A spatial-temporal prediction channel modeling method based on machine learning
By using graph attention network and gated loop unit in wireless channel modeling, the spatial and temporal characteristics of channel data are extracted, and the problems of insufficient performance and data correlation neglected by existing models in multi-dimensional change feature fusion are solved. It realizes high-precision channel statistical feature prediction and multi-scenario-applicable effects.
Patent Information
- Application Number
- CN202211189126.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-28
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-09-28
AI Technical Summary
The existing wireless channel prediction models have limited performance in multi-dimensional variability feature fusion, and ignore data sample correlation during training, resulting in complex model structure, redundant parameter and long training time.
The channel data set is constructed using graph attention network (GAT) and gated cycle unit (GRU), extract high correlation information through spatial features and fuse historical channel features in the time dimension, and construct space-time domain fusion features to achieve high-precision channels Statistical characteristics prediction.
It improves the accuracy and performance of channel statistical characteristics prediction, reduces the complexity of the model, and is suitable for channel modeling in multiple communication scenarios.
Smart Images

Figure CN115694697B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the cross-integration application field of machine learning (ML) and wireless channel modeling, and in particular to a space-time domain prediction channel modeling method based on machine learning. Background Art
[0002] Wireless channel modeling is the basis of wireless system design, evaluation and application. Practical wireless channel models help us understand the physical impact of different wireless channels on wireless signal transmission. With the development of new technologies and applications of the Sixth Generation Wireless Communications (6G), wireless communication network systems will face unprecedented complexity. Traditional channel modeling processes will encounter more and more problems in the new 6G frequency bands and new scenarios. New channel modeling methods are urgently needed to meet the new needs of predictive channel modeling.
[0003] With the continuous development of artificial intelligence (AI) and machine learning technologies and the continuous expansion of the scale of communication data that can be obtained, AI-based channel modeling is regarded as a powerful tool to make up for the shortcomings of traditional wireless channel modeling methods such as high complexity and slow operation speed, and is an indispensable part of wireless channel modeling for 6G. Since AI algorithms have excellent ability to mine hidden features in channel measurement data and can uniformly analyze and characterize different channel characteristics from a large amount of communication data, AI-based channel models can use channel databases based on known scenarios, past times, and known frequency bands to predict channel characteristics in unknown scenarios, future times, and location frequency bands. Therefore, AI-based channel modeling is also increasingly being used in the field of wireless communications and plays a huge role in it.
[0004] According to the different dimensions of channel features explored and learned, AI-based predictive channel modeling can be divided into: spatial, time and frequency domain predictive channel modeling. Among them, spatial domain predictive channel modeling realizes channel modeling of unknown scenes by exploring the relationship between the channel and the parameters of the communication physical scene; time domain predictive channel modeling realizes channel modeling of future moments by exploring the law of channel changes over time and using historical information; frequency domain predictive channel modeling realizes cross-band channel modeling by exploring the similarities and differences of channels between different frequency bands. However, the algorithms used by the current predictive channel models are similar to deep learning network algorithms such as feed-forward neural network (FNN), radial basis function neural network (RBF-NN), and convolutional neural network (CNN). They cannot effectively integrate and utilize the changing characteristics of the channel in multiple dimensions, which limits the performance upper limit of the prediction accuracy of the predictive channel model; in addition, most of the existing predictive channel models do not perform data batch preprocessing during model training, and use all data samples for training at the same time, ignoring the correlation between data samples, resulting in complex model structure, redundant model parameters, and long model training time. Summary of the invention
[0005] In view of this, the purpose of the present invention is to provide a space-time domain prediction channel modeling method and device based on machine learning, which can realize information extraction of highly correlated data points in the spatial dimension, and fusion of historical channel features in the time dimension to complete high-precision channel statistical characteristics prediction by constructing Graph Attention Networks (GAT) and Gated Recurrent Unit (GRU) to perform spatial feature-assisted space-time domain prediction channel modeling.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] A space-time domain prediction channel modeling method based on machine learning, the method comprising:
[0008] Step S1, constructing a channel data set, which includes: first obtaining channel data through channel measurement or simulation, and then preprocessing the collected channel data to generate a channel data set consisting of channel statistical characteristics and environmental parameters;
[0009] Step S2: for the channel data set constructed in step S1, divide the channel data at the same time into a channel data subset;
[0010] Step S3, constructing a graph channel data set, which includes: first calculating the channel data correlation, then setting the number of neighboring nodes, and based on the correlation and the number of neighboring nodes, constructing a subset of channel data at each moment into the graph channel data at the current moment, and finally constructing the graph channel data set from the graph channel data at different moments;
[0011] Step S4: for the graph channel data set constructed in step S3, divide it into a training set and a test set according to a certain ratio, then set the time series length Seq, construct the graph channel data into a sequence with a length of Seq+1, and the sequence set is the space-time graph channel data set;
[0012] Step S5, performing model training for a prediction channel model based on GAT-GRU, which includes: first constructing a prediction channel model based on GAT-GRU and initializing its model parameters, then taking the space-time graph channel data set as the input of the model, and using the model to extract channel spatial features and time features in turn, and forming channel space-time fusion features, and finally using the channel space-time fusion features to predict the channel statistical characteristics of each position at future times, calculating the error between the model output result and the actual measurement value, and updating the model parameters based on the results.
