MST radar power spectrum parameter calculation method and system based on deep learning, and electronic equipment

Abstract

The invention relates to an MST radar power spectrum parameter calculation method and system based on deep learning, and electronic equipment, and belongs to the field of MST radar information processing. The method comprises the following steps: acquiring an original MST radar power spectrum; calculating an initial MST radar power spectrum parameter; controlling the quality of an initial MST radar power spectrum parameter by using a beam symmetry spectrum picking and space-time continuity comparison method, and correcting an inaccurate power spectrum parameter; establishing an MST radar power spectrum parameter calculation model based on deep learning, and training the MST radar power spectrum parameter calculation model to obtain an optimal MST radar power spectrum parameter calculation model; and inputting a to-be-measured MST radar power spectrum into the trained optimal MST radar power spectrum parameter calculation model, and obtaining MST radar power spectrum parameters after calculation is completed. According to the method, the power spectrum parameters can be accurately extracted, and a data basis is provided for high-precision wind field inversion.

Description

Technical Field

[0001] The present invention relates to a method, a system and an electronic device for calculating MST radar power spectrum parameters based on deep learning, and belongs to the technical field of MST radar information processing. Background Art

[0002] The atmospheric horizontal wind field is an important parameter for studying atmospheric dynamics processes and understanding the coupling mechanism between the middle and upper atmosphere and the ionosphere and the lower atmosphere. A large number of atmospheric remote sensing instruments, such as wind profilers, medium frequency radars, meteor radars, lidars, etc., have been successively applied to the detection of middle and upper atmosphere parameters. At present, the maximum altitude range for detecting the atmosphere using balloons and airplanes is about 30 kilometers, and the on-orbit altitude of satellites is generally above 250 kilometers and they orbit the earth rapidly. Neither can effectively detect the middle and upper atmosphere of a certain area for a long time. Sounding rockets are the only means that can conduct in-situ detection of the middle and upper atmosphere, but sounding rockets are costly and unsustainable.

[0003] Ground-based remote sensing detection technology is currently the most important means for detecting the middle and upper atmosphere at home and abroad, and the MST (Mesosphere Stratosphere and Troposphere) radar is the observation device with the strongest detection ability for the neutral three-dimensional middle and upper atmosphere wind field. Compared with other high-altitude wind measurement systems, the MST radar has the characteristics of high time and space resolution, long uninterrupted continuous observation time, large detection altitude range, and strong timeliness. The MST radar applies phased array and digital beam synthesis technology, and successively emits five beams in the east, west, south, north, and vertical directions. When the electromagnetic wave emitted by the radar encounters an irregular body of the atmospheric refractive index to generate turbulent scattering, the scattered air mass moves with the wind, and the atmospheric wind field information is synthesized by receiving the scattered echoes of each beam.

[0004] After the MST radar antenna receives the signal, it is amplified by the transceiver module and then transmitted to the digital receiver. Beam forming is performed by DBF (digital beam forming), and the power spectral density is obtained after being processed by the signal processor. The data processing is based on the power spectral density, performs non-target signal suppression and target signal recognition, and then calculates the power spectrum parameters: signal-to-noise ratio (SNR), spectral width (w), and radial average velocity The target signals of the MST radar satisfy the characteristics of Gaussian distribution, and in the power spectrum diagram, they are the peak regions with the maximum power, the maximum peak value, and the maximum spectral width. The traditional method for processing MST radar data is to calculate the power spectrum parameters by the spectral moment method, which will be affected by precipitation, ground clutter (such as signals caused by buildings, mountains, trees, etc. on the ground), and intermittent clutter (such as signals caused by the flight of airplanes, birds, etc.), resulting in inaccurate calculation of the power spectrum parameters and unable to meet the needs of studying the fine wind field changes. Therefore, accurate data processing is beneficial to obtaining higher-quality and more accurate atmospheric parameters.

