An AI-based intelligent predictive channel modeling method for millimeter waves

Abstract

This invention discloses an artificial intelligence-based intelligent predictive channel modeling method for millimeter waves, belonging to the field of wireless communication and channel modeling technology. It addresses the limitations of traditional channel modeling methods in predicting channel characteristics in future times or unknown scenarios, and in meeting the high-precision requirements of 6G. The method comprises three core functions: 1. Channel characteristic analysis: standardizing imported measurement or simulation channel data and extracting unified multi-domain channel statistical characteristics as a benchmark; 2. Channel generation: optimizing parameters based on a geometric random channel model to generate high-fidelity channel data and supporting dataset expansion; 3. Channel prediction: integrating temporal convolutional networks, long short-term memory networks, and gated recurrent unit prediction algorithms to achieve intelligent channel state prediction for millimeter waves. This invention achieves high-precision channel prediction with millisecond-level latency, providing core technical support for the large-scale deployment of 6G networks and the design of intelligent communication systems.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of wireless communication and artificial intelligence, and in particular relates to a millimeter-wave intelligent predictive channel modeling method based on artificial intelligence. Background Technology

[0002] As sixth-generation (6G) mobile communication systems evolve towards an integrated air-space-ground-sea approach, the wireless channel environment is becoming increasingly complex and dynamic, with significant differences in channel statistical characteristics across different propagation scenarios. Existing wireless channel modeling methods primarily include empirical models based on measurement statistical characteristics and channel models based on geometric stochastic theory, such as the 3GPP model, WINNER model, and QuaDRiGa model. These methods typically rely on measurement data and manual parameter configuration for specific scenarios, making it difficult to achieve unified modeling and rapid adaptation in complex millimeter-wave propagation environments. Furthermore, in the modeling process based on geometric stochastic models, there are complex coupling relationships between large-scale and small-scale parameters, and parameter sensitivity varies considerably across different scenarios. Traditional manual parameter tuning, grid search, or heuristic optimization methods suffer from high computational complexity, low optimization efficiency, and insufficient global adaptability, making it difficult to achieve joint matching of multi-dimensional statistical characteristics in the time delay, angle, and spatial domains.

[0003] On the other hand, with the development of artificial intelligence technology, data-driven methods based on deep learning are gradually being applied to the field of wireless channel modeling and prediction. These methods, trained on historical channel data, can learn the evolutionary patterns of channels over time and space, exhibiting good predictive capabilities in complex propagation environments. However, most existing AI methods focus on channel prediction or sample generation, lacking constraints on the physical characteristics of channel propagation. This leads to a decline in generalization performance in unknown scenarios or under varying data distribution conditions. Furthermore, channel characteristic analysis, channel generation, and channel prediction in existing technologies are typically independent, lacking a unified data processing and collaborative optimization mechanism. This makes it difficult to achieve information sharing and joint optimization between different modules, thus limiting overall modeling efficiency and adaptability.

[0004] Therefore, there is an urgent need for a millimeter-wave intelligent channel modeling platform that integrates channel characteristic analysis, channel generation, and intelligent prediction functions. By combining geometric random channel models with artificial intelligence prediction methods, a unified modeling framework with statistical characteristic constraints and joint optimization capabilities can be constructed to achieve integrated collaborative modeling of channel data processing, channel generation, channel prediction, and adaptive updating of model parameters under complex propagation environments. This will improve the accuracy, generalization ability, and engineering application efficiency of channel modeling and prediction under multiple scenarios. Summary of the Invention

[0005] Objective: To address the challenges of existing channel modeling and prediction methods, such as difficulty in unified modeling under multiple scenarios, lack of dynamic prediction capabilities, and inability to simultaneously consider the joint matching of multi-dimensional statistical characteristics of the channel in the time, angular, and frequency domains, this invention provides an artificial intelligence-based intelligent prediction channel modeling method for millimeter waves. This method addresses the problems of traditional channel modeling methods relying on empirical modeling, difficulty in depicting channel evolution patterns under complex propagation environments, and the independence of channel modeling, channel generation, and channel prediction processes without a collaborative optimization mechanism. It utilizes a deep learning model to establish a mapping relationship between historical channel sequences and future channel states, and constructs a joint optimization mechanism based on channel statistical characteristic constraints. Under the condition of satisfying the physical characteristics of the channel, it achieves adaptive updating of model parameters, thereby improving the matching accuracy between the predicted channel and the real channel in statistical indicators such as delay spread, angular spread, and Doppler characteristics, and enhancing the accuracy and stability of channel modeling and prediction under complex millimeter wave propagation environments.

