CN114611388B - Wireless channel characteristic screening method based on artificial intelligence - Google Patents

Abstract

The invention relates to a wireless channel feature screening method based on artificial intelligence, and belongs to the technical field of wireless channel modeling. The present invention conducts feature screening of wireless channel model based on artificial intelligence by summarizing the full set of features involved in the wireless channel model. The efficient machine learning model depends on the strong correlation between the input variables and the problem target, so as to ensure the reliability of the wireless channel model. On the premise, the computational time complexity is reduced.

Description

A wireless channel feature screening method based on artificial intelligence

technical field

The invention belongs to the technical field of wireless channel modeling, and in particular relates to a wireless channel feature screening method based on artificial intelligence.

Background technique

The propagation process of radio waves is extremely complex. The plains, hills, oceans, forests, lakes, the curvature of the earth itself, atmospheric attenuation, building density and other factors on the propagation path will lead to complex transmission, diffraction, scattering and reflection of electromagnetic waves. , refraction, etc. In the process of wireless channel modeling, the essence of feature engineering is to convert the original data to obtain parameters that can best characterize the target problem, and make the dynamic range of each parameter within a relatively stable range. Therefore, it is necessary to filter appropriate features. Set to improve modeling accuracy.

The collection locations of this dataset are laboratory and field environments, and specifically include engineering parameter data, map data, and wireless channel RSRP label data of multiple cells. The training data set contains multiple files, and each file represents the data in a cell. Each line of the file represents the relevant data of the test area with a fixed size in the cell. The number of lines is uncertain (depending on the size of the cell, the larger the cell, the more lines, and vice versa), and the number of columns is a fixed 18 columns. The first 9 columns are the engineering parameter data of the site; the middle 8 columns are the map data; the last column is the RSRP label data. Table 1 is one row of data, as an example:

Table 1 Examples of training data

The following subsections describe the specific meaning of each part of the data.

3.1 Engineering parameter data

The engineering parameter data records the engineering parameter information of the site in a certain cell, and there are 9 fields in total. The corresponding meaning of each field is shown in Table 2:

Table 2 Field meanings of engineering parameter data

Field Name meaning unit Cell Index Cell Unique ID - Cell X The grid position of the site to which the cell belongs, X coordinate - Cell Y The grid location of the site to which the cell belongs, Y coordinate - Height Height of cell transmitter relative to ground m Azimuth Cell transmitter horizontal direction angle Deg Electrical Downtilt Cell transmitter vertical electrical downtilt Deg Mechanical Downtilt Cell transmitter vertical mechanical downtilt Deg Frequency Band Cell transmitter center frequency MHz RS Power Cell transmitter transmit power dBm

In order to facilitate data processing, the map is rasterized. Each grid represents an area of 5m × 5m (as shown in Figure 1 below), where (Cell X, Cell Y) records the coordinates of the upper left corner of the grid where the site is located. Other engineering parameters (Height, Azimuth, Electrical Downtilt, Mechanical Downtilt) are shown in Figure 1, where the mechanical downtilt angle is achieved by adjusting the bracket behind the antenna panel, which is a physical signal downtilt; The tilt angle Electrical Downtilt is achieved by adjusting the coil inside the antenna, which is a downward tilt of the electrical signal. The actual downtilt of the signal line is the sum of the mechanical downtilt and the electrical downtilt.

3.2 Map data description

The map data mainly includes information such as the terrain and height of the test site, and is divided into 8 fields of information. The corresponding meaning of each field is shown in Table 3. Considering the diversity and complexity of test locations in the map, the actual transmission environments such as urban areas, industrial areas, rural areas, and business districts are abstracted into numbers. In Table 4, you can see the actual feature types corresponding to the feature type name numbers.

Table 3 Field meanings of map data

Table 4 The number meaning of the name of the feature type

Like the engineering parameter data, the map data is also rasterized, and each grid represents an area of 5m × 5m, where (X, Y) records the coordinates of the upper left corner of the grid where the map is located.

