WO2023061303A1

WO2023061303A1 - Large-scale fading modeling and estimation method, system, and device, and storage medium

Info

Publication number: WO2023061303A1
Application number: PCT/CN2022/124172
Authority: WO
Inventors: 廖金龙; 汪波; 许靖; 吕星哉; 芮华
Original assignee: 中兴通讯股份有限公司
Priority date: 2021-10-13
Filing date: 2022-10-09
Publication date: 2023-04-20
Also published as: CN115967457A

Abstract

Embodiments of the present application relate to the technical field of communications, and disclose a large-scale fading modeling and estimation method, system, and device, and a storage medium. The channel large-scale fading modeling method comprises: obtaining path loss parameters and large-scale fading data in a target area, the target area being an area jointly covered by several base stations; performing ensemble learning according to the path loss parameters and the large-scale fading data, and obtaining a path loss prediction model; according to the path loss prediction model, decoupling path loss and shadow fading in the large-scale fading data, and obtaining shadow fading data; gridding the target area to obtain multiple grids, and dividing the shadow fading data into corresponding grids; and performing ensemble learning according to the shadow fading data in each grid, and obtaining a shadow fading prediction model of each grid.

Description

Modeling and estimation method, system, device and storage medium of large-scale fading

Cross References to Related Applications

This application claims priority to Chinese Patent Application No. 202111194300.4 filed on October 13, 2021, the entire content of which is hereby incorporated by reference.

technical field

The embodiments of the present application relate to the field of communication technologies, and in particular to a method, system, device and storage medium for modeling and estimating large-scale fading.

Background technique

Modeling based on large-scale fading of wireless channels is of great significance for studying channel characteristics and performance to optimize wireless communication systems. Among them, large-scale fading includes two aspects of path loss and shadow fading, and path loss reflects the long-distance received power. , while shadow fading reflects changes in power at the obstacle scale. In the traditional modeling method, the path loss is generally modeled by empirical formulas, but the empirical formulas contain many non-deterministic calculation parameters, the parameter values are strongly correlated with the scene, and the robustness and generalization of the model are poor. The practicability is not strong; for shadow fading, because it obeys the Gaussian distribution, the correlation coefficient is generally corrected on the basis of the Gaussian distribution to obtain the model, but the initial value of the correlation coefficient in this method has a large randomness when setting , the calculation accuracy of the correlation coefficient is also low, which cannot accurately reflect the actual correlation size, so the overall calculation error is difficult to control. Therefore, with the rise of artificial intelligence algorithms, various modeling methods based on historical data have emerged. The common idea is: in the modeling process, ignore shadow fading and treat large-scale fading as path loss; Or, after obtaining the historical data of path loss and historical data of shadow fading, the historical data of path loss and historical data of shadow fading are used to perform model training respectively to obtain the prediction model of path loss and the prediction model of shadow fading.

However, in the actual process, shadow fading has a greater impact, and ignoring shadow fading will lead to large errors in the large-scale fading predicted by the model, and will also lead to poor generalization of the model, making it impossible to obtain a reliable large-scale fading model; In actual measurement, the loss and shadow fading cannot be completely distinguished, and it is difficult to measure independently. Therefore, it is impossible to separately obtain the historical data of path loss and shadow fading for training, and the feasibility is low.

Contents of the invention

The main purpose of the embodiments of the present application is to propose a large-scale fading modeling and estimation method, system, device, and storage medium, so as to realize the two-way measurement of path loss and shadow fading without independent measurement of path loss and shadow fading. On the one hand, the large-scale fading is modeled, so that an accurate, reliable, and highly achievable large-scale fading prediction model can be obtained, and then the large-scale fading of the channel can be accurately estimated.

To achieve the above purpose, an embodiment of the present application provides a modeling method for channel large-scale fading, including: acquiring path loss parameters and large-scale fading data in a target area, where the target area is an area covered by several base stations; performing integrated learning according to the path loss parameter and the large-scale fading data to obtain a path loss prediction model; decoupling path loss and shadow fading in the large-scale fading data according to the path loss prediction model to obtain shadow fading data ; performing gridding on the target area, obtaining several grids and dividing the shadow fading data into corresponding grids; performing integrated learning according to the shadow fading data in each grid, A shadow fading prediction model of each grid is obtained.

In order to achieve the above purpose, the embodiment of the present application also proposes a method for estimating large-scale channel fading, including: obtaining the path loss parameters of the channel to be measured; inputting the path loss parameters into the path loss prediction model and the shadow fading prediction model respectively , to obtain the path loss prediction value output by the path loss prediction model and the shadow fading prediction value output by the shadow fading prediction model, wherein the path loss prediction model and the shadow fading prediction model are obtained through the channel as described above obtained by a modeling method of large-scale fading; the sum of the predicted value of path loss and the predicted value of shadow fading is used as the estimated value of large-scale fading of the channel to be measured.

In order to achieve the above purpose, an embodiment of the present application also proposes a modeling system for large-scale channel fading, including: a first acquisition module, configured to acquire path loss parameters and large-scale fading data in a target area, the target area It is an area covered by several base stations; the first training module is used for performing integrated learning on the path loss parameter and the large-scale fading data as training data to obtain a path loss prediction model; the decoupling module is used for according to the The path loss prediction model decouples the path loss and shadow fading in the large-scale fading data to obtain shadow fading data; the processing module is used to grid the target area, obtain several grids and convert the shadow fading The fading data is divided into the corresponding grids; the second training module is configured to perform integrated learning according to the shadow fading data in each grid to obtain a shadow fading prediction model for each grid.