[0013] Furthermore, the step S1 specifically includes:
[0014] Step S101, using the software Wireless Insite to build two indoor office environments: a dense office environment and a sparse office environment;
[0015] Step S102: for the two indoor office environments, define the environment parameter vector and the channel statistical characteristic vector, which includes: defining the environment parameter vector as the receiving end antenna coordinates and the transceiver end spacing; defining the channel statistical characteristic vector as the receiving power, root mean square delay spread, root mean square azimuth arrival angle spread, root mean square azimuth departure angle spread, root mean square elevation arrival angle spread, and root mean square elevation departure angle spread; wherein the channel environment parameter vector and the channel statistical characteristic vector together constitute a set of channel data;
[0016] Step S103: perform data normalization processing on each parameter in the channel statistical characteristic vector and the channel environment parameter vector, so that the value of each parameter is distributed between [0, 1].
[0017] Furthermore, the step S3 specifically includes:
[0018] Step S301: when calculating the correlation of channel data, the channel data in the channel data subset at each moment is calculated, and the reciprocal of the Euclidean distance between channel environment parameter vectors is used to characterize the correlation of the channel data;
[0019] Step S302, setting the value of the number of neighbor nodes N, then selecting the N groups of channel data with the highest correlation for each group of channel data to establish a unidirectional edge, and constructing the channel data subset at each moment as the graph channel data at the current moment;
[0020] Step S303: Construct a graph channel data set from graph channel data at different times.
[0021] Furthermore, the step S4 specifically includes:
[0022] Step S401, setting the ratio of the training set to the test set to γ=0.8, and dividing the obtained graph channel data set into a training set and a test set according to the ratio of training set:test set=γ:1-γ;
[0023] Step S402, setting the time sequence length Seq, where the time sequence length is to use the channel characteristics of the previous Seq consecutive moments to predict the channel characteristics of the next moment; wherein, in a dense office environment, Seq takes the values of 4, 6 and 8, respectively, and in a sparse office environment, Seq takes the values of 4, 5 and 6, respectively;
[0024] Step S403, arranging the graph channel data in the training set and the test set in time sequence into a graph channel data sequence of length Seq+1;
[0025] Step S404: construct a set of graph channel data sequences with a length of Seq+1 into a space-time graph channel data set.
[0026] Furthermore, in step S5, firstly constructing a prediction channel model based on GAT-GRU and initializing its model parameters, which specifically includes:
[0027] Step S5011: Use a multi-head GAT model with 32 graph attention layers to form a GAT module in the GAT-GRU model to extract channel space features, followed by a ReLU activation function, whose expression is:
[0028]
[0029] Step S5012: Use a single-layer GRU model with 32 hidden units to construct a GRU module in the GAT-GRU model to extract channel time features.
[0030] Further, in the step S5, the space-time graph channel data set is then used as the input of the model, and the channel spatial features and time features are extracted in sequence by using the model, and the channel space-time domain fusion features are formed, which specifically includes:
[0031] Step S5021, after the space-time graph data set is input into the GAT-GRU model, the GAT module extracts spatial features from the graph channel data with a dimension of [12, 4] in the sequence, and after feature mapping and aggregation, outputs the spatial feature information of the channel with a dimension of [12, 32] to the GRU module;
[0032] Step S5022, the GRU module inputs the hidden state of dimension [12, 32] output by the GRU module at the previous moment and the channel spatial feature information extracted by the GAT module into the GRU module, extracts the channel time feature, and outputs the channel space-time fusion feature of dimension [12, 32], which contains the hidden state of the channel space feature and time feature information at the current moment.