[0006] Technical solution: The present invention provides a millimeter-wave intelligent predictive channel modeling method based on artificial intelligence, comprising the following steps:

[0007] Step S1: Determine the target propagation scenario, operating frequency band, and system configuration parameters, and construct a unified channel modeling environment;

[0008] Step S2: Acquire channel data in the target scenario and perform preprocessing;

[0009] Step S3: Extract the channel statistical characteristics of the preprocessed data and construct the channel feature sequence;

[0010] Step S4: Construct a channel generation model based on the geometric random channel model and generate a simulated channel;

[0011] Step S5: Construct an artificial intelligence-based channel prediction model. The model takes channel feature sequences and scenario conditions as input to predict future channel states and complete missing data.

[0012] Step S6: Construct a joint optimization objective function based on the difference between prediction error and channel statistical characteristics, and perform collaborative optimization on the channel generation model and prediction model;

[0013] Step S7: Output the optimized channel results and perform closed-loop update.

[0014] Furthermore, step S1 specifically includes:

[0015] S101. Select a propagation scenario, wherein the propagation scenario includes one of the following: indoor, outdoor, air-to-ground, space-to-ground, or sea-based scenarios;

[0016] S102, select millimeter wave band, also compatible with Sub-6GHz or terahertz band;

[0017] S103. Configure system parameters, including bandwidth, antenna array structure, antenna height, transceiver position relationship and propagation conditions;

[0018] S104. Construct a channel modeling environment based on the propagation scenario and system parameters.

[0019] Furthermore, step S2 specifically includes:

[0020] S201. Obtain channel impulse response data or channel transfer function data;

[0021] S202. Preprocess the channel data, including noise reduction, normalization, and time or frequency alignment.

[0022] The denoising process employs a wavelet threshold denoising method to suppress high-frequency noise components in the channel data.

[0023] The normalization process uses the Z-score normalization method, whose expression is:

[0024]

[0025] Where H represents the original channel data, and X represents the normalized channel data. This represents the mean of the original channel data H. This represents the standard deviation of the raw channel data H;

[0026] The time alignment process adopts an alignment method based on the maximum correlation peak. By calculating the cross-correlation function of the channel response between different snapshots, the position corresponding to the maximum correlation peak is determined, and time synchronization alignment is completed accordingly.

[0027] Frequency alignment processing achieves frequency domain unification by compensating for frequency offsets in the channel responses at different frequency points;

[0028] S203. Organize the processed data into standardized channel input data.

[0029] Furthermore, step S3 specifically includes:

[0030] S301. Extract statistical characteristics such as delay spread, angle spread, and Doppler spread from the preprocessed channel data;

[0031] S302. Construct the channel feature vector, which is expressed as follows:

[0032]

[0033] in, A vector representing the statistical characteristics of the target channel; This represents the preprocessed channel data; This represents the channel feature extraction function; Indicates delay spread characteristics; Indicates angular extension features; Indicates Doppler extended features;

[0034] S303. Use the channel feature vector as input for the subsequent model.

[0035] Furthermore, step S4 specifically includes:

[0036] S401. Construct a channel generation module based on a geometric random channel model;

[0037]

[0038] in, Indicates time and latency Channel impulse response under the following conditions; Indicates the number of multipath components; Indicates the first The amplitude coefficient of the path; Indicates the first The phase of the path; Indicates the first The Doppler frequency shift corresponding to each path; Indicates the first The propagation delay of each path; Represents the Dirac impulse function; Represents the imaginary unit;

[0039] S402. Define the channel model parameter vector, which is expressed as follows:

[0040]

[0041] in, Represents the channel model parameter vector; Represents a large-scale parameter vector used to describe statistical properties such as time delay spread, angle spread, and shadow fading; Represents a small-scale parameter vector used to describe the propagation characteristics of multipath components, such as path power, phase, Doppler shift, and time delay distribution; superscript This represents the matrix transpose operation;

[0042] S403. Generate the channel matrix based on the parameter vector, which is expressed as follows:

[0043]

[0044] in, This indicates large-scale decay. This represents small-scale fading, which is generated by a multipath propagation model.

[0045] S404. Extract statistical characteristics such as delay spread, angle spread, and Doppler spread from the generated channel and match them with the corresponding statistical characteristics of the target channel for subsequent joint optimization objective function construction and model parameter update.