Table 4 gives the number meaning of the name of the feature type, in which the feature type contains a lot of height information. Although there is a large amount of data, there are contents that contradict the description of the feature type index in the data. Therefore, data cleaning is required. For example, when the feature type index is 10, the building at the grid point is still less than 60m. For example, the building height of the observation point with the coordinates (411170, 3395480) of the cell number 2461901 is 12m, but the feature type index is 10, which contradicts the corresponding building height higher than 60m. Similarly, when the feature type index is 13, the height of the buildings at this grid point must be less than 20m, but according to the table data, there are still buildings larger than 20m, so the index according to the feature type corresponds to the actual feature height. The abnormal data in the process needs to be cleaned.

3.3 RSRP Tag Data

RSRP (Reference Signal Receiving Power, reference signal receiving power) tag data. RSRP is one of the key parameters that can represent the wireless signal strength and one of the physical layer measurement requirements in the cellular network, and refers to the average value of the signal power received on all REs (Resource Elements) carried by the reference signal. By comparing the measured received power with the known transmitted power, the attenuation of the radio wave signal by the radio wave transmission path can be obtained.

Reference Signal Received Power (RSRP) tag data is used as the actual measurement in supervised learning for comparison with the results predicted by the machine learning model. The data has a total of 1 field, and the corresponding meaning is shown in Table 5.

Table 5 Field meanings of the RSRP label data table

Field Name meaning unit RSRP Reference signal received power for grid (X, Y), label column dBm

Since wireless signals are mostly at the mW level, they are converted into dBm by polarizing them. dBm is the unit that represents the absolute value of power, and the conversion formula is:

0dBm=10lg(1mW) (1)

The above formula can also be understood as 1mW=0dBm, and the wireless signal less than 1mW means that dBm is a negative number. In the actual wireless signal transmission process, it is difficult for the signal receiver to reach the receiving power of 1mW, so the wireless signal dBm is all negative, and the maximum value is 0. dBm is only 0 under ideal conditions, that is, when the receiver receives all the signals transmitted by the transmitter. Generally speaking, the larger the dBm value, the higher the signal strength and the better the receiving effect, but considering the economic cost in practical applications, when the dBm value received in an area is between 0-50dBm, or between 0- When the value is between 70dBm, the signal value in this area is considered to be good. When the received wireless signal is less than -70dBm, the transmission is unstable and the speed is slow. At this time, the wireless network cannot be used normally. In this study, the value of the evaluation index weak coverage decision threshold P is set as -103dBm, that is, when the received dBm value in an area is between 0-103dBm, the signal value in the area is considered to be good.

According to the radio wave propagation process, if you want to design a wireless channel model, you first need to design a complete set of characteristics of radio wave transmission. The feature set mainly includes three types of wireless channel model features: features based on spatial position, features based on signal deflection angle, and features based on obstacles in the transmission environment.

SUMMARY OF THE INVENTION

(1) Technical problems to be solved

The technical problem to be solved by the present invention is: how to perform feature screening of the wireless channel model, and reduce the computational time complexity on the premise of ensuring the reliability of the wireless channel model.

(2) Technical solutions

In order to solve the above-mentioned technical problems, the present invention provides a wireless channel feature screening method based on artificial intelligence, which includes the following steps:

Step 1. According to the propagation characteristics of radio waves, establish a complete set of characteristics of the wireless channel model;

In step 2, a feature subset that can be used to establish a wireless channel model is screened by using an artificial intelligence related algorithm, and the feature subset is a variable in the wireless channel model.

(3) Beneficial effects

In the present invention, by summarizing the feature set involved in the wireless channel model, the feature screening of the wireless channel model based on artificial intelligence is performed, and the efficient machine learning model depends on the strong correlation between the input variables and the problem target, so as to ensure the wireless channel model. The computational time complexity is reduced under the premise of reliability.

Description of drawings

Figure 1 is a schematic diagram of the coordinates of rasterized ground objects;

Figure 2 shows the meaning of engineering parameter data;

Fig. 3 is the characteristic design flow chart of the present invention;

4 is a flow chart of feature screening and evaluation of the present invention;

FIG. 5 is a three-dimensional scene diagram of signal transmission between a signal transmitter and a receiver.

