A method and system for ground-ground collaborative anti-rain attenuation based on deep learning feature framework

Abstract

A space-ground collaborative anti-rainfall method and system based on a deep learning feature framework provides a high-precision rainfall attenuation prediction and compensation method that meets the actual weather conditions in the area through a space-ground collaborative anti-rainfall system composed of an image acquisition module, a satellite-ground link, and a collaborative control module. Through autonomous learning of the correlation mapping between the joint features of time-continuous meteorological observation image data and one-dimensional meteorological data and the attenuation of the satellite-ground link, the difficult problems of difficult to accurately predict rainfall attenuation and unstable satellite-ground link performance in engineering can be effectively solved.

Description

Technical Field

[0001] The present invention relates to a ground-ground collaborative anti-rain attenuation method and system based on a deep learning feature framework, belonging to the technical field of satellite communications. Background Art

[0002] With the development and evolution of broadband satellite communication technology, the spectrum resources of traditional low-frequency bands such as C and Ku are becoming increasingly scarce. In order to meet the growing demand for communication capacity and transmission rate of satellite systems, high-throughput satellites and very high-throughput satellites based on higher frequency bands such as Ka and Q / V have received widespread attention. Higher frequency bands can provide more spectrum resource bandwidth and effectively reduce the transmission cost per bit, but they are easily affected by weather factors such as rainfall.

[0003] The communication quality of the satellite-to-ground link is strongly correlated with the weather conditions in the area where the ground station is located. Taking the ground station in Beijing as an example, the cumulative rainfall since the beginning of the flood season in 2023 (June 1-August 30) has reached about 800 mm. Frequent rainfall has caused the satellite-to-ground link deployed by the ground station in Beijing to be seriously affected by rainfall attenuation. When the rainfall intensity is too high in a short period of time, it may even cause the link to be interrupted.

[0004] In the field of rainfall attenuation prediction, a large number of experiments have been carried out at home and abroad, and several rain attenuation prediction models have been established. Among them, the ITU-R model and the SAM (Simple Attenuation Model) model are the most widely used. The ITU-R model obtains the detailed empirical relationship between the power law function coefficients and the radio wave frequency and polarization mode based on experimental data, and then calculates the rainfall attenuation under arbitrary probability conditions based on the annual average single-point statistical rainfall with a probability of 0.01%; the SAM model gives a simpler relationship model between the power law coefficient and the radio wave frequency, and combines statistical data to obtain the horizontal distribution model of rain intensity under the conditions of stratiform cloud and convective cloud precipitation, and calculates the rainfall attenuation on this basis.

[0005] However, when the satellite-to-ground link propagates in the atmosphere, the macroscopic and microscopic characteristics such as the height of the rainfall area, the spectral distribution of rainfall, the type and shape of particles, etc. change in time and space in a complex manner. The above-mentioned traditional empirical model based on probability statistics is difficult to accurately calculate the attenuation characteristics caused by rainfall and cannot meet the needs of practical applications. Therefore, it is necessary to invent a satellite-to-ground link rainfall attenuation and link quality assessment method based on the actual weather conditions in the area where the ground station is located to improve the accuracy of rain attenuation forecasting. At the same time, according to the changes in the rain attenuation forecast level, the ground station and satellite power amplifier gears are adjusted in time to compensate for the impact of the rain attenuation level changes on the signal, and to ensure reliable and stable transmission of the satellite-to-ground link. Summary of the invention

[0006] The technical problem solved by the present invention is: In view of the problem that the empirical model based on probability statistics in the current existing technology is difficult to accurately calculate the attenuation characteristics caused by rainfall and cannot meet the needs of practical applications, a ground-ground collaborative anti-rainfall attenuation method and system based on a deep learning feature framework is proposed.

[0007] The present invention solves the above technical problems by the following technical solutions:

[0008] A ground-ground collaborative anti-rain attenuation method based on a deep learning feature framework, comprising:

[0009] Collect original image data and meteorological data through remote sensing satellites, and collect satellite-to-ground link signal data through ground stations;

[0010] The original image data, meteorological data, and satellite-to-ground link signal data are all preprocessed;

[0011] The rainfall attenuation level obtained after preprocessing of the satellite-to-ground link signal is used to establish a corresponding relationship with the preprocessed image data and meteorological data according to the timestamp, and a training set and a test set are constructed;

[0012] Establish the spatiotemporal convolution model structure and train it through the training set;

[0013] The trained spatiotemporal convolution model is tested by inputting the test set, and whether the output forecast accuracy meets the preset accuracy requirements is determined based on the output rainfall attenuation level;

[0014] The trained spatiotemporal convolutional model that meets the preset accuracy requirements is used for rain attenuation forecasting.