[0046] Furthermore, step S5 specifically includes:

[0047] S501. Construct a channel prediction model based on deep learning. The channel prediction model includes a temporal convolutional network (TCN) and a long short-term memory network (LSTM). The TCN is used to extract local temporal features from historical channel sequences, and the LSTM is used to learn the long-term correlation characteristics of channel changes over time. The output features of the TCN and the LSTM are fused to obtain the output result of the channel prediction model.

[0048] S502. Channel prediction based on historical channel sequences, expressed as follows:

[0049]

[0050] in, This represents the predicted future time-series channel matrix; The parameter is Channel prediction model; A sequence of channel features representing historical moments; Indicates scene condition parameters; Represents model parameters

[0051] S503. Based on the predicted channel results output by the channel prediction model, the missing channel data is completed. By utilizing the temporal correlation between channel features at adjacent times, the channel state information corresponding to the missing position is restored.

[0052] S504. The predicted channel matrix and its statistical characteristics are used as inputs to the joint optimization objective function to calculate the difference between the prediction error and the statistical characteristics, and the channel generation model parameters and channel prediction model parameters are updated according to the optimization results.

[0053] Furthermore, step S6 specifically includes:

[0054] S601. Construct the prediction error loss function, which is expressed as follows:

[0055]

[0056] in, This represents the prediction error loss function; Indicates the parameters of the channel prediction model; Indicates the number of training samples; Indicates the first A real target channel matrix; Indicates the first One prediction channel matrix; The squared Frobenius norm of a matrix is ​​equivalent to the sum of the squares of the moduli of all its elements;

[0057] S602. Construct a channel statistical characteristic constraint function to describe the difference in statistical characteristics between the generated channel and the target channel, expressed as follows:

[0058]

[0059] in, A function representing statistical differences; A vector representing the statistical characteristics of the target channel; This represents the statistical characteristic vector of the generated channel; Represents the L2 norm;

[0060] S603. Construct the joint optimization objective function, which is expressed as follows:

[0061]

[0062] in, This represents the statistical characteristic vector of the target channel. For parameter space; For prediction error loss; These are the weighting coefficients; This is a constraint function for statistical characteristic differences; and These are the statistical feature vectors of the target channel and the predicted channel, respectively.

[0063] S604. Update model parameters based on joint optimization objective function.

[0064] Furthermore, step S7 specifically includes:

[0065] S701. Output the predicted channel matrix and corresponding statistical characteristic parameters, including delay spread, angle spread and Doppler spread.

[0066] S702. The performance of the prediction results is evaluated based on the normalized mean square error (NMSE) and the KS test distance. The normalized mean square error is used to evaluate the error between the predicted channel and the target channel, and the KS test distance is used to evaluate the consistency of the statistical distribution between the generated channel and the target channel.

[0067] S703. Based on the performance evaluation results, the joint optimization objective function is iteratively updated, and the optimized model parameters are fed back to the channel generation model and the channel prediction model to update the channel generation parameters and prediction model parameters.

[0068] S704. When the input channel data or scenario conditions change, the channel generation and channel prediction process is re-executed based on the updated model parameters, forming a closed-loop optimization mechanism for channel generation and prediction.

[0069] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0070] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0071] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0072] The proposed artificial intelligence-based millimeter-wave channel prediction modeling method has the following advantages: First, this invention integrates channel statistical characteristic constraints and prediction errors into the optimization process, achieving coordinated matching of multi-dimensional statistical characteristics in the time delay domain, angle domain, and Doppler domain. This improves the matching accuracy between the predicted and real channels while ensuring channel physical consistency. Second, this invention constructs a unified modeling framework that integrates channel feature extraction, channel generation, and deep learning prediction. Through a joint optimization mechanism, it achieves coordinated updating of model parameters, effectively reducing the error accumulation problem caused by traditional independent modeling methods and improving channel modeling and prediction performance. Third, this invention is applicable to various propagation scenarios, including indoor, outdoor, air-to-ground, satellite-to-ground, and maritime environments, possessing excellent cross-scenario generalization and adaptability, and can meet the channel modeling and prediction needs in complex millimeter-wave propagation environments in 6G wireless communication systems. Attached Figure Description

[0073] Figure 1 Flowchart of an AI-based intelligent channel prediction modeling method for millimeter waves;

[0074] Figure 2 Overall framework diagram of an AI-based millimeter-wave intelligent channel prediction modeling platform;

[0075] Figure 3 : A real-world interface diagram of millimeter-wave intelligent channel generation and statistical characteristics in this embodiment of the invention;

[0076] Figure 4 : Interface diagram of millimeter-wave intelligent channel generation and statistical characteristics based on geometric random model in this embodiment of the invention;

[0077] Figure 5 The following is an interface diagram of the channel prediction and model training verification results based on deep learning in this embodiment of the invention. Detailed Implementation

[0078] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0079] Example 1: A millimeter-wave intelligent channel prediction modeling method based on artificial intelligence, comprising the following steps:

[0080] Step S1: The propagation scenario, frequency band, and system configuration parameters corresponding to an artificial intelligence-based millimeter-wave intelligent channel prediction modeling platform and method.

