A 6g dynamic channel map construction method based on machine learning and image processing

Abstract

The application discloses a 6G dynamic channel map construction method based on machine learning and image processing, comprising the following steps: 1) acquiring dynamic channel maps of different scenes and performing data analysis; 2) extracting building masks based on channel data; 3) constructing a dynamic channel map dataset according to the receiver signal strength as the main feature, and superimposing the building masks as heterogeneous data to enhance the prediction ability; 4) dividing the dataset into known and unknown scenes; 5) designing a pooling encoder to extract multi-scale information to obtain channel map features and introducing a self-attention mechanism for feature fusion, and finally using a full-scale decoder to realize the construction of the dynamic channel map. The dynamic channel map construction method provided by the application can capture the channel map changes in different scenes, has good performance in the construction of the channel map in the urban environment, and can be used for assisting in optimizing the channel estimation and assisting in the beam alignment and other key technologies in the 6G communication.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross-fusion application field of machine learning and wireless channel modeling, and particularly relates to a 6G dynamic channel map construction method based on machine learning and image processing. BACKGROUND

[0002] With the continuous development of 6G technology, communication systems are facing unprecedented challenges. These challenges include a large number of antennas and users, the deployment of large-scale multiple-input multiple-output systems, complex wireless propagation environments, and the uncertainty brought by the dynamic change of user terminal or base station positions. In addition, with the rapid growth of emerging application scenarios such as Internet of Things, Internet of Vehicles, and unmanned aerial vehicle communication, traditional channel information acquisition methods rely heavily on the transmission of pilot signals to obtain real-time channel state information. However, the substantial increase in pilot overhead and the improvement in system complexity limit the scalability and performance optimization space of future wireless communication systems.

[0003] Channel maps can greatly reduce pilot overhead and reduce system design complexity, thereby improving the overall performance of communication systems by pre-acquiring part of the channel state information. Channel maps can accurately model wireless channels by combining geographic information, environmental perception, and other data sources without relying entirely on real-time pilots.

[0004] Existing channel map construction methods are mainly divided into model-based methods and data-driven methods. Model-based channel map construction methods require the pre-acquisition of transmitter and receiver position information and combine mathematical models to calculate the propagation path to obtain channel state information within the entire scene. However, this method usually ignores the actual scattering and reflection characteristics in complex environments, limiting its modeling accuracy of the channel. When the transmitter position is unknown and dynamically changes over time, model-based methods need to rely on complex transmitter positioning algorithms, which greatly increases the difficulty of channel map construction and makes the final result heavily dependent on the accuracy of transmitter positioning. Compared with model-based methods, data-driven channel map construction methods rely on large-scale channel measurement data. This method uses data completion algorithms to utilize the correlation of channel characteristics in the spatial dimension to fill in the channel information of unmeasured positions. However, traditional data-driven methods usually only consider spatial channel characteristics and cannot effectively utilize environmental information. In the 6G era, with the widespread use of mobile communication devices and Internet of Things terminals, how to accurately and timely acquire dynamic channel information has become a pressing problem.

[0005] In order to solve the limitations of the existing channel map construction method, artificial intelligence (AI) technology provides a new solution. In recent years, AI technology has made significant progress in image processing, feature extraction, time series prediction and other fields, especially in data modeling and high-dimensional feature extraction in complex environments. Compared with traditional channel modeling methods, AI algorithms can effectively capture the correlation of channels in the spatial dimension, and can extract the implicit features in the channel through a deep learning network. Especially in known propagation environments, AI algorithms can adaptively predict channel information and extract effective information in the environment. However, the existing AI channel map construction does not well solve the dynamic changes of the time series model in predicting the channel state. In the case of unknown or rapidly changing transmitter positions, real-time tracking and prediction of channel state cannot be achieved. In 6G communication, the time trend of the channel map needs to be paid more attention to. SUMMARY

[0006] The present application provides a 6G dynamic channel map construction method based on machine learning and image processing, which can realize real-time tracking and prediction of channel information, and can obtain the channel map at future time in different scenarios. The method uses channel maps with time sequence relationship for training, improves the construction ability of 6G dynamic channel map, and obtains high construction accuracy.