Detailed ways

In order to make the purpose, content, and advantages of the present invention clearer, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

Wireless channel modeling to predict the propagation characteristics of radio waves is the basis for building wireless communication systems. With the rapid development and popularization of the fifth-generation mobile communication technology, the establishment of high-precision and low-latency wireless channel models is currently a hot research issue. In the process of establishing the wireless channel model, the model parameters can be regarded as wireless channel characteristic parameters. The artificial intelligence-based wireless channel model feature screening method proposed by the present invention improves the wireless channel prediction accuracy and reduces the time complexity. Among them, the radio wave propagation path is complex, and the wireless channel model parameters are dynamically variable. Determining the model parameters is the key task.

The overall scheme design idea of the present invention is as follows: on the premise of ensuring the reliability of the wireless channel, the feature set of the wireless channel model is established, and the feature engineering in deep learning is used to extract the better results from the original data, and then achieve the optimum prediction, thereby improving the The computational complexity of the wireless channel modeling accuracy provides a reference for the subsequent wireless channel model establishment. In the current wireless channel modeling process, due to the small amount of measured data, previous researchers have obtained the variables in the wireless channel model through experience, and then fitted the model formula.

The invention uses machine learning and deep learning to deeply mine the radio wave propagation characteristics based on a large amount of engineering measured parameter data, and selects the characteristics that are more relevant to the target label. On the basis of ensuring the reliability of the wireless channel model, the amount of calculation is reduced, thereby providing theoretical support and technical reference for mobile operators' base station deployment.

3, 4, the scheme of the present invention includes the following steps:

Step 1. According to the propagation characteristics of radio waves, establish a complete set of characteristics of the wireless channel model

Among them, the following wireless channel model feature construction method based on radio wave propagation is adopted:

Step 11. Data processing

The data used this time contains a total of 4000 cells, and each cell has different test points. There are a total of 12,011,832 test units, and there are no missing values.

In order to facilitate the subsequent construction of interpretable features, a wireless signal transmission model considering the three-dimensional spatial position information of the signal base station and the receiving point is designed, and then the features are constructed through the scene. At the same time, the validity of the structural features is visually verified by data visualization analysis. The scene model of the signal transmission between the signal transmitter and the receiving point is shown in Figure 5.

Based on the signal transmission scenario model of the target point and the base station point in general in Fig. 5, it is assumed that the plane where the sea level is located is H={(x,y,z)|z=0}, which is specifically expressed as follows:

In Fig. 5, θ is the actual emission angle of the signal line, and mathematically it is the sum of the mechanical downtilt (MechanicalDowntilt) and the electronic downtilt (ElectricalDowntilt), that is, θ _D =θ _MD +θ _ED ;

Cell Altitude of the grid (Cell X, Cell Y) where the cell site is located is expressed as h _ca ;

Cell Building Height of the grid (Cell X, Cell Y) where the cell site is located is expressed as h _cb ;

The effective height of the grid (Cell X, Cell Y) antenna where the cell site is located, that is, the effective antenna height of the mobile station is h _lc =h _b -h _cb ;

The height of the cell transmitter relative to the building is expressed as h _c ;

Cell transmitter base station antenna effective height h _rc =Cell Y+h _c ;

The height difference between the grid observation point and the base station antenna is expressed as Δh=h _cb +h _c ;

The altitude of the grid observation point is _ha ;

The horizontal distance between the current grid unit and the transmitter is expressed as _dh ;

The 3D Euclidean distance between the current grid point B and the signal antenna A is expressed as d;

The grid (X,Y) Altitude is expressed as h _b ;

The horizontal direction angle Azimuth of the signal transmitter is expressed as α;

The angle between the signal line and the AB connection line is denoted as the signal deflection angle and denoted as β.

The coordinates of A are (x ₀ , y ₀ , h ₀ ), and the coordinates of B are (x ₁ , y ₁ , h ₁ ), and the horizontal coordinates of the base station are (x ₀ , y ₀ ), h ₀ =h _c + h _cb +h _ca ; the horizontal coordinates of the receiving point are (x ₁ , y ₁ ), h ₁ =h _b +h _a .

The data processing steps are completed above, and the feature establishment steps are executed below.

Step 12. Design features based on the spatial location of radio wave transmission

According to the background knowledge of radio wave transmission, there is a direct relationship between the strength of the radio signal and the transmission distance. The distance d from the combined radio signals to A and B is determined by the relative height difference Δh and the horizontal distance d _h between A and B. Therefore, d, _Δh , and dh are regarded as one-dimensional features based on spatial locations.

