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

Abstract

The invention discloses a 6G dynamic channel map construction method based on machine learning and image processing. The method comprises the following steps: 1) obtaining dynamic channel maps of different scenes and carrying out data analysis; 2) extracting a building mask based on the channel data; (3) constructing a dynamic channel map data set by taking the signal intensity of a receiver as a main characteristic, and superposing a building mask to serve as heterogeneous data to enhance the predictive capability; 4) dividing the data set into a known scene and an unknown scene; and 5) a pooling encoder is designed to extract multi-scale information to obtain channel map features, a self-attention mechanism is introduced to carry out feature fusion, and finally, a full-scale decoder is used to realize construction of a dynamic channel map. The dynamic channel map construction method provided by the invention can capture channel map changes in different scenes, has better performance in the aspect of channel map construction in an urban environment, and can be used for assisting in key technologies such as optimizing channel estimation and assisting beam alignment in 6G communication.

Description

Technical Field

[0001] The present invention belongs to the cross-integration application field of machine learning and wireless channel modeling, and specifically relates to a 6G dynamic channel map construction method based on machine learning and image processing. Background Art

[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 MIMO systems, complex wireless propagation environments, and uncertainties caused by dynamic changes in the locations of user terminals or base stations. In addition, with the rapid growth of emerging application scenarios such as the Internet of Things, Internet of Vehicles, and UAV communications, 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 increase in system complexity limit the scalability and performance optimization space of future wireless communication systems.

[0003] Channel maps can obtain some channel status information in advance, which can not only significantly reduce pilot overhead, but also reduce the complexity of system design, thereby improving the overall performance of the communication system. Channel maps can accurately model wireless channels without relying entirely on real-time pilots by combining multiple data sources such as geographic information and environmental perception.

[0004] Existing channel map construction methods are mainly divided into model-based methods and data-driven methods. The model-based channel map construction method requires obtaining the location information of the transmitter and receiver in advance, and combining mathematical models to calculate the propagation path to obtain the channel state information in the entire scene. However, this method usually ignores the actual scattering and reflection characteristics in complex environments, resulting in limited accuracy in channel modeling. When the transmitter position is unknown and changes dynamically over time, the model-based method needs to rely on complex transmitter positioning algorithms, which greatly increases the difficulty of channel map construction, and the final result is heavily dependent on the accuracy of transmitter positioning. Compared with the model-based method, the data-driven channel map construction method relies on the support of large-scale channel measurement data. This method uses the correlation of channel features in the spatial dimension through a data completion algorithm to fill in the channel information of unmeasured locations. However, traditional data-driven methods usually only consider spatial channel features and cannot achieve effective use of environmental information. In the 6G era, with the widespread use of mobile communication devices and IoT terminals, how to accurately and real-time obtain dynamic channel information has become an urgent problem to be solved.

[0005] In order to solve the limitations of existing channel map construction methods, 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 spatial dimensions and extract implicit features in channels through deep learning networks. 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 solve the dynamic changes of channel states predicted by time series models. When the transmitter position is unknown or changes rapidly, real-time tracking and prediction of channel states cannot be achieved. In 6G communications, the temporal trend of channel maps needs to receive more attention. Summary of the invention

[0006] Purpose of the invention: The present invention provides a 6G dynamic channel map construction method based on machine learning and image processing, which is used to achieve real-time tracking and prediction of channel information, and can obtain channel maps at future times in different scenarios. The method uses channel maps with time series relationships for training, improves the construction capability of 6G dynamic channel maps, and obtains higher construction accuracy.

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

[0008] S1. Obtaining channel data through reconstruction simulation, including: firstly reconstructing the building vector data, then selecting a 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, extract and correct the obtained building mask, and process 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 areas and unknown areas according to proportion, and dividing the known areas into training sets and validation sets according to proportion, and using the unknown areas as test sets;

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

[0012] Firstly, a dynamic channel map network model is constructed and its model parameters are initialized;

[0013] Then, the dynamic channel map dataset is used as the input of the model, and after the data features are extracted by the model, the fusion features are generated by the attention module of the model;

[0014] Finally, the previously extracted data features and fused features are used to predict the change trend of the channel map, the error is calculated with the actual value, and the model parameters are updated according to the results.