[0015] The original image data is a two-dimensional image time series of the ground station area collected and photographed by meteorological satellites or remote sensing satellites and with time stamps; the meteorological data is a one-dimensional meteorological data time series of the ground station area collected by meteorological satellites or remote sensing satellites, including rainfall, temperature, humidity, wind direction, wind speed, and air pressure data; the satellite-to-ground link signal data is the satellite-to-ground link signal attenuation collected by the ground station and with time stamps.

[0016] The preprocessing of the original image data is specifically as follows:

[0017] Set the ground station as the center point of the image and determine the actual maximum rainfall height H at the location of the ground station;

[0018] Calculate the inclined path length L of radio wave propagation and the path length L of the actual highest rainfall height at the ground station and the horizontal projection S ;

[0019] Calculate the minimum frame width A of the image collected by the meteorological satellite or remote sensing satellite based on the obtained data min ;

[0020] The minimum frame width of the image is taken as the minimum frame side length after image cropping, and image cropping is performed frame by frame on the time-series two-dimensional image information of the original image data.

[0021] The minimum image width A min The calculation method is:

[0022]

[0023]

[0024] A min =2×L S

[0025] In the formula, is the minimum operating elevation angle of the earth station antenna, in degrees, and R is the effective radius of the earth in the presence of refractive index.

[0026] The meteorological data preprocessing method is:

[0027] The single-bit data of rainfall, temperature, humidity, wind direction, wind speed, and air pressure are concatenated row by row to construct a matrix, with each row of the matrix representing the value of a single parameter at different times, and each column representing the values ​​of all parameters at the same time;

[0028] Interpolation is performed on each row of matrix data, inserting n-1 data points between two adjacent sampling moments;

[0029] Divide the matrix into a specified number of sub-matrices by column, where each sub-matrix has the same number of rows as the original matrix and n number of columns;

[0030] Any sub-matrices are concatenated so that the size of the concatenated matrix is ​​the same as the size of the preprocessed two-dimensional image.

[0031] The signal attenuation of the satellite-to-ground link is graded and the maximum rainfall attenuation allowed for normal operation of the satellite-to-ground link is determined as R loss , the rainfall attenuation level is divided into five levels according to the attenuation amount, among which [80%R loss ,100%R loss ] The range is A, [60% R loss ,80%R loss ) interval is B level, [40%R loss ,60%R loss ) interval is C level, [20%R loss ,40%R loss ) interval is D level, [0%R loss ,20%R loss ) interval is E level.

[0032] The training set is the image data and meteorological data after the corresponding relationship is established. The spatiotemporal convolution model structure is realized by adding a convolution operation to the temporal neural network model. The input gate, forget gate and output gate structure are used as gate control units in the spatiotemporal convolution model structure. The information input in each gate structure adopts a convolution operation, the cell state is updated by a dot product operation, and the weight is set in the convolution kernel. The spatiotemporal convolution model structure is specifically as follows:

[0033] i t =σ(W xi · χ t +W hi · χ t-1 +W ci ·C t-1 +b i )

[0034] f t =σ(W xf · χ t +W hf ·H t-1 +W cf ·C t-1 +b f )

[0035] C t =f t ·C t-1 +i t tanh(W xc · χ t +W hc ·H t-1 +b c )

[0036] o t =σ(W xo · χ t +W ho ·H t-1 +W co ·C t-1 +b o )

[0037] H t =o t tanh(C t )

[0038] In the formula, i t represents the input gate, f t represents the forget gate, o t represents the output gate; c t Indicates the current state, c t-1 Indicates the state at the previous moment; h trepresents the final output; W represents the weight coefficient, b represents the corresponding bias coefficient; σ is the sigmoid function, * represents the convolution operation, ○ represents the Hadamard operation, that is, the corresponding elements of the matrix are multiplied;

[0039] The training set is input into the spatiotemporal convolution model to train the model, and the probability of forecasting five rainfall attenuation levels is output.

[0040] According to the rainfall attenuation level corresponding to the input data in the test set, the rainfall attenuation level of the output of the trained spatiotemporal convolution model is judged, and the output forecast accuracy y% is determined, and compared with the preset accuracy requirement x%. If y is greater than or equal to x, the accuracy of the trained spatiotemporal convolution model meets the preset accuracy requirement; if y is less than x, the accuracy does not meet the requirements, and the model is retrained using the training set.

[0041] A space-ground collaborative anti-rain fading system based on a deep learning feature framework for realizing a space-ground collaborative anti-rain fading method comprises an image acquisition module, a satellite-ground link, and a collaborative control module, wherein:

[0042] The image acquisition module collects raw image data and meteorological data through the satellite-to-ground link, collects satellite-to-ground link signal data at the same time, pre-processes all data, and sends all pre-processed data to the collaborative control module;

[0043] Satellite-to-ground link, including remote sensing satellites and ground stations. Original image data and meteorological data are collected through remote sensing satellites, and satellite-to-ground link signal data is collected through ground stations.

