An artificial intelligence-based ubiquitous channel model adaptive parameter optimization method

Abstract

The application discloses a kind of based on artificial intelligence's ubiquitous channel model adaptive parameter optimization method, it is related to wireless channel modeling technical field, comprising: 1) the parameter optimization environment for ubiquitous channel model is constructed, and channel model parameter is formed into the parameter vector to be optimized;2) the multi-domain channel statistical characteristics in channel measurement data are extracted as optimization target;3) the parameter optimization process is modeled as Markov decision process, and state, action and reward mechanism are designed;4) using deep reinforcement learning algorithm iteratively updates parameter, obtains the optimized channel model parameter configuration applicable to different scenarios.The application can improve the fitting accuracy of ubiquitous channel model to the statistical characteristics of multiple scenarios and the parameter optimization efficiency, and can be used for 6G wireless communication system simulation, performance evaluation and network optimization.

Description

Technical Field

[0001] This invention relates to the field of channel modeling technology, and in particular to an adaptive parameter optimization method for universal channel models based on artificial intelligence. Background Technology

[0002] In recent years, the development requirements of 6G wireless communication systems for full-band, full-coverage, and diversified service scenarios have been continuously increasing. Communication scenarios have gradually expanded from traditional terrestrial scenarios to integrated air-space-ground-sea coverage. Channel modeling, as an important foundation for wireless communication system design, performance evaluation, network optimization, and standardization research, plays an increasingly prominent role. Existing standardized or quasi-standardized channel models, such as 3GPP TR 38.901, IMT-2030, QuaDRiGa, and WINNER II, still have certain limitations in terms of frequency band coverage and scenario coverage, making it difficult to fully meet the requirements of 6G for unified modeling across multiple frequency bands and scenarios. To address these issues, existing research has proposed the 6G universal channel model 6GPCM, which provides a unified modeling framework for different frequency bands and global coverage scenarios.

[0003] Most existing channel models are based on the framework of geometrically random channel models. The proper configuration of model parameters directly affects the fitting accuracy of channel statistical characteristics and the applicability of the model. In general channel models, parameters typically include large-scale and small-scale parameters. Large-scale parameters describe statistical distribution characteristics such as delay spread, angle spread, and Rician K-factor, while small-scale parameters describe detailed characteristics such as cluster delay structure, cluster power fluctuations, cluster number, and the number of rays per cluster. Due to the strong coupling between large-scale and small-scale parameters, and the significant differences in parameter sensitivity across different frequency bands and scenarios, parameter optimization exhibits high dimensionality, complexity, and scenario-dependent characteristics. Especially in emerging propagation scenarios such as satellite-to-ground, air-to-ground, and maritime propagation, channel statistical characteristics vary significantly with satellite elevation angle, flight altitude, horizontal distance, and the sea surface propagation environment, further increasing the difficulty of parameter optimization.

[0004] Existing parameter configuration methods typically rely on channel measurement results, obtaining model parameters through multipath or scattering cluster extraction, statistical characteristic fitting, and manual parameter tuning. While these methods can improve the consistency between simulated and measured channels to some extent, they suffer from problems such as complex parameter extraction processes, heavy reliance on human experience, low optimization efficiency, and difficulty in obtaining globally optimal solutions in high-dimensional parameter spaces. Furthermore, existing methods often struggle to simultaneously consider the joint matching of statistical characteristics in the time delay and angle domains during parameter optimization, resulting in parameter configurations that perform well on a single statistical metric but still fall short in multi-domain joint fitting.

[0005] On the other hand, artificial intelligence (AI) technology has been gradually applied to wireless channel prediction and channel modeling in recent years. It can utilize data to learn the implicit relationships between channel characteristics and environmental factors, improving modeling and prediction capabilities under complex propagation conditions. However, existing AI methods are mostly focused on predictive channel modeling, channel sample generation, or specific module optimization. Directly applying AI to the optimization of universal channel model parameters, especially for the joint optimization of large-scale and small-scale parameters under multi-band and multi-scenario conditions, remains relatively unexplored. Meanwhile, traditional heuristic search methods or grid search methods have high computational overhead in high-dimensional mixed parameter spaces, making it difficult to balance optimization accuracy and efficiency. Existing universal channel models under multi-band and multi-scenario conditions suffer from problems such as parameter configuration dependence on experience, low optimization efficiency, and difficulty in simultaneously considering the joint matching of statistical characteristics across multiple domains.

