A scenario-adaptive radio wave propagation prediction method for spectrum sharing

Abstract

The present invention belongs to the technical field of radio propagation prediction, and in particular, to a method for predicting radio wave propagation based on scene adaptation for spectrum sharing. The purpose of the present invention is to propose a method for predicting radio wave propagation based on scene adaptation for spectrum sharing, which learns a correction model suitable for each scene for electromagnetic propagation scenes with different regional characteristics and different frequency bands, and can adaptively select a suitable correction model, so as to ultimately provide more accurate prediction capabilities for electromagnetic propagation in the scene. The present invention classifies the propagation scenes for which measurement data is collected, applies a neural network to each type of scene, and learns a correction model. The present invention adaptively selects the correction model for propagation prediction. That is, the propagation scenes to be predicted are classified, and the correction model corresponding to the classification to which it belongs is selected for prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio propagation prediction, and in particular relates to a method for predicting radio wave propagation based on scenario adaptation for spectrum sharing. Background Art

[0002] The use of spectrum resources is cross-regional, cross-frequency band, and cross-business. Therefore, if spectrum management is to achieve the purpose of spectrum sharing, it is necessary to have more accurate simulation predictions of the electromagnetic propagation conditions of different regional characteristics, different frequency bands, and different business applications.

[0003] The propagation characteristics of wireless channels are closely related to the propagation environment. Radio waves in different frequency bands are affected by the topography, artificial buildings, atmospheric environment, electromagnetic interference and movement speed in the propagation environment, and exhibit complex and changeable fading characteristics.

[0004] Traditional classical statistical propagation models can predict the mean received power and its changing trend and have good applicability. However, the prediction accuracy of these models is unsatisfactory in specific usage scenarios. The average absolute error between the predicted propagation loss value and the measured loss value is basically between 20 and 40 dB, and the root mean square error is between 5 and 15 dB.

[0005] Considering that the propagation characteristics of wireless channels are related to various factors of the propagation environment, if these factors can be measured and a large number of measurement data sets can be obtained, the prerequisites for using big data analysis methods can be met, that is, the data is authentic and reliable, the data reaches a certain amount of data, and the data is highly diversified. In this way, the machine learning method of big data analysis can be used to study, model, and modify the characteristics of the propagation model. In the measurement process, as many parameters that affect the propagation characteristics as possible should be included, and each parameter must cover a wide range, such as frequency, receiving height, coverage area, various landforms, weather conditions, wind speed and direction, etc. When all kinds of attribute parameters are complete and there are various segmented values, these parameters are used in the machine learning algorithm to establish a more accurate propagation model of the wireless channel.

[0006] The correction model trained by the machine learning algorithm is highly correlated with the specific scene. In specific usage scenarios, the scene can be classified according to parameters such as the scene's topography, buildings, temperature, humidity, and atmospheric pressure. When predicting in similar scenes, it can be assumed that the electromagnetic propagation channel has similar characteristics. Therefore, a correction model obtained by machine learning can be used in similar scenarios.

[0007] The present invention forms a classification model for each scene by analyzing the scene parameter data. The measured data under each scene type is trained by machine learning method to obtain the correction model of that type. When predicting radio wave propagation, the scene to be predicted is classified into its corresponding category according to the characteristic parameters and the corresponding correction model is selected for calculation to obtain a more accurate prediction result. Summary of the invention

[0008] The purpose of the present invention is to propose a method for predicting radio wave propagation based on scenario adaptation for spectrum sharing. According to electromagnetic propagation scenarios with different regional characteristics and different frequency bands, a correction model suitable for each scenario can be learned and a suitable correction model can be adaptively selected, thereby ultimately providing more accurate prediction capabilities for electromagnetic propagation in the scenario.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] The impact of the scene on the electromagnetic wave propagation channel includes many aspects, among which the main influencing factors considered in the present invention include landform type, water area ratio, temperature, humidity, atmospheric pressure, and average annual rainfall. Some of these factors characterize the topography and surface feature type of the scene, some factors characterize the temperature and atmospheric characteristics of the environment, and also include the rainfall characteristics of the scene environment.

[0011] For multiple scenes involved in the measurement data, these classification features are first extracted. The machine learning process of scene classification uses a clustering method to classify existing scenes. The present invention uses a k-means clustering method. This method is a centroid-based algorithm, or a distance-based algorithm, and each cluster is associated with a centroid. This method is an unsupervised learning process that can classify similar objects into a cluster. The center of the cluster is calculated by the mean of all samples belonging to the cluster. The distance from the sample to the center point of the cluster is calculated to determine which cluster the sample point belongs to. In this way, after iterative convergence, the sample points are divided into multiple cluster classifications, and the center point and sample variance of each cluster classification can also be calculated.

[0012] After using the clustering method to obtain the classification of the scene, for each scene cluster classification, the channel measurement data of all scenes under the classification can be merged together, and then the merged data can be trained by machine learning to obtain a corrected radio wave propagation model.

[0013] The ITU-R P.1546 model is designed and implemented based on Recommendation ITU-R P.1546-6, which is a recommendation issued by the International Telecommunication Union's Radio Sector (ITU-R): "Recommendation ITU-R P.1546-6 (08 / 2019) Method for point-to-area prediction of terrestrial services in the frequency range 30 MHz to 4 000 MHz". This recommendation describes a prediction method for point-to-area radio propagation for terrestrial services in the frequency range 30 MHz to 4 000 MHz. This method is applicable to land paths, sea paths and / or land-sea mixed paths with effective transmitting antenna heights less than 3 000 m and path lengths between 1 and 1 000 km for radio links. The basic part of the model is the propagation curves for different scenarios, different frequency bands, different time probabilities, and different transmitting antenna heights. For frequencies, time probabilities, and antenna heights other than the curves, interpolation / extrapolation is performed using given formulas.