[0044] The collaborative control module uses the rainfall attenuation level obtained after preprocessing the satellite-to-ground link signal according to the timestamp, establishes a corresponding relationship with the preprocessed image data and meteorological data according to the timestamp, constructs a training set and a test set, and constructs a spatiotemporal convolution model structure after adding a convolution operation through a temporal neural network model. After completing the training of the spatiotemporal convolution model with the training set, the trained spatiotemporal convolution model is tested through the test set input. According to the output rainfall attenuation level, it is determined whether the output forecast accuracy meets the preset accuracy requirement. If not, the training iteration is repeated until the preset accuracy requirement is met, and a trained spatiotemporal convolution model that meets the preset accuracy requirement is obtained; the trained spatiotemporal convolution model is applied to rainfall attenuation forecasting, and according to the rainfall forecast level and the power level of the remote sensing satellite or ground station obtained in real time, the power level of the remote sensing satellite or ground station is adjusted through the satellite-to-ground link.

[0045] The collaborative control module includes a rain attenuation prediction module and a power regulation module;

[0046] Rain attenuation forecast module, used to construct training sets and test sets, and complete the training of spatiotemporal convolutional models to output real-time rainfall forecast levels;

[0047] The power regulation module calculates and predicts the required power according to the real-time rainfall forecast level and the power level of the remote sensing satellite or ground station, and adjusts the power level of the remote sensing satellite or ground station through the satellite-to-ground link.

[0048] The advantages of the present invention compared with the prior art are:

[0049] (1) The present invention provides a method and system for ground-to-space collaborative anti-rainfall attenuation based on a deep learning feature framework. Aiming at the problem of poor prediction accuracy of satellite-to-ground link rainfall attenuation in traditional models, the present invention provides a high-precision rainfall attenuation prediction and compensation method that meets the actual weather conditions in the region by designing a ground-to-space collaborative rainfall attenuation prediction method. The present invention also provides an autonomous learning method for mapping the joint features of time-continuous meteorological observation image data and one-dimensional meteorological data with satellite-to-ground link attenuation. The system can effectively solve the problem of difficult accurate prediction of rainfall attenuation and unstable satellite-to-ground link performance in engineering.

[0050] (2) The present invention adopts a deep learning method to accurately predict the impact of rainfall on the satellite-to-ground link and compensates for it by increasing or decreasing the power amplifier gear. The actual attenuation of the satellite-to-ground link is controlled within the range of 2 dB, thereby effectively reducing the impact of rain attenuation on the signal strength of the satellite-to-ground link.

[0051] (3) Aiming at the actual operation system requirements, the present invention inputs the real-time collected meteorological data, including meteorological data obtained by sky meteorological radar or remote sensing images and ground meteorological observation stations, into the trained network to obtain the real-time rain attenuation evaluation prediction, and drives the satellite power amplifier gear control adjustment device according to the rain attenuation prediction result, so that the on-board power amplifier state can adapt to the link rain attenuation loss and maintain the quality of the satellite-to-ground link. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the ground-ground collaborative anti-rain attenuation method and system provided for the invention;

[0053] Figure 2 A network training flow chart based on a spatiotemporal combined deep learning framework provided for the invention;

[0054] Figure 3 A schematic diagram of network training provided for the invention;

[0055] Figure 4 A schematic diagram of input parameters for the satellite-to-earth path rainfall attenuation prediction process provided by the invention;

[0056] Figure 5 A schematic diagram of the structure of the spatiotemporal convolution model provided for the invention;

[0057] Figure 6 A schematic diagram of the working process of the power regulation module provided for the invention;

[0058] Figure 7 An example diagram of the earth-sky collaborative anti-rain attenuation process provided for the invention;

[0059] Figure 8 Link attenuation compensation diagram provided for the invention; DETAILED DESCRIPTION

[0060] A space-ground collaborative anti-rainfall method and system based on a deep learning feature framework provides a high-precision rainfall attenuation prediction and compensation method that meets the actual weather conditions in the area through a space-ground collaborative anti-rainfall system composed of an image acquisition module, a satellite-ground link, and a collaborative control module. Through autonomous learning of the correlation mapping between the joint features of time-continuous meteorological observation image data and one-dimensional meteorological data and the attenuation of the satellite-ground link, it can effectively solve the difficult problems of accurate prediction of rainfall attenuation in engineering and unstable performance of the satellite-ground link.