[0015] Furthermore, S1 includes:

[0016] S101, Reconstructing the urban environment based on open source architectural vectors;

[0017] S102, for the reconstructed urban environment, evenly spaced receivers are selected on a known path, and the receivers are evenly arranged in a matrix shape in a random area including the path to perform channel simulation, where the channel characteristic is the received power;

[0018] According to the channel map formed by the relative positions of the receivers, each interval of receivers is regarded as moving in the time dimension, and is stacked into tensors in the form of matrix data in sequence to form a set of dynamic channel map data.

[0019] Furthermore, S2 includes:

[0020] Extracting the channel building vector of the receiver area according to the constructed dynamic channel map data, extracting the location where the transmitter does not receive the signal according to the channel map obtained in the same area and setting it as a building to obtain a building mask;

[0021] For the part where the extraction fails due to the simulation error, the simulation is repeated and extracted separately; for the obtained channel map, the data with the receiving power less than the threshold is individually set to a uniform value.

[0022] Furthermore, S3 includes:

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

[0024] The channel maps at the first three moments plus the building mask of the area to which the channel map belongs are used as input, and the channel map at the last moment is used as the training label to form the dynamic channel map training data set of the neural network;

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

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

[0027] Two convolution modules are used, followed by a ReLU activation function, as convolution blocks to extract channel space features. The expression of the ReLU activation function is:

[0028]

[0029] Use ViT as the feature fusion and motion information extraction module, where the multi-head attention is 4 and the hidden layer dimension is 4×input dimension;

[0030] A fully connected decoder using linear interpolation as upsampling and max pooling as downsampling is used to decode the compressed information and obtain the desired output;

[0031] Mask pooling is used for feature extraction, which performs 2×2 maximum pooling on the building mask and 2×2 average pooling on the channel map after removing the mask position data.

[0032] Furthermore, S4 includes:

[0033] S401, using a pooling encoder to extract multi-scale information to obtain channel map features;

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

[0035] S403: predict the change trend of the channel map, calculate the error with the actual value, and update the model parameters.

[0036] Further, S401 uses a pooling encoder to extract multi-scale information to obtain channel map features, including:

[0037] After the dynamic channel map dataset is input into the model, the encoder module performs feature extraction on the data with a dimension of [4, 256, 256]. The encoder module operation is:

[0038] The original data is convolved twice after mask pooling to extract features, and then features are extracted by max pooling, and then the features extracted by both are input to the next layer of feature extraction;

[0039] Repeat the above feature extraction operation, and finally output it to the feature fusion module in the form of features with a dimension of [1024, 16, 16].

[0040] Furthermore, S402 uses a Transformer-based attention mechanism to fuse features, including:

[0041] The feature fusion module includes a self-attention mechanism module, which flattens the features of dimension [1024, 16, 16] extracted by the encoding module into a feature vector of dimension [1024, 256]. After passing through a two-layer attention module, it outputs channel spatiotemporal fusion features of dimension [1024, 256], and then restores them to fusion features of dimension [1024, 16, 16].

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

[0043] Use a fully connected decoder to upsample the fused features of dimension [1024, 16, 16]. After passing through four fully connected decoders, image features of the same matrix size as the channel map of dimension [320, 256, 256] are obtained. Finally, after passing through the convolutional layer, the channel map data of the next moment is obtained.

[0044] Calculate the mean square error and differential divergence regularization loss between the predicted value and the actual value of the channel map excluding the building mask at the next moment, and define the loss function L as

[0045]

[0046] in is the predicted channel map, P is the channel map obtained by actual simulation, and L reg is the differential divergence regularization loss, α is an adjustable parameter;

[0047] Select the Adam optimization algorithm to update the network parameters, set the learning rate and weight decay coefficient, and perform iterative training of the model until the model performance reaches the expected level.

