A method for modeling a space-time domain predicted channel based on a reinforced loss function

Abstract

The application discloses a kind of space-time domain prediction channel modeling methods based on reinforcement loss function, effectively improve the prediction accuracy of predictive channel modeling based on machine learning.Its implementation steps are as follows: first, pre-processing indoor corridor channel measurement data, obtain the channel statistical characteristics under the corresponding scene and frequency band;Then, according to the space-time correlation in channel statistical characteristics Design weight function based on space-time correlation function, and combine with traditional loss function to form weight loss function;Finally, by improving the loss function of gated recurrent unit network, the space-time correlation characteristics in channel data are perceived, more accurate prediction channel data is output, and the prediction performance of the prediction channel model is improved.The channel statistical characteristics predicted by the model clearly show that the application improves the prediction accuracy of the space-time domain prediction channel model, and the proposed loss function can be used to design more accurate machine learning-based prediction channel models.

Description

Technical Field

[0001] This invention belongs to the field of channel modeling technology, specifically relating to a spatiotemporal predictive channel modeling method based on a reinforcement loss function. Background Technology

[0002] In recent years, research on channel modeling for the Sixth Generation (6G) wireless communication system has played a crucial role in the development of 6G wireless communication technology. The 6G technology vision presents new challenges related to full coverage, all applications, and all frequency bands. However, current channel measurements cannot cover all scenarios, applications, and frequency bands. Furthermore, traditional channel modeling methods, such as geometry-based random channel models, primarily focus on describing the statistical characteristics of the channel and are currently unable to predict channel characteristics in future times and unknown scenarios.

[0003] Currently, machine learning has demonstrated strong capabilities in solving prediction problems, and its application in predictive channel modeling is continuously being optimized to improve the prediction accuracy of channel characteristics in diverse communication scenarios. Research has shown that machine learning model algorithms, such as Recurrent Neural Networks (RNNs) and Explainable Neural Networks (XNNs), can effectively align experimental channel measurement data while keeping prediction errors within acceptable thresholds. Furthermore, machine learning techniques are also being used to optimize wireless propagation channel research, including channel characteristic analysis and antenna-channel optimization, as well as to improve the efficiency of channel prediction in multi-user multiple-input multiple-output (MIMO) systems.

[0004] While dataset size and model parameter settings significantly impact the prediction accuracy of machine learning channel models, the design of the loss function is equally crucial yet often overlooked. The loss function not only affects the accuracy of the predicted sequence but also plays a decisive role in adjusting network parameters. In the spatiotemporal channel, the correlation between spatial and temporal channel characteristics directly influences the algorithm's training efficiency and convergence speed.

[0005] Therefore, designing a loss function that can fully consider the characteristics of the spatiotemporal channel and improve the prediction accuracy of machine learning-based predictive channel models is particularly important. Summary of the Invention

[0006] The problem to be solved by this invention is to provide a spatiotemporal predictive channel modeling method based on reinforcement loss function, which can improve the prediction accuracy of machine learning-based predictive channel models and design more accurate machine learning-based predictive channel models.

[0007] This invention adopts the following technical solution: a spatiotemporal predictive channel modeling method based on an enhancement loss function, comprising the following steps:

[0008] Step S1: Obtain channel measurement data and determine the channel measurement environment parameters of the wireless channel and the antenna location;

[0009] Step S2: Introduce a data augmentation module based on Generative Adversarial Network (GAN) to generate a channel dataset similar to the channel measurement data by using channel characteristics as conditional vectors.

[0010] Step S3: Introduce the prediction channel model architecture of Gated Recurrent Unit (GRU), use the GRU network to capture the dependencies in the sequence data, input the statistical characteristics of the channel dataset generated in step S2 into the network, and predict the channel characteristics at unknown locations in the future.

[0011] Step S4: Construct a space-time aware weighted loss function, consider the correlation between spatial and temporal channel characteristics, and integrate the space-time correlation function to enhance the accuracy of the predicted channel model.