[0006] Therefore, there is an urgent need to propose a new universal channel model parameter optimization method. Taking the statistical characteristics of channel measurement as the target, this method can efficiently optimize the parameters of universal channel models in multiple frequency bands and scenarios while ensuring the physical constraints of the model. This will improve the matching accuracy between the simulated channel and the measured channel in the time delay domain and angle domain statistical characteristics, thereby better serving the simulation, performance evaluation and network optimization of 6G wireless communication systems. Summary of the Invention

[0007] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide an adaptive parameter optimization method for universal channel models based on artificial intelligence. This invention addresses the problems of strong coupling between large-scale and small-scale parameters, high parameter space dimension, and difficulty in optimizing mixed continuous and discrete parameters in universal channel models. It utilizes deep reinforcement learning to establish an adaptive optimization mechanism between parameters and channel statistical characteristics, and realizes automatic parameter updates within the physical constraints of the model. This improves the matching accuracy between simulated and measured channels in terms of statistical characteristics such as root mean square delay spread and root mean square angle spread, and enhances the accuracy and efficiency of parameter optimization for universal channel models in multiple frequency bands and scenarios.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] According to the present invention, a universal channel model adaptive parameter optimization method based on artificial intelligence includes:

[0010] Step S1: Determine the propagation scenario, frequency band, and system configuration corresponding to the universal channel model parameter optimization, and construct the universal channel model parameter optimization environment based on the propagation scenario, frequency band, and system configuration;

[0011] Step S2: Based on the universal channel model parameter optimization environment constructed in step S1, preprocess the channel measurement data in the propagation scenario, and extract the measurement channel statistical characteristics as the parameter optimization target from the preprocessed channel measurement data.

[0012] Step S3: Construct the parameter vector of the channel model to be optimized for the universal channel model. The parameter vector of the channel model to be optimized includes a large-scale parameter vector and a small-scale parameter vector. After inputting the parameter vector of the channel model to be optimized into the universal channel model, the simulated channel statistical characteristics are obtained. Set the value range for the large-scale parameter vector and the small-scale parameter vector.

[0013] Step S4: Model the parameter optimization process as a Markov decision process, and design the state space, action space, and error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics based on the value range of each large-scale parameter vector and small-scale parameter vector.

[0014] Step S5: Based on the measured channel statistical characteristics extracted in step S2 as the parameter optimization target, the simulated channel statistical characteristics obtained in step S3, and the state space, action space, and error vector in step S4, construct a reward function to evaluate the degree of matching between the simulated channel statistical characteristics and the measured channel statistical characteristics.

[0015] Step S6: Based on the Markov decision process established in Step S4 and the reward function constructed in Step S5, a deep reinforcement learning algorithm is used to iteratively update the parameter vector of the channel model to be optimized, so as to obtain the optimized channel model parameter vector.

[0016] As a further optimization scheme of the universal channel model adaptive parameter optimization method based on artificial intelligence described in this invention, step S1 includes:

[0017] S101. Select a propagation scenario and a carrier frequency band as the parameter optimization object. The propagation scenario includes one or more of the following: indoor, outdoor, satellite-to-ground, air-to-ground, and maritime scenarios.

[0018] S102. Determine the system configuration parameters under the propagation scenario. The system configuration parameters include bandwidth, antenna array type of the transmitting and receiving ends, antenna height, motion trajectory, and line-of-sight or non-line-of-sight propagation conditions.

[0019] S103. Construct a universal channel model parameter optimization environment based on propagation scenarios and system configuration parameters.

[0020] As a further optimization scheme of the universal channel model adaptive parameter optimization method based on artificial intelligence described in this invention, step S2 includes:

[0021] S201. Based on the propagation scenario, frequency band, and system configuration parameters determined in step S1, obtain channel measurement data and perform noise reduction, normalization, and statistical processing.

[0022] S202. Calculate the delay power spectral density and angle power spectral density based on the channel measurement data, and extract the delay spread DS and angle spread AS samples.

[0023] S203. Construct the cumulative distribution function of the delay-spread DS sample and the angle-spread AS sample to obtain the measurement channel statistical characteristics as the parameter optimization target.

[0024] As a further optimization scheme of the universal channel model adaptive parameter optimization method based on artificial intelligence described in this invention, step S3 includes:

[0025] S301. Define the parameter vector of the channel model to be optimized, specifically including: Based on the prior knowledge of the Geometric Random Channel Model (GBSM), define the parameter vector of the channel model to be optimized as follows:

[0026]

[0027] in, and These are the large-scale parameter vector and the small-scale parameter vector, respectively. Let be the parameter vector of the channel model to be optimized. It is a transpose operation;

[0028] S302. Define the large-scale parameter vector and the small-scale parameter vector; the large-scale parameter vector and the small-scale parameter vector together constitute the parameter vector of the channel model to be optimized. The large-scale parameter vector and the small-scale parameter vector are defined as follows:

[0029]

[0030]

[0031] Among them, in the large-scale parameter vector, and These represent the mean and standard deviation of the logarithmic field of the delay spread, respectively. and These represent the mean and standard deviation of the logarithmic domain with angular extension, respectively. and Let represent the mean and standard deviation of the Rician K factor, respectively; in the small-scale parameter vector, This is the time delay distribution scaling factor. The standard deviation of cluster shadow fading. The number of scattering clusters, Define the number of rays within each cluster; and set the value ranges for the large-scale parameter vector and the small-scale parameter vector.