In order to achieve the above purpose, the embodiment of the present application also proposes a system for estimating large-scale channel fading, including: a second acquisition module, used to obtain the path loss parameter of the channel to be measured; a prediction module, used to convert the path loss parameter The parameters are respectively input into the path loss prediction model and the shadow fading prediction model, and the path loss prediction value output by the path loss prediction model and the shadow fading prediction value output by the shadow fading prediction model are obtained, wherein the path loss prediction model and the shadow fading prediction model output The shadow fading prediction model is obtained by the above-mentioned channel large-scale fading modeling method; the result generation module is used to use the sum of the path loss prediction value and the shadow fading prediction value as the channel to be tested Large-scale fading estimates for .

In order to achieve the above purpose, an embodiment of the present application also proposes an electronic device, the device includes: at least one processor; and a memory connected to the at least one processor in communication; wherein, the memory stores information that can be Instructions executed by the at least one processor, the instructions are executed by the at least one processor, so that the at least one processor can execute the above-mentioned method for modeling channel large-scale fading, or execute the above-mentioned The estimation method of channel large-scale fading described above.

In order to achieve the above purpose, the embodiment of the present application also proposes a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned method for modeling channel large-scale fading is implemented, or, The method for estimating channel large-scale fading as described above is realized.

Description of drawings

One or more embodiments are exemplified by pictures in the accompanying drawings, and these exemplifications are not intended to limit the embodiments.

FIG. 1 is a flowchart of a modeling method for channel large-scale fading provided by an embodiment of the present application;

FIG. 2 is a flowchart of a method for estimating channel large-scale fading provided by another embodiment of the present application;

FIG. 3 is a schematic structural diagram of a modeling system for channel large-scale fading provided by another embodiment of the present application;

FIG. 4 is a schematic structural diagram of a system for estimating large-scale channel fading provided by another embodiment of the present application;

Fig. 5 is a schematic structural diagram of an electronic device provided by another embodiment of the present application.

Detailed ways

It can be seen from the background technology that large-scale fading includes two aspects of path loss and shadow fading. The current modeling methods for large-scale fading either treat large-scale fading as path loss to obtain an inaccurate model, or only focus on shadow Fading modeling has not yet provided a method of how to obtain shadow fading data that is difficult to measure directly, and the feasibility is low.

In order to solve the above problems, an embodiment of the present application provides a modeling method for channel large-scale fading, including: acquiring path loss parameters and large-scale fading data in a target area, where the target area is an area covered by several base stations; performing integrated learning according to the path loss parameter and the large-scale fading data to obtain a path loss prediction model; decoupling path loss and shadow fading in the large-scale fading data according to the path loss prediction model to obtain shadow fading data ; performing gridding on the target area, obtaining several grids and dividing the shadow fading data into corresponding grids; performing integrated learning according to the shadow fading data in each grid, A shadow fading prediction model of each grid is obtained.

The modeling method for channel large-scale fading proposed in the embodiment of this application, after using the large-scale fading data and path loss parameter training to obtain the path loss prediction model, decouples the path loss and shadow in the large-scale fading data through the path loss prediction model Obtain the shadow fading data, and use the shadow fading data to train the shadow fading prediction model, so that the path loss prediction model and the shadow fading prediction model can be integrated to obtain a complete large-scale fading prediction model. By using the path loss prediction model to decouple the path loss and shadow fading in large-scale fading data, avoiding direct measurement of shadow fading data, reducing the difficulty of obtaining shadow fading data, improving the feasibility of shadow fading modeling, and due to training The shadow fading data of the shadow fading prediction model is obtained according to the path loss prediction model. Therefore, even if there is a certain error in the path loss prediction model, this error will be accumulated to the shadow fading part for modeling processing through the shadow fading data. When the shadow fading When the model is constructed accurately, the superposition calculation results of path loss and shadow fading will be infinitely close to the real large-scale fading, ensuring the overall accuracy of large-scale fading, that is, using the path loss prediction model and shadow fading prediction model obtained above The true large-scale fading value of the channel can be accurately estimated.

In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that in each embodiment of the application, many technical details are provided for readers to better understand the application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can also be realized. The division of the following embodiments is for the convenience of description, and should not constitute any limitation to the specific implementation of the present application, and the embodiments can be combined and referred to each other on the premise of no contradiction.

The implementation details of the distributed system control method described in this application will be described in detail below in conjunction with specific embodiments. The following content is only the implementation details provided for easy understanding, and is not necessary for implementing this solution.

The first aspect of the embodiments of the present application provides a modeling method for large-scale channel fading. Referring to FIG. 1 , the flow of the modeling method for large-scale channel fading includes the following steps.

Step 101, acquiring path loss parameters and large-scale fading data in a target area, where the target area is an area covered by several base stations.

Specifically, the area covered by several base stations is selected as the target area according to actual needs. If it is necessary to model the large-scale fading of base station A, base station B, and base station C, the common coverage of base station A, base station B, and base station C can be selected. The area of the target area can also be selected as the target area in the area covered by base station A, base station B, and base station C, such as a rectangular area, and then each base station corresponding to the target area and these base stations in the target area The user equipment (UE) in the area collects, and uses the collected data to obtain path loss parameters and large-scale fading data.

More specifically, a rectangular area covered by multiple base stations is selected as the target area, and base station work parameter information of the multiple base stations and UE measurement data in the target area are collected, wherein the base station work parameter information includes: base station name eNB _n , base station longitude Lon _n , base station latitude Lat _n , base station antenna height he _n , base station transmit power TX _n , the number of base stations that jointly cover the target area n=1, 2, ..., N; UE measurement data include: base station eNB The UE height hun _m covered by _n , the longitude Lon _n ^m of the UE, the latitude Lat _n ^m of the UE, the received power rsrp _n ^m of the UE, and the center frequency point f _n ^m , the number n=1 of the base stations ^that jointly cover the target area, 2, ..., N, m=1, 2, ..., Mn, m is the mth UE covered by the base station numbered n, Mn is the total number of UEs covered by the base station numbered n, and Lon _n ^m , Lat _n , he _n , f _nm ^, hun ^m , Lon _n _m ^, and Lat _n ^m k can be used as path loss parameters, and TX _n -rsrp _n ^m can be used as large-scale fading data.