[0005] In recent years, the popularity of deep learning has been gradually increasing, and it is widely used in various fields. Deep learning enables computer systems to simulate the learning methods and thinking patterns of humans and improves learning efficiency and accuracy through a large amount of data training, and has important applications in fields such as image processing, big data analysis, and speech recognition. At present, the vast majority of research results of deep learning (DeepLearning, DL) in the fields of radar target recognition and radar echo extrapolation focus on using deep learning technology to process the obtained radar image data, and do not use the processing of radar power spectra. The convolutional neural network has extremely strong feature extraction and processing capabilities, and the GRU neural network performs outstandingly in sequence problems, which is exactly consistent with the characteristics of radar power spectra. Therefore, using the method of deep learning can solve the problems of inaccurate and inefficient calculation of radar power spectrum parameters. Summary of the Invention

[0006] The technical problem solved by the present invention is: The present invention provides a method, system, and electronic device for calculating MST radar power spectrum parameters based on deep learning, which are used to solve the problems of inaccurate and inefficient calculation of radar power spectrum parameters in traditional technologies; the present invention calculates more accurate MST radar power spectrum parameters in the case of precipitation, ground clutter, and intermittent clutter.

[0007] The technical solution of the present invention is: A method for calculating MST radar power spectrum parameters based on deep learning, the method includes:

[0008] Step1: Obtain the original MST radar power spectrum;

[0009] Step2: First perform ground clutter suppression, three-point median filtering, and data normalization on the original MST radar power spectrum, and then calculate the initial MST radar power spectrum parameters;

[0010] Step3: Use the beam symmetry spectrum selection and spatio-temporal continuity comparison method to control the quality of the initial MST radar power spectrum parameters and correct inaccurate power spectrum parameters;

[0011] Step 4. Establish a calculation model for MST radar power spectrum parameters based on deep learning, and input the original MST radar power spectrum obtained in Step 1 into the model;

[0012] Step 5. Set the hyperparameters, total number of training epochs, adjustable learning rate, and loss function of the MST radar power spectrum parameter calculation model, so that the mean square loss between the power spectrum parameters calculated by the MST radar power spectrum parameter calculation model and the power spectrum parameters obtained after being processed in Step 3 reaches the minimum. Through repeated adjustment and comparative experiments, obtain the optimal parameters to get the optimal MST radar power spectrum parameter calculation model;

[0013] Step 6. Input the MST radar power spectrum to be measured into the trained optimal MST radar power spectrum parameter calculation model, and after the calculation is completed, obtain the MST radar power spectrum parameters of each beam in the low mode.

[0014] Furthermore, in the said Step 2, the initial MST radar power spectrum parameters include signal-to-noise ratio SNR, radial average velocity spectral width w.

[0015] Furthermore, in the said Step 2, the calculation of the initial MST radar power spectrum parameters includes: calculating the signal-to-noise ratio SNR of the power spectrum using the segmented averaging method, calculating the radial average velocity using the spectral moment method calculating the spectral width w using the Gaussian spectrum model; specifically including:

[0016] (1). Calculating the signal-to-noise ratio SNR of the power spectrum using the segmented averaging method includes:

[0017] Based on the noise obeying the χ 2 distribution with degrees of freedom 2N / k, where N is the number of points of FFT, and χ 2 is the chi-square distribution in mathematical statistics. Divide the power spectrum signal into k segments on average, calculate the average value of the power spectrum density of each segment respectively, and then select the smallest average value P N from the average values of each segment as the noise level value of the entire spectrum, and then calculate the signal-to-noise ratio SNR of the power spectrum;

[0018] (2). Calculating the radial average velocity using the spectral moment method includes:

[0019] First, for the power spectrum of each beam and each range bin, calculate the zero-order moment M 0 and the first-order moment M 1 of the signal power spectrum:

[0020]

[0021] N is the number of points of FFT, M 0is the zero - order moment, M 1 is the first - order moment, S i and f i are respectively the power value and the frequency value of the i - th point, and Δf is the frequency resolution;

[0022] Calculate the parameters of the atmospheric wind field according to the moment method. The parameters of the atmospheric wind field include the signal power estimation P S and the radial mean velocity The signal power estimation P S and the radial mean velocity The calculation formulas are expressed as:

[0023] P S = M 0 ;

[0024]

[0025] where Δv is the Doppler velocity resolution, λ is the wavelength of the electromagnetic wave emitted by the radar;

[0026] (3) Use the Gaussian spectrum model to calculate the spectral width w; where the echo power S′ fitted by the Gaussian spectrum is expressed as:

[0027]

[0028] where P r is the echo intensity, is the radial mean velocity obtained by the least - squares fitting method, w is the spectral width, v r represents the radial velocity of a single spectral point, and S′ represents the echo power fitted by the Gaussian spectrum; the fitting process is to fit the Gaussian spectrum model to the power spectrum line of each range gate of the MST radar, and determine the echo intensity P r and the radial mean velocity spectral width w; finding a set of P r and w such that the mean square error between the Gaussian spectrum model spectrum line and the power spectrum line of each range gate of the MST radar is minimized is expressed as: P r , where argmin is a function whose role is to make the variable value that minimizes the objective function, S refers to the actual echo power value, and P N is the noise level value.

