Wireless channel twin modeling method based on mechanism knowledge and meta-learning

Abstract

The application provides a wireless channel twin modeling method based on mechanism knowledge and meta learning, and belongs to the technical field of wireless communication.The application comprises the following steps: constructing a channel model based on mechanism knowledge; constructing a deep learning model, performing offline training on the deep learning model, and learning the mapping relationship between channel measurement data and a model parameter set; when a new scene outside the training task range is encountered, using unknown channel environment data to perform online fine tuning on the trained deep learning model, and estimating the mechanism model parameter set in a real-time scene by using the deep learning model after fine tuning; and obtaining the channel coefficients of each cluster and each path by using the mechanism model parameter set and the channel model based on mechanism knowledge.The method uses a meta learning strategy to overcome the problems that a traditional machine learning model has a high requirement for data quantity and computing resources and weak model generalization ability when facing unknown channel scenes, and can realize fast and accurate modeling of a dynamically changing complex channel environment.

Description

TECHNICAL FIELD

[0001] The present application relates to wireless communication technology, and particularly to a wireless channel twin modeling method based on mechanism knowledge and meta-learning. BACKGROUND

[0002] As the medium for transmitting electromagnetic wave signals in a wireless communication system, the propagation characteristics of a wireless channel have a decisive effect on the performance of a communication system, such as the bit error rate, channel capacity, energy efficiency, and spectrum efficiency. A channel model that reflects the actual propagation environment is of great significance for the design, development, performance evaluation, and actual deployment of a communication system. In addition, with the rapid development of digital twin technology, the combination of digital twin and wireless communication technology has attracted more and more attention. Utilizing digital twin to empower wireless networks can effectively model, simulate, verify, operate, and optimize real physical networks in a virtual environment, which can help networks to achieve low-cost trial and error, intelligent decision-making, high-efficiency innovation, and automated optimization, and thus realize self-optimization, self-evolution, and self-growth. Wireless communication network twinning has become a key research direction for B5G and 6G technologies, and is likely to drive a leap-forward development of wireless communication technology. In the construction of a wireless communication twin network, the twin modeling of a wireless channel is a key to network twinning and directly affects the accuracy of network twin modeling. Therefore, in order to better support the accurate mapping of a digital twin network to a physical network, a more real-time and accurate wireless channel twin modeling method is needed.

[0003] However, the complex and heterogeneous nature of a wireless channel poses a great challenge to accurate channel modeling, especially when the communication nodes are highly mobile and the geographical conditions are poor, resulting in a time-varying wireless channel with diverse characteristics. Moreover, network deployment is often on-the-fly, and limited channel data are available.

[0004] Existing channel twin modeling methods can be divided into mechanism models and data-driven models. The mechanism model usually includes deterministic ray tracing method and random model. Deterministic ray tracing method relies on accurate information of propagation environment. According to the determined propagation environment, the physical equation such as Maxwell equation set or approximate propagation equation is used to analyze and predict the radio wave propagation characteristics. This kind of method depends on the accurate information of geographical environment, and has high computational complexity. Generally, it is only suitable for channel modeling of specific scenes in a small range, and it is difficult to meet the real-time and high efficiency requirements of channel modeling in complex heterogeneous environment. The random model describes the channel parameters through a determined probability distribution. It summarizes the statistical characteristics of the channel, can adapt to different scenes, has universality and low complexity, but the modeling accuracy is lower than that of the deterministic ray tracing method. Random model is divided into non-geometric random model and geometry-based stochastic model (GBSM, Geometry-Based Stochastic Model). The non-geometric random model focuses on measurement. The geometry-based stochastic model abstracts the scatterers in the environment as clusters with a certain probability distribution, and then generates the multipath component parameters between clusters and within clusters. This kind of model can intuitively reflect the physical meaning of the channel due to the introduction of clusters, and can generate channel impulse response through the statistical information of channel parameters. The geometry-based stochastic model can better reflect the propagation characteristics of the channel in a certain environment in a statistical sense, and is widely used in statistical evaluation of system simulation. For example, the channel model adopted by 3GPP and other standardization organizations is a typical GBSM.