[0061] The ground-ground collaborative anti-rain attenuation method based on the deep learning feature framework has the following specific steps:

[0062] Collect original image data and meteorological data through remote sensing satellites, and collect satellite-to-ground link signal data through ground stations;

[0063] The original image data, meteorological data, and satellite-to-ground link signal data are all preprocessed;

[0064] The rainfall attenuation level obtained after preprocessing of the satellite-to-ground link signal is used to establish a corresponding relationship with the preprocessed image data and meteorological data according to the timestamp, and a training set and a test set are constructed;

[0065] Establish the spatiotemporal convolution model structure and train it through the training set;

[0066] The trained spatiotemporal convolution model is tested by inputting the test set, and whether the output forecast accuracy meets the preset accuracy requirements is determined based on the output rainfall attenuation level;

[0067] The trained spatiotemporal convolutional model that meets the preset accuracy requirements is used for rain attenuation forecasting.

[0068] The original image data is a two-dimensional image time series of the ground station area collected and shot by meteorological satellites or remote sensing satellites and with time stamps; the meteorological data is a one-dimensional meteorological data time series of the ground station area collected by meteorological satellites or remote sensing satellites, including rainfall, temperature, humidity, wind direction, wind speed, and air pressure data; the satellite-to-ground link signal data is the satellite-to-ground link signal attenuation collected by the ground station and with time stamps.

[0069] The preprocessing of the original image data is as follows:

[0070] Set the ground station as the center point of the image and determine the actual maximum rainfall height H at the location of the ground station;

[0071] Calculate the inclined path length L of radio wave propagation and the path length L of the actual highest rainfall height at the ground station and the horizontal projection S ;

[0072] Calculate the minimum frame width A of the image collected by the meteorological satellite or remote sensing satellite based on the obtained data min ;

[0073] The minimum frame width of the image is taken as the minimum frame side length after image cropping, and image cropping is performed frame by frame on the time-series two-dimensional image information of the original image data.

[0074] The meteorological data preprocessing method is:

[0075] The single-bit data of rainfall, temperature, humidity, wind direction, wind speed, and air pressure are concatenated row by row to construct a matrix, with each row of the matrix representing the value of a single parameter at different times, and each column representing the values ​​of all parameters at the same time;

[0076] Interpolation is performed on each row of matrix data, inserting n-1 data points between two adjacent sampling moments;

[0077] Divide the matrix into a specified number of sub-matrices by column, where each sub-matrix has the same number of rows as the original matrix and n number of columns;

[0078] Any sub-matrices are concatenated so that the size of the concatenated matrix is ​​the same as the size of the preprocessed two-dimensional image.

[0079] The signal attenuation of the satellite-to-ground link is graded and the maximum rainfall attenuation allowed for normal operation of the satellite-to-ground link is determined as R loss , the rainfall attenuation level is divided into five levels according to the attenuation amount.

[0080] The training set is the image data and meteorological data after the corresponding relationship is established. The spatiotemporal convolution model structure is realized by adding the convolution operation to the temporal neural network model. The input gate, forget gate and output gate structure are used as gating units in the spatiotemporal convolution model structure. The information input in each gate structure adopts the convolution operation, the cell state is updated by the dot product operation, and the weight is set in the convolution kernel. After the training set is input into the spatiotemporal convolution model, the model is trained and the output forecast probability of five rainfall attenuation levels is predicted.

[0081] According to the rainfall attenuation level corresponding to the input data in the test set, the rainfall attenuation level of the output of the trained spatiotemporal convolution model is judged, and the output forecast accuracy y% is determined, and compared with the preset accuracy requirement x%. If y is greater than or equal to x, the accuracy of the trained spatiotemporal convolution model meets the preset accuracy requirement; if y is less than x, the accuracy does not meet the requirements, and the model is retrained using the training set.

[0082] The specific architecture of the ground-ground collaborative anti-rain fade system based on the deep learning feature framework is as follows:

[0083] The image acquisition module collects raw image data and meteorological data through the satellite-to-ground link, collects satellite-to-ground link signal data at the same time, pre-processes all data, and sends all pre-processed data to the collaborative control module;

[0084] Satellite-to-ground link, including remote sensing satellites and ground stations. Original image data and meteorological data are collected through remote sensing satellites, and satellite-to-ground link signal data is collected through ground stations.

[0085] The collaborative control module uses the rainfall attenuation level obtained after preprocessing the satellite-to-ground link signal according to the timestamp, establishes a corresponding relationship with the preprocessed image data and meteorological data according to the timestamp, constructs a training set and a test set, and constructs a spatiotemporal convolution model structure after adding a convolution operation through a temporal neural network model. After completing the training of the spatiotemporal convolution model with the training set, the trained spatiotemporal convolution model is tested through the test set input. According to the output rainfall attenuation level, it is determined whether the output forecast accuracy meets the preset accuracy requirement. If not, the training iteration is repeated until the preset accuracy requirement is met, and a trained spatiotemporal convolution model that meets the preset accuracy requirement is obtained; the trained spatiotemporal convolution model is applied to rainfall attenuation forecasting, and according to the rainfall forecast level and the power level of the remote sensing satellite or ground station obtained in real time, the power level of the remote sensing satellite or ground station is adjusted through the satellite-to-ground link.