[0007] Technical scheme: In order to achieve the purpose of the present application, the technical scheme adopted by the present application is: a 6G dynamic channel map construction method based on machine learning and image processing, comprising the following steps:

[0008] S1, obtaining channel data by reconstruction simulation, including: first reconstructing the scene of the building vector data, then selecting the receiver on the path of the reconstructed scene, and generating a dynamic channel map composed of channel characteristics;

[0009] S2, for the dynamic channel map constructed in S1, extracting and correcting the obtained building mask, and processing the channel map, the dynamic channel map and the building mask are combined to form a data set;

[0010] S3, dividing the dynamic channel map data set constructed in S2 into known area and unknown area according to the proportion, and dividing the known area into training set and verification set according to the proportion, and taking the unknown area as test set;

[0011] S4, constructing a dynamic channel map network model based on machine learning and image processing and executing model training, including:

[0012] First, construct a dynamic channel map network model and initialize its model parameters;

[0013] Then take the dynamic channel map dataset as the input of the model, and use the model to extract data features, and then generate fusion features through the attention module of the model;

[0014] Finally, the change trend of the channel map is predicted by using the previously extracted data features and fusion features, the error is calculated with the actual value, and the model parameters are updated according to the result.

[0015] Further, S1 comprises:

[0016] S101, reconstructing the urban environment according to the open source building vector;

[0017] S102, for the reconstructed urban environment, uniformly interval receivers on the known path, and uniformly arrange receivers in the form of a matrix in the random area containing the path for channel simulation, and the channel characteristics are received power;

[0018] According to the channel map composed of the relative positions of the receivers, each interval receiver is regarded as moving in the time dimension, and is stacked in the form of matrix data in sequence to form a tensor, thereby forming a set of dynamic channel map data.

[0019] Further, S2 comprises:

[0020] According to the extracted channel building vector of the arranged receiver area according to the constructed dynamic channel map data, the position of the transmitter not receiving the signal is set as a building according to the obtained channel map in the same area, and a building mask is obtained;

[0021] For the part that fails to extract due to simulation error, repeat the simulation and extract it separately; for the obtained channel map, the data with received power less than the threshold value is set to a uniform value.

[0022] Further, S3 comprises:

[0023] For the dynamic channel map data set, every four consecutive channel maps in the same area are taken as a group of data;

[0024] The channel maps of the first three time points and the building mask of the area to which the channel map belongs are taken as the input, and the channel map of the next time point is taken as the training label, thereby constructing the dynamic channel map training data set of the neural network;

[0025] The data of the channel map is normalized, so that each parameter value is distributed between [0, 1].

[0026] Further, initialize the dynamic channel map network model:

[0027] Two convolution modules followed by a ReLU activation function are used as convolution blocks for extracting channel spatial features, and the expression of the ReLU activation function is:

[0028]

[0029] ViT is used as a feature fusion and moving information extraction module, where the multi-head attention is 4 and the hidden layer dimension is 4 times the input dimension;

[0030] Linear interpolation is used as upsampling, and max pooling is used as down-sampling full connection decoder for decoding compressed information to obtain the required output;

[0031] Masked pooling is used for feature extraction operation, and masked pooling is 2x2 max pooling on the building mask, and 2x2 average pooling on the channel map after removing the mask position data.

[0032] Further, S4 comprises:

[0033] S401, obtaining channel map features by extracting multi-scale information using a pooling encoder;

[0034] S402, using a Transformer-based attention mechanism for feature fusion;

[0035] S403, predicting the trend of the channel map, calculating the error with the actual value, and updating the model parameters.

[0036] Further, S401 obtains channel map features by extracting multi-scale information using a pooling encoder, comprising:

[0037] After the dynamic channel map dataset is input into the model, the encoder module extracts features from data with a dimension of [4, 256, 256], and the encoder module operates as follows:

[0038] The original data is extracted by two masked pooling and convolution, and the features are extracted by max pooling, and then the extracted features are input into the next layer of feature extraction together with the extracted features.

[0039] The above feature extraction operation is repeated, and finally the feature with a dimension of [1024, 16, 16] is output to the feature fusion module.

[0040] Further, S402 uses a Transformer-based attention mechanism to fuse the features, comprising:

[0041] The feature fusion module includes a self-attention mechanism module, which flattens the extracted features with a dimension of [1024, 16, 16] obtained by the encoding module into a feature vector with a dimension of [1024, 256], and after passing through 2 layers of attention module, outputs channel space fusion features with a dimension of [1024, 256], and then restores the fusion features with a dimension of [1024, 16, 16].