[0014] The present invention mainly considers the basic propagation curve part of the correction model. Under the conditions of a fixed typical transmitting antenna height, a fixed transmission frequency and a certain time probability, the basic propagation curve is expressed as the relationship between the transmission distance and the transmission loss or the scene. In the present invention, the transmission distance (unit km) is used to take the log value with a base of 10. The ITU-R P.1546 model is used to calculate the correction term part excluding the basic propagation curve. According to the actual measurement results, the measured basic loss value corresponding to the measured value can be obtained. Using the BP neural network model, the transmission distance (unit km) is used to take the log value with a base of 10 as the learning independent variable, and the basic loss value calculated using the measured data is used as the learning target value. Through learning, the basic loss fitting curve under the conditions of the scene and frequency, typical antenna height, and time variability is obtained. This curve can be used to predict the basic loss value in similar scenarios. Combined with the correction term part of the ITU-R P.1546 model excluding the basic propagation curve, the comprehensive loss result of channel propagation can be obtained.

[0015] BP neural network is the most widely used network model in artificial neural network, and many improved models have been derived for different fields. The specific algorithm of BP neural network was proposed by David Rumelhart, Geoffrey Hinton and Williams in the 1980s, which solved the learning and error propagation problems of multiple neural networks and realized the concept of multi-layer neural network. BP neural network can solve nonlinear problems well and can fit all non-continuous functions. From the perspective of the composition of neural network, BP neural network mainly consists of the following parts: input layer, hidden layer, output value, activation function, error calculation, and iterative process.

[0016] During the BP neural network training process, information is forward propagated, that is, the output of this layer is used as the input of the next layer, and the weight and bias calculations are performed on the layer. The calculation results are used as the output of this layer, and the calculations continue to the next layer until the output layer. The results of the output layer are compared with the actual sample values ​​to calculate the error. The error is reversely corrected through methods such as gradient calculation until the target error or training times are reached, and the training is terminated.

[0017] For an environmental parameter feature to be classified, the cluster classification to which the scene belongs can be obtained by calculating the distance between the feature and each center point. Select the corresponding correction model for calculation to obtain a more accurate prediction result.

[0018] A method for predicting radio wave propagation based on scenario adaptation for spectrum sharing, wherein the method comprises the following steps:

[0019] Step 1: Collect N radio wave propagation measurement data samples and obtain classification feature data of the sample collection scene

[0020] The radio wave propagation measurement data is power values ​​of multiple groups of input and output signals in different scenarios;

[0021] The classification characteristic data of the sample collection scene includes the following characteristic factors: landform type: the landform characteristics of the radio wave propagation scene, and the landform types include: water, mixed water and land, mountain, forest, dense urban area, suburb, and rural area;

[0022] Water area ratio: the percentage of water area in the measurement scene to the total area;

[0023] Temperature: The air temperature of the measured environment, expressed in degrees Celsius;

[0024] Humidity: The humidity of the test environment, expressed as relative humidity;

[0025] Atmospheric pressure: The atmospheric pressure of the measuring environment, expressed in hectopascals;

[0026] Precipitation: The annual average daily rainfall of the measurement environment, expressed in mm / day;

[0027] Step 2: Clean the classification feature data of the scene

[0028] Clean the extracted scene classification data and remove unreasonable scene sample data;

[0029] Step 3: Discretely encode the measurement scene classification feature data

[0030] Discretely encode the classification feature data of the measurement scene, including landform type, water area ratio, temperature, humidity, atmospheric pressure and rainfall; that is, according to the encoding rules, correspond different landform types in the classification feature data to a set of encoding values, and correspond different value ranges of water area ratio, temperature, humidity, atmospheric pressure and rainfall in the classification feature data to a set of encoding values;

[0031] After encoding the classification feature data of the measurement scene, a set of classification feature data of the measurement scene represented by discrete coding values ​​is obtained;

[0032] Step 4: Classify the collected scenes and process them using the k-means clustering method; this includes the following steps:

[0033] Step 4.1, select the number of classification clusters K and the maximum number of iterations m;

[0034] Step 4.2: Select K sample points that are different in one or more dimensions from the sample as the center points of cluster classification;

[0035] Step 4.3, calculate the distances from other sample points to the K center points to determine which category the sample data belongs to;

[0036] Step 4.4: Re-evaluate the mean of all dimensions according to the classification as the new center point;

[0037] Step 4.5, iterate the above two steps until the center point does not change or the maximum number of iterations m is reached;

[0038] Step 4.6, obtain the center point of each classification cluster and calculate the variance of each classification cluster, and store them in the database for persistence;

[0039] In steps 4.1 to 4.6, the sample is a collection scene that needs to be classified, and the sample point is the classification feature data of the measurement scene represented by discrete coding values ​​obtained in step 3;

[0040] Step 5: Merge the radio wave propagation measurement data of multiple scenarios under one category

[0041] In the case where there are measurement data of multiple scenes under one classification cluster, the radio wave propagation measurement data in the same scene category are merged to obtain comprehensive radio wave propagation measurement data for use in subsequent processing;

[0042] Step 6: Clean the radio wave propagation measurement data to remove abnormal points in the data

[0043] Step 7: For each scene category, perform variability processing on the cleaned radio wave propagation measurement data

[0044] Step 7.1, according to the regional resolution requirements of the cleaned data, the data in each resolution grid is screened, and the measurement data of the grid is obtained according to the probability exceeding a certain level, that is, the signal strength that 50% of the measurement values ​​exceed is taken;

[0045] Step 7.2, using the center position of the grid as the position of the grid data, and performing a subsequent learning process of model correction and a prediction process using the correction model;

[0046] Step 8: For each scene category, extract the feature data of the model correction based on Recommendation ITU-R P.1546, where the feature data is the propagation distance and propagation loss data.