[0032] As a further optimization scheme of the universal channel model adaptive parameter optimization method based on artificial intelligence described in this invention, step S4 includes:

[0033] S401. Model the parameter optimization process as a Markov decision process; a Markov decision process is defined as... ;in, Representing the state space, Represents the action space, Represents the state transition probability. Represent the space of reward functions;

[0034] S402. Design the state space; the state space consists of state vectors; the state vectors include large-scale parameter vectors, small-scale parameter vectors, and the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics; the state vectors are defined as follows:

[0035]

[0036] in, This represents the number of interaction steps between the agent and the universal channel model parameter optimization environment. Indicates the first A state vector with 1 interaction step. Indicates the first A large-scale parameter vector of interaction steps. Indicates the first A small-scale parameter vector of interaction steps. This represents the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics. Indicates the first The error vector for each interaction step;

[0037] S403. Design the motion space; the motion space consists of motion vectors; motion vectors, specifically, include those used for... The parameter vector of the channel model to be optimized is incrementally updated for each interaction step; the action vector is defined as follows:

[0038]

[0039] in, Indicates the first The action vector of each interaction step is used to update the parameter vector of the channel model to be optimized. The update method is as follows:

[0040]

[0041] in, Indicates the first A channel model parameter vector with 1 interaction step. This represents the maximum value vector of the channel model parameter vector update. It represents the Hadamardi (or Hadama) stack;

[0042] S404. Design the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics. Specifically, the error vector comprises the normalized root mean square error (NRMSE) between the cumulative distribution functions (CDF) of the simulated and measured channel statistical characteristics and the Kolmogorov-Smirnov (KS) distance, used to characterize the degree of matching of the current parameter configuration. The NRMSE and KS distance are defined as follows:

[0043]

[0044] ;

[0045] in, This represents the channel characteristics, specifically the delay spread (DS) and angle spread (AS). and Let the cumulative distribution function (CDF) represent the channel statistical characteristics measured and simulated, respectively. The NRMSE represents the relationship between simulated channel statistical characteristics and measured channel statistical characteristics. The KS distance represents the distance between the simulated channel statistical characteristics and the measured channel statistical characteristics. This represents the domain of the cumulative distribution function. Given the mathematical expectation, therefore, the error vector The definition is as follows:

[0046]

[0047] in, To improve the NRMSE between the delay spread CDF of the simulated channel and the measurement channel, To measure the delay spread (CDF) of the channel, To simulate the delay spread (CDF) of the channel, The KS distance between the CDF delay spread of the simulated channel and the measurement channel is calculated. To extend the NRMSE between the simulated channel and the measurement channel angles of the CDF, To measure the angle spread CDF of the channel, To simulate the angle spread CDF of the channel, The KS distance between the CDF is extended to the angle of the simulated channel and the measurement channel.

[0048] As a further optimization scheme of the universal channel model adaptive parameter optimization method based on artificial intelligence described in this invention, step S5 includes:

[0049] S501, Successful Construction Rewards;

[0050] A positive reward is given to the agent when the error vector is less than a preset success threshold, and the reward decays as the number of interaction steps increases; the success reward term is defined as follows:

[0051]

[0052]

[0053] in, Indicates the number of interaction steps Successful rewards at that time It is a preset success threshold vector. Indicates the number of interaction steps Successful reward events, It is the success reward coefficient. It is the maximum number of interaction steps in each decision round;

[0054] S502. Construct a penalty term, including imposing a penalty when the error vector exceeds a preset penalty judgment threshold; the penalty term is defined as follows:

[0055]

[0056] in, Indicates the number of interaction steps The penalty items at that time Indicates the number of interaction steps Error at time, and These are the penalty coefficients for the error vector and the number of interaction steps, respectively. It is a preset penalty judgment threshold;

[0057] S503. Construct parameter-guided reward items, which are defined as follows:

[0058]

[0059] in, Indicates the first The parameter for each interaction step guides the reward item. Indicates the first Error in the number of interaction steps, It is the error reduction guiding coefficient. and These are the optimization guiding coefficients for the large-scale parameter vector and the small-scale parameter vector, respectively. and They are respectively in the 1st The action vector consists of a large-scale parameter vector and a small-scale parameter vector for each interaction step;

[0060] S504. Construct the reward function, in the... Reward function for each interaction step The definition is as follows:

[0061] .

[0062] As a further optimization scheme of the universal channel model adaptive parameter optimization method based on artificial intelligence described in this invention, step S6 includes:

[0063] S601. A deep reinforcement learning agent based on the soft actor critic algorithm (SAC) is used for training. The deep reinforcement learning agent includes a policy network, a dual critic network, a target network, and an experience replay cache.