[0012] Step S5: Execute GRU prediction channel model network training, initialize network parameter configuration, set different prediction distances and different training set sample sizes as model training variables, and test the prediction accuracy of the model under different prediction distances and different training set sample sizes.

[0013] Step S6: Evaluate the performance of the enhanced prediction channel model. Compare the differences in prediction accuracy between the space-time aware weight loss function constructed in step S4 and the traditional loss function under different prediction distances and different training set sample sizes.

[0014] Furthermore, step S1 specifically includes:

[0015] S101. Select an indoor corridor scenario for channel measurement and determine the basic parameters of the channel measurement, including frequency, transmit and receive antenna positions, measurement distance d, etc.

[0016] S102. Under line-of-sight conditions, capture the channel impulse response (CIR), preprocess the collected channel measurement data to obtain the corresponding channel statistical characteristics, so as to facilitate subsequent model training and prediction.

[0017] Furthermore, step S2 specifically includes:

[0018] S201. Define the condition vector: The condition vector v is calculated as follows:

[0019] v=[PL(d),τ(d),K R (t)]

[0020] Where PL(d) is the path loss, τ(d) is the delay spread, and K... R (t) is the Rician K factor;

[0021] S202, Generator Design: Use a GAN network to expand the channel measurement data, and take the random Gaussian noise vector s and the conditional vector v as input to generate synthetic data;

[0022] S203, Discriminator Design: Take real or synthetic channel data and the corresponding condition vector v as input, and learn to distinguish between channel measurement data and generated channel data through training;

[0023] S204. Data Synthesis and Augmentation: By training the GAN network, the discriminator is unable to distinguish between channel measurement data and generated channel data, which can be extracted independently to expand the training dataset.

[0024] Furthermore, step S3 specifically includes:

[0025] S301, GRU Predictive Channel Model Network Construction: Treating the spatial-temporal channel prediction problem as a sequence prediction task, a GRU network is constructed to capture the sequence correlation in channel measurement data of continuous antenna locations.

[0026] S302, Channel Data Input: The enhanced channel dataset from step S2 is divided into training, validation, and test sets according to a certain ratio and used as input to the prediction network.

[0027] Furthermore, step S4 specifically includes:

[0028] S401. Spatial-temporal correlation analysis: Based on the channel dataset, the spatial-temporal correlation function is defined as follows:

[0029] R h (Δr)=E{h(r)h*(r+Δr)}

[0030] Where r is the distance between the receiving antenna Rx and the transmitting antenna, Δr is the distance difference between the two Rx channel measurement positions, h(r) represents the channel impulse response at position r, the superscript * indicates conjugate, and E{·} represents the expected value.

[0031] S402. Construct a weight allocation mechanism: Based on spatial-temporal correlation, assign a weight to each sample in the channel measurement dataset. The weight calculation function is as follows:

[0032] W(r,Δr)=f(R h (Δr))

[0033] Where f is the expression describing r and R h The weighted correlation function of (Δr) depends on the specific channel measurement scenario.

[0034] S403. Constructing a normalized weight function: Normalize the calculated weights to ensure a reasonable weight distribution and prevent extreme values ​​from affecting model training. The weight function is calculated as follows:

[0035]

[0036] Where, ∑ Δr W(r,Δr) represents the sum of weights in the training sample set.

[0037] S404. Integrating weights into the loss function: The calculated weights are applied to traditional loss functions (such as mean squared error (MSE) or log-cosh loss) to form a new space-time-aware weighted loss function.

[0038] Furthermore, step S5 specifically includes:

[0039] S501. Initialize network parameter configuration. This initialization specifically includes setting the number of hidden layers and the number of units in each hidden layer, where the hidden layers use the PReLU activation function. Then, set the training batch size to 128, the training epochs to 2500, the initial learning rate to 0.0001, and the learning rate decay to 0.9.

[0040] S502. Set the model input according to different prediction distances Δr and different training set sample sizes N, and compare the prediction accuracy. Prediction accuracy specifically includes obtaining the root mean square error (RMSE) of the output predicted channel data and the input channel data, thereby enabling the exploration of the optimal prediction distance, best prediction range, and optimal training set sample size values ​​for the model.