[0047] For the localized modeling of the ITU-R P.1546 model, the channel parameter extraction process is divided into two parts: basic parameter extraction and basic transmission curve parameter extraction;

[0048] Basic parameters refer to the most basic information in the measurement. The basic parameters that need to be extracted include path type scenario, measurement frequency f, horizontal propagation distance d, and actual measured value of electromagnetic propagation. When only the median field strength is considered, the time probability t prob 50%;

[0049] In some cases, the radio propagation measurement data is field strength, and the equivalent conversion from field strength to transmission loss under typical receiving parameters (impedance 50Ω, gain 0dBi) is achieved according to Formula 1;

[0050] P r = E-77.22-20log f MHz Formula 1

[0051] In formula 1, E is the electric field strength at the receiving end, in dBuv / m; P r is the transmission loss, in dBm; f MHz is the measured frequency in MHz;

[0052] In the basic transmission curve part, the basic parameters involved are the measurement frequency f MHz , horizontal propagation distance d km , the effective height h1 of the transmitting antenna and the path type scenario; at the measurement frequency f MHz , the transmitting antenna effective height h1 and the path type scenario are certain, the basic propagation curve is the independent variable d km and the dependent variable received field strength E b Function curve of

[0053] According to the calculation method for removing the correction items other than the basic propagation loss in Recommendation ITU-R P.1546, combined with the radio wave propagation measurement data, the measurement value of the basic propagation loss part is calculated; according to the location of the radio wave propagation measurement data and the parameters of the transmitting antenna and the receiving antenna, a model is built to calculate the characteristic value of the radio wave propagation distance;

[0054] Among them, the measured value of the basic propagation loss part PL b (f MHz ,d km ) is calculated according to the method of Recommendation ITU-R P.1546:

[0055] Total path loss PL all Includes basic propagation loss and related correction terms

[0056] PL all =PL b (f MHz ,d km )+Corr(h2)+Corr(Clu t )+Corr(θ tca ) Formula 2

[0057] In this way, the basic propagation loss is obtained by the following formula

[0058] PL b (f MHz ,d km )=PL all -Corr(h2)-Corr(Clu t )-Corr(θ tca ) Formula 3

[0059] In formula 2 and formula 3,

[0060] PL b (f MHz ,d km ) is the basic propagation loss, unit: dB,

[0061] PL all is the total path loss, unit: dB,

[0062] Corr(h2) is the height correction term at the receiving end, unit: dB,

[0063] Corr(Clu t ) is the clutter correction item at the transmitting end, unit: dB,

[0064] Corr(Clu t ) is the receiving end terrain clearance angle (θ tca ) correction term, unit: dB;

[0065] Among them, the receiving end height correction term, the transmitting end clutter correction term, and the receiving end terrain clearance angle correction term are calculated according to the method given in Recommendation ITU-R P.1546;

[0066] The total path loss PL in Formula 2 all The received power and other transmit and receive parameters measured by radio wave propagation are calculated as:

[0067] P r =P t +G t -PL all +G r -L f -L o Formula 4

[0068] PL all =P t +G t +G r -L f -L o -P r Formula 5

[0069] In formula 4 and formula 5, P r is the received power, unit: dBm,

[0070] P t is the power of the transmitted signal, unit: dBm,

[0071] G t is the gain of the transmitting antenna, unit: dB,

[0072] PL all is the total path loss, unit: dB,

[0073] G r is the receiving antenna gain, unit: dB,

[0074] L f is the feeder loss, unit: dB,

[0075] L o is other loss, unit: dB;

[0076] The calculation of the characteristic value of the radio wave propagation distance is carried out by calculating the radio wave propagation distance according to the geographical geometric relationship through the longitude, latitude and transmitting antenna height of the signal transmitter and the longitude, latitude and receiving antenna height of the receiving data;

[0077] Step 9: For each scene category, use the BP neural network to train the feature data extracted in step 8 to obtain the correction model and save it.

[0078] Using machine learning methods, a calculation model for obtaining propagation loss by inputting distance is obtained, and a set of models for calculating basic propagation loss corresponding to different scene categories are obtained;

[0079] Use BP neural network to train the feature data and target data extracted in step 8 to obtain the network parameter results; the feature data refers to the log value of the transmission distance with base 10, and the target data is the measured value of the basic propagation loss part; and save the network parameter information to the database, and save the corresponding scene type at the same time; repeat steps 5 to 9, train each scene category to obtain the corresponding model parameters, so as to obtain a series of model parameters corresponding to each scene, that is, a series of correction models, and persist them in the database;

[0080] In the BP neural network, the network is set to 1 input layer, 2 hidden layers, and 1 output layer. The input layer has 4 nodes, the hidden layers have 70 nodes each, and the output layer has 1 node. The input parameter is the measurement frequency f MHz , horizontal propagation distance d km , the vector composed of the effective height h1 of the transmitting antenna and the path type scenario, and the output result is the electric field strength E at the receiving end b ;

[0081] Step 10: Extract parameters for the prediction scenario

[0082] Extract the landform type, water area ratio, temperature, humidity, atmospheric pressure and rainfall characteristic parameters of the scene to be predicted, and perform discretization encoding to form a feature vector of the scene characteristic parameters;

[0083] Step 11: Classify the parameters of the scene to be predicted and select the machine learning result correction model of the corresponding scene obtained in step 9

[0084] Calculate the distance between the feature vector of the scene to be predicted and the center point of the classification cluster, select the classification coarse that is closest to it, and classify the scene into this classification cluster; select the machine learning result correction model of the corresponding scene obtained in step 9;

[0085] Step 12: Predict the radio wave propagation loss in the target scenario

[0086] Find the BP neural network parameters of the correction model corresponding to the classification cluster to which the target scene belongs; extract the log value with base 10 of the transmission distance of the current scene to be predicted as the input of the network, and obtain the basic loss value under the scene through forward calculation of the network; use the recommended method given in Recommendation ITU-R P.1546 to remove the correction items other than the basic propagation loss, and calculate the remaining correction items; combine the results of the two to obtain the final loss prediction value.

[0087] in,

[0088] In step 2, the extracted scene classification data is cleaned including:

[0089] Landform type: if the landform data is not in the candidate scene, the data is considered unreasonable and is manually modified to meet the requirements or deleted; the candidate scenes include: water, mixed water and land, mountain, forest, dense urban area, suburb, and rural area;

[0090] Water area ratio: If the water area ratio is less than 0 or greater than 100, it is considered illegal data and will be removed;

[0091] Temperature: If the temperature is outside the range of -30℃~60℃, it is considered as illegal data, and the data must be manually confirmed and modified or the sample data must be deleted;

[0092] Humidity: If the relative humidity exceeds the range of 30 to 80, it is considered as illegal data and the data is manually confirmed and modified or the sample data is deleted;

[0093] Atmospheric pressure: If the atmospheric pressure data exceeds the value range of 800-1100, the data is considered illegal and the data must be manually confirmed and modified or the sample data must be deleted;

[0094] Precipitation: If the rainfall data exceeds the value range of 0 to 1000 mm / day, the data is considered illegal and the data must be manually confirmed and modified or the sample data must be deleted.