[0064] S602. In each training round, the policy network determines the training outcome based on the first... The state vector of each interaction step and action vectors Update the parameter vector of the channel model to be optimized, and calculate the first... The reward function value for each interaction step And store the updated state vector. To the experience replay cache;

[0065] S603. Randomly extract batch samples from the experience replay buffer, iteratively update the dual critic network and policy network, and synchronously update the target network; after the soft actor critic algorithm is trained until the channel model parameter vector to be optimized converges, output the optimized channel model parameter vector.

[0066] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0067] First, this invention uses channel measurement statistical characteristics as the optimization objective, incorporating the matching results of time delay and angle domain statistical characteristics into the parameter optimization process. This improves the fitting accuracy of the universal channel model to the channel statistical characteristics of the target scenario. Second, this invention models the universal channel model parameter optimization process as a Markov decision process and employs deep reinforcement learning to jointly optimize large-scale and small-scale parameters. This effectively reduces the computational burden of traditional manual parameter tuning and grid search methods in high-dimensional parameter spaces, improving parameter optimization efficiency. Third, this invention is applicable to various propagation scenarios, including indoor, outdoor, satellite-to-ground, air-to-ground, and maritime environments, as well as multiple frequency bands such as sub-6 GHz, centimeter waves, and millimeter waves, demonstrating good scenario expansion and frequency band adaptability. Finally, the optimized parameter configuration obtained by this invention enables the simulated channel generated by the universal channel model to more closely approximate the actual measured channel, providing effective support for 6G wireless communication system simulation, performance evaluation, network optimization, and channel modeling research, and possessing significant application value. Attached Figure Description

[0068] Figure 1 This is a flowchart of an adaptive parameter optimization method for a universal channel model based on artificial intelligence.

[0069] Figure 2 This is a general framework diagram of a universal channel model parameter optimization method based on artificial intelligence;

[0070] Figure 3 This is a framework diagram of a universal channel model parameter optimization algorithm based on deep reinforcement learning.

[0071] Figure 4 This is a diagram showing the matching results of the statistical characteristics of latency spread in an outdoor scenario according to an embodiment of the present invention;

[0072] Figure 5 This is a diagram showing the matching results of angle extension statistical characteristics in an outdoor scene according to an embodiment of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0074] Example 1

[0075] Step S1: Determine the propagation scenario, frequency band, and system configuration corresponding to the universal channel model parameter optimization, and construct the universal channel model parameter optimization environment based on the propagation scenario, frequency band, and system configuration parameters.

[0076] Specifically, in this embodiment, step S1 includes:

[0077] S101. Select the outdoor ultra-large-scale MIMO line-of-sight propagation scenario as the target for optimization of universal channel model parameters, and set the carrier frequency to 5.3 GHz.

[0078] S102. Determine the system configuration parameters, where the bandwidth is set to 160 MHz, the transmitter uses 8 omnidirectional antennas, the receiver uses a 4×32 uniform linear array, and the antenna heights of the transmitter and receiver are 1.5 m and 20 m, respectively.

[0079] S103. Based on the above propagation scenario and system configuration parameters, construct a universal channel model parameter optimization environment. The receiving antenna is installed at a high-level site, with a horizontal distance of approximately 130 m from the measurement route.

[0080] Step S2: Based on the propagation scenario, frequency band, and system configuration parameters determined in step S1, preprocess the channel measurement data under the propagation scenario and extract the measurement channel statistical characteristics as the parameter optimization target.

[0081] Specifically, in this embodiment, step S2 includes:

[0082] S201. Based on the outdoor ultra-large-scale MIMO line-of-sight propagation scenario, 5.3 GHz carrier frequency band, and corresponding system configuration parameters determined in step S1, obtain the channel measurement data under this condition, and perform noise reduction, normalization, and statistical processing.

[0083] S202. Calculate the delay power spectral density and angle power spectral density based on the channel measurement data, and extract the delay spread (DS) and angle spread (AS) samples.

[0084] S203. Construct the cumulative distribution function (CDF) of the delay-spread DS sample and the angle-spread AS sample to obtain the measurement channel statistical characteristics as the parameter optimization target.

[0085] Step S3: Construct the channel model parameter vector to be optimized for the universal channel model. The channel model parameter vector to be optimized includes large-scale parameters and small-scale parameters, and set a range of adjustable values ​​for each parameter.

[0086] Specifically, in this embodiment, step S3 includes:

[0087] S301. Define the parameter vector of the channel model to be optimized. The parameter vector of the channel model to be optimized is defined as follows:

[0088]

[0089] in, and These are the large-scale parameter vector and the small-scale parameter vector, respectively. Let be the parameter vector of the channel model to be optimized. It is a transpose operation.

