Multi-frequency communication perception integrated channel modeling method based on generative adversarial network

Abstract

The invention discloses a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network, which comprises the following steps of: 1, constructing a multi-frequency communication perception integrated channel model, namely, distributing a channel from a transmitter to a receiver, a base station and mobile user equipment in a city macro region; 2, simplifying the model established in the step 1; step 3, the base station and the mobile user equipment carry out deployment model assumption and are randomly placed in an environment according to different base station intervals and densities, then path parameters between the base station and the mobile user equipment are generated, and an MIMO channel is determined through the path parameters; 4, ray tracing simulation is further included, and training data are generated through ray tracing simulation; according to the scheme, the channel propagation characteristics and implicit laws are learned through machine learning algorithm training by using the channel parameter data set, and the channel model can be more intelligently constructed to adapt to various scenes and achieve better performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network. Background Art

[0002] Modern cellular systems increasingly rely on communicating simultaneously in multiple non-contiguous frequency bands to achieve macrodiversity and increased bandwidth. Multi-frequency communication is particularly important in millimeter wave and terahertz frequencies, as these bands are often coupled with lower frequencies to maintain robustness. Since millimeter wave and terahertz systems operate with very high bandwidths and highly directional beams, statistical models must capture the full bidirectional characteristics of the channel, i.e., angles of arrival and departure, delays, and gains for all paths. Factors that affect propagation, such as transmission and reflection losses, vary widely at different frequencies, and these parameters can have complex relationships.

[0003] Chinese invention patent CN113381828B discloses a sparse code multiple access random channel modeling method based on conditional generative adversarial network, which takes advantage of the good capture and distribution capabilities of CGAN. CGAN is used to model channel distribution to represent channel effects. The received DNN-SCMA signal corresponding to the pilot symbol is used as a component of conditional information. The method uses a model-free learning method to accurately learn different types of random channel models and effectively obtain dynamic channel information in actual communication scenarios. In existing communication systems, only one-way multipath channels between the signal transmitter and the receiver are modeled. However, in the synaesthesia integrated system, since perception needs to be based on the target reflection signal, a two-way multipath channel needs to be constructed. The machine learning method is an algorithm that can mine implicit rules from a large amount of historical data and make predictions or classifications. It is represented by a data-driven approach without the need for a strictly defined model, and can optimize the system in an end-to-end manner. Many key issues in wireless channel modeling can be viewed as regression, clustering, and classification problems in machine learning. These characteristics of wireless channel modeling are consistent with the capabilities of machine learning methods. Therefore, by using channel parameter datasets to learn channel propagation characteristics and implicit laws through machine learning algorithm training, channel models can be constructed more intelligently to adapt to various scenarios and achieve better performance. Summary of the invention

[0004] In response to the problems raised by the above background technology, the purpose of the present invention is to provide a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network, which uses a channel parameter data set to learn the channel propagation characteristics and implicit laws through machine learning algorithm training, and can more intelligently construct a channel model to adapt to various scenarios and achieve better performance.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A multi-frequency communication perception integrated channel modeling method based on a generative adversarial network includes a transmitter to a receiver, a transmitter to a receiver channel, a base station, a mobile user device and clustering data, wherein the base station and the mobile user device are distributed in a macro area of ​​a city, and the clustering data includes the angle and delay of the clustering, and the vector of random received power, angle and delay propagation at different frequencies, and also includes the following steps:

[0007] Step 1: Build a multi-frequency communication perception integrated channel model, i.e., the channel from transmitter to receiver, base stations and mobile user devices distributed in an urban macro area;

[0008] Step 2: Simplify the model established in step 1;

[0009] Step 3: The base station and the mobile user equipment are deployed model-assuming that they are randomly placed in an environment according to different base station spacings and densities, and then path parameters between the base station and the mobile user equipment are generated, and the MIMO channel is determined by the path parameters;

[0010] Step 4: also includes ray tracing simulation, applying ray tracing simulation to generate training data;

[0011] Step 5: Cluster the collected paths, and the rays of each frequency generated by ray tracing are pre-processed to identify clusters and determine clustering parameters;

[0012] Step 6: Generate random cluster vector x by using neural network.

[0013] It is further defined that there is reciprocity between the base station and the mobile user equipment, so the base station and the mobile user equipment can be interchangeable.