[0005] In recent years, with the rapid development of artificial intelligence and machine learning, data-driven methods have also been applied to channel modeling. Machine learning can learn the nonlinear characteristics of wireless channels, effectively mine the channel propagation characteristics and implicit rules from a large amount of historical data, and represent the channel through a data-driven way without a strictly defined model. In actual wireless channel environment modeling, data-driven models have good performance, but machine learning models usually need a large amount of training data to ensure the inference effect of the model.

[0006] In the face of complex heterogeneous wireless communication environment, the deterministic channel modeling method in pure mechanism model has high complexity and is not suitable for large-scale scenarios, the random model has many assumption conditions, the characterization ability of channel characteristics is limited, the characteristic parameters of the actual channel environment need to be estimated, in the case of difficult to obtain a large number of channel data samples, the accuracy of modeling is limited. The performance of pure data-driven method depends too much on high-quality data samples, and the cost of collecting channel data is high in actual application, especially in the mobile environment, the channel state changes dynamically, the number and type of samples that can be collected are limited, which greatly weakens the real-time fidelity and accuracy of the channel model. At the same time, the generalization ability of the data-driven method is weak, and a model can only be applied to a certain channel scene, and a large amount of computing resources and time cost are needed to update the model in the environment with various scenes. SUMMARY

[0007] In view of the above problems, the application provides a wireless channel twin modeling method based on mechanism knowledge and meta-learning. The method has the ability to explain and depict the characteristics of the channel based on the mechanism knowledge, and uses the meta-learning strategy to overcome the problems of high data volume and computing resource requirements and weak model generalization ability of traditional machine learning models when facing unknown channel scenes, thereby improving the flexibility and environmental adaptability of the model and realizing rapid and accurate modeling of dynamically changing complex channel environments.

[0008] The technical scheme adopted by the application is as follows:

[0009] The wireless channel twin modeling method based on mechanism knowledge and meta-learning comprises the following steps:

[0010] Step 1, constructing a channel model based on mechanism knowledge;

[0011] Step 2, constructing a deep learning model, and performing offline training on the deep learning model to learn the mapping relationship between channel measurement data and model parameter sets;

[0012] Step 3, when a new scene outside the training task range is encountered, the deep learning model trained in step 2 is fine-tuned online using unknown channel environment data, and the mechanism model parameter set P in the real-time scene is estimated using the deep learning model after fine-tuning; the mechanism model parameter set P includes delay spread, angle of departure spread, angle of arrival spread, shadow fading, and K factor;

[0013] Step 4, using the mechanism model parameter set P obtained in step 3 and the channel model based on mechanism knowledge constructed in step 1, the channel coefficients of each cluster and each path are obtained.

[0014] Further, in step 1, a random channel modeling method based on geometry is used to establish a channel model based on mechanism knowledge.

[0015] Further, the deep learning model in step 2 is a convolutional neural network or a back propagation neural network, and the specific manner of step 2 is as follows:

[0016] Step 2.1, generate a channel matrix by using simulation data or field measurement data to construct a training data set; in the training data set, the channel matrix is taken as a training sample, and the mechanism model parameters corresponding to the channel are taken as a sample label;

[0017] Step 2.2, construct a task set by using channel matrices in different channel scenarios, and the type of model parameters to be estimated in each channel scenario is the same, and the channel matrix and the mechanism parameters to be estimated in each scenario are regarded as a task set;

[0018] The task set composed of U channel matrices in different scenarios is T = {T (1) ,...T (u) ,...,T (U)}, wherein G is the number of sample sets of the u-th task, that is, the number of samples of the channel matrix and the mechanism parameters corresponding to the channel matrix in the scenario, is the g-th channel matrix of the u-th task, is the mechanism parameter set corresponding to the g-th channel matrix of the u-th task; each task T (u) is divided into two sample sets of the same size, namely a support set and a query set