[0090] S302. Define the large-scale parameter vector and the small-scale parameter vector. The large-scale parameter vector and the small-scale parameter vector together constitute the parameter vector of the channel model to be optimized. Specifically, the large-scale parameter vector and the small-scale parameter vector are defined as follows:

[0091]

[0092]

[0093] Among them, in the large-scale parameter vector, and These represent the mean and standard deviation of the logarithmic field of the delay spread, respectively. and These represent the mean and standard deviation of the logarithmic domain with angular extension, respectively. and Let represent the mean and standard deviation of the Rician K factor, respectively. In the small-scale parameter vector, This is the time delay distribution scaling factor. The standard deviation of cluster shadow fading. The number of scattering clusters, This defines the number of rays within each cluster. It also sets the range of values ​​that can be retrieved for both the large-scale and small-scale parameter vectors.

[0094] Step S4: Model the parameter optimization process as a Markov decision process, and design the state space, action space, and error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics according to the range of values ​​that each parameter can take.

[0095] Specifically, in this embodiment, step S4 includes:

[0096] S401. Model the parameter optimization process as a Markov decision process; a Markov decision process is defined as... .in, Representing the state space, Represents the action space, Represents the state transition probability. This represents the reward function space.

[0097] S402. Construct the state vector, which includes the large-scale parameters, small-scale parameters, and the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics at the current moment. The state vector is defined as follows:

[0098]

[0099] in, This represents the number of interaction steps between the agent and the universal channel model parameter optimization environment. Indicates the first A state vector with 1 interaction step. Indicates the first LSP vectors with interaction steps, Indicates the first An SSP vector with a number of interaction steps. This represents the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics. Indicates the first The error vector for each interaction step;

[0100] S403. Design the motion space; the motion space is composed of motion vectors. Specifically, the motion vectors include those used for... The parameter vector of the channel model to be optimized is incrementally updated for each interaction step. The action vector is defined as follows:

[0101]

[0102] in, Indicates the first The action vector of each interaction step is used to update the parameter vector of the channel model to be optimized. The update method is as follows:

[0103]

[0104] in, Indicates the first A channel model parameter vector with 1 interaction step. This represents the maximum value vector of the channel model parameter vector update. It represents the Hadamardi (or Hadama) stack;

[0105] S404. Design the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics. Specifically, the error vector comprises the normalized root mean square error (NRMSE) between the cumulative distribution functions (CDF) of the simulated and measured channel statistical characteristics and the Kolmogorov-Smirnov (KS) distance, used to characterize the degree of matching of the current parameter configuration. The NRMSE and KS distance are defined as follows:

[0106]

[0107]

[0108] in, This represents the channel characteristics, specifically the delay spread (DS) and angle spread (AS). and Let the cumulative distribution function (CDF) represent the channel statistical characteristics measured and simulated, respectively. The NRMSE represents the relationship between simulated channel statistical characteristics and measured channel statistical characteristics. The KS distance represents the distance between the simulated channel statistical characteristics and the measured channel statistical characteristics. This represents the domain of the cumulative distribution function. Given the mathematical expectation, therefore, the error vector The definition is as follows:

[0109]

[0110] in, To improve the NRMSE between the delay spread CDF of the simulated channel and the measurement channel, To measure the delay spread (CDF) of the channel, To simulate the delay spread (CDF) of the channel, The KS distance between the CDF delay spread of the simulated channel and the measurement channel is calculated. To extend the NRMSE between the simulated channel and the measurement channel angles of the CDF, To measure the angle spread CDF of the channel, To simulate the angle spread CDF of the channel, The KS distance between the CDF is extended to the angle of the simulated channel and the measurement channel.

[0111] Step S5: Based on the measured channel statistical characteristics extracted in step S2 as the parameter optimization target, the simulated channel statistical characteristics obtained in step S3, and the state space, action space, and error vector in step S4, construct a reward function to evaluate the degree of matching between the simulated channel statistical characteristics and the measured channel statistical characteristics.

[0112] Specifically, in this embodiment, step S5 includes:

[0113] S501. Construct a success reward program. The success reward program includes a positive reward given to the agent when the error vector is less than a preset success threshold, which decays with the number of interaction steps. The success reward program is defined as follows:

[0114]

[0115]

[0116] in, Indicates the first Successful interaction steps reward items It is a preset success threshold vector. Indicates the first A successful reward event for each interactive step. It is the success reward coefficient. This is the maximum number of interaction steps in each decision round. In this embodiment, Set to [0.05, 0.10, 0.05, 0.10]. Set to 100, Set it to 100.

[0117] S502. Construct a penalty term, including imposing a penalty when the error vector exceeds a preset penalty judgment threshold; the penalty term is defined as follows:

[0118]

[0119] in, Indicates the number of interaction steps The penalty items at that time Indicates the first Error in the number of interaction steps, and These are the penalty coefficients for the error vector and the number of interaction steps, respectively. This is a preset penalty threshold. In this embodiment, and Set them to 15 and 0.5 respectively. Set it to 0.2.