[0095] in,

[0096] In step 6, the radio wave propagation measurement data is cleaned to remove abnormal points in the data, including:

[0097] Filter the data whose position information in the radio wave propagation measurement data is not within the range;

[0098] Filter out the radio wave propagation measurement data that is not within a reasonable range; the reasonable range of radio wave propagation measurement data is required to be above the sensitivity of the signal receiver, and the reasonable range of radio wave propagation measurement data is required to be between -130dBm and -30dBm; filter out the data that exceeds the range.

[0099] Wherein, in step 7.1, the median value of the probability is 50%.

[0100] Compared with the prior art, the present invention has the following beneficial effects:

[0101] The present invention can learn a correction model suitable for each electromagnetic propagation scenario with different regional characteristics and different frequency bands, and can adaptively select a suitable correction model, thereby providing a more accurate prediction capability for electromagnetic propagation in the scenario. Similarly, the processing idea of ​​this method can also be applied to the needs of 6G air, space, sea and land integration as a functional support for the underlying electromagnetic propagation prediction.

[0102] The present invention classifies the propagation scenarios of collected measurement data, applies a neural network to each category of scenarios, and learns a correction model.

[0103] The present invention adaptively selects the corrected model for propagation prediction, that is, classifies the propagation scenarios to be predicted, and selects the corrected model corresponding to the classification to perform prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 A flowchart of a method for predicting radio wave propagation based on scenario adaptation for spectrum sharing according to the present invention;

[0105] Figure 2 This is a schematic diagram of the network structure of the BP neural network constructed in the present invention. DETAILED DESCRIPTION

[0106] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0107] A method for predicting radio wave propagation based on scenario adaptation for spectrum sharing comprises the following steps:

[0108] Step 1: Collect N radio wave propagation measurement data samples and obtain classification feature data of the sample collection scene

[0109] The measurement system for collecting radio wave propagation measurement data samples includes a signal transmitter, a signal receiver and auxiliary facilities and equipment.

[0110] The signal transmitting end includes a high-frequency signal generator, a control computer, a power amplifier, a feeder, a transmitting antenna and an antenna bracket.

[0111] The signal receiver includes a receiving antenna and an antenna bracket, a signal acquisition instrument (such as a spectrum analyzer), a control computer, measurement software, a portable computer and a GPS receiver system.

[0112] Auxiliary facilities and equipment include vehicles to assist in monitoring and temperature and humidity measurement sensors to collect climate and weather information.

[0113] The radio wave propagation measurement data are power values ​​of multiple groups of input and output signals in different scenarios.

[0114] The classification feature data of the sample collection scene includes the following characteristic factors: landform type: the landform characteristics of the radio wave propagation scene, and the landform types include: water, mixed water and land, mountain, forest, dense urban area, suburb, and rural area.

[0115] Water area ratio: measures the percentage of water area in the scene to the total area.

[0116] Temperature: The air temperature of the measurement environment, expressed in degrees Celsius.

[0117] Humidity: The humidity of the test environment, expressed as relative humidity.

[0118] Atmospheric Pressure: The atmospheric pressure of the measurement environment, expressed in hectopascals.

[0119] Precipitation: The annual average daily rainfall for the measurement environment, expressed in mm / day.

[0120] Step 2: Clean the classification feature data of the scene

[0121] Clean the extracted scene classification data and remove unreasonable scene sample data.

[0122] Landform type: If the landform data is not in the candidate scene, the data is considered unreasonable and is manually modified to meet the requirements or deleted. The candidate scenes include: water, mixed water and land, mountain, forest, dense urban area, suburb, and rural area.

[0123] Water area ratio: If the water area ratio is less than 0 or greater than 100, it is considered illegal data and will be removed.

[0124] Temperature: If the temperature is outside the range of -30℃~60℃, it is considered as illegal data and the data must be manually confirmed and modified or the sample data must be deleted.

[0125] Humidity: If the relative humidity exceeds the range of 30 to 80, it is considered to be illegal data and the data must be manually confirmed and modified or the sample data must be deleted.

[0126] Atmospheric pressure: If the atmospheric pressure data exceeds the value range of 800-1100, the data is considered illegal and the data must be manually confirmed and modified or the sample data must be deleted.

[0127] Precipitation: If the rainfall data exceeds the value range of 0 to 1000 mm / day, the data is considered illegal and the data must be manually confirmed and modified or the sample data must be deleted.

[0128] Step 3: Discretely encode the measurement scene classification feature data

[0129] The classification feature data of the measurement scene, including landform type, water area ratio, temperature, humidity, atmospheric pressure and rainfall, are discretized and encoded. That is, according to the encoding rules, different landform types in the classification feature data correspond to a set of encoding values, and different value ranges of water area ratio, temperature, humidity, atmospheric pressure and rainfall in the classification feature data correspond to a set of encoding values.

[0130] Preferably, the encoding rules are as shown in Table 1:

[0131] Table 1

[0132]

[0133]

[0134] After the classification feature data of the measurement scene are encoded, a set of classification feature data of the measurement scene represented by discrete encoding values ​​is obtained.

[0135] Preferably, the classification feature data of the measurement scene represented by discrete coding values ​​is as shown in Table 2:

[0136] Table 2

[0137] Scene Number Landform type Water area ratio Measuring ambient temperature Measuring the humidity of the environment Atmospheric pressure Rainfall 1 5 1 1 2 2 2 2 5 1 2 1 1 1 3 4 2 2 1 1 3 …… 60 1 7 3 2 1 2 61 2 3 2 2 3 3 62 3 1 2 1 2 2 63 6 1 4 2 3 2

[0138] Step 4: Classify the collected scenes and process them using the k-means clustering method. This includes the following steps:

[0139] Step 4.1. Select the number of classification clusters K and the maximum number of iterations m.

[0140] Step 4.2: Select K sample points in the sample that are different in one or more dimensions as the center points of cluster classification.