[0019] Step 2.3, offline train the deep learning model through mutual cooperation between the meta-learner and the base learner, the base learner trains on each task to learn the characteristics of the task, and the meta-learner summarizes, summarizes and refines the rules on all tasks and learns the commonness between tasks;

[0020] Suppose the initial parameters of the deep learning model on the base learner are θ, for the u-th task T (u) , first use the samples in the support set to train the deep learning model, and then the parameters of the trained deep learning model are updated as:

[0021]

[0022] wherein, is the deep learning model parameter trained by the base learner in the last task; is the base learner with the deep learning model parameter ; α is the learning rate of the base learner; is the input on the task T (u) ; loss function of the current time gradient of the loss function

[0023] inputting the sample in the query set into the trained deep learning model calculating the gradient of the new loss function the meta-learner updates the parameters of the deep learning model using the gradient of the loss function as follows:

[0024]

[0025] in formula (3), is the parameter of the deep learning model trained by the meta-learner in the last task, and β is the learning rate of the meta-learner;

[0026] In training the next task, the meta-learner takes as the initial value of the deep learning model of the next task, and performs retraining using the support set and the query set of the next task. After all the tasks are traversed in this way, the final model parameter after training u tasks is obtained.

[0027] Further, in step 2.1, if simulation data is used, a random disturbance is added to the simulation data to generate a channel matrix, instead of the original channel matrix generated based on a single standard parameter.

[0028] If field measurement data is used, first, the parameters of the multipath components are extracted from the channel measurement data using a high-resolution parameter estimation method, and then the large-scale parameters of the model are estimated by the maximum likelihood method to obtain the label of the training sample.

[0029] Further, in step 2.3, the loss function uses a least square estimation function or a linear minimum mean square error estimation function.

[0030] Further, the specific manner of step 3 is as follows:

[0031] constructing a data set in a new environment where H' f is a channel matrix, P' f is an estimated mechanism model parameter, F is the size of the fine-tuning data set, and f is an index used to distinguish the new environment data and the training data.

[0032] fine-tuning the deep learning model trained in step 2 using A FT , and the network parameter is updated to Ψ:

[0033]

[0034] where γ is the learning rate of the fine-tuning network.

[0035] Further, the specific manner of step 4 is:

[0036] Obtaining large-scale parameters in the mechanism model parameter set P, including delay spread, shadow fading, angle of departure spread and angle of arrival spread, and K factor of Rice;

[0037] Randomly calculating small-scale parameters according to the large-scale parameters and the probability distribution functions thereof, including the delay, power, angle of arrival and angle of departure of each cluster;

[0038] After obtaining the small-scale parameters of each cluster, assuming that there are M sub-paths in each cluster, the angle of departure and the angle of arrival of each path are randomly matched, and a random initial phase is generated; assuming that the antenna array is a uniform linear array, for n-2 weak power clusters, n=3, 4, …, N, the channel coefficient between the corresponding receiving antenna u and transmitting antenna s on the nth cluster is:

[0039]

[0040] Wherein, P n is the power of the nth cluster, which is generated according to the delay, and the delay is determined by the large-scale parameter delay spread; F tx,s,V , F tx,s,H are the field strengths of the transmitting antenna s in the vertical and horizontal directions, respectively, F rx,s,V , F rx,s,H are the field strengths of the receiving antenna u in the vertical and horizontal directions, respectively; κ n,m is the cross-polarization power ratio of the mth path of the nth cluster, which is subject to a lognormal distribution; is the initial phase in different polarization directions; β n,m and alpha n,m are the angle of arrival and the angle of departure of the mth path on the nth cluster, respectively, which are randomly generated according to the angle spread and the corresponding distribution; d u , d s are the array distances of the receiving antenna and the transmitting antenna, respectively, and lambda0 is the carrier wavelength, is the Doppler shift, which is related to the speed v and the moving angle theta v of the base station; the superscript T represents matrix transposition, and j is the imaginary unit.

[0041] The present application has the following beneficial effects:

[0042] 1. The present application is based on a mechanism model, and uses a neural network based on meta-learning to estimate channel model parameters. A small amount of real data is augmented by a generative adversarial network to increase the size of the neural network training set, a parameter estimation training task set is constructed under different channel scenarios, the common relationship between different channel scenario data and mechanism model parameters is learned through training of the task itself and cross-training between tasks, and an initial model with good generalization is obtained.