[0081] Specifically, in this embodiment, step S1 includes:

[0082] S101. Select the outdoor urban micro-cell millimeter wave propagation scenario as the channel prediction modeling object, set the carrier frequency to 28 GHz, and the propagation environment includes both line-of-sight (LOS) and non-line-of-sight (NLOS) propagation conditions.

[0083] S102. Determine the system configuration parameters, where the system bandwidth is set to 100 MHz, the base station uses an 8×8 uniform planar antenna array, the terminal uses a 2×2 uniform planar antenna array, the element spacing is set to half wavelength, the transmit power is set to 30 dBm, the terminal moving speed is set to 20 km / h, the base station antenna height is 10 m, and the terminal antenna height is 1.5 m.

[0084] S103. Based on the above propagation scenarios and system configuration parameters, a unified channel modeling environment is constructed. A millimeter-wave multipath propagation model is established according to the scenario propagation conditions. The channel impulse response data under the target scenario is generated by combining the ray tracing method, providing basic environmental support for subsequent channel statistical characteristic extraction, channel generation and deep learning prediction.

[0085] Step S2: Preprocess the channel impulse response data under the target scenario and extract the corresponding channel statistical characteristics as input for the subsequent channel generation and prediction model.

[0086] Specifically, in this embodiment, step S2 includes:

[0087] S201. Channel impulse response data in an outdoor urban micro-cell millimeter-wave scenario was acquired using the ray tracing method. A total of 10,000 channel snapshots were collected, and the raw channel data was denoised, time-aligned, and normalized.

[0088] S202. A wavelet thresholding denoising method is used to suppress high-frequency noise components in the channel impulse response, wherein the wavelet thresholding function is a soft thresholding function; a maximum correlation peak matching method is used to align the time offsets between different snapshots; and a Z-score normalization method is used to standardize the channel data, the expression of which is:

[0089]

[0090] Where μ and σ represent the mean and standard deviation of the channel data, respectively.

[0091] S203. Organize the preprocessed channel impulse response data into a three-dimensional tensor form [number of snapshots, number of antenna pairs, time delay taps], and further calculate statistical characteristics such as time delay power spectral density, angular power spectral density and Doppler spectrum. Extract samples of root mean square time delay spread, angular spread and Doppler spread under millimeter wave scenario, and construct the empirical cumulative distribution function of the corresponding statistical characteristics as the target statistical characteristics for subsequent channel generation model parameter optimization and deep learning prediction model training.

[0092] Step S3: Using 100 channel snapshots as a statistical window, extract statistical characteristics such as root mean square delay spread, angle spread, and Doppler spread to construct channel feature vectors, which serve as inputs for subsequent channel generation models and deep learning prediction models, as well as statistical characteristic matching benchmarks.

[0093] Specifically, in this embodiment, step S3 includes:

[0094] S301. Extract the time delay domain, spatial domain, and frequency domain statistical characteristics based on the preprocessed channel impulse response data. Specifically, the root mean square time delay spread is obtained through power delay spectrum calculation, the angle spread is obtained through statistical estimation of the spatial spectrum, and the Doppler spread is obtained through the snapshot correlation function.

[0095] S302. Construct the channel feature vector, which is expressed as follows:

[0096]

[0097] in, A vector representing the statistical characteristics of the target channel; This represents the preprocessed channel data; This represents the channel feature extraction function; Indicates delay spread characteristics; Indicates angular extension features; Indicates Doppler extended features;

[0098] Furthermore, the root mean square delay spread is calculated as follows:

[0099]

[0100] in, This represents the propagation delay of the i-th path. This indicates the received power for the corresponding path.

[0101] S303. Organize the feature vectors obtained under different statistical windows into a feature sequence in chronological order, and construct the empirical cumulative distribution function of the corresponding statistical characteristics as the statistical characteristic benchmark for subsequent optimization of channel generation model parameters and training of deep learning channel prediction model.

[0102] Step S4: Construct a channel generation module based on the 3GPP TR 38.901 geometric random channel model, define large-scale parameters and small-scale parameters, and generate the corresponding simulated channel matrix. Extract statistical characteristics from the generated channel for subsequent matching and optimization.