[0086] Rain attenuation forecast module, used to construct training sets and test sets, and complete the training of spatiotemporal convolutional models to output real-time rainfall forecast levels;

[0087] The power regulation module calculates and predicts the required power according to the real-time rainfall forecast level and the power level of the remote sensing satellite or ground station, and adjusts the power level of the remote sensing satellite or ground station through the satellite-to-ground link.

[0088] The following is further described in conjunction with the accompanying drawings and preferred embodiments of the specification:

[0089] In the current embodiment, if Figure 1As shown in the figure, it is the architecture of the space-ground collaborative anti-rain fade system based on the deep learning feature framework. The device system consists of three parts. The first is a meteorological satellite or remote sensing satellite and image acquisition station for collecting two-dimensional image information, and a meteorological data acquisition station for one-dimensional meteorological data such as rainfall, temperature, humidity, wind direction, wind speed, and air pressure, namely the image acquisition module; the second is a satellite-to-ground link system affected by rain fade composed of target satellites and ground stations; the third is a control center including a rain fade prediction module and a power regulation module, namely the collaborative control module. The collaborative control module also includes a rain fade prediction module and a power regulation module;

[0090] The network used in the rain attenuation prediction module, the specific model training process is as follows:

[0091] Step 1: Raw data collection.

[0092] The original data consists of two parts. The first is a time series of two-dimensional images of the ground station area taken by meteorological satellites or remote sensing satellites with time stamps, as well as a time series of one-dimensional meteorological data such as rainfall, temperature, humidity, wind direction, wind speed, and air pressure near the ground station area, which serve as input data for model training. The second is the time-stamped satellite-to-ground link signal attenuation collected by the ground station as a data label.

[0093] Step 2: Data preprocessing.

[0094] The time-series two-dimensional image information obtained in step 1 is preprocessed frame by frame, including image cropping and adding noise.

[0095] The image cropping process should follow two principles. First, the 2D image data should take the location of the ground station as the center point of the image. Second, the side length of the cropped image should be greater than or equal to the minimum side length A. min The geometric relationship of related parameters is as follows: Figure 4 shown.

[0096] Calculate the minimum frame side length A min The process is as follows:

[0097] 1) Obtain the actual maximum rainfall height H at the location of the earth station, in km.

[0098] 2) Calculate the length L of the inclined path for radio wave propagation in km.

[0099]

[0100] Where:

[0101] ——Minimum operating elevation angle of earth station antenna, °;

[0102] R is the effective radius of the Earth taking into account the refractive index, generally 8500km.

[0103] 3) Calculate the path length L of the horizontal projection S , km.

[0104]

[0105] 4) Calculate the minimum frame width A of the image collected by the meteorological satellite or remote sensing satellite min , km.

[0106] A min =2×L S .........(3)

[0107] The one-dimensional meteorological data such as rainfall, temperature, humidity, wind direction, wind speed, and air pressure obtained in step 1 are preprocessed frame by frame, including data splicing and conversion of one-dimensional data to two-dimensional data. The processing flow is as follows:

[0108] 1) One-dimensional data such as rainfall, temperature, humidity, wind direction, wind speed, and air pressure are concatenated row by row to construct a matrix, where each row of the matrix represents the value of a single parameter (such as rainfall, temperature) at different times, and each column represents the values ​​of all parameters at the same time;

[0109] 2) Perform interpolation processing on each row of matrix data, inserting n-1 data points between two adjacent sampling moments; n is greater than or equal to 2; the interpolation processing adopts a known method;

[0110] 3) Divide the matrix into several sub-matrices by columns, where each sub-matrix has the same number of rows as the original matrix and n number of columns;

[0111] 4) Concatenate multiple sub-matrices so that their size is the same as the size of the preprocessed two-dimensional image.

[0112] The rainfall attenuation is graded, assuming that the maximum rainfall attenuation allowed for normal operation of the satellite-to-ground link is R loss The rainfall attenuation level is divided into five levels according to the attenuation amount, among which [80%R loss ,100%R loss ] The range is A, [60% R loss ,80%R loss ) interval is B level, [40%R loss ,60%R loss ) interval is C level, [20%R loss ,40%R loss ) interval is D level, [0%R loss ,20%R loss ) interval is E level.