[0042] Further, S403 predicts the trend of the change of the channel map, updates the model parameters, including:

[0043] The fusion feature with the dimension of [1024, 16, 16] is up-sampled by using a full connection decoder, and an image feature with the dimension of [320, 256, 256] of the same matrix size as the channel map is obtained after passing through four full connection decoders, and finally the channel map data of the next moment is obtained after passing through a convolution layer;

[0044] The mean square error between the predicted value and the actual value of the channel map excluding the building mask at the next moment and the difference dispersion regularization loss are calculated, and the loss function L is defined as

[0045]

[0046] Wherein is the predicted channel map, P is the channel map obtained by real simulation, and L reg is the difference dispersion regularization loss, and a is an adjustable parameter;

[0047] The Adam optimization algorithm is selected to update the network parameters, the learning rate and the weight decay coefficient are set, and the iterative training of the model is performed until the performance of the model reaches the expectation.

[0048] Beneficial effects: compared with the prior art, the technical scheme of the present application has the following beneficial technical effects:

[0049] The dynamic channel map construction method based on machine learning and image processing can capture the change of the channel map in different scenes, obtain the channel characteristics at different moments through the channel space correlation data features, has good performance in the construction of the channel map in the urban environment, can obtain the channel information in advance in the 6G communication, optimizes the channel estimation and assists the beam alignment and other key technologies. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 It is a flowchart of the dynamic channel map construction method based on machine learning and image processing according to the present application.

[0051] Figure 2 It is a schematic diagram of data acquisition and data set processing provided by the embodiment of the present application.

[0052] Figure 3 It is a schematic diagram of network structure provided by the embodiment of the present application.

[0053] Figure 4 It is a schematic diagram of network structure encoder provided by the embodiment of the present application.

[0054] Figure 5A network structure feature fusion module schematic diagram provided according to an embodiment of the present application.

[0055] Figure 6 A prediction comparison diagram of dynamic channel map construction provided according to an embodiment of the present application.

[0056] Figure 7 A cumulative distribution comparison diagram of absolute error of dynamic channel map construction provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0058] A dynamic channel map construction method based on machine learning and image processing is provided in the present application, and the flow is as shown in the figure Figure 1 The dynamic channel map construction method based on machine learning and image processing includes the following steps:

[0059] S1, obtain channel data by reconstruction simulation, including: first, reconstruct the scene of building vector data, then select the receiver on the path of the reconstructed scene, and generate a dynamic channel map composed of channel characteristics.

[0060] S101, reconstruct the urban environment according to the open source building vector and import the software Wireless Insite.

[0061] S102, for the reconstructed urban environment, select the receiver on the known path at equal intervals, and arrange the receiver in a matrix shape in the random area containing the path for channel simulation, and the channel characteristics are the received power; wherein, according to the channel map composed of the relative positions of the receivers, each interval receiver is regarded as moving in the time dimension, and is stacked into a tensor in the form of matrix data in sequence to form a group of dynamic channel map data, and the data set construction process is as shown in the figure Figure 2 .

[0062] S2, for the dynamic channel map constructed in S1, extract and correct the obtained building mask, and process the channel map, and the dynamic channel map and the building mask are combined to form a data set.

[0063] According to the dynamic channel map data, the channel building vector of the arranged receiver area is extracted, and according to the obtained channel map in the same area, the position of the transmitter not receiving the signal is set as a building to obtain a building mask.

[0064] A 1 / 0 binary building mask map is constructed, and the existing building is set to 1 and the others are set to 0; for the part that fails to extract due to simulation error, repeat the simulation and extract it separately; for the obtained channel map, the data with received power less than the threshold value is set to a uniform value.

[0065] S3, divide the dynamic channel map dataset constructed in S2 into known areas and unknown areas according to a certain proportion, verify the generality of the model through the known areas, and divide the known areas into a training set and a verification set according to a certain proportion, and use the unknown areas as a test set, and test the precision of the model through the known areas.

[0066] Because there are many channel map frames on each path, it is necessary to divide each path according to the video prediction idea, and for the dynamic channel map dataset, every four consecutive channel maps in the same area are regarded as a group of data; the channel map at the previous three time points and the building mask of the area to which the channel map belongs are used as input, and the channel map at the next time point is used as the training label, to form the dynamic channel map training dataset of the neural network.

[0067] S4, construct a dynamic channel map network model based on machine learning and image processing and perform model training.

[0068] Initialize the dynamic channel map network model, and the constructed network model is as shown in Figure 3 , and input the dynamic channel map dataset.

[0069] Two convolution modules followed by ReLU activation functions are used as convolution blocks for extracting channel spatial features, and the expression of the ReLU activation function is:

[0070]

[0071] ViT is used as a feature fusion and moving information extraction module, where the number of attention heads is 4 and the hidden layer dimension is 4 times the input dimension.

[0072] Linear interpolation is used as upsampling, and max pooling is used as downsampling, and a fully connected decoder is used to decode compressed information to obtain the required output.