It should be noted that the base station industrial parameter information and UE measurement data can be collected by professional equipment, or can be obtained by collecting user terminal measurement reports and UE location distribution and other information on the base station side. This embodiment does not limit the acquisition method.

In one example, the rectangular area covered by 5 base stations in the field is selected as the target area, and the industrial parameter information of the 5 base stations and the measurement report (Measurement Report, MR) report of the UE covered by the 5 base stations are collected as UE measurement data . The obtained data includes: base station name eNB ₁ ~eNB ₅ ; base station latitude and longitude (Lon ₁ , Lat ₁ ), ..., (Lon ₅ , Lat ₅ ), base station antenna height he ₁ , ..., he ₅ , and base station transmit power TX ₁ ,...,TX ₅ ; covered UE height hu ₁ ^M ,...,hu ₅ ^M ; UE latitude and longitude (Lon ₁ ^M , Lat ₁ ^M ),..., (Lon ₅ ^M , Lat ₅ ^M ); and UE received power rsrp ₁ ^M ,..., rsrp ₁ ^M ; center frequency points f ₁ ^M ,..., f ₅ ^M , among which, a total of 1 million pieces of UE measurement data are collected for each base station, that is, M=1000000, and 5 million pieces of training data can be obtained .

Step 102, performing integrated learning according to path loss parameters and large-scale fading data to obtain a path loss prediction model.

Specifically, the path loss parameter and its corresponding large-scale fading data are used as a piece of training data to obtain a training set containing several training data, and then use the training set to train the preset integrated model, and the trained integrated model obtained is It is the path loss prediction model, in which the large-scale fading data is the training target and can be used as a supervisory signal in the training process.

More specifically, take Lon _n ^m , Lat _n , he _n , f _n ^m , hu _n ^m , Lon _n ^m and Lat _n ^m and TX _n -rsrp _n ^m as a piece of training data, n=1, 2, ... , N; m=1, 2,..., M _n , n is the number of the base stations that jointly cover the target area, m is the mth UE covered by the base station UE with the number n, and Mn is the number covered by the base station UE with the number n The total number of UEs, therefore, can be obtained at this time

According to the preset integrated model type, the training data is processed to obtain one or more training sets, so as to use the obtained training set to train the preset integrated model, and obtain the trained integrated model as path loss prediction Model.

This embodiment does not limit the type of integrated model, which may be a random forest model based on a tree model, or an integrated algorithm based on a neural network based on a bagging algorithm, a boosting algorithm, etc. The integrated model formed, etc., will not be described here one by one.

It is worth mentioning that this embodiment can implement modeling based on a variety of model types, so as to achieve the purpose of extending the modeling method, so that the modeling method provided by this embodiment can be applied to various situations, thereby enhancing the Adaptability and practicality of modeling methods.

It should be noted that in order to improve the accuracy of the path loss prediction model, the training set can also be preprocessed before training. Delete the entire training data, smooth the singular data, correct the abnormal data, and fill the missing data. Of course, the above is only a specific example of the preprocessing method of the training set, and other processing methods that can improve the data of the training set may also be used, and details will not be repeated here.

In order to facilitate those skilled in the art to better understand the above step 102, the path loss prediction model will be described below as an example of a random forest model composed of a classification and regression tree (Classification and Regression Tree, CRAT) as a basic model.

At this point, step 102 may include: generating an initial path loss training set comprising several training data according to the path loss parameters and large-scale fading data, wherein the large-scale fading data is the training target in the training data; from the initial path loss The training data is extracted with replacement in the training set until the extracted training data constitutes a preset number of path loss training sets. The capacity of the path loss training set is the same as that of the initial path loss training set; Assuming the number of CRAT, the first random forest model is obtained as the path loss prediction model.

In particular, if the base model is a tree model, no data standardization is required, but for other types of base models, the data needs to be standardized, and the standardization method can be Min-Max (minimum-maximum) standardization, etc.

More specifically, generating an initial path loss training set consisting of several training data according to path loss parameters and large-scale fading data may include: Lon _n ^m , Lat _n , he _n , f _n ^m , hu _n ^m , Lon _n ^m and Lat _n ^m and TX _n -rsrp _n ^m are used as a piece of training data, n=1, 2,..., N; m=1, 2,..., M _n , n is the number of base stations that jointly cover the target area , m is the mth UE covered by the base station UE numbered n, and Mn is the total number of UEs covered by the base station numbered n, therefore, it can be obtained at this time

pieces of training data.

Extract training data from the initial path loss training set with replacement until the extracted training data constitute a preset number of path loss training sets including: when training a random forest model, there is no need to divide the training set and verification set in advance, but in The data size is

In the initial data set S of , each time there is a random sampling with replacement, the total number of samples in the training set with the same initial path loss is

training data to form a path loss training set S _i , and repeat the extraction process for K rounds to form K path loss training sets S ₁ ,..., S _k , so as to construct K tree models for the random forest model. Path loss training set. Among them, a total of

According to the probability calculation, about 36.8% of the training data cannot be selected, and this part of the unselected training data is uniformly divided into the verification set Y _i generated during the i-th round of sampling. In this way, in each different training set S _i , a part of training data is randomly selected for training, that is, in K rounds of training, the input features used in each round are part of the training data randomly selected.

Use the path loss training set to train a preset number of CRATs respectively, and obtain the first random forest model as the path loss prediction model includes: using the mean square error (Mean-Square Error, MSE) function as the objective function, using K path loss training sets respectively Train K CRATs, and import the corresponding verification set Y _i into CRATs for verification to obtain verification errors after one training session. Model robustness verification can also be performed until the obtained verification errors are relatively stable, that is, until the random forest model has a relatively When the robustness is good, the training is stopped and the modeling is successful. Then K CARTs are combined by calculating the average value of the output results of K CARTs to obtain a random forest model, that is, the output result of the random forest model is the average value of the output results of K CARTs. Among them, the depth of CRAT generated by each round of training is not limited, and in order to speed up the training, parallel training can also be used for training.