[0014] Further defined, according to the multi-frequency communication perception integrated channel model described in step 1, a channel model is modeled by a set of common L clusters on M frequencies, wherein each cluster Described by a parameter vector:

[0015]

[0016] in is the minimum path delay within the cluster, are the arrival inclination angle and azimuth angle within the average cluster, are the departure tilt angle and azimuth within the average cluster, respectively;

[0017] Among them, the parameters m=1,…,M is given by the frequency-dependent components:

[0018]

[0019] in is the arrival rms tilt and azimuth, is the departure rms tilt and azimuth, is the rms delay propagation, is the total received power; where the multi-frequency channel can be described by a set of L clusters: x = [x (1) ,…x (L) ], where x is a clustering vector.

[0020] Further, the simplified method in step 2 is: fix the number of clusters to L=10, when there are less than L clusters at a certain frequency, if L 0 <L, that is, all clusters with frequency m Set the total receive power If a path exists in a cluster at one frequency m but not at another frequency m′, then the total received power is taken as and

[0021] The model x (l) Each cluster of has 6+6M parameters, and there are L clusters in total, so there are a total of (6+6M)L parameters in the cluster vector x.

[0022] Further defining, including the links, also includes the condition vector u set in each link, so the three-dimensional distance vector between the transmitter and the receiver is expressed as:

[0023] u=(d x ,d y ,d z )

[0024] The condition vector u may also include other parameters, including antenna height or base station type.

[0025] Further defined, according to step 3, the distribution of the clustering vector x is modeled as a function of the link condition p(x|u), that is, the conditional probability distribution p(x|u) is modeled, wherein the conditional distribution is represented by a mapping x=g(u,z), wherein z is a random vector with a known distribution, and x=g(u,z) is the basis for cellular area evaluation. Base stations and mobile user devices are randomly placed in an environment according to the deployment model assumption. The deployment determines the link condition u between each base station and the mobile user device. The path parameters between all base stations and mobile user devices are randomly generated by x=g(u,z). The path parameters are combined with the antenna array assumptions preset on both sides of the base station to determine all MIMO channels in the cellular network and calculate the signal-to-noise ratio or power distribution.

[0026] It is further defined that the step of applying ray tracing to generate training data in step 4 includes selecting a millimeter wave frequency band f 1 =28GHz and terahertz band f 2 =140GHz, import an area of ​​550×380 square meters into Wireless Insite, manually place 100 ground receiving base stations outside a 10m high building, and place transmitting equipment terminals at 320 locations on a 1.5m street. The deployment result is 100×320=32000 links. Then run independent ray tracing simulations for these two frequencies, set the maximum number of reflections to 6, the maximum number of diffractions to 1, and limit the maximum number of links to 20. In addition, ray tracing collects paths with a power of at least -250dB.

[0027] It is further defined that the clustering process of step five includes using a K-means clustering process for the arriving tilt angle and azimuth, the departing tilt angle and azimuth, and the delay of each link in the group, using a silhouette coefficient to determine the optimal number of clusters for each link, and after identifying each cluster, determining the angle and delay propagation and power in each cluster to generate a clustering vector x for each link.

[0028] Further defined, the process of using a neural network to generate a random clustering vector x in step 6 uses a conditional Wasserstein generative adversarial network CWGAN-GP with gradient penalty, CWGAN-GP includes a generator G and a discriminator C, the generator G accepts a conditional vector u and a random input z~p(z), and generates a random path data vector x=G(u,z), the vector z is a potential vector, the p(z) is a unit variance Gaussian vector, and also includes a potential dimension, the potential dimension is set to 20, and the generator G is ensured to match the data distribution by training the discriminator C, and the discriminator C is used to distinguish between generated samples and real samples;

[0029] The discriminator C and generator G network are trained through the loss function as follows:

[0030]

[0031] Where P g Is Implicitly defined generator distribution, P r is the distribution of data points.

[0032] Beneficial effects of the present invention:

[0033] 1. The present invention provides a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network. Compared with the prior art, the technical solution of the present invention has the least statistical assumptions and can learn complex statistical relationships between variables.