[0113] Step 3: Construct training set and test set.

[0114] A one-to-one correspondence is established between the cropped two-dimensional image in step 2 and the rainfall attenuation levels (five levels: A, B, C, D, and E) according to the timestamps, and a training set and a test set are constructed, where the timestamp error between the image information and the attenuation information should be less than or equal to 5 seconds.

[0115] Step 4: Train the neural network model using the training set

[0116] In order to better extract the spatiotemporal features of image data with spatiotemporal variations, the present invention takes into account both spatial structure information and temporal dependency information, and adds convolution operations on the basis of the temporal neural network model structure, so that the feature extraction of temporal images will be more effective. The input gate, forget gate and output gate structure are used as the gating unit. The information input in each gate structure adopts convolution operation, and the cell state update adopts dot product operation. The weight part is placed in the convolution kernel, and part is placed in the recurrent kernel of the recurrent layer. The model unit is as follows: Figure 5 shown.

[0117] The model unit formula is expressed as:

[0118] i t =σ(W xi · χ t +W hi · χ t-1 +W ci ·C t-1 +b i )

[0119] f t =σ(W xf · χ t +W hf ·H t-1 +W cf ·C t-1 +b f )

[0120] C t =f t ·C t-1 +i t tanh(W xc · χ t +W hc ·H t-1 +b c )

[0121] o t =σ(W xo · χ t +W ho ·H t-1 +W co ·C t-1 +b o )

[0122] H t =o t tanh(C t )

[0123] Among them, i t represents the input gate, f t represents the forget gate, o t represents the output gate; c t Indicates the current state, c t-1 Indicates the state at the previous moment; h t represents the final output; W represents the weight coefficient, b represents the corresponding bias coefficient; σ is the sigmoid function, * represents the convolution operation, and ○ represents the Hadamard operation, that is, the corresponding elements of the matrix are multiplied.

[0124] The training set constructed in step 3 is input into the above-mentioned spatiotemporal convolutional model, the model is trained, and the output is the probability of predicting five rain attenuation levels.

[0125] Step 5: Test the neural network model using the test set

[0126] Input the test set constructed in step 3 into the neural network model trained in step 4, evaluate the forecast accuracy, and output the forecast accuracy y%.

[0127] Step 6: Determine whether the accuracy meets the requirements

[0128] Compare the test accuracy y% obtained in step 5 with the preset accuracy requirement x%. If y is greater than or equal to x, the accuracy meets the requirement and proceed to step 7. If y is less than x, it means that the accuracy does not meet the requirement and it is necessary to jump to step 4 and retrain the model.

[0129] Step 7: Inject the trained network model into the rain attenuation prediction module

[0130] The trained network model that meets the test accuracy requirements is injected into the rain attenuation prediction module of the control center to complete the network training process.

[0131] Step 8: Complete power adjustment operation according to rain attenuation forecast

[0132] According to the level changes of the output results of the rain attenuation prediction module, the ground station or satellite power amplifier gear is adjusted in time to compensate for the impact of rain attenuation level changes on the signal and ensure stable transmission of the satellite-to-ground link.

[0133] The image acquisition function system consists of meteorological satellites or remote sensing satellites and image acquisition stations, and is mainly responsible for completing the function of raw data acquisition. The raw data mainly includes two parts: two-dimensional images with time stamps and one-dimensional meteorological data such as rainfall, temperature, humidity, wind direction, wind speed, and air pressure of the target area where the ground station is located obtained through the on-orbit meteorological satellite or remote sensing satellite, and the attenuation of the satellite-to-ground link signal power with time stamps collected by the ground station.

[0134] The satellite-to-ground link function system consists of the target satellite and the ground station. The transmission attenuation of the satellite-to-ground link changes in real time with the weather conditions in the area where the ground station is located. When the attenuation causes the receiving level value to be lower than the threshold, the link is interrupted, which greatly affects the service provision capability.

[0135] The control center functional system consists of a rain attenuation prediction module and a power regulation module. Among them, the rain attenuation prediction module is mainly implemented based on a neural network model trained by a time-space combined deep learning framework. After the original data collection is completed and preprocessing is done, a correspondence is established between the two-dimensional image and the rainfall attenuation level according to the timestamp, the neural network model is trained through the training set, and the accuracy of the neural network model's data interpretation is evaluated through the test set. It is continuously iterated until a neural network model that meets the interpretation accuracy is found. The power regulation module mainly corresponds the five levels of rainfall attenuation A, B, C, D, and E to the five power amplifier adjustment gears of PA, PB, PC, PD, and PE respectively. According to the rainfall forecast level probability given by the rain attenuation prediction module and the current gear of the satellite or ground station power amplifier, a decision is made in time whether the transmission section power amplifier gear needs to be adjusted to ensure the continuous transmission of satellite-to-ground link data. The working process of the power regulation module is as follows: Figure 6 shown.