The horizontal distance d _h between A and B is:

The distance d between A and B is:

Step 13. Design features based on signal deflection angle

Considering that the height of the receiving antenna relative to the ground is not given, it is assumed that the average received signal strength is measured on the ground of the grid where the test point is located.

(1) The total downtilt angle of the base station is θ _D :

θ _D = θ _MD + θ _ED (4)

where _θMD is the vertical mechanical downtilt angle, and _θED is the vertical electrical downtilt angle.

In the formula, the coordinate position of the grid test point is (x, y);

The signal emitted by the base station antenna has a certain degree of concentration. For example, the signal on the back of the antenna is usually weaker than the signal at the same angle and distance on the front of the antenna, that is, the signal strength is related to the angle between the signal line and the AB connection. When other conditions are the same, the smaller the angle β between the AB connection and the signal line, the stronger the signal at point B. Make a circle with the base station as the center. The smaller the angle between the point on the circumference and the connection line between the base station and the signal line, the stronger the signal.

Since the horizontal direction angle Azimuth of the transmitter is given in the data, that is, the angle α increases clockwise from the positive direction of the Y-axis. Increased α′:

α′=(2π-α)+π/2 (5)

AB连线之间的向量为

信号线

与A-B点连线对应的

在3D环境里夹角β可以通过余弦表示：Then the signal line vector

The vector between the lines AB is

signal line

Corresponding to the line connecting point AB

In a 3D environment the angle β can be expressed as a cosine:

与

在水平面上的夹角可以通过两向量的x,y分量用上述相同方法算式计算。将3D传输环境下的夹角cosα与水平面夹角cosβ都记为基于信号的水平偏转角特征，该部分特征为训练模型的重要特征。at this time,

and

The angle on the horizontal plane can be calculated using the same formula as above by using the x and y components of the two vectors. The angle cosα and the angle cosβ of the horizontal plane in the 3D transmission environment are recorded as the horizontal deflection angle feature based on the signal, and this part of the feature is an important feature of the training model.

Step 14. Design a feature set based on the radio wave transmission environmental barrier

Through the analysis of the Cost231-Hata model, it can be seen that the transmission of wireless signals is not only related to the spatial position relationship of the transceiver, but also to the signal transmission environment. The radio wave transmission environment is usually complex and changeable. Uncertain factors such as mountain height, building density, and lake reflection area on the transmission path will cause radio waves to no longer transmit in a single path. This is also an important factor that the traditional transmission model cannot accurately describe the signal transmission at a finer granularity. Therefore, designing the characteristics of the transmission environment and realizing a reasonable representation of the transmission environment are the key and difficult points in the establishment of feature engineering. The map data about the environment given in the dataset includes building height (Height), altitude (Altitude), and feature type (ClutterIndex). Therefore, robust extraction of transmission environment features should be achieved from three aspects: building distribution and height, cell topography, and feature types.

(1) Building distribution and height characteristics M _b :

In the formula, A _b , B _b , and C _b are the coordinate values representing the X, Y, and Z (building heights at different measurement points) axis directions in the building point cloud map, respectively. It can be known from the properties of the covariance matrix that the covariance matrix of the building point cloud is a symmetric matrix, so Cov(A _b ,B _b )=Cov(B _b ,A _b ), Cov(A _b ,C _b )=Cov(C b ) _b , A _b ), Cov(B _b , C _b ) = Cov(C _b , B _b ). Therefore, the covariance matrix M _b of the map point cloud has a total of 6 effective element parameters: Cov(A _b , B _b ), Cov(A _b , C _b ), Cov(B _b , C _b ) and their respective variances, for For coordinate column vectors A and B, the covariance Cov(A _b , B _b ) is calculated as:

是列向量A_b的均值，

是列向量B_b的均值。In the formula, A _bi represents the ith element value of the column vector A _b , B _bi represents the ith element value of the column vector B _b ,

is the mean of the column vector A _b ,

is the mean of the column vector B _b .