[0103] Specifically, in this embodiment, step S4 includes:

[0104] S401. A channel generation module is constructed based on the geometric random channel model in the 3GPP TR 38.901 standard. The channel model includes both large-scale fading characteristics and small-scale multipath fading characteristics, and is used to describe the high path loss, sparse multipath, and directional propagation characteristics in the millimeter wave propagation environment.

[0105] S402. Define the channel model parameter vector, where the parameter vector to be optimized includes a large-scale parameter vector and a small-scale parameter vector, and its expression is as follows:

[0106]

[0107] in, Represents a large-scale parameter vector. This represents a small-scale parameter vector.

[0108] Furthermore, large-scale parameters include statistical parameters such as time delay spread, angular spread, Rician K factor, and shadow fading standard deviation; small-scale parameters include propagation parameters such as the number of scattering clusters, the number of rays per cluster, cluster time delay spread, and cluster angular spread.

[0109] S403. Generate a channel matrix based on the parameter vector, the expression of which is as follows:

[0110]

[0111] in, Represents the large-scale fading component. This represents the small-scale fading component.

[0112] Furthermore, small-scale fading can be represented as a superposition of multipath components, as shown in the following expression:

[0113]

[0114] in, Indicates the number of multipath clusters. This indicates the number of rays within each cluster. This represents the complex magnitude of the k-th path in the l-th cluster. Indicates Doppler frequency shift, Indicates a random phase.

[0115] S404. Based on the generated channel matrix, further calculate the statistical characteristics such as time delay power spectrum, angle power spectrum and Doppler spectrum, and extract statistical parameters such as root mean square time delay spread, angle spread and Doppler spread as target constraints in the subsequent statistical characteristic matching and joint optimization process.

[0116] Step S5: Construct a deep learning channel prediction model that includes a temporal convolutional network (TCN) and a long short-term memory network (LSTM). Use historical channel sequences and scenario conditions as input to achieve prediction of future channel states and completion of missing channel data.

[0117] Specifically, in this embodiment, step S5 includes:

[0118] S501. Construct a channel prediction network based on deep learning. The prediction network consists of a Temporal Convolutional Network (TCN) branch and a Long Short-Term Memory (LSTM) branch. The TCN is used to extract local temporal features in the channel sequence, and the LSTM is used to learn the long-term temporal correlation of the channel.

[0119] S502. Construct the historical channel input sequence and scenario condition input vector. The historical window length is set to L=10, and the historical channel sequence is defined as follows:

[0120]

[0121] The scene condition vector includes the millimeter-wave scene number, terminal moving speed, operating bandwidth, and propagation condition parameters.

[0122] Furthermore, based on historical channel sequences and scenario conditions, the future channel state is predicted, and its expression is as follows:

[0123]

[0124] in, This represents the predicted channel state at a future time. C represents the deep learning prediction model, and C represents the scene condition information.

[0125] S503. Configure the structure of the deep learning prediction network. The TCN contains 3 one-dimensional convolutional layers, each with a kernel size of 3 and 64 channels; the LSTM contains 2 hidden layers, each with 128 hidden units; the outputs of the TCN and LSTM are weighted and fused to obtain the final prediction result.

[0126] Furthermore, the Adam optimizer was used to update parameters during model training, with the learning rate set to 0.001, the batch size set to 64, and the number of training epochs set to 100.

[0127] S504. In the missing channel data completion mode, the missing location mask and historical channel data are input into the prediction network. The missing snapshots are reconstructed using the model output to obtain the complete channel sequence. The prediction result and the completion result are used as inputs for the subsequent joint optimization process.

[0128] Step S6: Calculate the prediction mean square error loss function and the statistical characteristic difference function, construct a joint optimization objective function, and perform collaborative optimization on the channel generation model parameters and the deep learning prediction model parameters.

[0129] Specifically, in this embodiment, step S6 includes:

[0130] S601. Calculate the prediction error loss function based on the predicted channel results and the actual target channel results. This function is used to measure the channel prediction accuracy of the deep learning prediction model. Its expression is as follows:

[0131]

[0132] Where N represents the number of samples, This represents the i-th real target channel sample. This represents the corresponding predicted channel sample.

[0133] S602. Construct a statistical characteristic difference function based on the difference between the statistical characteristics of the generated channel and the target channel. This function is used to measure the consistency of the statistical distribution between the generated channel and the real channel. Its expression is as follows:

[0134]

[0135] in, Represents the target channel statistical characteristic vector. This represents the generation of a channel statistical characteristic vector.