In one example, after obtaining 5 million pieces of training data, it is preprocessed to remove training data that does not meet the requirements, and an initial path loss training set containing 4.8 million pieces of training data is obtained. Then, each time in the initial training set Random sampling with replacement, a total of 4.8 million pieces of training data are selected to form a path loss training set S _i , and this process is repeated for a total of 20 rounds to form a total of 20 training path loss training sets S ₁ ,..., S ₂₀ , In order to obtain a training set of 20 CART trees, a total of 4.8 million samples are taken in the i-th round of sampling to generate 4.8 million data to form a path loss training set. According to the probability calculation, there are about 36.8% of the number of about 1.7 million training data Unable to be selected, and do not exist in the path loss training set, such unselected training data are uniformly divided into the verification set Y _i generated during the i-th round of extraction. Then in each different path loss training set S _i , randomly select 2/3 features from each piece of training data, that is, 4 features for training, such as in [Lo _n n, Lat _n , he _n , f _n ^m , hu _n ^m , Lon _n ^m , Lat _n ^m ] randomly select 4 features Lo _n n, Lat _n , he _n and f _n ^m . During the 20 rounds of training, the input features used in each round of training may be different, or some rounds may randomly get the same input features. Then use the path loss training set S _i to train a CART respectively, in which, each time the left and right subtrees are divided, traverse the feature j and the corresponding division point s, and find j and s that minimize the sum of the loss function calculation values of the left and right subtrees :

Min _{j, s} (Min _c1 Loss(y ₁ -c ₁ )+Min _c2 Loss(y ₂ -c ₂ ))

Among them, y1 is all the training data points divided into the left subtree according to the division point s, c1 is the large-scale fading average value of all the training data points divided into the left subtree, and y2 is the data points divided into the right subtree according to the division point s All training data points, c2 is the large-scale fading average of all training data points divided into the right subtree. Loss represents the mean square error MSE function, which can be:

Among them, X is the number of all training data points of y1, and y1 _x is the large-scale fading value of the xth training data point. The 4.8 million pieces of training data in the S _i- th path loss training set are cyclically divided according to the above method until they are all divided into the leaf nodes of the CART, that is, the training of a single CART is completed. The 20 training path loss training sets are trained to generate 20 CARTs, and finally the average of the output results of the 20 CARTs is used as the output of the random forest model. At this time, it is also necessary to import 1.7 million training data in each round of verification set Y _i into the model for verification. If the average absolute percentage error is used to measure the final error, that is, to calculate the average value of the absolute value of the percentage error of 1.7 million training data, then 20 error results can be obtained for the path cycle training set generated by 20 rounds of sampling, as shown in Table 1 Show.

Table 1

It can be seen from Table 1 that the average value of 20 error results is 6.6%. If the large-scale fading modeling is directly based on the neural network, when 3.1 million pieces of training data constitute the training set and 1.7 million pieces of training data constitute the verification set, the error result obtained is 18.4%. Apparently, the method provided in this embodiment has smaller errors than the existing modeling methods.

In addition, in order to prevent the contingency of the results caused by the randomness contained in the model, the model robustness test was carried out, and the above steps were repeated for 30 tests. The overall average absolute percentage error obtained was all about 6.8%, and the error distribution range was [7.1%, 6.5%], it can also be seen that the established random forest model has good robustness.

It is worth mentioning that large-scale fading can be regarded as shadow fading superimposed on the path loss. Compared with path loss, the statistical law of shadow fading obeys the Gaussian distribution with a mean value of 0. Therefore, large-scale fading can be regarded as Gaussian noise The path loss of , where Gaussian noise is the shadow fading. The random forest model can have a good ability to resist noise interference due to the introduction of feature randomness and training data randomness in the training process. On the one hand, the random forest does not use the gradient descent method to iterate. On the other hand, due to the introduction of randomness in the selection of training data and the randomness of feature selection in the training data, it can greatly reduce the generation of such training data in the training set. In addition, the model built based on the idea of ensemble learning can well prevent overfitting. At this time, the shadow fading part in the large-scale fading data can be processed as Gaussian noise, so as to better fit the corresponding relationship between the path loss feature quantity and the real path loss, that is, obtain an accurate path loss prediction model.

It should be noted that the variables in the commonly used path loss empirical calculation model include the three-dimensional (3-dimension, 3D) distance from the UE to the base station, the center frequency, etc., and the empirical model contains many parameters that need to be corrected and calculated based on actual data. Therefore, it is difficult to identify accurate empirical model parameters. Therefore, this embodiment selects parameters related to path loss as path loss parameters, including: longitude Lon _n and latitude Lat _n of the base station; base station antenna height he _n , center frequency point f _n ^m , UE height hu _n ^m , UE longitude Lon _n ^m and latitude Lat _n ^m , a total of 7-dimensional feature quantities [Lo _n n, Lat _n , he _n , f _n ^m , hu _n ^m , Lon _n ^m , Lat _n ^m ], since large-scale fading includes path loss and shadow fading, both of which are difficult to distinguish and cannot be measured separately. Therefore, the complete large-scale fading can be used as the target quantity TX _n -rsrp _n ^m first, and a training data set is constructed as : [Lo _n n, Lat _n , he _n , f _n ^m , hu _n ^m , Lon _n ^m , Lat _n ^m , TX _n -rsrp _n ^m ], each piece of training data contains 8-dimensional data, and the first 7 dimensions are features The eighth dimension is the target quantity. Using this data set, all base stations can be directly included in the modeling together, without having to model each base station separately.