Using the cell coordinates, building height data, and RSRP label data in the transmission environment, the three-dimensional point cloud of the cell building can be obtained. The feature extraction of residential buildings is transformed into the problem of feature extraction of three-dimensional rigid body targets to achieve robust extraction of target point cloud features. The covariance matrix of the residential building point cloud reflects the distribution and height characteristics of the residential buildings. Therefore, in this step, the set feature is the number of obstacles (Obstacle_num) on the connection line between the observation point and the base station point, and the transmission loss of different observation point densities and building density on the transmission path is calculated.

Step 15: Summarize all the features that have been established

Through the analysis of traditional wireless channel path loss models such as the Cost231-Hata model, the distance information used in the traditional wireless channel path loss model is all logarithmic with the base 10, and the geometric angle can be represented by trigonometric functions. The features and label columns in the original data set and the new features constructed by the present invention are shown in Table 5:

Table 5 All feature and label columns given in the dataset and new features constructed

Step 2: Use artificial intelligence correlation algorithms to screen feature subsets that can be used to establish a wireless channel model. The feature subset is a feature set that is highly correlated with the target label in the feature set, and the feature subset is the feature set in the wireless channel model. variable.

Step 21. Use the feature screening method based on machine learning to perform feature screening, in which two methods are used for screening to improve feature reliability

Method 1: Filtered Feature Screening Method

The filtering principle of feature selection is that, first, according to the agreed rules, and then score according to the feature divergence or correlation, set the threshold to select the features of the data set, and then train the learner. The feature selection process is independent of subsequent learners, so the computation is fast. That is, the features are first "filtered" using the feature selector, and then the neural network model is trained with the filtered features. Filtered feature screening can use statistical methods to score and rank features, and the higher the ranking, the greater the value of the feature. Common filtering methods include T-test (T-test), chi-square test (χ ² -test), Pearson Correlation Coefficient (PCC), and Maximum Information Coefficient (MIC).

The present invention uses two filtering methods of Pearson correlation coefficient and maximum information coefficient to evaluate feature scores according to different indexes.

Method 2: Embedding-based feature screening method

The main idea of embedded feature screening is to select the best features that can improve the accuracy in the model that has been trained for re-learning. In the process of determining the model, the features that are meaningful to the training model are selected, and the model training process and the feature screening process are integrated.

Two embedded feature evaluation methods are Random Forest (RF) and XGBoost boosted tree model. Random forest is a supervised learning method, which is trained based on bagging method. The so-called bagging method, also known as bootstrap aggregating, adopts the method of randomly selecting and putting back the training data, then constructing the classifier, establishing multiple unrelated decision tree models, and finally merging the learned models together to obtain more accurate Stable prediction effect.

The XGBoost boosted tree uses the sum of the number of splits in each tree for all features to calculate the feature score. For example, this feature is split once in the first tree, twice in the second tree, n times in the nth tree..., then the score of this feature is (1+2+...+n). The parameter selection process of XGBoost is very important, so the operation of parameter shuffle can be performed. Finally, based on the features and socre obtained by XGBoost with different combinations of the above parameters, the score averaging operation can be performed to filter out the features with high scores.

The advantage of the filtering feature screening method is that the traditional statistical measurement method can be used to quickly judge the importance of the feature, but in this process, the correlation between each feature and the target variable needs to be independently examined, and the difference between different features is ignored. associated information and combined effects. Wrapped and embedded feature evaluation methods can take into account the effects of different feature combinations to better evaluate the importance of features, but the computational time is high.

Table 6 Scores of different features under each evaluation index

Step 22. Obtain a single feature screening result based on machine learning, and perform simulation analysis

For the correlation between features and targets, two types of filter evaluation and CV are used for scoring. Regarding the correlation between the feature and the target, it is defined as a weighted evaluation of the three evaluation indicators (the sum of the weights is 1), and a comprehensive target-related score is given to the feature by integrating the feature results obtained by different evaluation methods. If the weights are evenly distributed, the target relevance scoring formula is:

Score=ω ₁ S _PCC +ω ₂ S _MIC +ω ₃ S _XGBoost (9)

Using the above three evaluation indicators, the scoring results of the constructed single feature are shown in Table 6. The scores have been scaled appropriately for observation, due to the conditions under which the data was collected. The absolute value of each evaluation index has a total score of 100 points, and the total score is allocated to each feature to complete the scoring. The higher the number of scores a feature receives, the greater the correlation.