[0136] like Figure 7 As shown, the process of a satellite's ground-to-space collaborative anti-rain attenuation is taken as an example.

[0137] 1. Collect meteorological image data for a complete natural year and meteorological station data in the ground station area, combine the communication rain attenuation and link quality data recorded within a year as training data, and complete data sampling to form 10,000 groups of time series data combinations. The time series interval is 1 second and the time series length is 18 seconds. The time series data combination is used as training data, and the communication quality data at the last frame in the time series is used as the training label to constitute the model training and test data.

[0138] 2. 8,000 of the 10,000 sets of data are randomly selected as training data and 2,000 sets are selected as test data to train the temporal convolutional network model.

[0139] 3. Use the trained model to take the real-time collected meteorological data as the model input. The model outputs the probability of rain attenuation level. The link quality is judged according to the rain attenuation probability. The satellite-ground coordinated rain attenuation prediction and the satellite control center device are used to complete the automatic adjustment of the satellite power amplifier gear.

[0140] 4. Evaluate the quality of the satellite-to-ground link.

[0141] 5. After using the proposed method to compensate for link attenuation, the effect is as follows Figure 8 As shown;

[0142] It can be seen that in a rainfall event, the natural attenuation of the satellite-to-ground link fluctuates between 4dB and 16dB. Using the deep learning method proposed in this patent, the impact of rainfall on the satellite-to-ground link can be accurately predicted, and the actual attenuation of the satellite-to-ground link can be controlled within 2dB by increasing or decreasing the power amplifier gear, thereby effectively reducing the impact of rain attenuation on the signal strength of the satellite-to-ground link.

[0143] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

[0144] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.

Claims

1. A ground-ground collaborative anti-rain attenuation method based on a deep learning feature framework, characterized in that include: Collect original image data and meteorological data through remote sensing satellites, and collect satellite-to-ground link signal data through ground stations; The original image data, meteorological data, and satellite-to-ground link signal data are all preprocessed; The rainfall attenuation level obtained after preprocessing of the satellite-to-ground link signal is used to establish a corresponding relationship with the preprocessed image data and meteorological data according to the timestamp, and a training set and a test set are constructed; Establish the spatiotemporal convolution model structure and train it through the training set; The trained spatiotemporal convolution model is tested by inputting the test set, and whether the output forecast accuracy meets the preset accuracy requirements is determined based on the output rainfall attenuation level; The trained spatiotemporal convolutional model that meets the preset accuracy requirements is used for rain attenuation forecasting; The meteorological data preprocessing method is: The one-dimensional data of rainfall, temperature, humidity, wind direction, wind speed, and air pressure are concatenated row by row to construct a matrix, with each row of the matrix representing the value of a single parameter at different times, and each column representing the values ​​of all parameters at the same time; Interpolation is performed on each row of matrix data, and n-1 data points are inserted between two adjacent sampling moments; n is greater than or equal to 2; Divide the matrix into a specified number of sub-matrices by column, where each sub-matrix has the same number of rows as the original matrix and n number of columns; Any sub-matrices are concatenated so that the size of the concatenated matrix is ​​the same as the size of the preprocessed two-dimensional image.

2. The method for ground-ground collaborative anti-rainfall based on deep learning feature framework according to claim 1, characterized in that: The original image data is a two-dimensional image time series of the ground station area collected and photographed by meteorological satellites or remote sensing satellites and with time stamps; the meteorological data is a one-dimensional meteorological data time series of the ground station area collected by meteorological satellites or remote sensing satellites, including rainfall, temperature, humidity, wind direction, wind speed, and air pressure data; the satellite-to-ground link signal data is the satellite-to-ground link signal attenuation collected by the ground station and with time stamps.

3. The method for ground-ground collaborative anti-rainfall based on deep learning feature framework according to claim 2 is characterized in that: The preprocessing of the original image data is specifically as follows: Set the ground station as the center point of the image and determine the actual maximum rainfall height H at the location of the ground station; Calculate the inclined path length L of radio wave propagation and the path length L of the actual highest rainfall height at the ground station and the horizontal projection S ; Calculate the minimum frame width A of the image collected by the meteorological satellite or remote sensing satellite based on the obtained data min ; The minimum frame width of the image is taken as the minimum frame side length after image cropping, and image cropping is performed frame by frame on the time-series two-dimensional image information of the original image data.

4. The method for ground-ground collaborative anti-rainfall based on deep learning feature framework according to claim 3 is characterized by: The minimum image width A min The calculation method is: A min =2×L S In the formula, is the minimum operating elevation angle of the earth station antenna, in degrees, and R is the effective radius of the earth in the presence of refractive index.