[0120] S503. Construct parameter-guided reward items, including the following definitions:

[0121]

[0122] in, Indicates the first The parameter for each interaction step guides the reward item. Indicates the first Error in the number of interaction steps, It is the error reduction guiding coefficient. and These are the optimized guiding coefficients for the LSP vector and the SSP vector, respectively. and They are in the number of interaction steps The action vectors of the LSP vector and SSP vector at that time. In this embodiment, Set to 0.5. and Set them to 1 and 0.5 respectively.

[0123] S504. Construct the reward function, including, in the... Reward function for each interaction step The definition is as follows:

[0124]

[0125] Step S6: Based on the Markov decision process established in step S4 and the reward function constructed in step S5, a deep reinforcement learning algorithm is used to iteratively update the parameter vector to be optimized, and the optimized parameter vector is obtained.

[0126] Specifically, in this embodiment, step S6 includes:

[0127] S601. A deep reinforcement learning agent based on the soft actor critic algorithm (SAC) is used for training. The agent includes a policy network, a dual critic network, a target network, and an experience replay cache.

[0128] S602. In each training round, the policy network determines the training outcome based on the first... The state vector of each interaction step and action vectors Update the parameter vector of the channel model to be optimized, and calculate the first... The reward function value for each interaction step And store the updated state vector. Added to the experience replay cache.

[0129] S603. Randomly select a small batch of samples from the experience replay cache to iteratively update the dual critic network, policy network, and temperature parameters, and simultaneously update the target network. In this embodiment, the number of hidden layers in the network is set to 4, the number of hidden units in each layer is 256, the batch size is 256, and the experience replay cache size is... The learning rate is After training, the optimized channel model parameter vector is output.

[0130] Figure 1 This is a flowchart of an adaptive parameter optimization method for a universal channel model based on artificial intelligence. For this invention, the propagation scenario and system configuration for parameter optimization are first determined, and a universal channel model parameter optimization environment is constructed. Next, the channel measurement data for this scenario is preprocessed to extract the measured channel statistical characteristics. A parameter vector for the channel model to be optimized is constructed, and the range of parameter vector values ​​is set. The parameter optimization process is modeled as a Markov decision process, designing the state space, action space, and error vector. Then, a reward function is constructed to evaluate the degree of matching between the simulated and measured channel statistical characteristics, and a deep reinforcement learning algorithm is used to iteratively update the parameter vector to be optimized. Finally, the error between the simulated and measured channel statistical characteristics is compared, and the optimized channel model parameters for the current propagation scenario are output.

[0131] Figure 2This is a general framework diagram of a universal channel model parameter optimization method based on artificial intelligence. First, the propagation scenario and the corresponding channel model parameters to be optimized are set. Then, the channel model parameters are iteratively updated through interaction between an intelligent agent algorithm and the universal channel model parameter optimization environment. Finally, the optimized channel model parameters are output, and simulated channel statistical characteristics are generated using the universal channel model.

[0132] Figure 3 This is a schematic diagram illustrating the framework of a universal channel model parameter optimization algorithm based on deep reinforcement learning. The scenario is set as a universal channel model parameter optimization environment providing the channel model parameters to be optimized, with errors evaluated using KS-distance and NRMSE. The agent employs the SAC algorithm; the Actor network outputs the action policy, and the Critic network, through a dual-Q network and the target network, performs value evaluation and parameter updates. Markov decision processes are stored in an experience replay pool, completing closed-loop interactive training between the agent and the universal channel model parameter optimization environment.

[0133] Figure 4 The results show that, in outdoor scenarios, after inputting the optimized parameter configuration obtained by the method of this invention into the universal channel model, the generated simulated channel most closely matches the measured channel results in terms of root mean square delay spread statistical characteristics. Compared with 3GPP default parameter configuration, grid search, PPO, and TD3 methods, the delay spread cumulative distribution function obtained by the method of this invention has higher consistency with the measurement results, indicating that this invention can better achieve adaptive matching of delay domain statistical characteristics in outdoor scenarios.

[0134] Figure 5 The results show that, in outdoor scenarios, after inputting the optimized parameter configuration obtained by the method of this invention into the universal channel model, the generated simulated channel most closely resembles the measured channel results in terms of root mean square angle spread statistical characteristics. Compared with the 3GPP default parameter configuration, grid search, PPO, and TD3 methods, the angle spread cumulative distribution function obtained by the method of this invention has higher consistency with the measurement results, indicating that this invention can not only optimize the time delay domain statistical characteristics but also achieve better adaptive matching of the angle domain statistical characteristics.