[0141] Step 4.3: Calculate the distances from other sample points to the K center points to determine which category the sample data belongs to.

[0142] Step 4.4: Re-evaluate the mean of all dimensions according to the classification as the new center point.

[0143] Step 4.5, iterate the above two steps until the center point does not change or the maximum number of iterations m is reached.

[0144] Step 4.6: Get the center point of each classification cluster and calculate the variance of each classification cluster, and store them in the database for persistence.

[0145] In steps 4.1 to 4.6, the samples are acquisition scenes that need to be classified, and the sample points are classification feature data of the measurement scenes obtained in step 3 and represented by discrete coding values.

[0146] Step 5: Merge the radio wave propagation measurement data of multiple scenarios under one category

[0147] In the case where there are measurement data of multiple scenes under one classification cluster, the radio wave propagation measurement data in the same scene category are merged to obtain comprehensive radio wave propagation measurement data for use in subsequent processing.

[0148] Step 6: Clean the radio wave propagation measurement data to remove abnormal points in the data

[0149] Filter the data whose position information is not within the range in the radio wave propagation measurement data. For example, if the longitude of the measurement range is 120°~121° and the latitude position is 30°~32°, filter the data out of the range.

[0150] Filter out the radio wave propagation measurement data that is not within the reasonable range. The reasonable range of radio wave propagation measurement data is required to be above the sensitivity of the signal receiver. For example, taking the received power value as an example, the receiving sensitivity of the signal receiver is -130dBm, and the reasonable range of radio wave propagation measurement data is required to be between -130dBm and -30dBm. Filter out the data that exceeds the range.

[0151] Step 7: For each scene category, perform variability processing on the cleaned radio wave propagation measurement data

[0152] Since there are certain fluctuations in the radio wave propagation measurement data in the receiving location and the measurement time, a certain range of measurement data is taken in the geographical location and measurement time, and the signal strength exceeding a certain probability (the median is 50%) is statistically calculated. Processing the radio wave propagation measurement data in this way makes it easier to study the laws of radio wave propagation.

[0153] In this application, the fixed-size area is considered as a grid, so that the data on the continuous area can be discretized according to the grid. At the same time, in terms of acquisition time, since the data in a grid area corresponds to a piece of acquired data, the acquisition data falling on the grid is processed to handle the variability in area and time at the same time. The regional resolution is the size of the grid.

[0154] The method comprises the following steps:

[0155] Step 7.1: According to the regional resolution requirements of the cleaned data, the data in each resolution grid is screened, and the measurement data of the grid is obtained according to a certain probability (median 50%), that is, the signal strength exceeded by 50% of the measurement values ​​is taken.

[0156] Step 7.2: Use the center position of the grid as the position of the grid data to carry out the subsequent model correction learning process and the prediction process using the correction model.

[0157] Step 8: For each scene category, extract the feature data of the model correction based on Recommendation ITU-R P.1546, where the feature data is the propagation distance and propagation loss data.

[0158] The ITU-R P.1546 model is designed and implemented based on Recommendation ITU-R P.1546-6, which is a recommendation issued by the International Telecommunication Union's Radio Sector (ITU-R): "Recommendation ITU-R P.1546-6 (08 / 2019) Method for point-to-area prediction of terrestrial services in the frequency range 30 MHz to 4 000 MHz". This recommendation describes a prediction method for point-to-area radio propagation for terrestrial services in the frequency range 30 MHz to 4 000 MHz. This method is applicable to land paths, sea paths and / or land-sea mixed paths with effective transmitting antenna heights less than 3 000 m and path lengths between 1 and 1 000 km for radio links. The basic part of the model is the propagation curves for different scenarios, different frequency bands, different time probabilities, and different transmitting antenna heights. For frequencies, time probabilities, and antenna heights other than the curves, interpolation / extrapolation is performed using given formulas.

[0159] Channel parameter extraction refers to extracting parameters from processed data to form vectors in preparation for machine learning. For localized modeling of the ITU-R P.1546 model, the channel parameter extraction process is divided into two parts: basic parameter extraction and basic transmission curve parameter extraction.

[0160] Basic parameters refer to the most basic information in the measurement. The basic parameters that need to be extracted include path type scenario, measurement frequency f, horizontal propagation distance d, and actual measurement value of electromagnetic propagation. When only the median field strength is considered, the time probability t prob is 50%.

[0161] In some cases, the radio propagation measurement data is field strength, and the equivalent conversion from field strength to transmission loss under typical receiving parameters (impedance 50Ω, gain 0dBi) is achieved according to Formula 1.

[0162] P r =E-77.22-20logf MHz Formula 1

[0163] In formula 1, E is the electric field strength at the receiving end, in dBuv / m; P r is the transmission loss, in dBm; f MHz is the measured frequency in MHz.

[0164] In the basic transmission curve part, the basic parameters involved are the measurement frequency fMHz , horizontal propagation distance d km , the effective height of the transmitting antenna h1 and the path type scenario. At the measurement frequency f MHz , the transmitting antenna effective height h1 and the path type scenario are certain, the basic propagation curve is the independent variable d km and the dependent variable received field strength E b The function curve of .

[0165] According to the calculation method for removing the correction items other than the basic propagation loss in Recommendation ITU-R P.1546, combined with the radio wave propagation measurement data, the measurement value of the basic propagation loss is calculated; according to the location of the radio wave propagation measurement data and the parameters of the transmitting antenna and the receiving antenna, a model is built to calculate the characteristic value of the radio wave propagation distance.

[0166] Among them, the measured value of the basic propagation loss part PL b (f MHz ,d km ) is calculated according to the method of Recommendation ITU-R P.1546:

[0167] Total path loss PL all Includes basic propagation loss and related correction terms

[0168] PL all =PL b (f MHz ,d km )+Corr(h2)+Corr(Clu t )+Corr(θ tca ) Formula 2

[0169] In this way, the basic propagation loss is obtained by the following formula

[0170] PL b (f MHz ,d km )=PL all -Corr(h2)-Corr(Clu t )-Corr(θ tca ) Formula 3

[0171] In formula 2 and formula 3,

[0172] PL b (f MHz ,d km ) is the basic propagation loss, unit: dB,

[0173] PL all is the total path loss, unit: dB,

[0174] Corr(h2) is the height correction term at the receiving end, unit: dB,

[0175] Corr(Clu t ) is the clutter correction item at the transmitting end, unit: dB,

[0176] Corr(Clu t ) is the receiving end terrain clearance angle (θ tca ) correction term, unit: dB.