5. The method for ground-ground collaborative anti-rainfall based on deep learning feature framework according to claim 1, characterized in that: The signal attenuation of the satellite-to-ground link is graded and the maximum rainfall attenuation allowed for normal operation of the satellite-to-ground link is determined as R loss , the rainfall attenuation level is divided into five levels according to the attenuation amount, among which [80%R loss ,100%R loss ] The range is A, [60% R loss ,80%R loss ) interval is B level, [40%R loss ,60%R loss ) interval is C level, [20%R loss ,40%R loss ) interval is D level, [0%R loss ,20%R loss ) interval is E level.

6. The method for ground-ground collaborative anti-rain attenuation based on deep learning feature framework according to claim 5 is characterized in that: The training set is the image data and meteorological data after the corresponding relationship is established. The spatiotemporal convolution model structure is realized by adding a convolution operation to the temporal neural network model. The input gate, forget gate and output gate structure are used as gate control units in the spatiotemporal convolution model structure. The information input in each gate structure adopts a convolution operation, the cell state is updated by a dot product operation, and the weight is set in the convolution kernel. The spatiotemporal convolution model structure is specifically as follows: I t =σ(W xi ·h t +W hi ·h t-1 +W ci ·C t-1 +b i ) f t =σ(W xf ·χ t +W hf ·H t-1 +W cf ·C t-1 +b f ) C t =f t ·C t-1 +i t ·tanh(W xc ·χ t +W hc ·H t-1 +b c ) the t =σ(W xo ·h t +W ho ·H t-1 +W co ·C t-1 +b o ) H t =o t ·tanh(C t ) In the formula, i t represents the input gate, f t represents the forget gate, o t represents the output gate; c t Indicates the current state, c t-1 Indicates the state at the previous moment; h t represents the final output; W represents the weight coefficient, b represents the corresponding bias coefficient; σ represents the sigmoid function; The training set is input into the spatiotemporal convolution model to train the model, and the probability of forecasting five rainfall attenuation levels is output.

7. The method for ground-ground collaborative anti-rainfall based on deep learning feature framework according to claim 6, characterized in that: According to the rainfall attenuation level corresponding to the input data in the test set, the rainfall attenuation level of the output of the trained spatiotemporal convolution model is judged, and the output forecast accuracy y% is determined, and compared with the preset accuracy requirement x%. If y is greater than or equal to x, the accuracy of the trained spatiotemporal convolution model meets the preset accuracy requirement; if y is less than x, the accuracy does not meet the requirements, and the model is retrained using the training set.

8. A ground-to-earth collaborative anti-rain fade system based on a deep learning feature framework for implementing the ground-to-earth collaborative anti-rain fade method of claim 7, characterized in that: It includes image acquisition module, satellite-to-ground link, and collaborative control module, among which: The image acquisition module collects raw image data and meteorological data through the satellite-to-ground link, collects satellite-to-ground link signal data at the same time, pre-processes all data, and sends all pre-processed data to the collaborative control module; Satellite-to-ground link, including remote sensing satellites and ground stations. Original image data and meteorological data are collected through remote sensing satellites, and satellite-to-ground link signal data is collected through ground stations. The collaborative control module uses the rainfall attenuation level obtained after preprocessing the satellite-to-ground link signal according to the timestamp, establishes a corresponding relationship with the preprocessed image data and meteorological data according to the timestamp, constructs a training set and a test set, and constructs a spatiotemporal convolution model structure after adding a convolution operation through a temporal neural network model. After completing the training of the spatiotemporal convolution model with the training set, the trained spatiotemporal convolution model is tested through the test set input. According to the output rainfall attenuation level, it is determined whether the output forecast accuracy meets the preset accuracy requirement. If not, the training iteration is repeated until the preset accuracy requirement is met, and a trained spatiotemporal convolution model that meets the preset accuracy requirement is obtained; the trained spatiotemporal convolution model is applied to rainfall attenuation forecasting, and according to the rainfall forecast level and the power level of the remote sensing satellite or ground station obtained in real time, the power level of the remote sensing satellite or ground station is adjusted through the satellite-to-ground link.

9. The ground-ground collaborative anti-rain fade system based on deep learning feature framework according to claim 8, characterized in that: The collaborative control module includes a rain attenuation prediction module and a power regulation module; Rain attenuation forecast module, used to construct training sets and test sets, and complete the training of spatiotemporal convolutional models to output real-time rainfall forecast levels; The power regulation module calculates and predicts the required power according to the real-time rainfall forecast level and the power level of the remote sensing satellite or ground station, and adjusts the power level of the remote sensing satellite or ground station through the satellite-to-ground link.

Images