[0041] Furthermore, step S6 specifically includes:

[0042] S601. Compare the prediction performance of different loss functions under the GRU prediction channel model. The different loss functions specifically include the Mean Square Error (MSE) loss function, the Huber loss function, the log-cosh loss function, the proposed space-time aware weighted MSE loss function, and the proposed space-time aware weighted Log-cosh loss function.

[0043] The MSE loss function formula is as follows:

[0044]

[0045] Among them, y i This represents the channel measurement data at position i. It is the predicted channel output at position i.

[0046] The Huber loss function formula is as follows:

[0047]

[0048] Among them, parameter δ is an indicator that affects the prediction accuracy of the Huber loss function.

[0049] The formula for the Log-cosh loss function is as follows:

[0050]

[0051] The formula for the space-time-aware weighted MSE loss function is as follows:

[0052]

[0053] The formula for the space-time-aware weighted Log-cosh loss function is as follows:

[0054]

[0055] S602. Analyze the advantages of the space-time aware weighted loss function under different prediction distances and training dataset sample sizes. Simulations are used to obtain the RMSE between channel measurement data at different Rx locations and prediction data generated by different models under different prediction channel models using different loss functions, as well as the RMSE between channel measurement data with different training set sample sizes and prediction data generated by different models under different prediction channel models using different loss functions.

[0056] The present invention also provides: an electronic device, comprising:

[0057] One or more processors;

[0058] A storage device on which one or more programs are stored;

[0059] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-described spatiotemporal predictive channel modeling methods based on reinforcement loss functions.

[0060] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in any of the above-mentioned spatiotemporal predictive channel modeling methods based on reinforcement loss functions.

[0061] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0062] 1. The spatiotemporal prediction channel modeling method based on reinforcement loss function proposed in this invention constructs a loss function by considering the correlation between the channel in the spatial and temporal dimensions, which can more accurately reflect the channel characteristics and improve the accuracy of the prediction model. At the same time, the weighting mechanism makes the model pay more attention to the channel samples that have a greater impact on the prediction performance during the training process, thereby enhancing the model's learning of key channel characteristics.

[0063] 2. Experimental results show that, compared with traditional loss functions, the loss function of this invention can significantly reduce prediction errors in various prediction scenarios, especially when the prediction distance is large or the training data is small, its performance improvement is more obvious; based on its good performance in channel prediction, the method of this invention can be used for key technologies such as 6G prediction channel modeling, network optimization and network planning. Attached Figure Description

[0064] Figure 1 This is a flowchart of the spatiotemporal predictive channel modeling method based on the reinforcement loss function of the present invention;

[0065] Figure 2 This is a schematic diagram of the predictive channel network architecture based on GAN-GRU of the present invention;

[0066] Figure 3 This is a schematic diagram of spatiotemporal channel prediction based on a GRU network according to an embodiment of the present invention;

[0067] Figure 4 This is an RMSE diagram of predicted channel data and channel measurement data using different loss functions at different prediction distances according to an embodiment of the present invention;

[0068] Figure 5 This is an RMSE diagram of predicted channel data and channel measurement data using different loss functions under different training set sample sizes in an embodiment of the present invention.

[0069] Figure 6 This is a comparison chart of channel prediction delay power spectral density under different loss functions calculated in the embodiments of the present invention. Detailed Implementation

[0070] To better understand the purpose, structure, and function of this invention, a spatiotemporal predictive channel modeling method based on an enhancement loss function will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention. Furthermore, the step numbers in the embodiments of this invention are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0071] This invention provides a spatiotemporal predictive channel modeling method based on an enhancement loss function, such as... Figure 1 As shown, it includes the following steps:

[0072] Step S1: Obtain channel measurement data and determine the channel measurement environment parameters of the wireless channel and the antenna location;

[0073] Step S2: Introduce a data augmentation module based on Generative Adversarial Network (GAN) to generate a channel dataset similar to the channel measurement data by using channel characteristics as conditional vectors.