[0177] The receiving end height correction term, the transmitting end clutter correction term, and the receiving end terrain clearance angle correction term are calculated according to the method given in Recommendation ITU-R P.1546.

[0178] The total path loss PL in Formula 2 all The received power and other transmit and receive parameters measured by radio wave propagation are calculated as:

[0179] P r =P t +G t -PL all +G r -L f -L o Formula 4

[0180] PL all =P t +G t +G r -L f -L o -P r Formula 5

[0181] In formula 4 and formula 5, P r is the received power, unit: dBm,

[0182] P t is the power of the transmitted signal, unit: dBm,

[0183] G t is the gain of the transmitting antenna, unit: dB,

[0184] PL all is the total path loss, unit: dB,

[0185] G r is the receiving antenna gain, unit: dB,

[0186] L f is the feeder loss, unit: dB,

[0187] L ois other losses, unit: dB.

[0188] Among them, the calculation of the characteristic value of the radio wave propagation distance is to calculate the radio wave propagation distance according to the geographical geometric relationship through the longitude, latitude and transmitting antenna height of the signal transmitter and the longitude, latitude and receiving antenna height of the receiving data.

[0189] Step 9: For each scene category, use the BP neural network to train the feature data extracted in step 8 to obtain the correction model and save it.

[0190] Using machine learning methods, a calculation model for obtaining propagation loss by inputting distance is obtained, and a set of models for calculating basic propagation loss corresponding to different scenario categories are obtained.

[0191] use Figure 2 The BP neural network of the structure shown is trained on the characteristic data and target data extracted in step 8 to obtain the network parameter results. The characteristic data refers to the log value of the transmission distance (unit km) with a base of 10, and the target data is the measured value of the basic propagation loss part. The network parameter information is saved in the database, and the corresponding scene type is saved at the same time. Repeat steps 5 to 9 to train each scene category to obtain the corresponding model parameters, thereby obtaining a series of model parameters corresponding to each scene, that is, a series of correction models, and persisting them in the database.

[0192] Figure 2 In the BP neural network with the structure shown, the network is set to 1 input layer, 2 hidden layers, and 1 output layer. The input layer has 4 nodes, each hidden layer has 70 nodes, and the output layer has 1 node. The input parameter is the measurement frequency f MHz , horizontal propagation distance d km , the vector composed of the effective height h1 of the transmitting antenna and the path type scenario, and the output result is the electric field strength E at the receiving end b .

[0193] Step 10: Extract parameters for the prediction scenario

[0194] The landform type, water area ratio, temperature, humidity, atmospheric pressure and rainfall characteristic parameters of the scene to be predicted are extracted and discretized and encoded to form a feature vector of the scene characteristic parameters.

[0195] Step 11: Classify the parameters of the scene to be predicted and select the machine learning result correction model of the corresponding scene obtained in step 9

[0196] Calculate the distance between the feature vector of the scene to be predicted and the center point of the classification cluster, select the classification coarse with the closest distance, and classify the scene into this classification cluster. Select the machine learning result correction model of the corresponding scene obtained in step 9.

[0197] Step 12: Predict the radio wave propagation loss in the target scenario

[0198] Find the BP neural network parameters of the correction model corresponding to the classification cluster to which the target scene belongs. Extract the transmission distance (unit km) of the current scene to be predicted and take the log value with base 10 as the input of the network. After the network forward calculation, the basic loss value in this scene can be obtained. Use the recommended method given in Recommendation ITU-R P.1546 to remove the correction items other than the basic propagation loss and calculate the remaining correction items. Combine the results of the two to obtain the final loss prediction value.

[0199] The absolute error between the propagation model loss value predicted by the method of the present invention and the measured loss value is within 10 dB, and the root mean square error can reach between 1 and 5 dB.

[0200] Example

[0201] There are 5 different propagation measurement data, corresponding to 5 different scenarios.

[0202] Step 1: Obtain classification feature data of sample collection scenarios

[0203] For these five scenes, the classification feature data of each scene is obtained, and the data in the following format is obtained:

[0204]

[0205] Step 2: Clean the classified feature data of the scene. For example, if the temperature data of the seventh item is relatively abnormal, the data is deleted during the cleaning process.

[0206] Step 3: Encode the classification feature data of the scene and obtain the following results:

[0207] Scene Number Landform type Water area ratio Measuring ambient temperature Measuring the humidity of the environment Atmospheric pressure Rainfall 1 5 1 1 2 2 2 2 5 1 2 1 1 1 3 4 2 2 1 1 3 4 3 1 3 2 2 2 5 3 1 2 2 3 3

[0208] Step 4: Classify the five acquisition scenes, process them according to the number of classification clusters K = 3 target categories, select the maximum number of iterations m = 10000, and get the following results:

[0209] Scene Number Classification 1 1 2 1 3 2 4 3 5 3

[0210] Step 5: Merge the radio wave propagation measurement data under each scenario classification.

[0211] Taking scenario category 3 as an example, the radio wave propagation measurement data of scenarios 4 and 5 are merged together. The data includes the measured receiving position, altitude, received signal strength, transmitting station information, etc.

[0212] Step 6: Clean the combined radio wave propagation measurement data for each scene category

[0213] For example, in scenario category 3, the data that does not reach the receiver sensitivity in the merged data is deleted, and the data whose measurement distance range is just beyond the applicable range of the model is cleaned.

[0214] Step 7: For each scene category, the cleaned radio wave propagation measurement data is processed for variability.

[0215] A certain range of measurement data is taken at the geographical location and measurement time, and the signal strength exceeding a certain probability (the median value is 50%) is statistically calculated.

[0216] Step 8: For each scene category, extract feature data for model correction.

[0217] For scenario category 3, information such as the propagation distance and the measured received signal strength is extracted from the data obtained in step 6, and combined with the parameters of the transmitting station, the loss on the propagation path is extracted.