[0135] In summary, the universal channel model adaptive parameter optimization method established in this invention, based on artificial intelligence, takes the statistical characteristics of channel measurement as the optimization objective, models the parameter optimization process of the universal channel model as a Markov decision process, and utilizes deep reinforcement learning to achieve joint adaptive updates of large-scale and small-scale parameters. By inputting the optimized parameter configuration into the universal channel model to generate a simulated channel, and comparing it with the measured channel in terms of statistical characteristics such as root mean square delay spread and root mean square angle spread, this invention can improve the fitting accuracy of the universal channel model to the statistical characteristics of the target scenario channel, while balancing the accuracy and efficiency of parameter optimization. It has good scenario adaptability and application value, and can be used for 6G wireless communication system simulation, performance evaluation, network optimization, and channel modeling research.

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A universal channel model adaptive parameter optimization method based on artificial intelligence, characterized in that, include: Step S1: Determine the propagation scenario, frequency band, and system configuration corresponding to the universal channel model parameter optimization, and construct the universal channel model parameter optimization environment based on the propagation scenario, frequency band, and system configuration; Step S2: Based on the universal channel model parameter optimization environment constructed in step S1, preprocess the channel measurement data in the propagation scenario, and extract the measurement channel statistical characteristics as the parameter optimization target from the preprocessed channel measurement data. Step S3: Construct the parameter vector of the channel model to be optimized for the universal channel model. The parameter vector of the channel model to be optimized includes a large-scale parameter vector and a small-scale parameter vector. After inputting the parameter vector of the channel model to be optimized into the universal channel model, the simulated channel statistical characteristics are obtained. Set the value range for the large-scale parameter vector and the small-scale parameter vector. Step S4: Model the parameter optimization process as a Markov decision process, and design the state space, action space, and error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics based on the value range of each large-scale parameter vector and small-scale parameter vector. Step S5: Based on the measured channel statistical characteristics extracted in step S2 as the parameter optimization target, the simulated channel statistical characteristics obtained in step S3, and the state space, action space, and error vector in step S4, construct a reward function to evaluate the degree of matching between the simulated channel statistical characteristics and the measured channel statistical characteristics. Step S6: Based on the Markov decision process established in Step S4 and the reward function constructed in Step S5, a deep reinforcement learning algorithm is used to iteratively update the parameter vector of the channel model to be optimized, so as to obtain the optimized channel model parameter vector.

2. The adaptive parameter optimization method for a universal channel model based on artificial intelligence according to claim 1, characterized in that, Step S1 includes: S101. Select a propagation scenario and a carrier frequency band as the parameter optimization object. The propagation scenario includes one or more of the following: indoor, outdoor, satellite-to-ground, air-to-ground, and maritime scenarios. S102. Determine the system configuration parameters under the propagation scenario. The system configuration parameters include bandwidth, antenna array type of the transmitting and receiving ends, antenna height, motion trajectory, and line-of-sight or non-line-of-sight propagation conditions. S103. Construct a universal channel model parameter optimization environment based on propagation scenarios and system configuration parameters.

3. The adaptive parameter optimization method for a universal channel model based on artificial intelligence according to claim 2, characterized in that, Step S2 includes: S201. Based on the propagation scenario, frequency band, and system configuration parameters determined in step S1, obtain channel measurement data and perform noise reduction, normalization, and statistical processing. S202. Calculate the delay power spectral density and angle power spectral density based on the channel measurement data, and extract the delay spread DS and angle spread AS samples. S203. Construct the cumulative distribution function of the delay-spread DS sample and the angle-spread AS sample to obtain the measurement channel statistical characteristics as the parameter optimization target.

4. The adaptive parameter optimization method for a universal channel model based on artificial intelligence according to claim 3, characterized in that, Step S3 includes: S301. Define the parameter vector of the channel model to be optimized, specifically including: Based on the prior knowledge of the Geometric Random Channel Model (GBSM), define the parameter vector of the channel model to be optimized as follows: ； in, and These are the large-scale parameter vector and the small-scale parameter vector, respectively. Let be the parameter vector of the channel model to be optimized. It is a transpose operation; S302. Define the large-scale parameter vector and the small-scale parameter vector; the large-scale parameter vector and the small-scale parameter vector together constitute the parameter vector of the channel model to be optimized. The large-scale parameter vector and the small-scale parameter vector are defined as follows: ； ； Among them, in the large-scale parameter vector, and These represent the mean and standard deviation of the logarithmic field of the delay spread, respectively. and These represent the mean and standard deviation of the logarithmic domain with angular extension, respectively. and Let represent the mean and standard deviation of the Rician K factor, respectively; in the small-scale parameter vector, This is the time delay distribution scaling factor. The standard deviation of cluster shadow fading. The number of scattering clusters, Define the number of rays within each cluster; and set the value ranges for the large-scale parameter vector and the small-scale parameter vector.