[0043] 2、The neural network can infer the parameters of unknown channel scenes in real time by using a small amount of new channel environment data for online fine-tuning, and substituting the mechanism model to complete channel twin modeling, thereby effectively improving the training efficiency of the parameter estimation model and enhancing the generalization ability of the channel model in different scenes. Therefore, in practical applications, only a small amount of channel data can be used to complete fast and accurate modeling of actual unknown channel scenes. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a schematic diagram of a channel twin modeling method based on mechanism knowledge and meta-learning.

[0045] Figure 2 is a schematic diagram of a channel data augmentation model based on a generative adversarial network. DETAILED DESCRIPTION

[0046] The application will be further described below in conjunction with the drawings and specific embodiments.

[0047] The channel twin modeling method based on mechanism knowledge and meta-learning, as shown in Figure 1 includes the following steps:

[0048] 1. Construct a channel model based on mechanism knowledge.

[0049] Geometric-based random channel modeling is a widely used channel modeling method based on mechanism knowledge. Taking the establishment of the WINNER II model as an example, the parameters needed in the generation process of the channel coefficients are composed of two parts: large-scale parameters and small-scale parameters. The large-scale parameters mainly include delay spread, shadow fading, angle of departure spread and angle of arrival spread, and K factor. The small-scale parameters mainly calculate the delay, power, angle of arrival, and angle of departure of each cluster.

[0050] In the WINNER II model modeling process, each scene corresponds to a different set of channel model parameters. When modeling the actual wireless communication propagation environment, the channel parameters are usually estimated based on channel measurement data, such as the commonly used SAGE algorithm. However, with the increase in the number of antennas, the accuracy and speed of traditional channel parameter estimation methods cannot meet the needs of twin modeling. Therefore, the method constructs a channel model parameter set P, including delay spread, angle of departure spread, angle of arrival spread, shadow fading (dB), K factor, and other large-scale parameters. These parameters characterize the different dimensional characteristics of different channel scenes and are essential parameters for calculating channel coefficients. The method uses a neural network based on meta-learning to estimate the corresponding parameter set P based on the channel data in each scene.

[0051] 2. Construct a deep learning model, and perform offline training on the deep learning model to learn a mapping relationship between channel measurement data and a model parameter set.

[0052] The deep learning model, such as a convolutional neural network (CNN) or a back propagation neural network (BPNN), is used to learn the mapping relationship between the channel measurement data and the model parameter set. The training of the deep learning model usually requires a large amount of data, such as simulation data and actual measurement data. Figure 1 As shown in the figure, the training data sources of the model include simulation data and actual measurement data. However, it is often difficult to collect actual measurement data in practice. Therefore, the present application uses a generative adversarial network (GAN) to augment a small amount of actual measurement data.

[0053] For a channel model with mechanism knowledge, the model parameter set in different environments can be quite different, such as in cities, rural areas, and suburbs. For a traditional deep learning model, a model can only learn the parameter estimation task in a certain environment. In order to improve the adaptability of the neural network to different channel scenarios, the present application uses a model-agnostic meta-learning strategy to train the neural network. The parameter estimation plans in different scenarios are divided into different tasks, one city corresponds to one task, and the data in each task is evenly divided into a support set and a query set. The network is trained on the support set and the query set of each task in turn, the training on a single task is an inner loop, and the training between tasks is an outer loop. Through the iterative optimization of the inner and outer loops, good model initialization parameters that can be generalized to new cities are obtained. A small amount of data in a new city is used to fine-tune the model online, so as to quickly and accurately estimate the channel parameters.

[0054] In the training data set, the channel matrix is used as a training sample, and the model parameters corresponding to the channel are used as a sample label. The source of the data set is usually one of the following two:

[0055] The first one is to use simulation data sets generated by an empirical channel model. The channel parameters of this type of data are ideal parameters, and the empirical channel model in different scenarios can easily obtain the parameters. In order to ensure the diversity of the parameters, a tolerable random disturbance is added to the parameters to generate a channel matrix instead of the original channel matrix generated based on a single standard parameter.