[0048] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0049] The dynamic channel map construction method based on machine learning and image processing of the present invention can capture the changes in channel maps in different scenarios, obtain channel characteristics at different times through channel spatiotemporal correlation data features, and has good performance in channel map construction in urban environments. It can obtain channel information in advance in 6G communications, optimize channel estimation, and assist in key technologies such as beam alignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The figure is a flow chart of a method for constructing a dynamic channel map based on machine learning and image processing according to the present invention.

[0051] Figure 2 A schematic diagram of data acquisition and data set processing according to an embodiment of the present invention.

[0052] Figure 3 The present invention is a schematic diagram of a network structure provided according to an embodiment of the present invention.

[0053] Figure 4 A schematic diagram of a network structure encoder provided according to an embodiment of the present invention.

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

[0055] Figure 6 This is a prediction comparison chart constructed according to the dynamic channel map provided by an embodiment of the present invention.

[0056] Figure 7 This is a comparison diagram of the cumulative distribution of absolute errors of the dynamic channel map constructed according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0058] The present invention discloses a method for constructing a dynamic channel map based on machine learning and image processing, the process of which is as follows: Figure 1 As shown, the dynamic channel map construction method based on machine learning and image processing includes the following steps:

[0059] S1. Obtaining channel data through reconstruction simulation, including: firstly reconstructing the scene of the building vector data, then selecting a receiver on the path of the reconstructed scene, and generating a dynamic channel map composed of channel characteristics.

[0060] S101, reconstructed the urban environment based on open source architectural vectors and imported them into the software Wireless Insite.

[0061] S102. For the reconstructed urban environment, receivers are selected at uniform intervals on a known path, and the receivers are evenly arranged in a matrix shape in a random area containing the path for channel simulation. The channel characteristic is the received power. According to the channel map formed by the relative positions of the receivers, each interval of the receivers is regarded as moving in the time dimension, and is stacked in the form of matrix data in sequence as a tensor to form a set of dynamic channel map data. The data set construction process is as follows: Figure 2 shown.

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

[0063] The channel building vector of the receiver arrangement area is extracted according to the constructed dynamic channel map data, and the position where the transmitter does not receive the signal is extracted and set as the building according to the channel map obtained in the same area to obtain the building mask.

[0064] Construct a 1 / 0 binary building mask map, set the existing building to 1, and the others to 0; for the parts that failed to be extracted due to simulation errors, repeat the simulation and extract them separately; for the obtained channel map, the data with received power less than the threshold is set to a uniform value separately.

[0065] S3: Divide the dynamic channel map dataset constructed in S2 into known areas and unknown areas according to a certain ratio, and verify the generality of the model through the known areas. Divide the known areas into training sets and validation sets according to a certain ratio, use the unknown areas as test sets, and test the accuracy of the model through the known areas.

[0066] Because there are many channel map frames on each path, each path needs to be segmented according to the video prediction idea. For the dynamic channel map dataset, every four consecutive frames of channel maps in the same area are regarded as a set of data; the channel maps of the first three moments plus the building mask of the area to which the channel maps belong are used as input, and the channel map of the next moment is used as the training label to constitute the dynamic channel map training dataset of the neural network.

[0067] S4. Build 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 construct the network model as follows Figure 3 As shown, and input the dynamic channel map dataset.

[0069] Two convolution modules are used, followed by a ReLU activation function, as convolution blocks to extract channel space features. The expression of the ReLU activation function is:

[0070]

[0071] ViT is used as the feature fusion and motion information extraction module, where the multi-head attention is 4 and the hidden layer dimension is 4×input dimension.

[0072] A fully connected decoder using linear interpolation as upsampling and maximum pooling as downsampling is used to decode the compressed information and obtain the desired output.

[0073] Mask pooling is used for feature extraction, which performs 2×2 maximum pooling on the building mask and 2×2 average pooling on the channel map after removing the mask position data.

[0074] S401, using a 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 performs feature extraction on the data with a dimension of [4, 256, 256]. The encoder module operation is:

[0076] The original data is subjected to two mask pooling and convolution operations to extract features, and features are extracted by maximum pooling at the same time. The features extracted from both are then used as input for the next layer of feature extraction.

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

[0078] S402. Use the Transformer-based attention mechanism for feature fusion.