[0136] S603. Construct a joint optimization objective function based on the prediction error loss function and the statistical characteristic difference function, the expression of which is as follows:

[0137]

[0138] in, This represents the statistical characteristic constraint weight coefficient, which is set to 0.3 in this embodiment.

[0139] Furthermore, the Adam optimizer was used to update the network parameters during the training of the prediction model, with the learning rate set to 0.001 and the number of training epochs set to 100.

[0140] S604. The Covariance Matrix Adaptive Evolution Strategy (CMA-ES) algorithm is used to jointly optimize the parameters of the channel generation model and the deep learning prediction model. In each round of optimization, the loss value is calculated according to the joint optimization objective function, and the large-scale parameters, small-scale parameters, and prediction model weights are iteratively updated until the joint loss function converges or the maximum number of iterations is reached, thereby achieving joint optimization of the statistical characteristics of the generated channel and the predicted channel results.

[0141] Step S7: Input the jointly optimized parameter configuration into the channel generation model and the deep learning prediction model, output the predicted channel results and statistical characteristic matching results under the target scene and target frequency band, and construct a closed-loop optimization mechanism.

[0142] Specifically, in this embodiment, step S7 includes:

[0143] S701. Input the large-scale parameters, small-scale parameters and deep learning prediction model weights obtained by joint optimization into the channel generation module and the channel prediction module to generate the corresponding simulated channel and predicted channel results under the outdoor urban micro-cell millimeter wave propagation scenario and the 28 GHz working frequency band conditions.

[0144] S702. Extract statistical characteristics such as root mean square delay spread, root mean square angle spread, Doppler spread, and time delay power spectral density based on the channel generation and prediction results, and construct the empirical cumulative distribution function corresponding to the statistical characteristics.

[0145] S703. Compare the statistical characteristics of the generated channel and the predicted channel with the statistical characteristics of the target measurement channel. Use the normalized mean square error (NMSE) and the Kolmogorov-Smirnov distance (KS distance) as performance evaluation indicators to obtain the statistical characteristic matching results and prediction performance results.

[0146] S704, according to Figures 3 to 5 The results show that the statistical characteristics generated by the method of the present invention in the millimeter-wave scenario of an outdoor urban micro-cell have a high consistency with the target measurement channel results. The normalized mean square error (NMSE) of the predicted channel reaches −20 dB and the KS distance is less than 0.05. This indicates that the present invention can not only achieve good joint matching of statistical characteristics in the time delay domain, angle domain and Doppler domain, but also effectively improve the prediction accuracy of future channel state.

[0147] S705. The optimized parameter configuration is fed back to the channel generation model and the deep learning prediction model. When the measurement data or scene parameters change, the model is incrementally fine-tuned based on a small number of newly added channel samples, forming a closed-loop adaptive optimization mechanism of "measurement-generation-prediction-optimization-feedback", thereby realizing the continuous optimization of channel modeling and prediction capabilities in the millimeter wave band.

[0148] Figure 3 This demonstrates that, under the outdoor urban microcell millimeter-wave propagation scenario set in this embodiment of the invention, the constructed millimeter-wave channel characteristic analysis platform, configured with a 28 GHz operating frequency band and a 100 MHz system bandwidth, can effectively extract and visualize the channel statistical characteristics. After processing 10,000 channel impulse response snapshots with wavelet threshold denoising, time alignment, and Z-score normalization, the platform further outputs statistical characteristics such as the delay spread cumulative distribution function, Doppler spectrum, delay power spectral density, and angular power spectrum. The results show that the platform can effectively reflect the propagation characteristics of the channel in the delay, frequency, and spatial domains under millimeter-wave scenarios, providing reliable data support and statistical characteristic benchmarks for subsequent channel generation models and deep learning prediction models.

[0149] Figure 4 This demonstrates that, in the 28 GHz outdoor urban microcell propagation environment constructed in this embodiment of the invention, the channel generation module based on the 3GPPTR 38.901 geometric random channel model can effectively simulate and generate channel statistical characteristics. By jointly configuring large-scale and small-scale parameters, and combining propagation characteristics such as multipath clusters, ray number, Doppler shift, and random phase, the platform can generate a simulated channel that conforms to the statistical laws of the target scenario, and output results such as the time delay power spectrum and the time delay spread cumulative distribution function. The results show that the simulated channel generated by this invention has high consistency with the target measurement channel in terms of time delay spread statistical characteristics, and can well reflect the multipath propagation characteristics and channel fading laws in millimeter-wave scenarios, verifying the effectiveness of this invention in channel generation and statistical characteristic fitting.