[0029] Furthermore, the said Step3 includes:

[0030] Using the symmetry test of five-beam echoes and the principle of spatio-temporal continuity to determine the picking of spectral peaks and control the quality of MST radar power spectrum parameters, and manually correcting inaccurate power spectrum parameters.

[0031] Further, in the Step4, the deep learning-based MST radar power spectrum parameter calculation model includes a one-dimensional convolutional layer, a max pooling layer, a GRU layer, and a fully connected layer.

[0032] Further, in the Step4, two convolutional layers with a convolution kernel size of 3 and a stride of 1 and two max pooling layers with a pooling kernel size of 5 and a stride of 2 are used to extract the features of the MST radar power spectrum, one GRU layer is used to process the features extracted by the convolutional layer, and two fully connected layers are used to output the final power spectrum parameters.

[0033] Further, the Step4 includes:

[0034] (1) Normalize the original MST radar power spectrum for each height-range bin, and the normalized output X normalization is expressed as:

[0035]

[0036] When performing data normalization at each height, X normalization is the normalized output, X min and X max refer to the minimum and maximum echo powers, and X refers to the power spectrum at each height;

[0037] (2) Pass the normalized power spectrum through the first part of the convolutional layer, then input it into the max pooling layer, change the data dimension of the output of the max pooling layer and use it as the input of the second part of the GRU layer, and finally obtain the power spectrum parameter calculation result through the fully connected layer;

[0038] The calculation formula for each GRU layer is as follows:

[0039] Update gate z t : z t =σ(W z x t +U z h t-1 +b z ), W z and U z are the weight matrices of the update gate, b z is the bias term, σ is the Sigmoid activation function, and x t represents the input at the current time step;

[0040] Reset gate r t : rt = σ(W r x t + U r h t-1 + b r ), where W r and U r are the weight matrices of the reset gate, b r is the bias term, and σ is the Sigmoid activation function;

[0041] Candidate hidden state b h is the bias term, h t-1 represents the past hidden state, ⊙ represents element-wise multiplication, and tanh is used as the non-linear activation function. W h and U h are both weight parameters;

[0042] Calculate the final hidden state: h t represents the final hidden state.

[0043] The present invention also provides a system for calculating MST radar power spectrum parameters based on deep learning, which includes: a module for executing the method for calculating MST radar power spectrum parameters based on deep learning.

[0044] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the method for calculating MST radar power spectrum parameters based on deep learning.

[0045] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the method for calculating MST radar power spectrum parameters based on deep learning.

[0046] The beneficial effects of the present invention are as follows:

[0047] 1. The present invention solves the problem that under the influence of precipitation, ground clutter, and intermittent clutter, traditional calculation methods may lead to inaccurate calculation of MST radar power spectrum parameters, resulting in non-satisfaction of the research on fine wind fields. The present invention can calculate the power spectrum parameters of the MST radar power spectrum efficiently and accurately without relying on hardware conditions, and combines the advantages of the spectral moment method and the Gaussian spectrum for calculating power spectrum parameters; no additional hardware facilities are required on the hardware, saving costs; in terms of data processing, it can calculate the power spectrum parameters of the MST radar power spectrum quickly and accurately, and saves a lot of time, having good application prospects;

[0048] 2. In the present invention, only a small number of hyperparameters of the deep learning model need to be adjusted in the experimental stage. According to the samples of the power spectra of different modes of the MST radar input, the deep learning model captures the relevant information of the internal features of the radar power spectrum, accurately identifies the target signal and calculates the required parameters; automatically saves the best model parameters. After obtaining the best model, only by calling the model can the power spectrum parameters required for the calculation based on the MST radar power spectrum be realized: signal-to-noise ratio, spectral width, and radial mean velocity, which is convenient for obtaining more accurate atmospheric parameters and meeting the needs of studying the fine wind field changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a schematic flow chart of the method of the present invention;

[0050] Figure 2 is a schematic flow chart of the calculation of the MST radar power spectrum parameters in this method;

[0051] Figure 3 is a schematic diagram of the MST radar power spectrum for which the power spectrum parameters need to be manually corrected in this method;