[0079] The feature fusion module includes a self-attention mechanism module. The self-attention mechanism module flattens the features of dimension [1024, 16, 16] extracted by the encoding module into a feature vector of dimension [1024, 256]. After passing through the two-layer attention module, the channel spatiotemporal fusion features of dimension [1024, 256] are output, and then restored to the fusion features of dimension [1024, 16, 16]. The fusion process is as follows: Figure 5 shown.

[0080] S403: predict the change trend of the channel map, calculate the error with the actual value, and update the model parameters.

[0081] A fully connected decoder is used to upsample the fused features of dimension [1024, 16, 16]. After passing through four fully connected decoders, image features of the same matrix size as the channel map with dimension [320, 256, 256] are obtained. Finally, after passing through the convolutional layer, the channel map data of the next moment is obtained.

[0082] Calculate the mean square error and differential divergence regularization loss between the predicted value and the actual value of the channel map excluding the building mask at the next moment, and define the loss function L as

[0083]

[0084] in is the predicted channel map, P is the channel map obtained by actual simulation, and L reg is the differential divergence regularization loss, α is an adjustable parameter, and in this embodiment, α is set to 0.1.

[0085] The Adam optimization algorithm was selected to update the network parameters, the learning rate was set to 0.0001, the weight attenuation coefficient was set to 0.0005, and the sum of the mean square error and the differential divergence regularized loss was used as the loss function to train the model for 1000 iterations, and multiple iterations were performed until the model performance reached the expected level.

[0086] Whether the model performance meets expectations is measured by comparing the difference between the output predicted channel statistical characteristics and the actual values.

[0087] Specifically, in this embodiment, the channel map obtained by using the dynamic channel map to construct the 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. The simulation results refer to Figure 6 and Figure 7 . Figure 6 In the urban scenario, the channel map visualization of the channel map prediction model obtained by training is compared with the traditional algorithm and other AI algorithms under the same input and output conditions. Figure 6 (af) are, respectively, the real value obtained by simulation, model-based construction, optical flow method, ViT network, UNet network and the algorithm proposed by the present invention. It can be seen from the simulation figure that the channel map obtained by construction is similar to the simulation result, indicating that the network model has good prediction performance; Figure 7 It is the cumulative distribution diagram of the absolute errors of different schemes. The error curve of the constructed result obtained by training is above all the algorithms, indicating that the error predicted by the network model is the smallest and has good prediction ability.

[0088] In summary, the dynamic channel map construction method based on machine learning and image processing established in the present invention is a construction method that utilizes the temporal and spatial correlation of channels. It uses machine learning to extract the features of historical channel maps at multiple scales, and then uses a self-attention mechanism model to perform feature fusion and process information in the time dimension. Finally, a fully connected decoder is used to realize channel map prediction. It is suitable for the construction of dynamic channel maps in urban scenarios, and has reference significance for the study of channel feature extraction in dynamic electromagnetic environments.

[0089] The matters not described in detail in the present invention are all known technologies to those skilled in the art.

[0090] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A 6G dynamic channel map construction method based on machine learning and image processing, characterized in that: The following steps are involved: S1. Obtaining channel data through reconstruction simulation, including: firstly reconstructing the 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, extract and correct the obtained building mask, and process 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 proportion, and dividing the known areas into training sets and validation sets according to proportion, and using the unknown areas as test sets; S4. Build a dynamic channel map network model based on machine learning and image processing and perform model training, including: Firstly, a dynamic channel map network model is constructed and its model parameters are initialized; Then, the dynamic channel map dataset is used as the input of the model, and after the data features are extracted by the model, the fusion features are generated by the attention module of the model; Finally, the previously extracted data features and fused features are used to predict the change trend of the channel map, the error is calculated with the actual value, and the model parameters are updated according to the results.

2. According to a method for constructing a 6G dynamic channel map based on machine learning and image processing according to claim 1, it is characterized in that: S1 includes: S101, Reconstructing the urban environment based on open source architectural vectors; S102, for the reconstructed urban environment, evenly spaced receivers are selected on a known path, and the receivers are evenly arranged in a matrix shape in a random area including the path to perform channel simulation, where the channel characteristic is the received power; According to the channel map formed by the relative positions of the receivers, each interval of receivers is regarded as moving in the time dimension, and is stacked into tensors in the form of matrix data in sequence to form a set of dynamic channel map data.