[0074] Step S3: Introduce the prediction channel model architecture of Gated Recurrent Unit (GRU), use the GRU network to capture the dependencies in the sequence data, input the statistical characteristics of the channel dataset generated in step S2 into the network, and predict the channel characteristics at unknown locations in the future.

[0075] Step S4: Construct a space-time aware weighted loss function, consider the correlation between spatial and temporal channel characteristics, and integrate the space-time correlation function to enhance the accuracy of the predicted channel model.

[0076] Step S5: Execute GRU prediction channel model network training, initialize network parameter configuration, set different prediction distances and different training set sample sizes as model training variables, and test the prediction accuracy of the model under different prediction distances and different training set sample sizes.

[0077] Step S6: Evaluate the performance of the enhanced prediction channel model. Compare the differences in prediction accuracy between the space-time aware weight loss function constructed in step S4 and the traditional loss function under different prediction distances and different training set sample sizes.

[0078] In one embodiment of the present invention, step S1 specifically includes the following sub-steps:

[0079] S101. Select an indoor corridor scenario for channel measurement and determine the basic parameters of the channel measurement, including frequency f, transmit and receive antenna positions, measurement distance d, etc.

[0080] S102. Under line-of-sight conditions, CIR is captured, and the collected channel measurement data is preprocessed to obtain the corresponding channel statistical characteristics, so as to facilitate subsequent model training and prediction.

[0081] In this embodiment, step S2 specifically includes the following sub-steps:

[0082] S201. Define the condition vector. The condition vector v is calculated as follows:

[0083] v=[PL(d),τ(d),K R (t)]

[0084] Where PL(d) is the path loss, τ(d) is the delay spread, and K... R (t) is the Rician K factor.

[0085] S202. Generator Design. The generator design specifically includes using a GAN network to expand the channel measurement data, taking a random Gaussian noise vector s and a conditional vector v as inputs to generate synthetic data.

[0086] S203. Discriminator Design. The discriminator design specifically includes taking real or synthetic channel data and the corresponding condition vector v as input, and learning to distinguish between channel measurement data and generated channel data through training.

[0087] S204. Data synthesis and enhancement. Specifically, data synthesis and enhancement include training a GAN network so that the discriminator cannot distinguish between channel measurement data and generated channel data, thus enabling independent extraction and expansion of the training dataset.

[0088] In this embodiment, step S3 specifically includes the following sub-steps:

[0089] S301. Construction of the GRU predictive channel model network. Specifically, the GRU predictive channel model network treats the space-time domain channel prediction problem as a sequence prediction task. It constructs a GRU network to capture sequence correlations in channel measurement data from consecutive antenna locations.

[0090] S302, Channel data input. The channel data specifically includes the enhanced channel dataset from step S2, which is divided into a training set, a validation set, and a test set according to a certain ratio, and used as input to the prediction network.

[0091] Specifically, the data augmentation module architecture and the prediction channel module network architecture in this embodiment are as follows: Figure 2 As shown, this is a schematic diagram of space-time domain channel prediction based on a GRU network. Figure 3 As shown.

[0092] In this embodiment, step S4 specifically includes the following sub-steps:

[0093] S401. Spatial-temporal correlation analysis. Specifically, the spatial-temporal correlation includes defining a spatial-temporal correlation function based on the channel dataset as follows:

[0094] R h (Δr)=E{h(r)h*(r+Δr)}

[0095] Where r is the distance between the receiving antenna Rx and the transmitting antenna, Δr is the distance difference between the two Rx channel measurement positions, h(r) represents the channel impulse response at position r, the superscript * indicates conjugate, and E{·} represents the expected value.

[0096] S402. Construct a weight allocation mechanism. Specifically, the weight allocation mechanism includes assigning a weight to each sample in the channel measurement dataset based on spatial-temporal correlation. The weight calculation function is as follows:

[0097] W(r,Δr)=f(R h (Δr))

[0098] Where f is the expression describing r and R h The weighted correlation function of (Δr) depends on the specific channel measurement scenario.

