, a CSI compensation method based on AP clustering is proposed. Clustering is performed according to the similarity between different data, and the weighting coefficient is calculated according to the data similarity. The entropy value is used to determine the attribute value of the missing data, so as to achieve orderly filling of the distorted data and improve the fidelity of the measured channel data.

[0100] In order to solve the problem that effective CSI data samples measured in complex shielding environments are scarce and it is difficult to obtain channel state information at different global locations, the present invention constructs a lightweight digital cousin channel with similar distribution characteristics and time domain, frequency domain and spatial domain correlation to the real channel through a cousin channel construction method based on CrossDGAN, which can generate channel synthetic data covering all points in the global domain.

[0101] In order to solve the problem that the channel state is difficult to accurately predict due to the rapid time-varying characteristics of the channel, a channel prediction method with space-time memory function is adopted. The virtual and real channel data are fully integrated through the progressive spiral fusion strategy based on GCN to obtain the space-frequency channel feature matrix that can accurately reflect the global channel state. Then, a prediction model based on convolutional long short-term memory network is used, combined with the time domain correlation of channel data, to achieve high-precision CSI prediction.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A time-space-frequency multi-dimensional channel prediction method based on digital cousins, characterized in that: include: Receive original channel state information data from the sensor terminal; Filling missing values ​​of original channel state information data by affinity propagation clustering to obtain local measured channel state information data; Using a digital cousin-based channel generation model to obtain global approximate channel state information data through historical channel state information data; fusing global approximate channel state information data and local measured channel state information data to obtain fused channel state information data; A channel prediction network is used to obtain a global channel state information prediction value based on the fused channel state information data.

2. The method for predicting multi-dimensional channels based on digital cousins ​​in time, space and frequency according to claim 1, characterized in that: Filling missing values ​​of original channel state information data by affinity propagation clustering includes: Normalizing each attribute value of the original channel state information data; The Euclidean distance between data targets is calculated based on the normalized attribute values, and the similarity matrix between data targets is constructed; Dividing the original channel state information data set into a complete data subset and an incomplete data subset; Calculate density indicators of data targets based on complete data subsets and incomplete data subsets; The data target with the highest density is taken as the cluster center, and the relevant radius of the cluster neighborhood is obtained based on the similarity matrix between the data targets, and affinity propagation clustering is performed; When the maximum number of iterations is reached or when the cluster center does not change, the similarity between the incomplete data subset and the complete data subset in the same cluster is calculated; Get the entropy value of the complete data subset; The entropy value of the complete data subset and the similarity between the incomplete data subset and the complete data subset in the same cluster are combined to obtain the attribute value of the incomplete data subset, thereby completing the orderly filling of the missing values ​​of the original channel state information data.

3. The method for predicting multi-dimensional channels based on digital cousins ​​in time, space and frequency according to claim 1, characterized in that: The method of using the digital cousin-based channel generation model to obtain global approximate channel state information data through historical channel state information data includes: The local historical channel state information data is mapped into a real potential sequence through a segmented embedding layer, and a second-order attention layer based on an attention mechanism is introduced to capture the cross-dimensional dependencies between the time domain, frequency domain, and spatial domain; A random noise sequence is introduced into the simulation environment to simulate random factors, and a synthetic potential sequence is generated based on the random noise sequence through a sequence generator; reconstructing a synthetic time-frequency spatial channel coefficient matrix based on the synthetic latent sequence; Denormalizing the composite channel coefficient matrix to obtain global approximate channel state information data; The weights of the embedding network, restoration network, sequence generator, and sequence discriminator are updated according to the reconstruction loss, unsupervised loss, and supervised loss.

4. The method for predicting multi-dimensional channels based on digital cousins ​​in time, space and frequency according to claim 3, characterized in that: The sequence generator and the sequence discriminator are subjected to adversarial training to optimize the quality of the synthesized potential sequence, including: The sequence generator aims to maximize the score given by the sequence discriminator and updates the parameters by minimizing its loss function; The sequence discriminator aims to maximize the score of the true latent sequence while minimizing the score of the synthetic latent sequence, and updates its parameters by minimizing the loss function.

5. The method for predicting multi-dimensional channels based on digital cousins ​​in time, space and frequency according to claim 1, characterized in that: The channel prediction network includes: a convolutional long short-term memory network.

6. The time-space-frequency multi-dimensional channel prediction method based on digital cousin according to claim 1, characterized in that: The global approximate channel state information data and the local measured channel state information data are fused through a graph convolutional neural network, including: Using sensor terminals as nodes, construct an initial graph of local measured channel state information data and an initial graph of global approximate channel state information data; Obtaining a number of measured node embedding matrices and approximate node embedding matrices based on an initial graph of local measured channel state information data and an initial graph of global approximate channel state information data through a data fusion model; Adaptive complementary weighted fusion method is used to obtain the comprehensive node matrix based on the measured node embedding matrix and the approximate node embedding matrix; Concatenate all comprehensive node embedding matrices to obtain the global node matrix; The decoder is used to obtain the fused channel state information data feature map based on the global node matrix.

7. The time-space-frequency multi-dimensional channel prediction method based on digital cousin according to claim 6, characterized in that: The data fusion model: in, The graph normalized Laplacian matrix representing the local measured channel state information data, represents the measured node embedding matrix, , The graph minimum-maximum normalized node feature matrix representing the local measured channel state information data, represents a multi-layer perceptron; A graph normalized Laplacian matrix representing the global approximate channel state information data; represents the approximate node embedding matrix, , A graph min-max normalized node feature matrix representing the global approximate channel state information data; is the sum function; Indicates the number of iterations.

8. The method for predicting multi-dimensional channels based on digital cousins ​​in time, space and frequency according to claim 6, characterized in that: The method of using the adaptive complementary weighted fusion method to obtain a comprehensive node matrix based on the measured node embedding matrix and the approximate node embedding matrix includes: With the minimum total mean square error as the constraint, the corresponding optimal weighting factors are assigned to the measured node embedding matrix and the approximate node embedding matrix.

9. The method for predicting multi-dimensional channels based on digital cousins ​​in time, space and frequency according to claim 6, characterized in that: The method of obtaining a global channel state information prediction value based on the fused channel state information data by using a channel prediction network includes: A multi-dimensional convolution kernel is used to extract the spatiotemporal features of the fused channel state information data feature map to obtain output features of a preset length; A layer of convolutional long short-term memory network is used as a predictor to update the output features; The convolutional neural network is used to perform feature recovery based on the updated output features to obtain the global channel state information prediction value.

Images