3. The 6G dynamic channel map construction method based on machine learning and image processing according to claim 1, characterized in that S2 include: Extracting the channel building vector of the receiver area according to the constructed dynamic channel map data, extracting the location where the transmitter does not receive the signal according to the channel map obtained in the same area and setting it as a building to obtain a building mask; For the parts that failed to be extracted due to simulation errors, repeat the simulation and extract them separately; For the obtained channel map, data with received power less than the threshold value is individually set to a uniform value.

4. The method for constructing a 6G dynamic channel map based on machine learning and image processing according to claim 1, characterized in that: S3 includes: For dynamic channel map data sets, every four consecutive frames of channel maps in the same area are regarded as a set of data; The channel maps at the first three moments plus the building mask of the area to which the channel map belongs are used as input, and the channel map at the last moment is used as the training label to form the dynamic channel map training data set of the neural network; The channel map is normalized so that the value of each parameter is distributed between [0, 1].

5. The method for constructing a 6G dynamic channel map based on machine learning and image processing according to claim 1, characterized in that: Construct a dynamic channel map network model: Two convolution modules followed by ReLU activation function are used as convolution blocks to extract channel space features; Use ViT as the feature fusion and motion information extraction module, where the multi-head attention is 4 and the hidden layer dimension is 4×input dimension; A fully connected decoder using linear interpolation as upsampling and max pooling as downsampling is used to decode the compressed information and obtain the desired output; Mask pooling is used for feature extraction, which performs 2×2 maximum pooling on the building mask and 2×2 average pooling on the channel map after removing the mask position data.

6. A 6G dynamic channel map construction method based on machine learning and image processing according to any one of claims 1 to 5, characterized in that S4 include: S401, using a pooling encoder to extract multi-scale information to obtain channel map features; S402, feature fusion using Transformer-based attention mechanism; S403: predict the change trend of the channel map, calculate the error with the actual value, and update the model parameters.

7. The method for constructing a 6G dynamic channel map based on machine learning and image processing according to claim 6, characterized in that: S401 uses a pooling encoder to extract multi-scale information to obtain channel map features, including: After the dynamic channel map dataset is input into the model, the encoder module performs feature extraction on the data with a dimension of [4, 256, 256]. The encoder module operation is: The original data is convolved twice after mask pooling to extract features, and then the features are extracted by maximum pooling, and then the features extracted by both are used as the input of the next layer of feature extraction; Repeat the above feature extraction operation, and finally output it to the feature fusion module in the form of features with a dimension of [1024, 16, 16].

8. The method for constructing a 6G dynamic channel map based on machine learning and image processing according to claim 6, characterized in that: S402 uses the Transformer-based attention mechanism to fuse features, including: The feature fusion module includes a self-attention mechanism module, which flattens the features of dimension [1024, 16, 16] extracted by the encoding module into a feature vector of dimension [1024, 256]. After passing through a two-layer attention module, it outputs channel spatiotemporal fusion features of dimension [1024, 256], and then restores them to fusion features of dimension [1024, 16, 16].

9. The method for constructing a 6G dynamic channel map based on machine learning and image processing according to claim 6, characterized in that: S403 predicts the change trend of the channel map and updates the model parameters, including: Use a fully connected decoder to upsample the fused features of dimension [1024, 16, 16]. After passing through four fully connected decoders, image features of the same matrix size as the channel map of dimension [320, 256, 256] are obtained. Finally, after passing through the convolutional layer, the channel map data of the next moment is obtained. Calculate the mean square error and differential divergence regularization loss between the predicted value and the actual value of the channel map excluding the building mask at the next moment, and define the loss function L as in is the predicted channel map, P is the channel map obtained by actual simulation, and L reg is the differential divergence regularization loss, α is an adjustable parameter; Select the Adam optimization algorithm to update the network parameters, set the learning rate and weight decay coefficient, and perform iterative training of the model until the model performance reaches the expected level.

Images