[0033] Furthermore, in step S5, the channel spatial-temporal fusion features are finally used to predict the channel statistical characteristics of each position at the future time, the error between the model output result and the actual measurement value is calculated, and the model parameters are updated based on the feedback, which specifically includes:
[0034] Step S5031: Use a fully connected layer with a dimension of [32, 4] to map the channel space-time fusion features, and finally obtain the same dimension as the node channel statistical characteristics, that is, obtain the channel statistical characteristics prediction value with a dimension of [12, 4] at each position at the next moment;
[0035] Step S5032: Calculate the mean square error between the predicted value and the actual value of the channel statistical characteristics of each node at the next moment, and select the Adam optimization algorithm to update the network parameters, setting the learning rate to 0.002 and the weight attenuation coefficient to 5×10 -5 The model was trained for 1000 iterations using the mean square error as the loss function, and iterated multiple times until the model performance reached the expected level.
[0036] Furthermore, the channel modeling method also includes a performance evaluation step, which includes:
[0037] The performance of the predicted channel model is measured by comparing the difference between the output predicted channel statistical characteristics and the actual values.
[0038] The beneficial effects of the present invention are:
[0039] The present invention introduces a machine learning algorithm on the basis of traditional wireless channel modeling, and proposes a novel predictive channel modeling method based on GAT and GRU. The present invention is an auxiliary space-time domain predictive channel modeling method based on spatial and temporal features. By extracting highly correlated channel information in space and fusing historical channel information in time, it realizes the acquisition of channel information in both spatial and temporal dimensions, while improving the channel statistical characteristics prediction performance, it has an acceptable model complexity. In addition, by evaluating the performance of the predictive channel model trained using channel data in different communication scenarios, it is shown that this method is suitable for predictive channel modeling in a variety of scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 A schematic diagram of the flow of the predictive channel modeling method based on GAT and GRU provided in Example 1;
[0041] Figure 2 A three-dimensional propagation scene diagram of the channel simulation environment provided in Example 1;
[0042] Figure 3 The channel simulation environment transceiver position layout diagram provided in Example 1;
[0043] Figure 4 The flow chart of space-time channel characteristic prediction provided in Example 1;
[0044] Figure 5 The prediction result of the predicted channel model for the root mean square azimuth angle of arrival spread and the channel simulation data curve in the dense office environment provided in Example 1;
[0045] Figure 6 The following are the prediction results of the channel model for the root mean square azimuth angle of arrival spread and the channel simulation data curve in the sparse office environment provided in Example 1. DETAILED DESCRIPTION
[0046] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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 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 are within the scope of protection of the present invention.
[0047] Example 1
[0048] See also Figure 1-Figure 6 This embodiment provides a space-time domain prediction channel modeling method based on machine learning, which specifically includes:
[0049] Step S1, constructing a channel data set, which includes: first obtaining channel data through channel measurement or simulation, and then preprocessing the collected channel data to generate a channel data set consisting of channel statistical characteristics and environmental parameters;
[0050] Specifically, in this embodiment, step S1 specifically includes:
[0051] Step S101: Use the software Wireless Insite to build two indoor office environments: a dense office environment and a sparse office environment. The three-dimensional propagation scene diagrams are shown in Figure 2 . Figure 2 The dotted box in the middle represents the environmental objects deleted in the sparse office environment and the dense office environment. The layout of the transmitter (Receiver, Rx) and the receiver (Transmitter, Tx) in these two simulation environments can be found in Figure 3 ;
[0052] Step S102: Since the Rx coordinate and the carrier frequency f in the simulation m The environmental parameter vector is defined as the Tx antenna coordinate (x Tx ,y Tx ,z Tx ) and the distance between the transmitting and receiving ends d m ; Six channel statistical characteristics, including received power P, root mean square delay spread (Root Mean Square Delay Spread, RMSDS), root mean square azimuth angle spread (Root Mean Square Azimuth Angle Spread of Arrival, RMS AASA), root mean square azimuth angle spread (Root Mean Square Azimuth Angle Spread of Departure, RMS AASD), root mean square elevation angle spread (Root Mean Square Elevation Angle Spread of Arrival, RMS EASA) and root mean square elevation angle spread (Root Mean Square Elevation Angle Spread of Departure, RMS EASD), are recorded as channel statistical characteristic vectors; the channel environment parameter vector and the channel statistical characteristic vector together constitute a set of channel data;
[0053] Step S103: perform data normalization processing on each parameter in the channel statistical characteristic vector and the channel environment parameter vector, so that the value of each parameter is distributed between [0, 1].