[0099] S403. Construct a normalized weight function. Specifically, the normalized weight function involves normalizing the calculated weights to ensure a reasonable weight distribution and prevent extreme values ​​from affecting model training. The weight function is calculated as follows:

[0100]

[0101] Where, ∑ Δr W(r,Δr) represents the sum of weights in the training sample set;

[0102] S404. Integrating weights into the loss function. The weight loss function specifically includes applying the calculated weights to a traditional loss function (such as mean squared error (MSE) or log-cosh loss) to form a new space-time-aware weight loss function.

[0103] In this embodiment, step S5 specifically includes the following sub-steps:

[0104] S501. Initialize network parameter configuration.

[0105] Initialize the network parameter configuration, specifically by setting the number of hidden layers to 2, with each hidden layer having 256 units, and using the PReLU activation function. Then set the training batch size to 128, the training epochs to 2500, the initial learning rate to 0.0001, and the learning rate decay to 0.9.

[0106] S502. Set the model input according to different prediction distances Δr and different training set sample sizes N, and compare the prediction accuracy.

[0107] Among them, prediction accuracy includes obtaining the root mean square error (RMSE) of the output prediction channel data and the input channel data, thereby enabling the exploration of the optimal prediction distance, the best prediction range, and the optimal training set sample size of the model prediction.

[0108] In this embodiment, step S6 specifically includes the following sub-steps:

[0109] S601. Compare the prediction performance of different loss functions under the GRU prediction channel model. The different loss functions specifically include the MSE loss function, Huber loss function, log-cosh loss function, the proposed space-time aware weighted MSE loss function, and the proposed space-time aware weighted Log-cosh loss function.

[0110] The MSE loss function formula is as follows:

[0111]

[0112] Among them, y i This represents the channel measurement data at position i. It is the predicted channel output at position i.

[0113] The Huber loss function formula is as follows:

[0114]

[0115] Among them, parameter δ is an indicator that affects the prediction accuracy of the Huber loss function.

[0116] The formula for the Log-cosh loss function is as follows:

[0117]

[0118] The formula for the space-time-aware weighted MSE loss function proposed in this embodiment is as follows:

[0119]

[0120] The formula for the space-time-aware weighted Log-cosh loss function proposed in this embodiment is as follows:

[0121]

[0122] S602. Analyze the advantages of the space-time aware weighted loss function under different prediction distances and training dataset sample sizes. Simulations are used to obtain the RMSE between channel measurement data at different Rx locations and prediction data generated by different models under different prediction channel models using different loss functions, as well as the RMSE between channel measurement data with different training set sample sizes and prediction data generated by different models under different prediction channel models using different loss functions.

[0123] Specifically, in this embodiment, the RMSE of predicted channel data and channel measurement data using different loss functions at different prediction distances is as follows: Figure 4 As shown; the RMSE of predicted channel data and channel measurement data using different loss functions under different training set sample sizes, such as... Figure 5 As shown; further, the channel prediction delay power spectral density calculated under different loss functions in this embodiment is compared, as follows: Figure 6 As shown.

[0124] Figure 4 This study demonstrates the performance of prediction channel models using different loss functions (including mean squared error (MSE), Huber loss, Log-cosh loss, weighted MSE loss, and weighted Log loss) at different receiver locations (Rx locations). Performance is measured by the root mean square error (RMSE); a lower RMSE indicates higher prediction model accuracy. Specifically, the RMSE of all loss functions increases with increasing prediction point distance. However, the model using the space-time aware weighted loss function exhibits a lower RMSE at larger prediction distances, indicating better prediction accuracy when handling space-time channel characteristics.

[0125] Figure 5 This demonstrates the performance of prediction channel models using different loss functions on training datasets of varying sizes. The training dataset sizes range from 180 to 250 samples, and performance is measured by RMSE. The figure shows that the RMSE of all models decreases as the training dataset size increases. In particular, the model using the space-time-aware weighted loss function achieves a low RMSE on a training dataset of 250 samples, indicating that this loss function is more effective in improving the model's prediction performance when training with limited data. Specifically, the space-time-aware weighted loss function significantly improves the model's prediction accuracy when the dataset size is larger. This is because the weighted loss function becomes more sensitive to spatiotemporal channel characteristics with a larger training set.