[0218] Step 9: For each scene category, use the BP neural network to train the feature data extracted in step 8.

[0219] For example, for scene category 3, training can obtain a set of parameters (70*70) for the first hidden layer, parameters (70*70) for the second hidden layer, and parameters (70*1) for the output layer. Similarly, after steps 5 to 9, such correction models can be obtained for scene category 1 and scene category 3. Save these correction models.

[0220] Scene Classification Calibration model 1 Calibration model 1 2 Calibration Model 2 3 Correction Model 3

[0221] Step 10: Now there is a scene to be predicted for electromagnetic propagation, for example, scene 11. The classification feature data of the scene is obtained in the following format:

[0222] Data No. Scene Number Data Types Feature data 1 11 Landform type Mountain 2 11 Water area ratio 0 3 11 Measuring ambient temperature 12℃ 4 11 Measuring ambient temperature 15℃ 5 11 Measuring ambient temperature 14℃ 6 11 Measuring ambient temperature 16℃ 7 11 Measuring ambient temperature 20℃ 8 11 Measuring ambient humidity 51 9 11 Measuring ambient humidity 52 10 11 Atmospheric pressure 1010hPa 11 11 Atmospheric pressure 1011hPa 12 11 Rainfall 33mm

[0223] Step 11: Encode and classify the scene to be predicted to obtain the scene data scene category 3

[0224] Scene Number Landform type Water area ratio Measuring ambient temperature Measuring the humidity of the environment Atmospheric pressure Rainfall 11 3 1 2 2 3 3

[0225] Step 12: Use the correction model 3 to predict the correction propagation situation in the prediction scenario to obtain a prediction result.

Claims

1. A method for predicting radio wave propagation based on scenario adaptation for spectrum sharing, characterized by: The method comprises the following steps: Step 1: Collect N radio wave propagation measurement data samples and obtain classification feature data of the sample collection scene The radio wave propagation measurement data is power values ​​of multiple groups of input and output signals in different scenarios; Classification feature data of sample collection scenarios, The following characteristic factors are included: landform type: landform characteristics of the radio wave propagation scene, the landform types include: water, mixed water and land, mountain, forest, dense urban area, suburb, and rural area; Water area ratio: the percentage of water area in the measurement scene to the total area; Temperature: The air temperature of the measured environment, expressed in degrees Celsius; Humidity: The humidity of the test environment, expressed as relative humidity; Atmospheric pressure: The atmospheric pressure of the measuring environment, expressed in hectopascals; Precipitation: The annual average daily rainfall of the measurement environment, expressed in mm / day; Step 2: Clean the classification feature data of the scene Clean the extracted scene classification data and remove unreasonable scene sample data; Step 3: Discretely encode the measurement scene classification feature data Discretely encode the classification feature data of the measurement scene, including landform type, water area ratio, temperature, humidity, atmospheric pressure and rainfall; that is, according to the encoding rules, correspond different landform types in the classification feature data to a set of encoding values, and correspond different value ranges of water area ratio, temperature, humidity, atmospheric pressure and rainfall in the classification feature data to a set of encoding values; After encoding the classification feature data of the measurement scene, a set of classification feature data of the measurement scene represented by discrete coding values ​​is obtained; Step 4: Classify the collected scenes and process them using the k-means clustering method; this includes the following steps: Step 4.1, select the number of classification clusters K and the maximum number of iterations m; Step 4.2: Select K sample points that are different in one or more dimensions from the sample as the center points of cluster classification; Step 4.3, calculate the distances from other sample points to the K center points to determine which category the sample data belongs to; Step 4.4: Re-evaluate the mean of all dimensions according to the classification as the new center point; Step 4.5, iterate the above two steps until the center point does not change or the maximum number of iterations m is reached; Step 4.6, obtain the center point of each classification cluster and calculate the variance of each classification cluster, and store them in the database for persistence; In steps 4.1 to 4.6, the sample is a collection scene that needs to be classified, and the sample point is the classification feature data of the measurement scene represented by discrete coding values ​​obtained in step 3; Step 5: Merge the radio wave propagation measurement data of multiple scenarios under one category In the case where there are measurement data of multiple scenes under one classification cluster, the radio wave propagation measurement data in the same scene category are merged to obtain comprehensive radio wave propagation measurement data for use in subsequent processing; Step 6: Clean the radio wave propagation measurement data to remove abnormal points in the data Step 7: For each scene category, perform variability processing on the cleaned radio wave propagation measurement data Step 7.1, according to the regional resolution requirements of the cleaned data, the data in each resolution grid is screened, and the measurement data of the grid is obtained according to the probability exceeding a certain level, that is, the signal strength that 50% of the measurement values ​​exceed is taken; Step 7.2, using the center position of the grid as the position of the grid data, and performing a subsequent learning process of model correction and a prediction process using the correction model; Step 8: For each scene category, extract the feature data of the model correction based on the ITU-RP.1546 recommendation, the feature data is the propagation distance and propagation loss data. For the localized modeling of the ITU-RP.1546 model, the channel parameter extraction process is divided into two parts: basic parameter extraction and basic transmission curve parameter extraction; Basic parameters refer to the most basic information in the measurement. The basic parameters that need to be extracted include path type scenario, measurement frequency f, horizontal propagation distance d, and actual measured value of electromagnetic propagation. When only the median field strength is considered, the time probability t prob 50%; According to Formula 1, an equivalent transformation from field strength to transmission loss under typical receiving parameters is achieved, wherein the typical receiving parameters are impedance 50Ω and gain 0dBi; P r =E-77.22-20logf MHz Formula 1 In formula 1, E is the electric field strength at the receiving end, in dBuv / m; P r is the transmission loss, in dBm; f MHz is the measured frequency in MHz; In the basic transmission curve part, the basic parameters involved are the measurement frequency f MHz , horizontal propagation distance d km , the effective height h1 of the transmitting antenna and the path type scenario; at the measurement frequency f MHz , the transmitting antenna effective height h1 and the path type scenario are certain, the basic propagation curve is the independent variable d km and the dependent variable received field strength E b Function curve of According to the calculation method of the correction item except the basic propagation loss in Recommendation ITU-RP.1546, combined with the radio wave propagation measurement data, the measurement value of the basic propagation loss part is calculated; according to the location of the radio wave propagation measurement data and the parameters of the transmitting antenna and the receiving antenna, the model is built to calculate the characteristic value of the radio wave propagation distance; Among them, the measured value of the basic propagation loss part PL b (f MHz ,d km ) is calculated according to the method of Recommendation ITU-RP.1546: Total path loss PL all Includes basic propagation loss and related correction terms PL all = PL b (f MHz , d km ) + Corr(h2) + Corr(Clu t ) + Corr(θ tca ) Formula 2 In this way, the basic propagation loss is obtained by the following formula PL b (f MHz ,d km ) = PL all -Corr(h2)-Corr(Clu t )-Corr(θ tca ) Equation 3 In formula 2 and formula 3, PL b (f MHz ,d km ) is the basic propagation loss, unit: dB, PL all is the total path loss, unit: dB, Corr(h2) is the height correction term at the receiving end, unit: dB, Corr(Clu t ) is the clutter correction item at the transmitting end, unit: dB, Corr(Clu t ) is the receiving end terrain clearance angle (θ tca ) correction term, unit: dB; Among them, the receiving end height correction term, the transmitting end clutter correction term, and the receiving end terrain clearance angle correction term are calculated according to the method given in Recommendation ITU-RP.1546; The total path loss PL in Formula 2 all The received power and other transmit and receive parameters measured by radio wave propagation are calculated as follows: P r =P t +G t -PL all +G r -L f -L o Formula 4 PL all =P t +G t +G r -L f -L o -P r Formula 5 In formula 4 and formula 5, P r is the received power, unit: dBm, P t is the power of the transmitted signal, unit: dBm, G t is the gain of the transmitting antenna, in dB, PL all is the total path loss, unit: dB, G r is the receiving antenna gain, unit: dB, L f is the feeder loss, unit: dB, L o is other loss, unit: dB; The calculation of the characteristic value of the radio wave propagation distance is carried out by calculating the radio wave propagation distance according to the geographical geometric relationship through the longitude, latitude and transmitting antenna height of the signal transmitter and the longitude, latitude and receiving antenna height of the receiving data; Step 9: For each scene category, use the BP neural network to train the feature data extracted in step 8 to obtain the correction model and save it. Using machine learning methods, a calculation model for obtaining propagation loss by inputting distance is obtained, and a set of models for calculating basic propagation loss corresponding to different scene categories are obtained; Use BP neural network to train the feature data and target data extracted in step 8 to obtain the network parameter results; the feature data refers to the log value of the transmission distance with base 10, and the target data is the measured value of the basic propagation loss part; and save the network parameter information to the database, and save the corresponding scene type at the same time; repeat steps 5 to 9, train each scene category to obtain the corresponding model parameters, so as to obtain a series of model parameters corresponding to each scene, that is, a series of correction models, and persist them in the database; In the BP neural network, the network is set to 1 input layer, 2 hidden layers, and 1 output layer. The input layer has 4 nodes, the hidden layers have 70 nodes each, and the output layer has 1 node. The input parameter is the measurement frequency f MHz , horizontal propagation distance d km , the vector composed of the effective height h1 of the transmitting antenna and the path type scenario, and the output result is the electric field strength E at the receiving end b ; Step 10: Extract parameters for the prediction scenario Extract the landform type, water area ratio, temperature, humidity, atmospheric pressure and rainfall characteristic parameters of the scene to be predicted, and perform discretization encoding to form a feature vector of the scene characteristic parameters; Step 11: Classify the parameters of the scene to be predicted and select the machine learning result correction model of the corresponding scene obtained in step 9 Calculate the distance between the feature vector of the scene to be predicted and the center point of the classification cluster, select the classification cluster with the closest distance, and classify the scene into this classification cluster; select the machine learning result correction model of the corresponding scene obtained in step 9; Step 12: Predict the radio wave propagation loss in the target scenario Find the BP neural network parameters of the correction model corresponding to the classification cluster to which the target scene belongs; extract the log value with base 10 of the transmission distance of the current scene to be predicted as the input of the network, and obtain the basic loss value under the scene through forward calculation of the network; use the recommended method given in Recommendation ITU-R P.1546 to remove the correction items other than the basic propagation loss, and calculate the remaining correction items; combine the results of the two to obtain the final loss prediction value.