[0150] Figure 5This demonstrates that, in the embodiments of the present invention, the deep learning channel prediction model constructed based on a Temporal Convolutional Network (TCN) and a Long Short-Term Memory Network (LSTM) can effectively achieve future channel state prediction and statistical characteristic verification. By using 10 consecutive historical snapshots as input and combining them with scene condition information for training, the platform can output performance indicators such as capacity verification results, PSD verification results, group RMSE, delay spread CDF, and Doppler spectrum. The results show that the statistical characteristic results predicted by the method of the present invention have high consistency with the actual measured channel results, with the predicted channel normalized mean square error (NMSE) reaching −20 dB and the KS test distance less than 0.05. Compared with traditional prediction methods, the present invention can effectively reduce prediction errors, improve the accuracy of future channel state prediction in complex millimeter-wave scenarios, and verify the effectiveness and practical value of the present invention in intelligent channel prediction and closed-loop optimization in multiple scenarios.

[0151] In summary, this invention achieves high-precision unification of channel modeling and prediction by constructing a unified modeling framework that integrates channel characteristic analysis, channel generation, and artificial intelligence prediction, and by introducing a joint optimization mechanism based on multi-domain statistical characteristic constraints. This method not only significantly improves channel prediction accuracy and statistical characteristic matching capability, but also enables adaptive updating of model parameters and cross-scenario generalization. It solves the problems of parameter dependence on experience, low optimization efficiency, and insufficient generalization capability in traditional methods, demonstrating good engineering application value and promising prospects for widespread adoption.

[0152] Any aspects of this invention not described in detail are well-known to those skilled in the art.

[0153] 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 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 fall within the protection scope defined by the claims of the present invention.

Claims

1. A millimeter-wave intelligent predictive channel modeling method based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Determine the target propagation scenario, operating frequency band, and system configuration parameters, and construct a unified channel modeling environment; Step S2: Acquire channel data in the target scenario and perform preprocessing; Step S3: Extract the channel statistical characteristics of the preprocessed data and construct the channel feature sequence; Step S4: Construct a channel generation model based on the geometric random channel model and generate a simulated channel; Step S5: Construct an artificial intelligence-based channel prediction model. The model takes channel feature sequences and scenario conditions as input to predict future channel states and complete missing data. Step S6: Construct a joint optimization objective function based on the difference between prediction error and channel statistical characteristics, and perform collaborative optimization on the channel generation model and prediction model; Step S7: Output the optimized channel results and perform closed-loop update.

2. The millimeter-wave intelligent predictive channel modeling method based on artificial intelligence according to claim 1, characterized in that, Step S1 specifically includes: S101. Select a propagation scenario, wherein the propagation scenario includes one of the following: indoor, outdoor, air-to-ground, space-to-ground, or sea-based scenarios; S102, select millimeter wave band, also compatible with Sub-6GHz or terahertz band; S103. Configure system parameters, including bandwidth, antenna array structure, antenna height, transceiver position relationship and propagation conditions; S104. Construct a channel modeling environment based on the propagation scenario and system parameters.

3. The millimeter-wave intelligent predictive channel modeling method based on artificial intelligence according to claim 1, characterized in that, Step S2 specifically includes: S201. Obtain channel impulse response data or channel transfer function data; S202. Preprocess the channel data, including noise reduction, normalization, and time or frequency alignment. The denoising process employs a wavelet threshold denoising method to suppress high-frequency noise components in the channel data. The normalization process uses the Z-score normalization method, whose expression is: ； Where H represents the original channel data, and X represents the normalized channel data. This represents the mean of the original channel data H. This represents the standard deviation of the raw channel data H; The time alignment process adopts an alignment method based on the maximum correlation peak. By calculating the cross-correlation function of the channel response between different snapshots, the position corresponding to the maximum correlation peak is determined, and time synchronization alignment is completed accordingly. Frequency alignment processing achieves frequency domain unification by compensating for frequency offsets in the channel responses at different frequency points; S203. Organize the processed data into standardized channel input data.

4. The millimeter-wave intelligent predictive channel modeling method based on artificial intelligence according to claim 1, characterized in that, Step S3 specifically includes: S301. Extract statistical characteristics such as delay spread, angle spread, and Doppler spread from the preprocessed channel data; S302. Construct the channel feature vector, which is expressed as follows: ； in, A vector representing the statistical characteristics of the target channel; This represents the preprocessed channel data; This represents the channel feature extraction function; Indicates delay spread characteristics; Indicates angular extension features; Indicates Doppler extended features; S303. Use the channel feature vector as input for the subsequent model.