[0126] Figure 6 This demonstrates a comparison of the delay power spectrum between predicted channel models trained with different loss functions and actual channel measurements at a specific receiver location (e.g., Rx15). The model employing a space-time-aware weighted loss function more closely approximates the actual measured data in predicting the delay power spectrum, especially at the 35ns time point. However, significant fluctuations occur in the prediction results within the 150ns to 200ns range, which may be due to noise caused by insufficient training data generated by the GAN data augmentation model.

[0127] In summary, the space-time domain predictive channel modeling method based on the reinforcement loss function proposed in this invention considers the correlation between the spatial and temporal dimensions of the channel. The proposed space-time aware weighted loss function can more accurately reflect channel characteristics, thereby improving the accuracy of the prediction model. Secondly, the weighting mechanism allows the model to pay more attention to channel samples that have a greater impact on prediction performance during training, enhancing the model's learning of key channel characteristics. This results in high accuracy, moderate complexity, and better universality, enriching the methods for predictive channel modeling.

[0128] Simulation results show that, compared with traditional loss functions, the loss function of this invention can significantly reduce prediction errors in various prediction scenarios, verifying the effectiveness and application potential of the method. Therefore, this invention has excellent performance in channel prediction and can be used for key technologies such as 6G predictive channel modeling, network optimization, and network planning.

[0129] Any aspects of this invention not described in detail are well-known to those skilled in the art.

[0130] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A spatiotemporal predictive channel modeling method based on reinforcement loss function, characterized in that, Includes the following steps: Step S1: Obtain channel measurement data and determine the channel measurement environment parameters of the wireless channel environment and antenna location; Step S2: Introduce a data augmentation module based on generative adversarial networks, using channel statistical characteristics as conditional vectors to generate a channel dataset similar to the channel measurement data; Step S3: Construct a prediction channel model based on gated cyclic units. Use the gated cyclic unit network to capture the dependencies of sequence data. Input the statistical characteristics of the channel dataset generated in step S2 into the prediction channel model to predict the channel characteristics at unknown locations in the future. Step S4: Construct a space-time aware weighted loss function, considering the correlation between spatial and temporal channel characteristics, and integrate the space-time correlation function to enhance the accuracy of the predicted channel model. Specifically, this includes: S401. Perform spatial-temporal correlation analysis. Based on the channel dataset, define the spatial-temporal correlation function as follows: R h (Δr)=E{h(r)h * (r+Δr)}; Where r is the distance between the receiving antenna Rx and the transmitting antenna, Δr is the distance difference between the two Rx channel measurement positions, h(r) represents the channel impulse response at position r, the superscript * indicates conjugate, and E{·} represents the expected value; S402. Construct a weight allocation mechanism. Based on the spatial-temporal correlation, assign a weight to each sample in the channel dataset. The weight calculation function is as follows: W(r,Δr)=f(R h (Δr)); Where f is the expression describing r and R h The weighted correlation function of (Δr) depends on the specific channel measurement scenario; S403. Construct a normalized weight function to normalize the calculated weights. The weight function is calculated as follows: Where, ∑ Δr W(r,Δr) represents the sum of weights in the training sample set; S404. Integrating weights into loss functions: Applying the calculated weights to traditional loss functions, including mean squared error (MSE) and log-cosh loss, to form new space-time-aware weighted loss functions, including space-time-aware weighted MSE loss function and space-time-aware weighted Log-cosh loss function. Step S5: Execute the prediction channel model training, set different prediction distances and training set sample sizes as model training variables, and test the prediction accuracy of the prediction channel model under different prediction distances and training set sample sizes. Step S6: Evaluate the performance of the enhanced prediction channel model. Under different prediction distances and different training set sample sizes, compare the differences in prediction accuracy between the space-time aware weight loss function constructed in step S4 and the traditional loss function.

2. The spatiotemporal predictive channel modeling method based on reinforcement loss function according to claim 1, characterized in that, Step S1 acquires channel measurement data, specifically including: S101. Select an indoor corridor scenario for channel measurement and determine the basic parameters of the channel measurement, including: frequency, transmit and receive antenna positions, and measurement distance d. S102. Capture the channel impulse response under line-of-sight conditions, preprocess the channel measurement data, and obtain the channel statistical characteristics under the corresponding scenario and frequency band.