[0054] Step S2: for the channel data set constructed in step S1, divide the channel data at the same time into a channel data subset;
[0055] Specifically, in this embodiment, the two simulation environments generate channel data at 104 time points respectively, and each time point contains channel data of 12 measurement points, so the channel data at the same time point in the same simulation environment is divided into 104 channel data subsets, and the size of each subset is 12.
[0056] Step S3, constructing a graph channel data set, which includes: first calculating the channel data correlation, then setting the number of neighboring nodes, and based on the correlation and the number of neighboring nodes, constructing a subset of channel data at each moment into the graph channel data at the current moment, and finally constructing the graph channel data set from the graph channel data at different moments;
[0057] Specifically, in this embodiment, step S3 specifically includes:
[0058] Step S301: when calculating the correlation of channel data, the channel data in the channel data subset at each moment is calculated, and the reciprocal of the Euclidean distance between channel environment parameter vectors is used to characterize the correlation of the channel data;
[0059] Step S302, setting the value of the number of neighbor nodes N, then selecting the N groups of channel data with the highest correlation for each group of channel data to establish a unidirectional edge, and constructing the channel data subset at each moment as the graph channel data at the current moment;
[0060] Step S303: Construct a graph channel data set from graph channel data at different times.
[0061] Step S4: for the graph channel data set constructed in step S3, divide it into a training set and a test set according to a certain ratio, then set the time series length Seq, construct the graph channel data into a sequence with a length of Seq+1, and the sequence set is the space-time graph channel data set;
[0062] Specifically, in this embodiment, step S4 specifically includes:
[0063] Step S401, set the ratio of training set to test set to γ=0.8, and divide the obtained graph channel data set into training set and test set according to: training set: test set=γ:1-γ, so that 83 graph data are divided into training set and 21 graph data are divided into test set;
[0064] Step S402, setting the time sequence length Seq, where the time sequence length is to use the channel characteristics of the previous Seq consecutive moments to predict the channel characteristics of the next moment; wherein, in a dense office environment, Seq takes the values of 4, 6 and 8, respectively, and in a sparse office environment, Seq takes the values of 4, 5 and 6, respectively;
[0065] Step S403, arranging the graph channel data in the training set and the test set in time sequence into a graph channel data sequence of length Seq+1;
[0066] Step S404: construct a set of graph channel data sequences with a length of Seq+1 into a space-time graph channel data set.
[0067] Step S5, initializing the training parameters of the prediction channel model based on GAT-GRU, and using the space-time graph channel data set as the data set of the model for training;
[0068] Specifically, in this embodiment, step S5 specifically includes:
[0069] Step S501: Use a multi-head GAT model with 32 graph attention layers to form a GAT module in the GAT-GRU model to extract channel space features, followed by a ReLU activation function, whose expression is:
[0070]
[0071] Step S502: Use a single-layer GRU model with 32 hidden units to construct a GRU module in the GAT-GRU model to extract channel time features;
[0072] Step S503: input the space-time graph channel data set into the GAT-GRU model for training.
[0073] Step S6: Use the prediction channel model based on GAT-GRU to extract channel spatial features and temporal features in sequence, and form channel spatial and temporal fusion features;
[0074] Specifically, in this embodiment, step S6 specifically includes:
[0075] Step S601, after the space-time graph data set is input into the GAT-GRU model, the GAT module extracts spatial features from the graph channel data with a dimension of [12, 4] in the sequence, and after feature mapping and aggregation, outputs the spatial feature information of the channel with a dimension of [12, 32] to the GRU module;
[0076] Step S602, the GRU module inputs the hidden state of dimension [12, 32] output by the GRU module at the previous moment and the channel spatial feature information extracted by the GAT module into the GRU module, extracts the channel time feature, and outputs the channel space-time fusion feature of dimension [12, 32], which contains the hidden state of the channel space feature and time feature information at the current moment.