[0034] 2. The model represents the channel in a compact way that can capture the key commonalities and differences between different frequencies. Specifically, the channel is represented by a set of clusters whose average angles of arrival, departure, and delay are common in frequency, reflecting that the propagation path at the macro level is independent of frequency; on the other hand, power, angular propagation, and delay propagation are modeled as frequency-dependent to capture the changes in scattering and transmission loss over frequency, which is especially important for modeling in the millimeter wave and terahertz bands. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention can be further illustrated by the non-limiting examples given in the accompanying drawings;

[0036] Figure 1 A ray tracing simulation area diagram of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention;

[0037] Figure 2 It is an architecture diagram of a K-Means path clustering generator and a generative adversarial network model of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention;

[0038] Figure 3 A cumulative distribution function diagram of received power of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention;

[0039] Figure 4 It is a kernel density estimation diagram of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention;

[0040] Figure 5 A cumulative distribution function diagram of the minimum path delay in a cluster of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention;

[0041] Figure 6 A cumulative distribution function diagram of the root mean square clustering delay spread of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention;

[0042] Figure 7 A graph showing the relationship between azimuth and tilt angles of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network according to the present invention;

[0043] Figure 8It is a root mean square extended cumulative distribution function diagram of the arrival tilt angle and azimuth angle of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention;

[0044] Fig. 9 This is a root mean square expanded cumulative distribution function diagram of the departure tilt angle and azimuth angle of an embodiment of a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments. The technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.

[0046] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0047] In addition, the descriptions of "first", "second", etc. in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0048] like Figure 1 As shown, a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network of the present invention includes a transmitter to a receiver, and also includes a transmitter to a receiver channel, a base station, a mobile user device and clustering data, wherein the base station and the mobile user device are distributed in a macro area of ​​a city, and the clustering data includes the angle and delay of the clustering, and the vector of random received power, angle and delay propagated at different frequencies, and also includes the following steps:

[0049] Step 1: Build a multi-frequency communication perception integrated channel model, i.e., the channel from transmitter to receiver, base stations and mobile user devices distributed in an urban macro area;

[0050] Step 2: Simplify the model established in step 1;

[0051] Step 3: The base station and the mobile user equipment are deployed model-assuming that they are randomly placed in an environment according to different base station spacings and densities, and then path parameters between the base station and the mobile user equipment are generated, and the MIMO channel is determined by the path parameters;

[0052] Step 4: also includes ray tracing simulation, applying ray tracing simulation to generate training data;

[0053] Step 5: Cluster the collected paths, and the rays of each frequency generated by ray tracing are pre-processed to identify clusters and determine clustering parameters;

[0054] Step 6: Generate random cluster vector x by using neural network.

[0055] It is further defined that there is reciprocity between the base station and the mobile user equipment, so the base station and the mobile user equipment can be interchangeable.

[0056] Preferably, according to the multi-frequency communication perception integrated channel model described in step 1, a channel model is modeled by a set of common L clusters on M frequencies, wherein each cluster Described by a parameter vector:

[0057]

[0058] in is the minimum path delay within the cluster, are the arrival inclination angle and azimuth angle within the average cluster, are the departure tilt angle and azimuth within the average cluster, respectively;

[0059] Among them, the parameters m=1,…,M is given by the frequency-dependent components:

[0060]

[0061] in is the arrival rms tilt and azimuth, is the departure rms tilt and azimuth, is the rms delay propagation, is the total received power; where the multi-frequency channel can be described by a set of L clusters: x = [x (1) ,…x (L)], where x is a clustering vector.

[0062] Preferably, the simplified method in step 2 is: fix the number of clusters to L=10, when there are less than L clusters at a certain frequency, if L 0 <L, that is, all clusters with frequency m Set the total receive power

[0063] If a path exists in a cluster at one frequency m but not at another frequency m′, then the total received power is and This could happen, for example, where a path can penetrate materials at lower frequencies but is blocked at mmWave frequencies.

[0064] The model x (l) Each cluster of has 6+6M parameters, and there are L clusters in total, so there are a total of (6+6M)L parameters in the cluster vector x. For example, if M=2 and L=10, there are a total of p=(6+6M)L=180 parameters.

[0065] Preferably, the link is included, and the condition vector u set in each link is also included, so the three-dimensional distance vector between the transmitter and the receiver is expressed as:

[0066] u=(d x ,d y ,d z )

[0067] The condition vector u may also be added with other parameters, including antenna height or base station type.