[0073] Masked pooling is used for feature extraction, and masked pooling is 2x2 max pooling on the building mask, and 2x2 average pooling on the channel map after removing the mask position data.

[0074] S401, use the pooling encoder to extract multi-scale information to obtain channel map features.

[0075] After the dynamic channel map dataset is input into the model, the encoder module extracts features from data with a dimension of [4, 256, 256], and the operation of the encoder module is:

[0076] The original data is extracted through two times of masked pooling and convolution, and the features are extracted through max pooling, and then the extracted features are used as the input of the next layer of feature extraction.

[0077] The above feature extraction operation is repeated, and finally the feature with a dimension of [1024, 16, 16] is output to the feature fusion module. The encoder structure is as shown in Figure 4

[0078] S402, using the attention mechanism based on the Transformer for feature fusion.

[0079] The feature fusion module includes a self-attention mechanism module, which flattens the extracted feature with a dimension of [1024, 16, 16] obtained by the encoding module into a feature vector with a dimension of [1024, 256], and outputs a channel space fusion feature with a dimension of [1024, 256] after passing through 2 layers of attention modules, and then restores it to a fusion feature with a dimension of [1024, 16, 16]. The fusion process is as shown in Figure 5

[0080] S403, predicting the trend of the channel map, calculating the error with the actual value, and updating the model parameters.

[0081] The fusion feature with a dimension of [1024, 16, 16] is upsampled using a fully connected decoder, and an image feature with a dimension of [320, 256, 256] is obtained after passing through four fully connected decoders. Finally, the channel map data of the next time is obtained after passing through a convolution layer.

[0082] The mean square error and the difference dispersion regularization loss between the predicted value and the actual value of the channel map of the next time except the building mask are calculated, and the loss function L is defined as

[0083]

[0084] wherein is the predicted channel map, P is the channel map obtained by actual simulation, L reg is the difference dispersion regularization loss, and a is an adjustable parameter, which is set to 0.1 in this embodiment.

[0085] The Adam optimization algorithm is selected to update the network parameters, the learning rate is set to 0.0001, the weight decay coefficient is set to 0.0005, and the sum of the mean square error and the difference dispersion regularization loss is used as the loss function for 1000 times of iteration training of the model. The model performance is iterated multiple times until it reaches the expected performance.

[0086] Whether the model performance reaches the expected performance is measured by comparing the difference between the statistical characteristics of the output predicted channel and the actual value.

[0087] ​​Specifically, in the present embodiment, the channel map obtained by using the dynamic channel map construction network is compared with the channel map obtained by other methods, and the model performance of the network in constructing the dynamic channel map is analyzed, and the simulation results are referred to Figure 6 and Figure 7 . Figure 6 is the visualization of the channel map under the urban scene, the comparison of the channel map prediction model obtained by training under the same input and output with the traditional algorithm and other AI algorithms shows from Figure 6 (a-f) are in turn, the simulation results of the true value, the model construction, the optical flow method, the ViT network, the UNet network and the algorithm proposed in the present application. From the simulation diagram, it can be seen that the constructed channel map is similar to the simulation result, which shows that the network model has good prediction performance; Figure 7 is the cumulative distribution diagram of the absolute error of different schemes, the error curve of the construction result obtained by training is above all the algorithms, which shows that the error predicted by the network model is the smallest, and the network model has good prediction ability.

[0088] In summary, the dynamic channel map construction method based on machine learning and image processing established by the present application is a construction method using the spatiotemporal correlation of the channel, which uses machine learning to extract the features of the historical channel map in multiple scales, then uses the self-attention mechanism to fuse and process the time dimension information, and finally uses the full connection decoder to realize the prediction of the channel map, and is suitable for the construction of the dynamic channel map under the urban scene, and has reference significance for the channel feature extraction research of the dynamic electromagnetic environment.

[0089] The details of the present application not described are well known to those skilled in the art.

[0090] It can be understood that the present application is described by some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to these features and embodiments without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, these features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of the present application.