[0056] The second one is field measurement data. This type of data only has channel matrix information, and does not have model parameter information. For this type of sample, the parameters of the multipath components are first extracted from the channel measurement data by using a high-resolution parameter estimation technology such as the SAGE algorithm, and then the large-scale parameters of the model are estimated by using a maximum likelihood method to obtain the label of the training sample.

[0057] In actual applications, especially in dynamically changing environments, the actual channel data samples that can be collected are often limited. The present application uses a generative adversarial network to augment a small amount of collected data in a certain scenario, as shown in the figure.Figure 2 As shown, the generative adversarial network is composed of a generator G and a discriminator D, the generator takes a randomly sampled noise vector z as input and outputs a synthetic channel matrix The real collected channel matrix H is input to the discriminator together with the output of the generator. The training goal of the discriminator is to distinguish the output of the generator from the training samples as much as possible, and the objective function is The training goal of the generator is to deceive the discriminator as much as possible, and the objective function is When the training of the generator and the discriminator reaches the minimum game Nash equilibrium, the network converges, and the generator generates The same distribution as the real channel matrix H, that is, the generator can be used to augment a small amount of measured data. Then use the SAGE algorithm to estimate the channel parameters corresponding to the channel matrix, so as to achieve the purpose of expanding the training set and improve the training efficiency and robustness of the neural network.

[0058] Construct a task set using channel matrices under different channel scenarios. The type of model parameters to be estimated under each channel scenario is the same. The channel matrix under each scenario and the mechanism parameters to be estimated are regarded as a task set. Assume that the task set composed of channel matrices under U different scenarios is T = {T (1) ,...T (u) ,...,T (U)}, where G is the number of sample sets of the u-th task, that is, the number of samples of the channel matrix and the corresponding mechanism parameters under this scenario, is the g-th channel matrix of the u-th task, is the mechanism parameter set corresponding to the g-th channel matrix of the u-th task. Each task T (u) can be divided into two sample sets of the same size, namely the support set and the query set

[0059] The training of the neural network is completed through the mutual cooperation between the meta-learner and the base learner. The base learner is trained on each task to learn the characteristics of the task. The meta-learner summarizes, summarizes and refines the regularity of all tasks based on the training results of many tasks, learns the commonality between tasks, and determines the generalization ability of the model.

[0060] Assume that the initial parameters of the neural network on the base learner are θ, for the u-th task T (u) , first use the samples in its support set to train the network, and then the parameters of the trained network are updated as:

[0061]

[0062] wherein, is the network parameter trained by the previous task in the base learner; is the base learner with the network parameter ; and a is the learning rate of the base learner; is the loss function gradient of the base learner when the input is (u) ; and is the new loss function gradient calculated for the sample in the query set of the trained network ; and is the network parameter updated by the meta-learner using the gradient value:

[0063]

[0064] In formula (3), is the network parameter trained by the previous task in the meta-learner, and b is the learning rate of the meta-learner. When training the next task, the meta-learner takes as the initial value of the neural network of the next task, and performs retraining using the support set and the query set of the next task. After all the tasks are traversed in this way, the final model parameter after training U tasks is obtained, which reflects the commonality between the U tasks and embodies the generalization ability of the model on the given task set.

[0065] The loss function can be obtained by using the least square estimation or the linear least mean square error estimation method. The following formula is the loss function obtained by the least square method, wherein is the estimated mechanism parameter:

[0066]

[0067] 3. When a new scene outside the training task range is encountered, the deep learning model trained in step 2 is fine-tuned online using unknown channel environment data, and the mechanism model parameter set P in the real-time scene is estimated using the deep learning model after fine-tuning.

[0068] The neural network trained by the meta-learning method learns the common mapping relationship between the channel matrix and the mechanism parameter in different scenes, and has good generalization performance. When online inference encounters a new scene outside the training task range, only a small amount of unknown channel environment data is needed to fine-tune the network, which greatly improves the convergence speed of the network.