It should also be noted that when the verification set is put into the model for verification, it is found that the error still exists, but it is much lower than the model with poor anti-noise ability such as neural network. Wherein, the error includes the prediction error of the path loss and the shadow fading value. After the robustness verification is completed, all data sets S _i are imported into the random forest model, and the obtained predicted value is regarded as the path loss, and then the difference between the corresponding large-scale fading and the path loss is exported as shadow fading data , used for the next step of shadow fading model construction. Although the path loss prediction model has an error caused by shadow fading, this error will be accumulated to the shadow fading part for modeling processing through the shadow fading data. When the shadow fading model is constructed accurately, it will make the superposition calculation of path loss and shadow fading The result is infinitely close to the real large-scale fading, which ensures the overall accuracy of large-scale fading.

In step 103, the path loss and shadow fading in the large-scale fading data are decoupled according to the path loss prediction model to obtain shadow fading data.

Specifically, the path loss parameters are input into the path loss prediction model, so that the path loss prediction model can be used to predict the path loss according to the path loss parameters, and then the path loss value output by the path loss prediction model is compared with the large-scale fading data, and the The result is used as the shadow fading value, so as to realize the decoupling of path loss and shadow fading in large-scale fading data.

Step 104, performing gridding on the target area to obtain several grids and dividing the shadow fading data into corresponding grids.

Specifically, divide the target area according to the preset grid shape to obtain several grids, determine the position of the user terminal corresponding to the shadow fading data in the target area, and determine the grid to which the position belongs as the shadow fading data The grid to which it belongs. This embodiment does not limit the shape of the preset grid, which may be a rectangle, a square, a regular hexagon, etc., and details will not be repeated here.

In one example, starting from the lower left vertex of the rectangular target area, the target area is sequentially divided and numbered according to a square grid with a side length of 50 meters, and is divided into T grids, and each grid is represented as gridt, Wherein, grid numbers t=1, 2, 3, . . . , T. Classify the shadow fading data calculated in the third step according to the landing point of the UE in the grid area to which the UE measurement data belongs, and express the shadow fading data of each base station in the grid gridt as Dt ₁ ,..., Dt ₅ , 1- 5 is the base station number.

It should be noted that, according to the characteristic that the shadow fading data obeys the normal distribution, the path loss prediction model can be tested by using the shadow fading data: calculate the average value of the shadow fading data in each grid, and check whether there is an average value of at least one grid value exceeds the preset threshold, if so, retrain the path loss prediction model. Specifically, it is to find the statistical distribution parameters of the shadow fading Dt _n of each base station in each grid t, assuming that it obeys the Gaussian distribution, and calculate the mean value and variance, then the shadow fading distribution of the base station eNB _n in the grid t is expressed as

Judging whether the mean value exceeds a preset threshold, such as judging whether it exceeds 0.01. Generally speaking, the closer the mean value is to 0, the more reliable the path loss prediction model is. In particular, the above-mentioned

steps

103 and 104 and the above-mentioned mean value verification process can be performed during the verification process of the training model in step 103 .

It is worth mentioning that, according to the characteristic that shadow fading obeys normal distribution, the accuracy of the model is tested, which improves the accuracy of the path loss prediction model.

Step 105, performing integrated learning according to the shadow fading data in each grid to obtain a shadow fading prediction model for each grid.

Specifically, the shadow fading value of each base station located in the same grid is used as a piece of training data to obtain a training set containing several training data, and then use the training set to train the preset ensemble model, and the trained ensemble The model is the shadow fading prediction model, in which the shadow fading value of a base station is sequentially selected as the training target and as the supervisory signal in the training process.

More specifically, for the shadow fading data of N base stations in a certain grid, N data sets of multi-base station shadow fading feature input and single base station shadow fading feature output are respectively derived. Taking the shadow fading Dt _N of the predicted target base station number N in the grid gridt as an example to illustrate specifically, the characteristic input is: the shadow fading [Dt1 , Dt2,...DtN-1], the feature output is: the shadow fading Dt _N of the base station number N, therefore, its shadow fading training set is [Dt ₁ , Dt ₂ ,...Dt _N-1 , Dt _N ], the first (N-1) dimension is the input, and the Nth dimension is the output. Similarly, for predicting the shadow fading Dt _n of target base station number n, the data set is [Dt ₁ , Dt ₂ , ..., Dt _N , Dt _n ], so that N shadow fading training sets are constructed for a certain grid. For the shadow fading of each base station in the grid with sufficient data, a shadow fading training set is constructed for model training, and the shadow fading prediction model of each grid is obtained.

In order to facilitate those skilled in the art to better understand the above step 105, the shadow fading prediction model is a random forest model based on the CRAT model as an example for illustration. At this time step 105 includes: sequentially select a base station as the target base station, take the shadow fading data of the target base station in each grid as the training target, generate a shadow fading training set based on the target base station for each grid, and use the data corresponding to the same The shadow fading training set of the grid trains a second preset number of CRATs respectively to obtain a second random forest model as the shadow fading prediction model of the corresponding grid, and the second preset number is the total number of base stations.

In an example, for the shadow fading of 5 base stations in the grid, 5 data sets of multi-base station shadow fading feature input and single base station shadow fading feature output are respectively derived. Taking the base station whose prediction number is 1 in the grid gridt as the target base station as an example for specific illustration, the corresponding feature input in the shadow fading training set is: the shadow fading [Dt2 , Dt3, Dt4, Dt5], the supervision signal is the shadow fading Dt1 of the target base station, therefore, the training data in the shadow fading training set is [Dt2, Dt3, Dt4, Dt5, Dt1], the first 4 dimensions are input, and the fifth dimension is output.