[0068] Preferably, according to step 3, the distribution of the clustering vector x is modeled as a function of the link condition p(x|u), that is, the conditional probability distribution p(x|u) is modeled, wherein the conditional distribution is represented by a mapping x=g(u,z), wherein z is a random vector with a known distribution, and x=g(u,z) is the basis for cellular area evaluation. Base stations and mobile user devices are randomly placed in an environment (assuming that they are placed in urban micro-cells with a certain inter-station distance and density) according to the deployment model assumption. The deployment determines the link condition u between each base station and the mobile user device. The path parameters between all base stations and mobile user devices are randomly generated by x=g(u,z). The path parameters are combined with the antenna array assumptions preset on both sides of the base station to determine all MIMO channels in the cellular network and calculate the signal-to-noise ratio or power distribution.

[0069] Preferably, the step of applying ray tracing to generate training data in step 4 includes selecting a millimeter wave frequency band f 1 =28GHz and terahertz band f 2=140GHz, import an area of ​​550×380 square meters into Wireless Insite, manually place 100 ground receiving base stations outside a 10m high building, and place transmitting equipment terminals at 320 locations on a 1.5m street. The deployment result is 100×320=32000 links. Then run independent ray tracing simulations for these two frequencies, set the maximum number of reflections to 6, the maximum number of diffractions to 1, and limit the maximum number of links to 20. In addition, ray tracing collects paths with a power of at least -250dB.

[0070] Specifically, the key to modeling frequency dependence is the effect of diffuse scattering, which is particularly prominent in the millimeter wave and terahertz bands. As part of the simulation, the technicians enabled the scattering coefficient S in the ray tracing tool, which is the ratio of the scattered electric field to the incident electric field; the scattering coefficient of 28GHz was set to S = 0.273, and the scattering coefficient of 140GHz was set to S = 0.4, so that the 140GHz ray has a greater scattered power on the front lobe of the diffuse scattering, and the diffuse scattering interaction is limited to the last interaction along the path before reaching the receiver, which helps to reduce the running time of the diffuse scattering calculation.

[0071] Preferably, the clustering process of step five includes using a K-means clustering process for the arrival tilt angle and azimuth, the departure tilt angle and azimuth, and the delay of each link in the group, using a silhouette coefficient to determine the optimal number of clusters for each link, and after identifying each cluster, determining the angle and delay propagation and power in each cluster to generate a clustering vector x for each link.

[0072] Specifically, the collected paths are clustered, and the rays of each frequency generated by ray tracing are preprocessed to identify clusters and determine clustering parameters; in order to limit the size of the clustering vector, L = 10 is used as the maximum number of clusters; at the same time, a data record is obtained for each link, and each link also has a conditional vector u.

[0073] Preferably, the process of generating a random clustering vector x using a neural network in step 6 uses a conditional Wasserstein generative adversarial network CWGAN-GP with a gradient penalty, wherein the CWGAN-GP includes a generator G and a discriminator C, wherein the generator G accepts a conditional vector u and a random input z~p(z), and generates a random path data vector x=G(u,z), wherein the vector z is a potential vector, and the p(z) is a unit variance Gaussian vector, and also includes a potential dimension, wherein the potential dimension is set to 20, and the generator G is ensured to match the data distribution by training the discriminator C, and the discriminator C is used to distinguish between generated samples and real samples;

[0074] The discriminator C and generator G network are trained through the loss function as follows:

[0075]

[0076] Where P g Is Implicitly defined generator distribution, P r is the distribution of data points.

[0077] Specifically, the conditional vector u and cluster data x have heterogeneous data types, so a unified format is required. To overcome this challenge, the angle of each path is aligned to the line-of-sight direction, and the excess signal delay relative to the line-of-sight distance delay is measured, and then all values ​​are scaled using the minimum and maximum values ​​or a standard scalar.

[0078] The specific description of the embodiments of the present invention and its working principle are as follows:

[0079] This system simulation is constructed using Python. In this example, the millimeter wave frequency band f is selected 1 =28GHz and terahertz band f 2 = 140GHz, using ray tracing software: WirelessInsite developed by Remcom. The input is an area of ​​550×380 square meters, such as Figure 1 As shown, and using parameters similar to the 3GPP report, 100 ground receiving base stations are manually placed outside a 10m high building, and the transmitting equipment terminal is placed at 320 locations on a 1.5m street. The deployment result is 100×320=32000 links. Run independent ray tracing simulations for these two frequencies, set the maximum number of reflections to 6, the maximum number of diffractions to 1, and limit the maximum number of links to 20. In addition, ray tracing only collects paths with a power of at least -250dB. In the following example, 80% of the links will be used for training and 20% for testing, and the generated network is implemented in Tensorflow2.