5. The millimeter-wave intelligent predictive channel modeling method based on artificial intelligence according to claim 1, characterized in that, Step S4 specifically includes: S401. Construct a channel generation module based on a geometric random channel model; ； in, Indicates time and latency Channel impulse response under the following conditions; Indicates the number of multipath components; Indicates the first The amplitude coefficient of the path; Indicates the first The phase of the path; Indicates the first The Doppler frequency shift corresponding to each path; Indicates the first The propagation delay of each path; Represents the Dirac impulse function; Represents the imaginary unit; S402. Define the channel model parameter vector, which is expressed as follows: ； in, Represents the channel model parameter vector; Represents a large-scale parameter vector used to describe the statistical characteristics of time delay spread, angle spread, and shadow fading; Represents a small-scale parameter vector used to describe the propagation characteristics of multipath components, such as path power, phase, Doppler shift, and time delay distribution; superscript This represents the matrix transpose operation; S403. Generate the channel matrix based on the parameter vector, which is expressed as follows: ； in, This indicates large-scale decay. This represents small-scale fading, which is generated by a multipath propagation model. S404. Extract the statistical characteristics of delay spread, angle spread, and Doppler spread from the generated channel and match them with the corresponding statistical characteristics of the target channel for subsequent joint optimization objective function construction and model parameter update.

6. The millimeter-wave intelligent predictive channel modeling method based on artificial intelligence according to claim 1, characterized in that, Step S5 specifically includes: S501. Construct a channel prediction model based on deep learning. The channel prediction model includes a temporal convolutional network (TCN) and a long short-term memory network (LSTM). The TCN is used to extract local temporal features from historical channel sequences, and the LSTM is used to learn the long-term correlation characteristics of channel changes over time. The output features of the TCN and the LSTM are fused to obtain the output result of the channel prediction model. S502. Channel prediction based on historical channel sequences, expressed as follows: ； in, This represents the predicted future time-series channel matrix; The parameter is Channel prediction model; A sequence of channel features representing historical moments; Indicates scene condition parameters; Indicates the parameters of the channel prediction model; S503. Based on the predicted channel results output by the channel prediction model, the missing channel data is completed. By utilizing the temporal correlation between channel features at adjacent times, the channel state information corresponding to the missing position is restored. S504. The predicted channel matrix and its statistical characteristics are used as inputs to the joint optimization objective function to calculate the difference between the prediction error and the statistical characteristics, and the channel generation model parameters and channel prediction model parameters are updated according to the optimization results.

7. The millimeter-wave intelligent predictive channel modeling method based on artificial intelligence according to claim 1, characterized in that, Step S6 specifically includes: S601. Construct the prediction error loss function, which is expressed as follows: ； in, This represents the prediction error loss function; Indicates the parameters of the channel prediction model; Indicates the number of training samples; Indicates the first A real target channel matrix; Indicates the first One prediction channel matrix; The squared Frobenius norm of a matrix is ​​equivalent to the sum of the squares of the moduli of all its elements; S602. Construct a channel statistical characteristic constraint function to describe the difference in statistical characteristics between the generated channel and the target channel, expressed as follows: ； in, A function representing statistical differences; A vector representing the statistical characteristics of the target channel; This represents the statistical characteristic vector of the generated channel; Represents the L2 norm; S603. Construct the joint optimization objective function, which is expressed as follows: ； in, This represents the statistical characteristic vector of the target channel. For parameter space; For prediction error loss; These are the weighting coefficients; This is a constraint function for statistical characteristic differences; and These are the statistical feature vectors of the target channel and the predicted channel, respectively. S604. Update model parameters based on joint optimization objective function.

8. The millimeter-wave intelligent predictive channel modeling method based on artificial intelligence according to claim 1, characterized in that, Step S7 specifically includes: S701. Output the predicted channel matrix and corresponding statistical characteristic parameters, including delay spread, angle spread and Doppler spread. S702. The performance of the prediction results is evaluated based on the normalized mean square error (NMSE) and the KS test distance. The normalized mean square error is used to evaluate the error between the predicted channel and the target channel, and the KS test distance is used to evaluate the consistency of the statistical distribution between the generated channel and the target channel. S703. Based on the performance evaluation results, the joint optimization objective function is iteratively updated, and the optimized model parameters are fed back to the channel generation model and the channel prediction model to update the channel generation parameters and prediction model parameters. S704. When the input channel data or scenario conditions change, the channel generation and channel prediction process is re-executed based on the updated model parameters, forming a closed-loop optimization mechanism for channel generation and prediction.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images