[0077] Step S7: Use the channel space-time fusion features to predict the channel statistical characteristics of each position at the future time, calculate the error between the model output result and the actual measurement value, and use the results to feedback and update the model parameters;
[0078] Specifically, in this embodiment, step S7 specifically includes:
[0079] Step S701: Use a fully connected layer with a dimension of [32, 4] to map the channel space-time fusion features, and finally obtain the same dimension as the node channel statistical characteristics, that is, obtain the channel statistical characteristics prediction value with a dimension of [12, 4] at each position at the next moment;
[0080] Step S702: Calculate the mean square error between the predicted value and the actual value of the channel statistical characteristics of each node at the next moment, and select the Adam optimization algorithm to update the network parameters, setting the learning rate to 0.002 and the weight attenuation coefficient to 5×10 -5 The model was trained for 1000 iterations using the mean square error as the loss function, and iterated multiple times until the model performance reached the expected level.
[0081] Step S8: Evaluate the performance of the model by comparing the difference between the output predicted channel statistical characteristics and the actual values to measure the performance of the obtained predicted channel model.
[0082] Specifically, in this embodiment, the six channel statistical characteristics predicted by the predicted channel model are compared with the channel statistical characteristics of the channel simulation data to analyze the performance of the predicted channel model. Figure 5 and Figure 6 . Figure 5 This is the prediction result of the trained prediction channel model for the RMS azimuth arrival angle spread and the channel simulation data curve in a dense office environment. From the simulation graph, it can be seen that the prediction result and the channel simulation result have a good fit, indicating that the prediction model has good prediction performance; Figure 6 The following are the prediction results of the trained prediction channel model for the RMS azimuth arrival angle spread and the channel simulation data curve in a sparse office environment. The good fit between the prediction results and the channel simulation results shows that the prediction model still has good prediction performance in different scenarios.
[0083] In summary, the space-time domain prediction channel model based on machine learning established in the present invention is a space-time domain prediction channel model assisted by spatial features. The GAT model is used to extract highly correlated data information in space, and then the GRU model is used to fuse it with historical channel information in time to form a channel space-time domain fusion feature containing channel spatial features and time feature information. Finally, the channel space-time domain fusion feature is used to achieve high-precision channel statistical characteristics prediction, and is suitable for predictive channel modeling in a variety of scenarios, which has reference significance for multi-dimensional predictive channel modeling research.
[0084] The matters not described in detail in the present invention are all known technologies to those skilled in the art.
[0085] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A spatial-temporal prediction channel modeling method based on machine learning, characterized in that: The method includes: Step S1, constructing a channel data set, which includes: first obtaining channel data through channel measurement or simulation, and then preprocessing the collected channel data to generate a channel data set consisting of channel statistical characteristics and environmental parameters; Step S2: for the channel data set constructed in step S1, divide the channel data at the same time into a channel data subset; Step S3, constructing a graph channel data set, which includes: first calculating the channel data correlation, then setting the number of neighboring nodes, and based on the correlation and the number of neighboring nodes, constructing a subset of channel data at each moment into the graph channel data at the current moment, and finally constructing the graph channel data set from the graph channel data at different moments; Step S4: for the graph channel data set constructed in step S3, divide it into a training set and a test set according to a certain ratio, then set the time series length Seq, construct the graph channel data into a sequence with a length of Seq+1, and the sequence set is the space-time graph channel data set; Step S5, performing model training for a prediction channel model based on GAT-GRU, which includes: first constructing a prediction channel model based on GAT-GRU and initializing its model parameters, then taking the space-time graph channel data set as the input of the model, and using the model to extract channel spatial features and time features in sequence, and forming channel space-time fusion features, and finally using the channel space-time fusion features to predict channel statistical characteristics at each position at a future time, calculating the error between the model output result and the actual measurement value, and updating the model parameters based on the results; The step S1 specifically includes: Step S101, using the software Wireless Insite to build two indoor office environments: a dense office environment and a sparse office environment; Step S102: for the two indoor office environments, define the environment parameter vector and the channel statistical characteristic vector, which includes: defining the environment parameter vector as the receiving end antenna coordinates and the transceiver end spacing; defining the channel statistical characteristic vector as the received power, root mean square delay spread, root mean square azimuth arrival angle spread, root mean square azimuth departure angle spread, root mean square elevation arrival angle spread and root mean square elevation departure angle spread; wherein the channel environment parameter vector and the channel statistical characteristic vector together constitute a set of channel data; Step S103, performing data normalization processing on each parameter in the channel statistical characteristic vector and the channel environment parameter vector, so that the value of each parameter is distributed between [0, 1]; The step S3 specifically includes: Step S301: when calculating the correlation of channel data, the channel data in the channel data subset at each moment is calculated, and the reciprocal of the Euclidean distance between channel environment parameter vectors is used to characterize the correlation of the channel data; Step S302, setting the value of the number of neighbor nodes N, then selecting the N groups of channel data with the highest correlation for each group of channel data to establish a unidirectional edge, and constructing the channel data subset at each moment as the graph channel data at the current moment; Step S303, forming a graph channel data set from graph channel data at different times; In step S5, a prediction channel model based on GAT-GRU is first constructed and its model parameters are initialized, which specifically includes: Step S5011: Use a multi-head GAT model with 32 graph attention layers to form a GAT module in the GAT-GRU model to extract channel space features, followed by a ReLU activation function, whose expression is: Step S5012: Use a single-layer GRU model with 32 hidden units to construct a GRU module in the GAT-GRU model to extract channel time features.
