A method for predicting electromagnetic wave rain attenuation based on deep learning

Abstract

The present invention relates to a method for predicting electromagnetic wave rain attenuation based on deep learning, and belongs to the technical field of electromagnetic wave rain attenuation prediction. The present invention collects data on unit distance rain attenuation, rainfall rate, equipment frequency band and polarization mode of an experimental site, and uses a deep learning model to fit the relationship between unit distance rain attenuation and rainfall rate, equipment frequency and polarization mode of a specific site. The input data includes rainfall rate, frequency and polarization mode, and the output is unit distance rain attenuation, k and α parameters based on the classic ITU-R rain attenuation model. This method can realize accurate modeling and prediction of electromagnetic wave rain attenuation characteristics, and can provide more accurate scene adaptability and optimization effects compared to traditional models. In addition, the model training process of the present invention adopts an early stopping mechanism to ensure that the optimal model parameters are saved after reaching the predetermined accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic wave rain attenuation prediction, and in particular to an electromagnetic wave rain attenuation prediction method based on deep learning. Background Art

[0002] When electromagnetic waves propagate, they will produce rain attenuation when encountering rainfall, especially in the microwave and millimeter wave frequency bands. This effect significantly affects the performance of communication systems. Existing rain attenuation models, such as the nonlinear rain attenuation model recommended by the International Telecommunication Union (ITU-R), are based on statistical empirical formulas and are widely used in engineering practice.

[0003] In the article "Research on Rain Attenuation of Sanya Ka-band Satellite Remote Sensing Data Signal" published in the journal "Radio Communication Technology" by Xin Jin and others from the Xi'an Satellite Tracking and Control Center in China, the measured data were analyzed in detail, and the rain attenuation of satellite remote sensing data signals was modeled based on domain knowledge and probability theory.

[0004] In the article "Modeling and Analysis of Scattering Characteristics of Rainfall Environment in Autonomous Driving Radar Frequency Band" published by Zou Mengfan and others from the Beijing University of Aeronautics and Astronautics in the journal "Journal of System Simulation", the premise assumptions of rainfall characteristics were set, and a scattering characteristic model of rainfall environment was constructed through theoretical analysis and mathematical modeling.

[0005] The national invention patent "A method and device for ensuring the reliability of satellite broadband communication links" applied for by Liu Zhixin and others from the Systems Engineering Institute of the Chinese People's Liberation Army Academy of Military Sciences, with authorization announcement number CN117294341B, directly adopts the International Telecommunication Union's rain attenuation model to ensure satellite broadband communication links.

[0006] However, traditional rain attenuation models based on probability theory and domain knowledge have limited prediction accuracy in specific scenarios, especially when rain attenuation is affected by site environment and equipment characteristics, the applicability of standard models is poor. Summary of the invention

[0007] The purpose of the present invention is to provide an electromagnetic wave rain attenuation prediction method based on deep learning to solve the problem that the rain attenuation model in the prior art has poor applicability in specific scenarios.

[0008] The technical solution adopted by the present invention is:

[0009] A method for predicting electromagnetic wave rain attenuation based on deep learning, comprising the following steps:

[0010] Step 1: Data rainfall data, environmental data, equipment data and measurement data of rain attenuation per unit distance are used to construct a data set;

[0011] Step 2: Construct a neural network model. The input data include rainfall rate, frequency and polarization mode. The output data are rain attenuation per unit distance and k and α parameters based on the classic ITU-R rain attenuation model.

[0012] Step 3: Construct loss function;

[0013] Step 4: Train and test the neural network model to obtain the trained neural network parameters;

[0014] Step 5: Load the trained neural network parameters and use the neural network model to predict the unit path length rain attenuation and the k parameter and α parameter based on the ITU-R rain attenuation model.

[0015] Furthermore, the specific method of step 1 is:

[0016] Step 1-1: Collect the rainfall rate R of the experimental site through the rain gauge, collect the environmental data k parameter and α parameter based on the ITU-R rain attenuation model through the sensor, set the frequency f and polarization mode v of the measuring equipment, and use the measuring equipment to obtain the measurement data of rain attenuation per unit distance ;

[0017] After T times of data collection, the rainfall rate dataset is constructed , electromagnetic wave rain attenuation dataset , k parameter data set , α parameter data set , frequency dataset and polarization data sets ;

[0018] Step 1-2: The data sets collected T times are proportional Randomly split into training data set and test data set, where the rainfall rate training data set is , the rainfall rate test data set is ; Electromagnetic wave rain attenuation training data set , electromagnetic wave rain attenuation test data set ; k parameter training data set , k parameter test data set ;α parameter training data set ,α parameter test data set ; Frequency training data set , frequency test data set ; Polarization training data set , polarization test data set .