Claims

1. A 6G dynamic channel map construction method based on machine learning and image processing, characterized in that, The method comprises the following steps: S1, obtaining channel data through reconstruction simulation, comprising: firstly reconstructing the scene of building vector data, then selecting a receiver on the path of the reconstructed scene, and generating a dynamic channel map composed of channel characteristics; S2, for the dynamic channel map constructed in S1, extracting and correcting the obtained building mask, and processing the channel map, the dynamic channel map and the building mask are combined to form a data set; S3, dividing the dynamic channel map data set constructed in S2 into known areas and unknown areas according to the proportion, and dividing the known areas into training set and verification set according to the proportion, and taking the unknown area as the test set; S4, constructing a dynamic channel map network model based on machine learning and image processing and performing model training, comprising: S401, constructing a dynamic channel map network model and initializing its model parameters, using a pooling encoder to extract multi-scale information to obtain channel map features; S402, taking the dynamic channel map data set as the input of the model, and using the model to extract data features, then passing through the attention module of the model, using the attention mechanism based on Transformer to fuse the features, and generating fused features; S403, using the previously extracted data features and fused features to predict the trend of the channel map, calculating the error with the actual value, and updating the model parameters according to the results, comprising: using a fully connected decoder to perform upsampling operation on the fused features with a dimension of [1024, 16, 16], obtaining image features with a dimension of [320, 256, 256] of the same matrix size as the channel map after passing through four fully connected decoders, and finally obtaining the channel map data of the next moment through the convolution layer; Calculate the mean square error and the difference dispersion regularization loss between the predicted value and the actual value of the channel map of the next moment except the building mask, and define the loss function L as wherein is the predicted channel map, P is the real simulation obtained channel map, L reg is the differential dispersion regularization loss, and a is a tunable parameter. Select Adam optimization algorithm to update network parameters, set learning rate and weight decay coefficient, and perform iterative training of the model until the model performance reaches the expected value.

2. The method of claim 1, wherein the method is based on machine learning and image processing. S1 includes: S101, reconstructing the urban environment according to the open source building vector; S102, for the reconstructed urban environment, uniformly interval receivers on the known path, and uniformly arrange receivers in the random area containing the path in the form of a matrix for channel simulation, and the channel characteristics are received power; According to the channel map composed of the relative positions of the receivers, each interval receiver is regarded as moving in the time dimension, and is stacked into a tensor in the form of matrix data in sequence, to form a group of dynamic channel map data.

3. The method of claim 1, wherein S2 It includes: According to the dynamic channel map data constructed, the channel building vector of the area where the receivers are arranged is extracted, and the position where the transmitter does not receive the signal is set as the building according to the channel map obtained in the same area to obtain the building mask; For the part that fails to extract due to simulation error, repeat the simulation and extract it separately; For the obtained channel map, the data with received power less than the threshold value is set to a uniform value.

4. The method of claim 1, wherein the method is based on machine learning and image processing. S3 includes: For the dynamic channel map data set, every four consecutive channel maps in the same area are taken as a group of data; The channel map of the previous three time instants and the building mask of the area to which the channel map belongs are taken as inputs, and the channel map of the next time instant is taken as a training label to form a dynamic channel map training data set of the neural network; Data normalization is performed on the channel map to make each parameter value distributed between [0, 1].

5. The 6G dynamic channel map construction method based on machine learning and image processing according to claim 1, characterized in that, A dynamic channel map network model is constructed: Two convolution modules followed by ReLU activation functions are used as convolution blocks to extract channel spatial features. ViT is used as a feature fusion and mobile information extraction module, with 4 multi-head attentions and a hidden layer dimension of 4 times the input dimension. Linear interpolation is used as upsampling, and max pooling is used as down-sampling in a fully connected decoder to decode compressed information and obtain the required output. Masked pooling is used for feature extraction, and masked pooling is performed on the building mask with 2x2 max pooling and on the channel map after removing the mask position data with 2x2 average pooling.

6. The 6G dynamic channel map construction method based on machine learning and image processing according to claim 1, characterized in that, S401 uses a pooling encoder to extract multi-scale information to obtain channel map features, including: After the dynamic channel map data set is input into the model, the encoder module extracts features from data with a dimension of [4, 256, 256], and the encoder module operates as follows: The original data is extracted through two masked pooling and convolution operations, and the features are extracted through max pooling. Then, the extracted features are taken as the input of the next layer of feature extraction. The above feature extraction operation is repeated, and finally the feature with a dimension of [1024, 16, 16] is output to the feature fusion module.

7. The 6G dynamic channel map construction method based on machine learning and image processing according to claim 1, characterized in that, S402 uses a Transformer-based attention mechanism to fuse features, including: The feature fusion module includes a self-attention mechanism module, which flattens the extracted features with a dimension of [1024, 16, 16] obtained by the encoding module into a feature vector with a dimension of [1024, 256]. After passing through 2 layers of attention modules, the channel space-time fusion features with a dimension of [1024, 256] are output, and then restored to the fusion features with a dimension of [1024, 16, 16].

Images