Table 7 Correlation scores and rankings of different features under each evaluation index

Step 23: Use the feature screening method based on deep learning for feature screening

Using artificial intelligence-related algorithms to combine feature subsets of wireless channel models The previous article constructed a total of 36 single features based on the radio wave transmission process. However, in the process of actually establishing the radio wave prediction model and the wireless channel, if you want to train the neural network efficiently Further data mining of the features is required. By using different feature screening methods in machine learning, the scales of the features constructed in combination with the radio wave transmission process are also different. Therefore, in order to fuse multi-scale features into a more discriminative ability, and also to create new features from existing features, this method uses Deep Feature Synthesis (DFS) to automatically complete this work.

Deep feature synthesis is a method that can quickly create feature combinations with different depths, and construct new, deeper features by decomposing complex massive data into numerical components. The deep feature synthesis process is a visualization process, and the new features constructed by it can be explained in practical terms. This method uses Featuretools for deep feature synthesis. Featuretools is an open source library for performing automatic feature engineering. By converting different relational datasets into feature matrices that can be used for machine learning, deep feature synthesis can be quickly advanced, allowing more time Focus on other aspects of machine learning model building.

Featuretools mainly involves three basic concepts: entity (entity), relationship (relationship) and operator (primitive), which actually provides a framework for converting from a single table and connecting multiple tables. After deep feature synthesis using Featuretools, several new features are generated. The new features are based on easy-to-understand feature primitives, which are aggregated from the child table to the parent table through the Primary key (Cell Index).

Table 8 Featuretools generates some new features

It can be seen that the engineering data set used in this method is the measured communication data set, and the practical meanings corresponding to each label are introduced one by one. According to the traditional wireless channel path loss model and the provided data information, select the appropriate features that can better describe the wireless channel transmission loss, and evaluate the scientific and practical engineering significance of the constructed features by calculating the score on the target tag. In the process of actually establishing a radio wave prediction model and a wireless channel, if you want to efficiently train the neural network, further data mining of features is required. Therefore, in order to merge multi-scale features into a more discriminative ability, new features are also created from existing features, and deep feature synthesis is used to automatically complete feature selection in wireless channels.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the technical principle of the present invention, several improvements and modifications can also be made. These improvements and modifications It should also be regarded as the protection scope of the present invention.

Claims (4)

1. A wireless channel characteristic screening method based on artificial intelligence is characterized by comprising the following steps:

step one, establishing a feature complete set of a wireless channel model according to the radio wave propagation characteristics;

screening a characteristic subset which can be used for building a wireless channel model by using an artificial intelligence correlation algorithm, wherein the characteristic subset is a variable in the wireless channel model;

step one comprises the data processing steps: step 11, designing a wireless signal transmission model considering the three-dimensional space position information of the signal base station and the receiving point; and a feature establishing step performed based on the data processing step;

in the signal transmission scene model of the target point and the base station, which is established in the first step:

assuming that sea level is in a plane H = { (x, y, z) | z =0}, and theta is the actual emission angle of the signal line, and is mathematically expressed as the sum of a mechanical downward inclination angle and an electronic downward inclination angle, namely theta _D ＝θ _MD +θ _ED ；

In the formula, theta _D Is the total base station downtilt angle, θ _MD At a vertical mechanical down-tilt angle, theta _ED Is a vertical electrical downtilt;

the altitude of the grid (Cell X, cell Y) where the Cell site is located is represented as h _ca ；

The building height of the grid (Cell X, cell Y) in which the Cell site is located is denoted by h _cb ；

The effective antenna height of Cell X, cell Y grid, i.e. the effective antenna height of mobile station is h _lc ＝h _b -h _cb ；

The Height of the cell transmitter relative to the building, height, is denoted h _c ；

Effective height h of cell transmitter base station antenna _rc ＝Cel Y+h _c ；

Grid observation point and base station antennaThe height difference therebetween is expressed as Δ h = h _cb +h _c ；

The elevation of the grid observation point is h _a ；

The horizontal distance of the current grid cell from the transmitter is denoted d _h ；