[0019] Furthermore, the neural network model is a deep learning model constructed in series using a deep neural network fully connected layer, with a total of 3 layers. The first layer has an input dimension of 3, corresponding to rainfall rate, frequency and polarization mode, and an output dimension of 64; the second layer has an input dimension of 64 and an output dimension of 32; the third layer has an input dimension of 32 and an output dimension of 3, using a ReLU activation function, corresponding to unit distance rain attenuation, k parameter and α parameter; the relationship between the input and output of the neural network model is expressed as:

[0020]

[0021] Among them, h represents the operation of the neural network model on the input data, , and is the unit distance rain attenuation value obtained by the neural network model based on the observed value As well as the k parameter and the α parameter, , , .

[0022] Furthermore, the loss function is a mean square error function, which is expressed as:

[0023]

[0024] in, Indicates rounding up. is the number of samples in the training data set, , and .

[0025] Furthermore, the specific method of step 4 is:

[0026] Step 4-1: Initialize the neural network model and ensure that the gradient of the neural network parameters is calculated during back propagation;

[0027] Step 4-2: Use the Adam optimizer to pass in the parameters of the neural network and set the learning rate lr;

[0028] Step 4-3: Randomly take a batch of data from the training data set, clear the gradient value stored in the optimizer, calculate the loss function based on the neural network model, calculate the gradient of the loss function with respect to the neural network parameters, and then update the values ​​of the neural network parameters;

[0029] Step 4-4: Repeat step 4-3 until all the data in the training data set are taken, and then execute step 4-5;

[0030] Step 4-5: Set the neural network model to evaluation mode based on the test data set , and , the estimated electromagnetic wave rain attenuation is calculated , Parameters and Parameters, calculate estimated values , and With the test data set , and Correlation coefficient :

[0031]

[0032] Step 4-6: When the correlation coefficient When the threshold value is reached, the threshold value>0.9, the neural network parameters are saved; otherwise, return to step 4-3.

[0033] Furthermore, the specific method of step 5 is:

[0034] Step 5-1: Load the neural network model and import the trained neural network parameters;

[0035] Step 5-2: Input the rainfall rate, equipment usage frequency and polarization mode in the actual scene into the neural network model to obtain the electromagnetic wave rain attenuation per unit path length, k parameter and α parameter.

[0036] The beneficial effects of the present invention are:

[0037] 1. The present invention uses deep learning technology and scenario data to predict electromagnetic wave rain attenuation, which significantly improves the accuracy of rain attenuation prediction per unit distance, especially in complex scenarios, with stronger adaptability and robustness.

[0038] 2. The present invention adopts early stopping mechanism and automatic parameter tuning to ensure that the optimal model is obtained during the training process, reduce the risk of overfitting, improve training efficiency, and shorten the development cycle.

[0039] 3. The model input of the present invention is rainfall rate, device frequency and polarization mode, which is suitable for the electromagnetic wave rain attenuation prediction needs of different scenes, frequency bands and devices, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flowchart of an electromagnetic wave rain attenuation prediction method based on deep learning in an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of rainfall rate data.

[0042] Figure 3 This is a diagram of frequency data.

[0043] Figure 4 This is a schematic diagram of polarization data.

[0044] Figure 5 This is a schematic diagram of electromagnetic wave rain attenuation data per unit distance.

[0045] Figure 6 This is a schematic diagram of k parameter data.

[0046] Figure 7 This is a schematic diagram of α parameter data.

[0047] Figure 8 This is a diagram showing how the loss value changes with the number of training times.

[0048] Fig. 9 This is a schematic diagram of the correlation coefficient changing with the number of training times. DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0050] A deep learning-based electromagnetic wave rain attenuation prediction method is proposed. The method first collects environmental and equipment parameters such as rainfall rate, equipment frequency and polarization mode of the experimental site, as well as the measurement data of rain attenuation per unit distance. Then, the collected data is trained using a deep learning model to construct a mapping between rainfall rate, frequency and polarization mode and rain attenuation per unit distance, k and α parameters. Finally, the trained model is applied to actual scenarios to achieve accurate modeling and prediction of rain attenuation characteristics.