[0052] Figure 3 (a) is a schematic diagram of the MST radar power spectrum with random noise, and the dashed box indicates the area with obvious random noise;

[0053] Figure 3 (b) is a schematic diagram of the radar power spectrum with precipitation, and the dashed box indicates the area with obvious precipitation;

[0054] Figure 3 (c) is a schematic diagram of the radar power spectrum with intermittent clutter or system noise, and the dashed box indicates the area with noise;

[0055] Figure 3 (d) is a schematic diagram of the radar power spectrum with ground clutter, and the dashed box indicates the area with noise;

[0056] Figure 4 is the unit structure diagram of the GRU layer used in this method;

[0057] Figure 5 is a schematic diagram of the model structure for calculating the MST radar power spectrum parameters in this method;

[0058] Figure 6 is a schematic diagram of the result of the MST radar power spectrum calculation in this method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] Example 1: As Figures 1-6 shown, a method for calculating the MST radar power spectrum parameters based on deep learning, the method comprising:

[0060] Step1. Obtain the original MST radar power spectrum from the China Meridian Project Data Center ( https: / / data.meridianproject.ac.cn / ), where the original MST radar power spectrum is the power spectrum data of the five-beam echoes in the low mode (2.55 - 10.05 km);

[0061] Step2. According to the obtained power spectrum data, first perform ground clutter suppression, three-point median filtering, and data normalization on the original MST radar power spectrum, and then calculate the initial MST radar power spectrum parameters, as Figure 2 shown;

[0062] Furthermore, in Step2, the initial MST radar power spectrum parameters include signal-to-noise ratio SNR, radial mean velocity spectral width w.

[0063] Furthermore, in Step2, the calculation of the initial MST radar power spectrum parameters includes: calculating the signal-to-noise ratio SNR of the power spectrum using the segmented average method, calculating the radial mean velocity using the spectral moment method, and calculating the spectral width w using the Gaussian spectral model; specifically including:

[0064] (1). Calculating the signal-to-noise ratio SNR of the power spectrum using the segmented average method includes:

[0065] According to the research of Petitdidier et al. (1997), based on the noise following a χ 2 distribution with degrees of freedom 2N / k, where N is the number of FFT points, and χ 2 is the chi-square distribution in mathematical statistics, divide the power spectrum signal into k segments (usually k = 8 is selected in actual use), calculate the average value of the power spectrum density of each segment respectively, and then select the smallest average value P N from the segment-by-segment average values as the noise level value of the entire spectrum, and then calculate the signal-to-noise ratio SNR of the power spectrum;

[0066] (2). Calculating the radial mean velocity using the spectral moment method includes:

[0067] First, calculate the zero-order moment M 0 and the first-order moment M 1 of the signal power spectrum for the power spectrum of each beam and each range bin:

[0068]

[0069] N is the number of FFT points, M 0 is the zero-order moment, M 1 is the first-order moment, S i and f iThey are the power value and frequency value of the i-th point respectively, and Δf is the frequency resolution;

[0070] Calculate the parameters of the atmospheric wind field according to the method of moments. The parameters of the atmospheric wind field include the signal power estimate P S , the radial mean velocity The signal power estimate P S , the radial mean velocity The calculation formula is expressed as:

[0071] P S = M 0 ;

[0072]

[0073] where Δν is the Doppler velocity resolution, λ is the wavelength of the electromagnetic wave emitted by the radar;

[0074] (3) Use the Gaussian spectrum model to calculate the spectral width w; among them, the echo power S' fitted by the Gaussian spectrum is expressed as:

[0075]

[0076] where P r is the echo intensity, is the radial mean velocity obtained by the least squares fitting method, w is the spectral width, v r represents the radial velocity of a single spectral point, and S' represents the echo power fitted by the Gaussian spectrum; the fitting process is to fit the Gaussian spectrum model to the power spectrum line of each range gate of the MST radar, and determine the echo intensity P r , the radial mean velocity spectral width w; find a set of P r , w such that the root-mean-square (RMS) of the Gaussian spectrum model spectral line and the power spectrum line of each range gate of the MST radar is minimized, which is expressed as: P r , where argmin is a function whose role is to make the variable value that minimizes the objective function, S refers to the actual echo power value, and P N is the noise level value. Step 3: Use the beam symmetry spectrum selection and spatio-temporal continuity comparison method to control the quality of the initial MST radar power spectrum parameters and correct inaccurate power spectrum parameters. As Figure 3 shown, the incorrect spectral peaks are within the red frame, and the incorrect power spectrum parameters obtained need to be corrected manually;

[0077] Furthermore, the said Step 3 includes:

[0078] Use the symmetry test of the five-beam echoes and the principle of spatio-temporal continuity to determine the picking of spectral peaks and control the quality of MST radar power spectrum parameters, and manually correct inaccurate power spectrum parameters.