The 3D Euclidean distance between the current grid point B and the signal antenna A is represented as D;

the grid (X, Y) Altitude is denoted as h _b ；

The horizontal directional angle Azimuth of the signal transmitter is denoted as α;

the included angle between the signal line and the AB connecting line is recorded as a signal deflection angle and expressed as beta;

a has a coordinate of (x) ₀ ,y ₀ ,h ₀ ) The coordinates of B are (x) ₁ ,y ₁ ,h ₁ ) Wherein the base station horizontal coordinate is (x) ₀ ,y ₀ )，h ₀ ＝h _c +h _cb +h _ca (ii) a The horizontal coordinate of the receiving point is (x) ₁ ,y ₁ )，h ₁ ＝h _b +h _a ；

The characteristic establishing step specifically comprises the following steps:

step 12, designing characteristics based on the space position of radio wave transmission;

step 13, designing a characteristic based on a signal deflection angle;

step 14, designing a feature set based on the radio wave transmission environment barrier;

step 15: determining all features that have been established;

the step 12 specifically comprises:

the distance d from the radio signal to A and B is determined by the relative height difference Δ h and the horizontal distance d between A and B _h Determined jointly, therefore, d, Δ h, d _h As a one-dimensional feature based on spatial location;

A. horizontal distance d of B _h Comprises the following steps:

A. the distance d of B is:

step 13 specifically comprises:

assuming that the average received signal strength is measured on the ground of the grid where the test points are located;

(1) Total base station down tilt angle theta _D ：

θ _D ＝θ _MD +θ _ED (3)

In the formula, the coordinate position of the grid test point is (x, y);

given a transmitter horizontal angle, i.e., the angle α increases clockwise from the positive Y-axis, it is first changed to a' that increases counterclockwise from the positive X-axis, which is commonly used in two-dimensional planes:

α′＝(2π-α)+π/2 (4)

then the signal line vector

The vector between the AB lines is

Signal line

Corresponding to the connection of points A-B

The angle β is represented by the cosine in a 3D environment:

at this time, the process of the present invention,

and

calculating an included angle on a horizontal plane through x and y components of two vectors, and recording an included angle cos alpha and an included angle cos beta of the horizontal plane in a 3D transmission environment as horizontal deflection angle characteristics based on signals;

step 14 specifically comprises:

robust extraction of transmission environment characteristics is achieved from three aspects of building distribution and height, district terrain and ground feature types;

(1) Building distribution and height characteristics M _b ：

In the formula, A _b 、B _b 、C _b Coordinate values in X, Y and Z axis directions in the building point cloud map are respectively represented, and the covariance matrix of the building point cloud is a symmetric matrix, so that Cov (A) _b ,B _b )＝Cov(B _b ,A _b ),Cov(A _b ,C _b )＝Cov(C _b ,A _b ),Cov(B _b ,C _b )＝Cov(C _b ,B _b ) Covariance matrix M of map point clouds _b There are 6 valid element parameters: cov (A) _b ,B _b )、Cov(A _b ,C _b )、Cov(B _b ,C _b ) And the respective variances, cov (A) for the coordinate column vectors A and B _b ,B _b ) Calculated using the formula:

in the formula, A _bi Represents a column vector A _b The value of the ith element of (1), B _bi Representing a column vector B _b The value of the ith element of (a),

is in the column directionQuantity A _b The average value of (a) of (b),

is a column vector B _b The mean value of (a);

the three-dimensional point cloud of the cell building can be obtained by utilizing the cell coordinates, the building height data and the RSRP label data in the transmission environment, the feature extraction is carried out on the cell building, the problem of feature extraction on the three-dimensional rigid body target is solved, the robust extraction of the target point cloud feature is realized, the cell building point cloud covariance matrix reflects the distribution and height features of the cell building, so in the step 14, the set features are the number of barriers on the observation point-base station point connecting line, and the transmission loss of the density of the building degree on different observation points and transmission paths can be calculated.

2. The method of claim 1, wherein step two specifically comprises:

step 21, performing feature screening by using a feature screening method based on machine learning;

and step 22, obtaining a single feature screening result based on machine learning, and performing simulation analysis.

3. The method of claim 1, wherein the step 21 of feature screening is performed using a filtered feature screening method and an embedded feature screening method.

4. The method of claim 2, wherein step two further comprises step 23: and (4) performing feature screening by adopting a feature screening method based on deep learning.

Images