[0069] Suppose the data set in the newly constructed environment is wherein H' is the channel matrix, and P' is the mechanism parameter. f f ​​For the mechanism model parameters estimated by the traditional method, F « G is the size of the fine-tuning dataset. Using A FT For the fine-tuned network, the network parameters are updated as:

[0070]

[0071] where γ is the learning rate of the fine-tuned network, and the parameter Ψ is compared with the parameter The general characteristics of a new channel task outside the given task set range.

[0072] In the online inference stage, the mechanism model parameter set P, i.e., the large-scale parameters such as delay spread, angle of departure spread, angle of arrival spread, shadow fading (dB), and K factor, can be quickly estimated using the fine-tuned neural network in real-time scenarios.

[0073] 4. Using the mechanism model parameter set P and the channel model based on mechanism knowledge, the channel coefficients of each cluster and each path are obtained.

[0074] According to the large-scale parameters and their probability distribution functions, the small-scale parameters of each cluster are randomly calculated. Then, assuming that there are M sub-paths in each cluster, the angle of departure and the angle of arrival of each path are randomly matched, and a random initial phase is generated. Assuming that the antenna array is a uniform linear array, for n-2 weak power clusters, i.e., n = 3, 4, …, N, the channel coefficient between the corresponding receiving antenna u and transmitting antenna s on the nth cluster is:

[0075]

[0076] where P n is the power of the nth cluster, which is generated according to the delay, and the delay is determined by the large-scale parameter delay spread; F tx,s,V , F tx,s,H are the field strengths of the transmitting antenna s in the vertical and horizontal directions, respectively, F rx,s,V , F rx,s,H are the field strengths of the receiving antenna u in the vertical and horizontal directions, respectively; κ n,m is the cross-polarization power ratio of the mth path of the nth cluster, which follows a lognormal distribution; is the initial phase in different polarization directions; β n,m and α n,m are the angle of arrival and the angle of departure of the mth path on the nth cluster, respectively, which are randomly generated according to the angle spread and the corresponding distribution; d u , d s are the array distances of the receiving antenna and the transmitting antenna, respectively, λ0 is the carrier wavelength, is the Doppler shift, which is related to the speed v and the moving angle θ v of the base station.

[0077] So far, the wireless channel twin modeling based on mechanism knowledge and meta-learning is completed.

[0078] The following is a more specific example:

[0079] Consider that a small amount of urban macrocell, suburban macrocell, and rural channel data is collected in A, and it is assumed that the number of channel matrix samples collected in each scenario is 200. The goal is to complete high-precision modeling of the wireless channel in the suburban microcell scenario in A, and the specific steps are as follows:

[0080] 1. Construct a channel model based on mechanism knowledge.

[0081] According to equation (1), the mechanism channel model is constructed, and the parameters to be estimated are delay spread, angle of departure spread, angle of arrival spread, shadow fading (dB), and K factor. Next, a neural network is trained to estimate the parameter set of the model according to the channel data.

[0082] 2. Construct a deep learning model and perform offline training on the deep learning model.

[0083] The deep learning model uses a back propagation neural network (BPNN).

[0084] First, use the empirical channel WINNER II model in the MATLAB simulation platform to generate channel data samples of Q>>200 for three scenarios: urban macrocell, suburban macrocell, and rural. By querying the model in the simulation platform, the channel model parameters corresponding to each scenario can be obtained, so a total of 3Q training samples are obtained.

[0085] Merge the software simulation data and the measured data of the three scenarios, and input them into the generative adversarial network in units of scenarios. Train to obtain three generative adversarial networks corresponding to the urban macrocell, suburban macrocell, and rural scenarios, respectively. The generator in the trained generative adversarial network can be used for data augmentation, and I channel data with the same data distribution can be generated using it. Therefore, the total number of samples of the measured data + augmented data in each scenario is 200+I, and the total number of training samples of the simulation software data is G=200+Q+I.