[0079] Further, in the Step4, the deep learning-based MST radar power spectrum parameter calculation model includes a one-dimensional convolutional layer, a max pooling layer, a GRU layer, and a fully connected layer.

[0080] Further, in the Step4, two convolutional layers with a convolution kernel size of 3 and a stride of 1 and two max pooling layers with a pooling kernel size of 5 and a stride of 2 are used to extract the features of the MST radar power spectrum, one GRU layer is used to process the features extracted by the convolutional layer, and two fully connected layers are used to output the final power spectrum parameters.

[0081] Step4: Establish a deep learning-based MST radar power spectrum parameter calculation model, and input the original MST radar power spectrum obtained in Step1 into the model;

[0082] Further, the Step4 includes:

[0083] (1) Normalize the original MST radar power spectrum for each height-distance bin, and the normalized output X normalization is expressed as:

[0084]

[0085] When performing data normalization at each height, X normalization is the normalized output, X min and X max refer to the minimum and maximum echo powers, and X refers to the power spectrum at each height;

[0086] (2) Pass the normalized power spectrum through the first part of the convolutional layer, then input it into the max pooling layer, change the data dimension of the output of the max pooling layer and use it as the input of the second part of the GRU layer, and finally obtain the power spectrum parameter calculation result through the fully connected layer;

[0087] The calculation formula of each GRU layer is as follows:

[0088] Update gate z t : z t =σ(W z x t +U z h t-1 +b z ), W z and U z are the weight matrices of the update gate, b zis the bias term, σ is the Sigmoid activation function, and x t represents the input at the current time step;

[0089] Reset gate r t : r t = σ(W r x t + U r h t-1 + b r ), where W r and U r are the weight matrices of the reset gate, b r is the bias term, and σ is the Sigmoid activation function;

[0090] Candidate hidden state b h is the bias term, h t-1 represents the past hidden state, ⊙ represents element-wise multiplication, and tanh is used as the non-linear activation function. W h and U h are both weight parameters;

[0091] Calculate the final hidden state: h t represents the final hidden state.

[0092] Step 5, set the hyperparameters of the MST radar power spectrum parameter calculation model, the total number of training epochs, the adjustable learning rate, and the loss function, so that the mean square loss between the power spectrum parameters calculated by the MST radar power spectrum parameter calculation model and the power spectrum parameters obtained after Step 3 processing reaches the minimum. Through repeated adjustment and comparative experiments, the best parameters are obtained to obtain the best MST radar power spectrum parameter calculation model;

[0093] In this embodiment, the specific parameter settings of the model are as follows: the total number of training epochs epoch is 200 rounds, the initial value of the learning rate learningrate is 0.001, and the learning rate is dynamically updated based on the loss of the validation set. If the loss of the validation set does not improve within 5 rounds of training, the learning rate will be reduced. The core purpose is to help the training avoid stagnating at the local optimal solution and thus improve the final performance of the model; define MSE as the loss function of the model. The model contains 3 power spectrum parameters for calculating the MST radar power spectrum, so there are 3 loss values. The sum of the 3 loss values is averaged and applied in the forward propagation of the model; through repeated experiments and comparisons, the best parameters are obtained.

[0094] Step 6, input the MST radar power spectrum to be measured into the trained best MST radar power spectrum parameter calculation model. After the calculation is completed, the MST radar power spectrum parameters of each beam in the low mode are obtained, such as Figure 6As shown, the abscissa is the altitude (2.55 - 10.05 km). Among them, for the signal-to-noise ratio, the Original SNR is the signal-to-noise ratio generated by the existing 8-segment averaging method, and the Model SNR is the signal-to-noise ratio generated by the method of the present invention; for the spectral width w, the Original spectral width is the spectral width fitted by the Gaussian spectrum in Step2, and the Model spectral width is the spectral width generated by the method of the present invention; for the radial average velocity, the Original radial average velocity is the radial average velocity calculated by the spectral moment method in Step2, and the Model radial average velocity is the radial average velocity generated by the method of the present invention.