3. The spatiotemporal predictive channel modeling method based on reinforcement loss function according to claim 2, characterized in that, Step S2 introduces a data augmentation module based on generative adversarial networks, specifically including: S201. Define the space-time channel characteristic condition vector v, and calculate it as follows: v=[PL(d),τ(d),K R (t)] Where PL(d) is the path loss, τ(d) is the delay spread, and K... R (t) is the Rician K factor; S202, Generator Design: Use a generative adversarial network to expand the channel measurement data, and use a random Gaussian noise vector s and a conditional vector v as inputs to generate synthetic data; S203, Discriminator Design: Take real or synthetic channel data and the corresponding condition vector v as input, and learn to distinguish between channel measurement data and generated channel data through training; S204. Data Synthesis and Augmentation: By training a generative adversarial network, the discriminator is unable to distinguish between channel measurement data and generated channel data, which are extracted independently to expand the training dataset.

4. The spatiotemporal predictive channel modeling method based on reinforcement loss function according to claim 3, characterized in that, Step S3 involves constructing a prediction channel model network based on gated cyclic units, specifically including: S301. Construct a predictive channel model network: Treat the space-time domain channel prediction problem as a sequence prediction task, construct a gated cyclic unit network, and capture the sequence correlation in the channel measurement data of continuous antenna positions; S302. Input channel data: Based on the channel dataset obtained in step S2, it is divided into training set, validation set and test set according to a certain ratio, which are used as input to the prediction channel model network.

5. The spatiotemporal predictive channel modeling method based on reinforcement loss function according to claim 1, characterized in that, Step S5 executes the training of the gated recurrent unit channel prediction network, specifically including: S501. Initialize the network parameter configuration of the prediction channel model, including: setting the number of hidden layers and the number of units in each hidden layer, using the PReLU activation function for the hidden layers, setting the training batch size, training epochs, initial learning rate, and learning rate decay. S502. Set the model input according to different prediction distances Δr and different training set sample sizes N, and compare the prediction accuracy. The prediction accuracy method includes obtaining the root mean square error of the output prediction channel data and the input channel data, and obtaining the values ​​of the optimal prediction distance, the best prediction range and the optimal training set sample size of the model prediction.

6. The spatiotemporal predictive channel modeling method based on reinforcement loss function according to claim 5, characterized in that, Step S6 evaluates the performance of the prediction channel model, specifically including: S601. Compare the prediction performance of the space-time aware weighted loss function constructed in step S4 with that of the traditional loss function in the prediction channel model based on gated cyclic units. The traditional loss functions include: MSE loss function, Huber loss function, and log-cosh loss function; The space-time perceived weight loss function includes: the space-time perceived weight MSE loss function and the space-time perceived weight Log-cosh loss function; S602. Analyze the advantages of the space-time aware weighted loss function under different prediction distances and training dataset sample sizes. Through simulation, obtain the RMSE between channel measurement data at different Rx locations and prediction data generated by different models under prediction channel models using different loss functions, as well as the RMSE between channel measurement data with different training set sample sizes and prediction data generated by different models under prediction channel models using different loss functions.

7. The spatiotemporal predictive channel modeling method based on reinforcement loss function according to claim 6, characterized in that, In step S601, the MSE loss function is defined as follows: Among them, y i This represents the channel measurement data at position i. It is the predicted channel output at position i; The Huber loss function is defined as follows: Among them, parameter δ is an indicator that affects the prediction accuracy of the Huber loss function; The Log-cosh loss function is defined as follows: The formula for the space-time perceived weighted MSE loss function is as follows: The proposed space-time-aware weighted Log-cosh loss function formula is as follows:

8. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the spatiotemporal predictive channel modeling method based on the reinforcement loss function as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps in the spatial-temporal predictive channel modeling method based on the reinforcement loss function as described in any one of claims 1 to 7.

Images