5. The adaptive parameter optimization method for a universal channel model based on artificial intelligence according to claim 4, characterized in that, Step S4 includes: S401. Model the parameter optimization process as a Markov decision process; a Markov decision process is defined as... ;in, Representing the state space, Represents the action space. Represents the state transition probability. Represent the space of reward functions; S402. Design the state space; the state space consists of state vectors; the state vectors include large-scale parameter vectors, small-scale parameter vectors, and the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics; the state vectors are defined as follows: ； in, This represents the number of interaction steps between the agent and the universal channel model parameter optimization environment. Indicates the first A state vector with 1 interaction step. Indicates the first A large-scale parameter vector of interaction steps. Indicates the first A small-scale parameter vector of interaction steps. This represents the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics. Indicates the first The error vector of each interaction step; S403. Design the motion space; the motion space consists of motion vectors; motion vectors, specifically, include those used for... The parameter vector of the channel model to be optimized is incrementally updated for each interaction step; the action vector is defined as follows: ； in, Indicates the first The action vector of each interaction step is used to update the parameter vector of the channel model to be optimized. The update method is as follows: ； in, Indicates the first A channel model parameter vector with 1 interaction step. This represents the maximum value vector of the channel model parameter vector update. It represents the Hadamardi (or Hadama) stack; S404. Design the error vector between the simulated channel statistical characteristics and the measured channel statistical characteristics. Specifically, the error vector comprises the normalized root mean square error (NRMSE) between the cumulative distribution functions (CDF) of the simulated and measured channel statistical characteristics and the Kolmogorov-Smirnov (KS) distance, used to characterize the degree of matching of the current parameter configuration. The NRMSE and KS distance are defined as follows: ； ； in, This represents the channel characteristics, specifically the delay spread (DS) and angle spread (AS). and Let the cumulative distribution function (CDF) represent the channel statistical characteristics measured and simulated, respectively. The NRMSE represents the relationship between simulated channel statistical characteristics and measured channel statistical characteristics. The KS distance represents the distance between the simulated channel statistical characteristics and the measured channel statistical characteristics. This represents the domain of the cumulative distribution function. Given the mathematical expectation, therefore, the error vector The definition is as follows: ； in, To improve the NRMSE between the delay spread CDF of the simulated channel and the measurement channel, To measure the delay spread (CDF) of the channel, To simulate the delay spread (CDF) of the channel, The KS distance between the CDF delay spread of the simulated channel and the measurement channel is calculated. To extend the NRMSE between the simulated channel and the measurement channel angles of the CDF, To measure the angle spread CDF of the channel, To simulate the angle spread CDF of the channel, The KS distance between the CDF is extended to the angle of the simulated channel and the measurement channel.

6. The adaptive parameter optimization method for a universal channel model based on artificial intelligence according to claim 5, characterized in that, Step S5 includes: S501, Successful Construction Rewards; A positive reward is given to the agent when the error vector is less than a preset success threshold, and the reward decays as the number of interaction steps increases; the success reward is defined as follows: ； ； in, Indicates the number of interaction steps Successful rewards at that time It is a preset success threshold vector. Indicates the number of interaction steps Successful reward events, It is the success reward coefficient. It is the maximum number of interaction steps in each decision round; S502. Construct a penalty term, including imposing a penalty when the error vector exceeds a preset penalty judgment threshold; the penalty term is defined as follows: ； in, Indicates the number of interaction steps The penalty items at that time Indicates the number of interaction steps Error at time, and These are the penalty coefficients for the error vector and the number of interaction steps, respectively. It is a preset penalty judgment threshold; S503. Construct parameter-guided reward items, which are defined as follows: ； in, Indicates the first The parameter for each interaction step guides the reward item. Indicates the first Error in the number of interaction steps, It is the error reduction guiding coefficient. and These are the optimization guiding coefficients for the large-scale parameter vector and the small-scale parameter vector, respectively. and They are respectively in the 1st The action vector consists of a large-scale parameter vector and a small-scale parameter vector for each interaction step; S504. Construct the reward function, in the... Reward function for each interaction step The definition is as follows: 。 7. The adaptive parameter optimization method for a universal channel model based on artificial intelligence according to claim 6, characterized in that, Step S6 includes: S601. A deep reinforcement learning agent based on the soft actor critic algorithm (SAC) is used for training. The deep reinforcement learning agent includes a policy network, a dual critic network, a target network, and an experience replay cache. S602. In each training round, the policy network determines the training outcome based on the first... The state vector of each interaction step and action vectors Update the parameter vector of the channel model to be optimized, and calculate the first... The reward function value for each interaction step And store the updated state vector. To the experience replay cache; S603. Randomly extract batch samples from the experience replay buffer, iteratively update the dual critic network and policy network, and synchronously update the target network; after the soft actor critic algorithm is trained until the channel model parameter vector to be optimized converges, output the optimized channel model parameter vector.

Images