It should be noted that, due to insufficient training data and other reasons, there may be cases of training failure. At this time, it is necessary to generalize the shadow fading prediction model that has been successfully trained to obtain a shadow fading prediction model that has failed to train. Specifically, after step 105 is executed, it is also necessary to detect whether there is a grid that fails ensemble learning. If there is, the shadow fading prediction model of the grid that succeeds ensemble learning is generalized according to a preset method to obtain ensemble learning failure The shadow fading prediction model of the grid, wherein the preset method includes: screening the shadow fading prediction model of the grid whose distribution of shadow fading data is closest to the Gaussian distribution, or the ensemble learning of the grid adjacent to the ensemble learning failure A weighted combination of shadow-fading prediction models for successful grids. Specifically, it is realized through generalization: if the base station connected to the UE is independent of other base stations and has no correlation, then the shadow fading random value of the base station is generated according to an independent Gaussian distribution; if the base station connected to the UE is related to other base stations, other base stations The base station already has UE measurement data, and has established a shadow fading random forest model, then the UE measurement data of other base stations are decoupled from the path loss and shadow fading through the path loss random forest model, and then the decoupling The obtained shadow fading is input into the shadow fading random forest model to solve the shadow fading of the UE to be predicted; if the base station connected to the UE is related to other base stations, there is not enough shadow fading training set to train the shadow fading random forest model, but some UEs Measurement data: Select the applicable random forest model in other grids to solve according to the generalization method in step 7; if the base station connected to the UE is related to other base stations, but there is no shadow fading random forest model due to the lack of measurement data of many UEs, It cannot be solved without the initial value of shadow fading: first, select the existing shadow fading random forest model in other grids in the generalization method mentioned above, and then randomly assign the initial shadow fading to the UE with missing measurement data, and use the shadow fading random The forest model updates iterative shadow fading in sequence, and sets the number of iterations or the error with the last calculation result as the convergence condition until the iterative calculation is completed.

Among them, when the base station connected to the UE is related to other base stations, other base stations have UE measurement data, and have established a shadow fading random forest model, it is assumed that a certain grid in this situation is selected for verification and the to-be-predicted The shadow fading values and model prediction results obtained by decoupling 10 UEs are shown in Table 2.

Table 2

It can be seen from Table 2 that the predicted value is relatively close to the target value obtained by decoupling, and the average absolute error of predicted shadow fading is 4.5dB.

The base station connected to the UE is related to other base stations, there is not enough shadow fading training set to train the shadow fading random forest model, but there are some UE measurement data, assuming that a certain grid in this situation is selected for verification and the 10 to be predicted The shadow fading values and model prediction results obtained by decoupling UEs are shown in Table 3.

table 3

It can be seen from Table 3 that the mean value of the average absolute error of predicted shadow fading is 5.4dB.

However, in the case that it is related to other base stations, but the measurement data of many UEs is missing, so there is no random forest model of shadow fading, and there is no initial value of shadow fading, so it cannot be solved, choose a grid with sufficient data, assuming that the loss in the grid Multi-data, and then simulate the generation of data and then compare and check the error. At this time, we must first solve the problem of model selection. Since the statistical distribution of shadow fading cannot be obtained due to missing data, the second generalization method in step 7 is selected, that is, the random forest model of adjacent grids is selected for prediction. Then it is necessary to solve the input problem of the model data by randomly assigning initial values and then iteratively calculating until the results converge.

In this embodiment, assuming that the measurement data of base stations numbered eNB ₁ and eNB ₂ are complete, and the shadow fading data of base stations numbered eNB ₃ , eNB ₄ , and eNB ₅ are all missing, the shadow fading data of eNB4 and eNB5 base station data are initialized as 0dB, 1dB. Then use the shadow fading random forest model of adjacent grids to start cyclic and iterative calculations to update the shadow fading until the final predicted value is obtained. For example: first use the initial value of the base station [eNB ₁ , eNB2, eNB4, eNB5] to update the shadow fading of eNB3, and then use the updated eNB ₃ shadow fading value combined with other values to form [eNB ₁ , eNB ₂ , eNB ₃ , eNB ₅ ] Update eNB4, and then update _eNB5 sequentially until the change between the current update value and the last update value is less than 2dB, then it is determined as the final value. Using this calculation method can make full use of the correlation of base station shadow fading to limit the randomness of generating shadow fading, and can improve the accuracy of predicting shadow fading through the random forest model. Finally, compared with the target value obtained through decoupling, the results are shown in Table 4.

Table 4

It can be seen from Table 4 that the mean values of the average absolute errors of shadow fading predicted by the three base stations are 6.8dB, 7.1dB, and 6.2dB, respectively.

It is worth mentioning that, in the case that the shadow fading data corresponding to some grids is insufficient or incomplete, so that the shadow fading prediction model cannot be successfully trained, a relatively accurate and reliable shadow fading prediction model can still be generated for each grid.

In addition, it can be seen from the data in the above table that under various circumstances, the modeling method based on this embodiment can calculate and predict shadow fading with high precision. And since the path loss model errors are all accumulated in shadow fading, the absolute error of shadow fading calculated above is the overall large-scale fading error. The average absolute percentage error is less than 3%, and the accuracy is high.

On the other hand, the embodiments of the present application also provide a method for estimating large-scale channel fading. The flow of the method for estimating large-scale channel fading is shown in FIG. 2 , including the following steps.

Step 201, acquire the path loss parameter of the channel to be tested.

In this embodiment, the path loss parameters include base station operating parameter information and UE measurement data in the target area.

In an example, the path loss parameter may include: base station name, base station longitude, base station latitude, base station antenna height, UE height covered by the base station, UE longitude, UE latitude, and center frequency point.

In step 202, the path loss parameters are respectively input into the path loss prediction model and the shadow fading prediction model, and the path loss prediction value output by the path loss prediction model and the shadow fading prediction value output by the shadow fading prediction model are obtained.

It should be emphasized that the path loss prediction model and shadow fading prediction model are obtained through the modeling method of channel large-scale fading as described in the above embodiments.