2. The spatial-temporal prediction channel modeling method based on machine learning according to claim 1, characterized in that: In step S2, the two indoor office environments respectively generate channel data at 104 moments, and each moment contains channel data of 12 measurement points. The channel data at the same moment in the same simulation environment is divided into 104 channel data subsets, and the size of each subset is 12.
3. The spatial-temporal prediction channel modeling method based on machine learning according to claim 1, characterized in that: The step S4 specifically includes: Step S401, setting the ratio of the training set to the test set to γ=0.8, and dividing the obtained graph channel data set into a training set and a test set according to: training set: test set=γ:1-γ; Step S402, setting the time sequence length Seq, where the time sequence length is to use the channel characteristics of the previous Seq consecutive moments to predict the channel characteristics of the next moment; wherein, in a dense office environment, Seq takes the values of 4, 6 and 8, respectively, and in a sparse office environment, Seq takes the values of 4, 5 and 6, respectively; Step S403, arranging the graph channel data in the training set and the test set in time sequence into a graph channel data sequence of length Seq+1; Step S404: construct a set of graph channel data sequences with a length of Seq+1 into a space-time graph channel data set.
4. The spatial-temporal prediction channel modeling method based on machine learning according to claim 1, characterized in that: In step S5, the space-time graph channel data set is then used as the input of the model, and the channel spatial features and time features are extracted in sequence using the model, and the channel space-time domain fusion features are formed, which specifically includes: Step S5021, after the space-time graph data set is input into the GAT-GRU model, the GAT module extracts spatial features from the graph channel data with a dimension of [12, 4] in the sequence, and after feature mapping and aggregation, outputs the spatial feature information of the channel with a dimension of [12, 32] to the GRU module; Step S5022, the GRU module inputs the hidden state of dimension [12, 32] output by the GRU module at the previous moment and the channel spatial feature information extracted by the GAT module into the GRU module, extracts the channel time feature, and outputs the channel space-time fusion feature of dimension [12, 32], which contains the hidden state of the channel space feature and time feature information at the current moment.
5. The spatial-temporal prediction channel modeling method based on machine learning according to claim 4 is characterized in that: In step S5, the channel spatial-temporal fusion features are finally used to predict the channel statistical characteristics of each position at the future time, the error between the model output result and the actual measurement value is calculated, and the model parameters are updated based on the feedback, which specifically includes: Step S5031: Use a fully connected layer with a dimension of [32, 4] to map the channel space-time fusion features, and finally obtain the same dimension as the node channel statistical characteristics, that is, obtain the channel statistical characteristics prediction value with a dimension of [12, 4] at each position at the next moment; Step S5032: Calculate the mean square error between the predicted value and the actual value of the channel statistical characteristics of each node at the next moment, and select the Adam optimization algorithm to update the network parameters, setting the learning rate to 0.002 and the weight attenuation coefficient to 5×10 -5 The model was trained for 1000 iterations using the mean square error as the loss function, and iterated multiple times until the model performance reached the expected level.
6. The spatial-temporal prediction channel modeling method based on machine learning according to claim 5, characterized in that: The channel modeling method also includes a performance evaluation step, which includes: The performance of the predicted channel model is measured by comparing the difference between the output predicted channel statistical characteristics and the actual values.
Citation Information
Patent Citations
Heterogeneous agent trajectory prediction method, system, device and medium
CN113568410A
Prediction channel modeling method based on adversarial network and long short-term memory network
CN114584230A
Blind calibration method for wireless sensor network data drift
WO2020191980A1