[0086] Use the SAGE algorithm to estimate the corresponding labels for all channel data, i.e., the five parameters of the channel: delay spread, angle of departure spread, angle of arrival spread, shadow fading (dB), and K factor.

[0087] The channel matrix and the mechanism parameters to be estimated in each scenario are considered as a task set, so the three scenarios of urban macrocell, suburban macrocell, and rural form a task set T={T (1) ,T (2) ,T (3)}, where is the gth channel matrix of the u th task, is the mechanism parameter corresponding to the gth channel matrix of the u th task. Each task is further split into a support set and a query set

[0088] The neural network is trained using a meta-learning strategy, and is completed by mutual cooperation between the meta-learner and the base learner, to obtain a deep learning model with generalization ability.

[0089] 3. Online fine-tuning.

[0090] A small amount of channel data of the suburban microcell scenario in A is collected, the mechanism model parameters corresponding to the data sample are obtained as labels using the SAGE algorithm, and the neural network trained offline is fine-tuned online using the method proposed in the patent. Since the neural network has been trained offline based on the meta-learning method, it has fully learned the mapping relationship between the channel data and the model parameters in different scenarios, and has a certain generalization ability, so the neural network can quickly converge when fine-tuning in the new scenario of suburban microcell, and the training efficiency of the network is high.

[0091] During online inference, the channel data collected in the suburban microcell scenario is input, and the mechanism model parameter set P, i.e., the large-scale parameters of delay spread, angle of departure spread, angle of arrival spread, shadow fading (dB), K factor, etc., can be quickly estimated using the fine-tuned neural network.

[0092] 4. Online inference.

[0093] Using the large-scale parameters, combined with the support parameters, the channel coefficients of each cluster and each path can be calculated, and the wireless channel twin modeling based on mechanism knowledge and meta-learning is completed.

[0094] In summary, the present application is aimed at the wireless channel twin modeling demand in highly complex heterogeneous scenarios. In view of the problems that the pure mechanism model modeling accuracy is not high, and the pure data-driven model has high dependence on the data set and poor generalization performance in scene transformation, the present application proposes to use the meta-learning method to more accurately estimate the channel parameters in different scenarios based on the traditional mechanism knowledge modeling, and to achieve the purpose of fast and high-precision modeling of unknown channels under the condition of limited channel data.