It is worth mentioning that the above-mentioned channel large-scale fading modeling method uses the path loss prediction model to decouple the path loss and shadow fading in the large-scale fading data, so as to train the shadow fading prediction model to realize the path loss and shadow fading. Fading is modeled separately to avoid the problem of inaccurate models caused by single modeling. At the same time, since the shadow fading data for training the shadow fading prediction model is obtained according to the path loss prediction model, even if there is a certain error in the path loss prediction model, this The error will be accumulated to the shadow fading part through the shadow fading data for modeling processing. When the shadow fading model is constructed accurately, the superposition calculation result of the path loss and shadow fading will be infinitely close to the real large-scale fading, ensuring that the large-scale fading Overall accuracy, so that the path loss prediction model and shadow fading prediction model can be used to accurately estimate the real large-scale fading value of the channel, which is convenient for subsequent analysis and processing.

In step 203, the sum of the path loss prediction value and the shadow fading prediction value is used as the large-scale fading estimation value of the channel to be tested.

It should be noted that large-scale fading includes path loss and shadow fading. During the modeling process, due to the different characteristics of the two, for example, shadow fading obeys Gaussian distribution while path loss does not have this characteristic. Therefore, in order to accurately analyze large-scale Fading modeling needs to be carried out separately from path loss and shadow fading, but in the actual process of estimating large-scale fading, path loss and shadow fading need to be viewed as a whole, that is, the estimated value of large-scale fading needs to include path Loss and shadow fading.

In addition, it should be understood that the division of steps in the above methods is only for clarity of description, and may be combined into one step or split into multiple steps during implementation. As long as the same logical relationship is included, all Within the scope of protection of this patent; adding insignificant modifications or introducing insignificant designs to the algorithm or process, but not changing the core design of the algorithm and process are all within the scope of protection of the patent.

On the other hand, the embodiments of the present application also provide a modeling system for large-scale channel fading. The structure of the modeling system for large-scale channel fading is shown in FIG. 3 , including the following modules.

The first obtaining module 301 is configured to obtain path loss parameters and large-scale fading data in a target area, where the target area is an area covered by several base stations.

The first training module 302 is configured to use path loss parameters and large-scale fading data as training data for integrated learning to obtain a path loss prediction model.

The decoupling module 303 is configured to decouple path loss and shadow fading in the large-scale fading data according to the path loss prediction model to obtain shadow fading data.

The processing module 304 is configured to grid the target area, obtain several grids and divide the shadow fading data into corresponding grids.

The second training module 305 is configured to perform integrated learning according to the shadow fading data in each grid to obtain a shadow fading prediction model for each grid.

It is not difficult to find that this embodiment is a system embodiment corresponding to the embodiment of the modeling method for large-scale channel fading, and this embodiment can be implemented in cooperation with the embodiment of the modeling method for large-scale channel fading. The relevant technical details mentioned in the embodiment of the method for modeling large-scale channel fading are still valid in this embodiment, and will not be repeated here to reduce repetition. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the embodiment of the modeling method for channel large-scale fading.

It is worth mentioning that all the modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, or a part of a physical unit, or multiple physical units. Combination of units. In addition, in order to highlight the innovative part of the present application, units that are not closely related to solving the technical problem proposed in the present application are not introduced in this embodiment, but this does not mean that there are no other units in this embodiment.

On the other hand, the embodiments of the present application also provide a system for estimating large-scale channel fading. The structure of the system for estimating large-scale channel fading is shown in FIG. 4 , including the following modules.

The second acquiring module 401 is configured to acquire the path loss parameter of the channel to be tested.

The prediction module 402 is configured to input the path loss parameters into the path loss prediction model and the shadow fading prediction model respectively, and obtain the path loss prediction value output by the path loss prediction model and the shadow fading prediction value output by the shadow fading prediction model, wherein the path loss prediction Model and shadow fading prediction model is obtained through the modeling method of channel large-scale fading as described above.

The result generation module 403 is configured to use the sum of the path loss prediction value and the shadow fading prediction value as the large-scale fading estimation value of the channel to be tested.

It is not difficult to find that this embodiment is a system embodiment corresponding to the embodiment of the method for estimating large-scale channel fading, and this embodiment can be implemented in cooperation with the embodiment of the method for estimating large-scale channel fading. The relevant technical details mentioned in the embodiment of the method for estimating large-scale channel fading are still valid in this embodiment, and will not be repeated here to reduce repetition. Correspondingly, the related technical details mentioned in this embodiment can also be applied in the embodiment of the method for estimating large-scale channel fading.

On the other hand, the embodiment of the present application also provides an electronic device, as shown in FIG. 5 , including: at least one processor 501; and a memory 502 communicatively connected to the at least one processor 501; wherein, the memory 502 stores Instructions that can be executed by at least one processor 501. The instructions are executed by at least one processor 501, so that at least one processor 501 can execute the method for modeling channel large-scale fading described in any one of the above method embodiments.

Wherein, the memory 502 and the processor 501 are connected by a bus, and the bus may include any number of interconnected buses and bridges, and the bus connects one or more processors 501 and various circuits of the memory 502 together. The bus may also connect together various other circuits such as peripherals, voltage regulators, and power management circuits, all of which are well known in the art and therefore will not be further described herein. The bus interface provides an interface between the bus and the transceivers. A transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing means for communicating with various other devices over a transmission medium. The data processed by the processor 501 is transmitted on the wireless medium through the antenna, and further, the antenna also receives the data and transmits the data to the processor 501 .

Processor 501 is responsible for managing the bus and general processing, and may also provide various functions including timing, peripheral interface, voltage regulation, power management and other control functions. And the memory 502 may be used to store data used by the processor 501 when performing operations.

On the other hand, the embodiments of the present application also provide a computer-readable storage medium storing a computer program. The above method embodiments are implemented when the computer program is executed by the processor.

That is, those skilled in the art can understand that all or part of the steps in the method of the above-mentioned embodiments can be completed by instructing related hardware through a program, the program is stored in a storage medium, and includes several instructions to make a device ( It may be a single-chip microcomputer, a chip, etc.) or a processor (processor) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. .