[0080] We first evaluate the model’s ability to capture the joint distribution of total received power at two frequencies. The comparison can be done as follows: Let represents the learned generative model, let (u i ,x i ),i=1,…,N ts Represents the test data, where each sample contains the link condition u i and the corresponding cluster data vector x i For each test sample, we can calculate the vector v i =(v i1 ,v i2 )=(φ 1(x 1 ),φ 2 (x 2 )), where v ij is the omnidirectional received power of sample i at frequency j, where j = 1, 2. j (x) is a function that calculates the omnidirectional received power at frequency j based on cluster data x, where the omnidirectional received power is the received power experienced by the base station and the terminal device when they have omnidirectional antennas.

[0081] To compare these values ​​with the generated distribution, for each test sample i, we use the trained generator G and the trained generator u with the same conditions as the test data. i The random z i Generate a random sample Then a set of generated received powers is calculated Ideally, the sample v i and should have similar distributions.

[0082] Figure 3 The cumulative distribution functions of the generated and true power at two frequencies are plotted. The left plot shows the data v ij and generate samples CDF at frequency j = 1. The right figure shows the data and the resulting CDF for frequency j = 2. We see that in both cases the two marginal CDFs match almost exactly.

[0083] Figure 4 It shows that the generator can also capture joint statistics. The left figure shows the test data points (v i1 ,v i2 ), and the right figure shows the kernel density estimation plot of the trained model From the kernel density estimation diagram, we can see that the two joint distributions match well.

[0084] Figure 5 The ability of the model to generate latency was evaluated and the cumulative distribution function of the minimum path latency on the clusters was plotted. The latency between the clusters generated by the neural network and the true ray tracing values ​​was almost the same, with the maximum latency also being around 4000 nanoseconds.

[0085] Figure 6 It shows that the RMS delay spread at 140 GHz is generally smaller than that at 28 GHz, indicating that as the frequency increases, the RMS delay spread decreases due to the increase in power loss due to diffuse scattering.

[0086] Figure 7 The distribution of arrival tilt and azimuth angles is shown to be similar to that of real ray tracing.

[0087] Figure 8 and Fig. 9 The cumulative distribution function plots of the RMS spread of the arrival tilt angle and azimuth angle, respectively, show that the RMS angular spread at 28 GHz is slightly larger than that at 140 GHz. When there is only one cluster on a link at a given frequency, the RMS angular spread is zero. Figure 8 and Fig. 9 It can be seen that at 140 GHz, the RMS spread is higher when it is 0 than at 28 GHz. This is because the diffuse scattering power is higher and the path loss increases, reducing the number of paths that can be detected by the receiver.

[0088] In summary, compared with the prior art, the technical solution of the present invention has the least statistical assumptions and can learn complex statistical relationships between variables; the present solution also uses the channel parameter data set to learn the channel propagation characteristics and implicit laws through machine learning algorithm training, and can more intelligently construct channel models to adapt to various scenarios and achieve better performance.

[0089] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A multi-frequency communication perception integrated channel modeling method based on a generative adversarial network, including a transmitter to a receiver, characterized in that: Also included are a transmitter-to-receiver channel, a base station, a mobile user device, and clustering data, wherein the base station and the mobile user device are distributed in a city macro area, the clustering data includes angles and delays of clusters, and vectors of random received power, angles, and delays propagated at different frequencies, and also includes the following steps: Step 1: Build a multi-frequency communication perception integrated channel model, i.e., the channel from transmitter to receiver, base stations and mobile user devices distributed in an urban macro area; Step 2: Simplify the model established in step 1; Step 3: The base station and the mobile user equipment are deployed model-assuming that they are randomly placed in an environment according to different base station spacings and densities, and then path parameters between the base station and the mobile user equipment are generated, and the MIMO channel is determined by the path parameters; Step 4: also includes ray tracing simulation, applying ray tracing simulation to generate training data; Step 5: Cluster the collected paths, and the rays of each frequency generated by ray tracing are pre-processed to identify clusters and determine clustering parameters; Step 6: Generate random cluster vector x by using neural network.

2. According to claim 1, a multi-frequency communication perception integrated channel modeling method based on a generative adversarial network is characterized in that: The base station and the mobile user equipment have reciprocity, so the base station and the mobile user equipment can be interchangeable.