[0051] like Figure 1 As shown, the method specifically comprises the following steps:

[0052] Step 1: Data collection and dataset construction;

[0053] Step 1-1: Data collection:

[0054] These data are obtained by installing rain gauges, sensors and measuring equipment, collecting the rainfall rate R of the experimental site and the environmental data of k and α parameters based on the ITU-R rain attenuation model, using frequency f and polarization mode v equipment parameters, and measurement data of rain attenuation per unit distance ;

[0055] Repeat the above process T=1000 times to build a rainfall rate dataset , electromagnetic wave rain attenuation dataset , k parameter data set ,α parameter data set , frequency dataset and polarization data sets ,in , , , , and The data obtained are as follows Figures 2 to 7 shown.

[0056] Step 1-2: Dataset construction:

[0057] According to the proportion = 0.8, and randomly divide the data set obtained from T tests into a training data set and a test data set, the numbers of which are and ,in represents rounding up; the rainfall rate training data set is , the rainfall rate test data set is ; Electromagnetic wave rain attenuation training data set , electromagnetic wave rain attenuation test data set ; k parameter training data set , k parameter test data set ;α parameter training data set ,α parameter test data set ; Frequency training data set , frequency test data set ; Polarization training data set , polarization test data set .

[0058] Step 2: Design a neural network model;

[0059] The deep learning model h is built in series using a deep neural network with fully connected layers. The neural network has 3 layers. The first layer has an input dimension of 3, corresponding to rainfall rate, frequency, and polarization, and an output dimension of 64; the second layer has an input dimension of 64 and an output dimension of 32; the third layer has an input dimension of 32 and an output dimension of 3, using the ReLU activation function, corresponding to unit distance rain attenuation, k, and α. The relationship between input and output can be described as:

[0060] ,

[0061] in, , and The unit distance rain attenuation value obtained by the deep learning model based on the observed value with k and α environmental parameters, , and .

[0062] Step 3: Design loss function:

[0063] The mean square error is used as the loss function, and its formula is as follows:

[0064] ,

[0065] in, Indicates rounding up. is the number of samples in the training data set, , and .

[0066] Step 4: Neural network parameter optimization and verification:

[0067] Step 4-1: Initialize the deep neural network model and ensure that the gradients of the neural network parameters are calculated during backpropagation.

[0068] Step 4-2: Define the optimizer. Use the Adam optimizer, pass in the parameters of the neural network, and set the learning rate lr=0.001.

[0069] Step 4-3: Randomly take out batch=32 data from the training data set, clear the gradient values ​​stored in the optimizer, calculate the loss function based on the deep neural network model, calculate the gradient of the loss function with respect to the neural network parameters, and then update the values ​​of the neural network parameters.

[0070] Step 4-4: Repeat step 4-3 until all the data in the training data set are obtained, and then proceed to step 4-5.

[0071] Step 4-5: Validate the model. Set the model to evaluation mode, using a deep neural network, based on a test dataset. , and , the estimated electromagnetic wave rain attenuation is calculated , and Parameters, calculate estimated values , and With the test data set , and Correlation coefficient :

[0072] .

[0073] Step 4-6: Iterative optimization. When the correlation coefficient When the specified threshold value of 0.9999 is reached, the deep neural network parameters are saved; if the specified threshold is not reached, return to step 4-3. During the optimization and verification process, the loss value changes with the number of training times and the correlation coefficient changes with the number of training times respectively as follows Figure 8 and Fig. 9 shown.

[0074] Step 5: Predict the rain attenuation per unit path length with the k and α parameters based on the ITU-R rain attenuation model:

[0075] Step 5-1: Load the deep neural network model and import the neural network parameters;

[0076] Step 5-2: The observed rainfall rate, equipment usage frequency and polarization mode, i.e., 20.0 mm / h, 5.0 GHz and 1 (indicating vertical polarization mode), are input into the deep neural network model to obtain the electromagnetic wave rain attenuation per unit path length, k and α environmental parameters, with values ​​of 0.86762762 dB / km, 0.08914762 and 0.75572407.