Claims

1. A method for wireless channel twin modeling based on mechanism knowledge and meta-learning, characterized in that, The method comprises the following steps: Step 1, constructing a channel model based on mechanism knowledge; Step 2, constructing a deep learning model, and learning the mapping relationship between channel measurement data and the model parameter set through offline training of the deep learning model; wherein the deep learning model is a convolutional neural network or a back propagation neural network, and the specific manner of step 2 is as follows: Step 2.1, generating a channel matrix by using simulation data or field measurement data to construct a training data set; in the training data set, the channel matrix is used as a training sample, and the mechanism model parameters corresponding to the channel are used as a sample label; Step 2.2, constructing a task set by using channel matrices in different channel scenarios, and the type of model parameters to be estimated in each channel scenario is the same, and the channel matrix and the mechanism parameters to be estimated in each scenario are regarded as a task set; The task set composed of channel matrices in U different scenarios is T = {T (1) ,...T (u) ,...,T (U)}, wherein G is the number of sample sets of the u-th task, that is, the sample number of channel matrices and mechanism parameters corresponding to the u-th scenario, is the g-th channel matrix of the u-th task, is the mechanism parameter set corresponding to the g-th channel matrix of the u-th task; each task T (u) is divided into two sample sets of the same size, that is, a support set and a query set Step 2.3, offline training of the deep learning model through mutual cooperation between a meta-learner and a basic learner; the basic learner is trained on each task to learn the characteristics of the task, and the meta-learner summarizes, summarizes and refines the regularities on all tasks by comprehensively learning the training results of multiple tasks to learn the commonness between tasks; Assuming the initial parameters of the deep learning model on the base learner are θ, for the u-th task T (u) First, the samples of the support set are used to train the deep learning model, and the parameter update of the trained deep learning model is: wherein, is the deep learning model parameter trained by the previous task in the base learner; is the base learner with the deep learning model parameter ; a is the learning rate of the base learner; is the gradient of the loss function (u) when the input is ; and is the gradient of the loss function inputting samples in a query set into the trained deep learning model computing a new loss function gradient the meta-learner updates the deep learning model parameters using the loss function gradient In formula (3), is the deep learning model parameter trained by the previous task in the meta-learner, and β is the learning rate of the meta-learner. When training the next task, the meta-learner will As the initial value of the deep learning model of the next task, it is retrained with the support set and query set of the next task. After all tasks are traversed in this way, the final model parameters after training U tasks are obtained Step 3, when a new scenario outside the training task range is encountered, the deep learning model trained in step 2 is fine-tuned online by using unknown channel environment data, and the mechanism model parameter set P in the real-time scenario is estimated by using the deep learning model after fine-tuning; the mechanism model parameter set P includes delay spread, angle of departure spread, angle of arrival spread, shadow fading, and K factor; the specific manner is as follows: Constructing a dataset in a new environment where H' f is a channel matrix, P' f are estimated mechanism model parameters, F is the size of the fine-tuning dataset, and f is an index used to distinguish the new environment data from the training data. Utilizing A FT Fine-tune the deep learning model trained in Step 2, updating the network parameters to Ψ: Wherein, γ is the learning rate of the fine-tuning network; Step 4, obtaining the channel coefficients of each cluster and each path by using the mechanism model parameter set P obtained in step 3 and the channel model based on mechanism knowledge constructed in step 1; the specific manner is as follows: Obtain the large-scale parameters in the mechanism model parameter set P, including delay spread, shadow fading, angle of departure spread and angle of arrival spread, and K factor; Randomly calculate the small-scale parameters according to the large-scale parameters and their probability distribution functions, including the delay, power, angle of arrival and angle of departure of each cluster; After obtaining the small-scale parameters of each cluster, it is assumed that there are M sub-paths in each cluster, the angle of departure and the angle of arrival of each path are randomly matched, and a random initial phase is generated; assuming that the antenna array is a uniform linear array, for n-2 weak power clusters, n=3, 4, …, N, the channel coefficient between the corresponding receiving antenna u and the transmitting antenna s on the nth cluster is: where P n is the power of the nth cluster, which is generated according to the delay, which is determined by the large-scale parameter delay spread; F tx,s,V , F tx,s,H are the field strengths of the transmit antenna s in the vertical and horizontal directions, respectively; F rx,s,V , F rx,s,H are the field strengths of the receive antenna u in the vertical and horizontal directions, respectively; κ n,m is the cross-polarization power ratio of the mth ray of the nth cluster, which is subject to a lognormal distribution; is the initial phase in different polarization directions; β n,m and α n,m are the angle of arrival and angle of departure of the mth ray on the nth cluster, which are randomly generated according to the angle spread and the corresponding distribution; d u , d s are the array distances of the receive antenna and the transmit antenna, respectively, λ0is the carrier wavelength, is the Doppler shift, which is related to the speed v and the moving angle θ v of the base station; the superscript T represents matrix transposition, and j is the imaginary unit.

2. The mechanism knowledge and meta-learning based wireless channel twin modeling method of claim 1, wherein, In step 1, a random channel modeling method based on geometry is used to establish a channel model based on mechanism knowledge.

3. The mechanism knowledge and meta-learning based wireless channel twin modeling method of claim 1, wherein, In step 2.1, if simulation data is used, random perturbation is added to the simulation data to generate a channel matrix, instead of the original channel matrix generated based on a single standard parameter; If field measurement data is used, first, the parameters of the multipath components are extracted from the channel measurement data by using a high-resolution parameter estimation method, and then the large-scale parameters of the model are estimated by using a maximum likelihood method to obtain the label of the training sample.

4. The mechanism knowledge and meta-learning based wireless channel twin modeling method of claim 1, wherein, In step 2.3, the loss function adopts a least square estimation function or a linear minimum mean square error estimation function.

Images