3. According to claim 1, a multi-frequency communication perception integrated channel modeling method based on generative adversarial network is characterized by: According to the multi-frequency communication perception integrated channel model described in step 1, a channel model is built by a set of common L clusters on M frequencies, where each cluster Described by a parameter vector: in is the minimum path delay within the cluster, are the arrival inclination angle and azimuth angle within the average cluster, are the departure tilt angle and azimuth within the average cluster, respectively; Among them, the parameters is given by the frequency-dependent components: in is the arrival rms tilt and azimuth, is the departure rms tilt and azimuth, is the rms delay propagation, is the total received power; where the multi-frequency channel can be described by a set of L clusters: x = [x (1) ,…x (L) ], where x is a clustering vector.

4. According to claim 3, a multi-frequency communication perception integrated channel modeling method based on generative adversarial network is characterized in that: The simplified method in step 2 is: fix the number of clusters to L = 10. When there are less than L clusters at a certain frequency, if L0 < L, that is, all clusters of frequency m Set the total receive power If a path exists in a cluster at one frequency m but not at another frequency m′, then the total received power is taken as and The model x (l) Each cluster of has 6+6M parameters, and there are L clusters in total, so there are a total of (6+6M)L parameters in the cluster vector x.

5. The multi-frequency communication perception integrated channel modeling method based on generative adversarial network according to claim 4 is characterized by: Including the links, it also includes the condition vector u set in each link, so the three-dimensional distance vector between the transmitter and the receiver is expressed as: u=(d x ,d y ,d z ) The condition vector u may also include other parameters, including antenna height or base station type.

6. The multi-frequency communication perception integrated channel modeling method based on generative adversarial network according to claim 3 is characterized by: According to step 3, the distribution of the clustering vector x is modeled as a function of the link condition p(xu), that is, the conditional probability distribution p(xu) is modeled, wherein the conditional distribution is represented by a mapping x=g(u,z), wherein z is a random vector with a known distribution, and x=g(u,z) is the basis for cellular area evaluation. The base stations and mobile user devices are randomly placed in an environment according to the deployment model assumptions. The deployment determines the link condition u between each base station and the mobile user device. The path parameters between all base stations and mobile user devices are randomly generated by x=g(u,z). The path parameters are combined with the antenna array assumptions preset on both sides of the base station to determine all MIMO channels in the cellular network and calculate the signal-to-noise ratio or power distribution.

7. The multi-frequency communication perception integrated channel modeling method based on generative adversarial network according to claim 1 is characterized by: The step of applying ray tracing to generate training data in step 4 includes selecting the millimeter wave frequency band f1=28GHz and the terahertz frequency band f2=140GHz, importing an area of ​​550×380 square meters in Wireless Insite, manually placing 100 ground receiving base stations outside a 10m high building, and placing transmitting equipment terminals at 320 locations on a 1.5m street. The deployment result is 100×320=32000 links, and then running independent ray tracing simulations for these two frequencies, setting the maximum number of reflections to 6, the maximum number of diffractions to 1, and limiting the maximum number of paths of the link to 20. In addition, ray tracing collects paths with a power of at least -250dB.

8. The multi-frequency communication perception integrated channel modeling method based on generative adversarial network according to claim 1 is characterized by: The clustering process of step five includes using the K-means clustering process for the arrival tilt angle and azimuth, the departure tilt angle and azimuth, and the delay of each link in the group, using the silhouette coefficient to determine the optimal number of clusters for each link, and after identifying each cluster, determining the angle and delay propagation and power in each cluster to generate a clustering vector x for each link.

9. The multi-frequency communication perception integrated channel modeling method based on generative adversarial network according to claim 1 is characterized by: The process of generating a random clustering vector x using a neural network in step 6 uses a conditional Wasserstein generative adversarial network CWGAN-GP with gradient penalty. CWGAN-GP includes a generator G and a discriminator C. The generator G accepts a conditional vector u and a random input z~p(z), and generates a random path data vector x=G(u,z). The vector z is a potential vector, and the p(z) is a unit variance Gaussian vector. It also includes a potential dimension, and the potential dimension is set to 20. The discriminator C is trained to ensure that the generator G matches the data distribution. The discriminator C is used to distinguish between generated samples and real samples. The discriminator C and generator G network are trained through the loss function as follows: Where P g Is z~p(z) implicitly defines the generator distribution, P r is the distribution of data points.

Images