Those of ordinary skill in the art can understand that the above-mentioned embodiments are specific embodiments for realizing the present application, and in practical applications, various changes can be made to it in form and details without departing from the spirit and spirit of the present application. scope.

Claims

A modeling method for channel large-scale fading, including:

Acquiring path loss parameters and large-scale fading data in a target area, where the target area is an area covered by several base stations;

performing integrated learning according to the path loss parameter and the large-scale fading data to obtain a path loss prediction model;

Decoupling path loss and shadow fading in the large-scale fading data according to the path loss prediction model to obtain shadow fading data;

Perform gridding on the target area to obtain several grids and divide the shadow fading data into corresponding grids;

An integrated learning is performed according to the shadow fading data in each of the grids to obtain a shadow fading prediction model of each of the grids.
The modeling method for channel large-scale fading according to claim 1, wherein said integrated learning is performed according to the shadow fading data in each said grid, and after obtaining the shadow fading prediction model of each said grid, The method also includes:

If it is detected that there is a grid that failed in integrated learning, generalize the shadow fading prediction model of the grid that failed in integrated learning according to a preset method, and obtain the shadow of the grid that failed in integrated learning Fading prediction model, wherein the preset method includes: screening the shadow fading prediction model of the grid whose distribution of the shadow fading data is closest to a Gaussian distribution, or, for the grid that fails to integrate learning The shadow fading prediction models of the grids whose adjacent ensemble learning is successful are weighted and combined.
The method for modeling channel large-scale fading according to claim 1 or 2, wherein the target area is gridded to obtain several grids and the shadow fading data is divided into corresponding grids After, the method also includes:

calculating the average value of the shadow fading data in each of the grids;

If it is detected that the average value of at least one grid exceeds a preset threshold, the path loss prediction model is retrained.
The method for modeling channel large-scale fading according to claim 3, wherein the base model of the path loss prediction model and the shadow fading prediction model is a tree model or a neural network.
The method for modeling channel large-scale fading according to claim 4, wherein, when the base model of the path loss prediction model is a classification regression tree CRAT, the method is performed according to the path loss parameters and the large-scale fading data Integrated learning to obtain a path loss prediction model, including:

generating an initial path loss training set including several training data according to the path loss parameter and the large-scale fading data, wherein the large-scale fading data is a training target in the training data;

Extract the training data from the initial path loss training set with replacement until the extracted training data forms a preset number of path loss training sets, the capacity of the path loss training set is the same as the initial path loss The capacity of the training set is the same;

Using the path loss training set to respectively train the preset number of CRATs to obtain a first random forest model as the path loss prediction model.
The method for modeling channel large-scale fading according to claim 4, wherein when the base model of the shadow fading prediction model is CRAT, the integrated learning is performed according to the shadow fading data in each grid, Obtain the shadow fading prediction model of each grid, including:

Sequentially select one of the base stations as the target base station, use the shadow fading data of the target base station in each grid as a training target, and generate a shadow fading training set based on the target base station for each grid ;

Use the shadow fading training set corresponding to the same grid to train a second preset number of CRATs to obtain a second random forest model as the shadow fading prediction model of the corresponding grid, the second preset The quantity is the total quantity of the base stations.
The method for modeling channel large-scale fading according to claim 1, wherein said target area is gridded to obtain several grids and divide said shadow fading data into corresponding grids , including:

dividing the target area according to a preset grid shape to obtain several grids;

determining the location of the user terminal corresponding to the shadow fading data in the target area;

determining the grid to which the position belongs as the grid to which the shadow fading data belongs.
A method for estimating channel large-scale fading, comprising:

Obtain the path loss parameter of the channel to be tested;

Inputting the path loss parameters into the path loss prediction model and the shadow fading prediction model respectively, to obtain the path loss prediction value output by the path loss prediction model and the shadow fading prediction value output by the shadow fading prediction model, wherein the path The loss prediction model and the shadow fading prediction model are obtained by the modeling method of channel large-scale fading as described in any one of claims 1 to 7;

The sum of the path loss prediction value and the shadow fading prediction value is used as the large-scale fading estimation value of the channel to be tested.
A modeling system for channel large-scale fading, including:

The first acquisition module is configured to acquire path loss parameters and large-scale fading data in a target area, where the target area is an area covered by several base stations;

A first training module, configured to use the path loss parameter and the large-scale fading data as training data for integrated learning to obtain a path loss prediction model;

A decoupling module, configured to decouple path loss and shadow fading in the large-scale fading data according to the path loss prediction model, to obtain shadow fading data;

A processing module, configured to grid the target area, obtain several grids and divide the shadow fading data into corresponding grids;

The second training module is configured to perform integrated learning according to the shadow fading data in each of the grids to obtain a shadow fading prediction model of each of the grids.
An estimation system for channel large-scale fading, including:

The second obtaining module is used to obtain the path loss parameter of the channel to be tested;

A prediction module, configured to input the path loss parameters into a path loss prediction model and a shadow fading prediction model respectively, to obtain a path loss prediction value output by the path loss prediction model and a shadow fading prediction value output by the shadow fading prediction model, Wherein, the path loss prediction model and the shadow fading prediction model are obtained by the modeling method of channel large-scale fading as described in any one of claims 1 to 7;

A result generation module, configured to use the sum of the path loss prediction value and the shadow fading prediction value as the large-scale fading estimation value of the channel to be tested.
An electronic device comprising:

at least one processor; and,

a memory communicatively coupled to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, the instructions are executed by the at least one processor, so that the at least one processor can perform the operation described in any one of claims 1 to 7 The modeling method of the large-scale fading of the channel described above, or, the estimation method of the large-scale fading of the channel as described in claim 8 is performed.
A computer-readable storage medium storing a computer program, wherein, when the computer program is executed by a processor, the method for modeling channel large-scale fading according to any one of claims 1 to 7 is realized, or, The method for estimating channel large-scale fading as claimed in claim 8.