2. The method for predicting radio wave propagation based on scenario adaptation for spectrum sharing according to claim 1, characterized in that: In step 2, the extracted scene classification data is cleaned including: Landform type: if the landform data is not in the candidate scene, the data is considered unreasonable and is manually modified to meet the requirements or deleted; the candidate scenes include: water, mixed water and land, mountain, forest, dense urban area, suburb, and rural area; Water area ratio: If the water area ratio is less than 0 or greater than 100, it is considered illegal data and will be removed; Temperature: If the temperature is outside the range of -30℃~60℃, it is considered as illegal data, and the data must be manually confirmed and modified or the sample data must be deleted; Humidity: If the relative humidity exceeds the range of 30 to 80, it is considered illegal data and the data must be manually confirmed and modified or the sample data must be deleted; Atmospheric pressure: If the atmospheric pressure data exceeds the value range of 800-1100, the data is considered illegal and the data must be manually confirmed and modified or the sample data must be deleted; Rainfall: If the rainfall data exceeds the value range of 0 to 1000 mm / day, the data is considered illegal and the data must be manually confirmed and modified or the sample data must be deleted.

3. The method for predicting radio wave propagation based on scenario adaptation for spectrum sharing according to claim 1, characterized in that: In step 6, the radio wave propagation measurement data is cleaned to remove abnormal points in the data, including: Filter the data whose position information in the radio wave propagation measurement data is not within the range; Filter out the radio wave propagation measurement data that is not within a reasonable range; the reasonable range of radio wave propagation measurement data is required to be above the sensitivity of the signal receiver, and the reasonable range of radio wave propagation measurement data is required to be between -130dBm and -30dBm; filter out the data that exceeds the range.

4. The method for predicting radio wave propagation based on scenario adaptation for spectrum sharing according to claim 1, characterized in that: In step 7.1, the median value of the probability is 50%.

Images