[0077] In summary, the present invention collects data on rain attenuation per unit distance, rainfall rate, device frequency band and polarization mode of the experimental site, and uses a deep learning model to fit the relationship between rain attenuation per unit distance and rainfall rate, device frequency and polarization mode of a specific site. The input data includes rainfall rate, frequency and polarization mode, and the output is rain attenuation per unit distance, k and α parameters based on the classic ITU-R rain attenuation model. This method can achieve accurate modeling and prediction of electromagnetic wave rain attenuation characteristics, and can provide more accurate scene adaptability and optimization effects compared to traditional models. In addition, the model training process of the present invention adopts an early stopping mechanism to ensure that the optimal model parameters are saved after reaching the predetermined accuracy. The present invention can predict the attenuation value per unit path length, and can automatically obtain k and b values.

Claims

1. A method for predicting electromagnetic wave rain attenuation based on deep learning, characterized in that: The following steps are involved: Step 1: Data rainfall data, environmental data, equipment data and unit distance rain attenuation measurement data are used to construct a data set. The specific method is as follows: Step 1-1: Collect the rainfall rate R of the experimental site through the rain gauge, collect the environmental data k parameter and α parameter based on the ITU-R rain attenuation model through the sensor, set the frequency f and polarization mode v of the measuring equipment, and use the measuring equipment to obtain the measurement data of rain attenuation per unit distance ; After T times of data collection, the rainfall rate dataset is constructed , electromagnetic wave rain attenuation dataset , k parameter data set , α parameter data set , frequency dataset and polarization data sets ; Step 1-2: The data sets collected T times are proportional Randomly split into training data set and test data set, where the rainfall rate training data set is , the rainfall rate test data set is ; Electromagnetic wave rain attenuation training data set , electromagnetic wave rain attenuation test data set ; k parameter training data set , k parameter test data set ;α parameter training data set ,α parameter test data set ; Frequency training data set , frequency test data set ; Polarization training data set , polarization test data set ; Step 2: Construct a neural network model. The input data include rainfall rate, frequency and polarization mode. The output data are rain attenuation per unit distance and k and α parameters based on the classic ITU-R rain attenuation model. Step 3: Construct loss function; Step 4: Train and test the neural network model to obtain the trained neural network parameters; Step 5: Load the trained neural network parameters and use the neural network model to predict the unit path length rain attenuation and the k parameter and α parameter based on the ITU-R rain attenuation model.

2. The electromagnetic wave rain attenuation prediction method based on deep learning according to claim 1 is characterized in that: The neural network model is a deep learning model constructed by serially connecting the deep neural network fully connected layers. It has 3 layers. The first layer has an input dimension of 3, corresponding to the rainfall rate, frequency and polarization mode, and an output dimension of 64; the second layer has an input dimension of 64 and an output dimension of 32; the third layer has an input dimension of 32 and an output dimension of 3, and uses the ReLU activation function, corresponding to the unit distance rain attenuation, k parameter and α parameter; the relationship between the input and output of the neural network model is expressed as: Among them, h represents the operation of the neural network model on the input data, , and is the unit distance rain attenuation value obtained by the neural network model based on the observed value As well as the k parameter and the α parameter, , , .

3. The electromagnetic wave rain attenuation prediction method based on deep learning according to claim 2 is characterized in that: The loss function is the mean square error function, expressed as: in, Indicates rounding up. is the number of samples in the training data set, , and .

4. The electromagnetic wave rain attenuation prediction method based on deep learning according to claim 3 is characterized in that: The specific method of step 4 is: Step 4-1: Initialize the neural network model and ensure that the gradient of the neural network parameters is calculated during back propagation; Step 4-2: Use the Adam optimizer to pass in the parameters of the neural network and set the learning rate lr; Step 4-3: Randomly take a batch of data from the training data set, clear the gradient value stored in the optimizer, calculate the loss function based on the neural network model, calculate the gradient of the loss function with respect to the neural network parameters, and then update the values ​​of the neural network parameters; Step 4-4: Repeat step 4-3 until all the data in the training data set are taken, and then execute step 4-5; Step 4-5: Set the neural network model to evaluation mode based on the test data set , and , the estimated electromagnetic wave rain attenuation is calculated , Parameters and Parameters, calculate estimated values , and With the test data set , and Correlation coefficient : ； Step 4-6: When the correlation coefficient When the threshold is reached, the neural network parameters are saved; otherwise, return to step 4-3.

5. The electromagnetic wave rain attenuation prediction method based on deep learning according to claim 4 is characterized in that: The specific method of step 5 is: Step 5-1: Load the neural network model and import the trained neural network parameters; Step 5-2: Input the rainfall rate, equipment usage frequency and polarization mode in the actual scene into the neural network model to obtain the electromagnetic wave rain attenuation per unit path length, k parameter and α parameter.

Images