[0095] The present invention also provides a system for calculating MST radar power spectrum parameters based on deep learning, and the system includes:

[0096] An original power spectrum acquisition module, configured to acquire the original MST radar power spectrum;

[0097] An initial power spectrum parameter calculation module, configured to first perform ground clutter suppression, three-point median filtering, and data normalization on the original MST radar power spectrum, and then calculate the initial MST radar power spectrum parameters;

[0098] A power spectrum parameter quality control module, configured to use beam symmetry spectrum selection and spatio-temporal continuity comparison methods to control the quality of the initial MST radar power spectrum parameters and correct inaccurate power spectrum parameters;

[0099] A calculation model establishment module, configured to establish a calculation model for MST radar power spectrum parameters based on deep learning and input the obtained original MST radar power spectrum into the model;

[0100] A calculation model training module, configured to set the hyperparameters, total number of training epochs, adjustable learning rate, and loss function of the MST radar power spectrum parameter calculation model, so that the mean square loss between the power spectrum parameters calculated by the MST radar power spectrum parameter calculation model and the power spectrum parameters processed by the power spectrum parameter quality control module reaches the minimum, and obtain the best parameters through repeated adjustment and comparative experiments to obtain the best MST radar power spectrum parameter calculation model;

[0101] A power spectrum parameter calculation module, configured to input the MST radar power spectrum to be measured into the trained best MST radar power spectrum parameter calculation model, and obtain the MST radar power spectrum parameters of each beam in the low mode after the calculation is completed.

[0102] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the program, the method for calculating MST radar power spectrum parameters based on deep learning is implemented.

[0103] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method for calculating MST radar power spectrum parameters based on deep learning is implemented.

[0104] The present invention can calculate accurate MST radar power spectrum parameters under precipitation, ground clutter, and intermittent clutter, facilitating the obtaining of more accurate atmospheric parameters and meeting the needs of studying fine wind field changes. At present, the construction of MST radars and the upgrading of radar technologies in China are all steadily advancing, and more MST radar data are continuously being released, laying a foundation for expanding the dataset in the future and further improving the generalization ability of the model and the accuracy of calculations. The present invention can quickly obtain the calculation results of MST radar power spectrum parameters after the MST radar power spectrum is released, and thus has higher efficiency and higher reliability.

[0105] The specific embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those of ordinary skill in the art.

Claims

1. A method for calculating MST radar power spectrum parameters based on deep learning, characterized by: The method comprises: Step 1, obtain the original MST radar power spectrum; Step 2: The original MST radar power spectrum is first subjected to ground clutter suppression, three-point median filtering, and data normalization, and then the initial MST radar power spectrum parameters are calculated; Step 3, use the beam symmetry spectrum picking and time-space continuity comparison method to control the quality of the initial MST radar power spectrum parameters and correct inaccurate power spectrum parameters; Step 4: Establish a MST radar power spectrum parameter calculation model based on deep learning, and input the original MST radar power spectrum obtained in Step 1 into the model; Step 5, set the hyperparameters, total number of training rounds, adjustable learning rate and loss function of the MST radar power spectrum parameter calculation model, so that the mean square loss of the power spectrum parameters calculated by the MST radar power spectrum parameter calculation model and the power spectrum parameters obtained by Step 3 is minimized, and the optimal parameters are obtained through repeated adjustments and comparative tests to obtain the optimal MST radar power spectrum parameter calculation model; Step 6. Input the power spectrum of the MST radar to be tested into the trained optimal MST radar power spectrum parameter calculation model. After the calculation is completed, the power spectrum parameters of the MST radar for each beam in the low mode are obtained.

2. The method for calculating MST radar power spectrum parameters based on deep learning according to claim 1, characterized in that: In Step 2, the initial MST radar power spectrum parameters include signal-to-noise ratio SNR, radial average velocity Spectral width w.

3. The method for calculating MST radar power spectrum parameters based on deep learning according to claim 1, characterized in that: In the Step 2, the calculation of the initial MST radar power spectrum parameters includes: using the segmented average method to calculate the signal-to-noise ratio (SNR) of the power spectrum, and using the spectral moment method to calculate the radial average velocity. The spectral width w is calculated using the Gaussian spectrum model; specifically, it includes: (1) The signal-to-noise ratio (SNR) of the power spectrum calculated using the segmented averaging method includes: Based on the noise obeying the χ2 with 2N / k degrees of freedom 2 Distribution, where N is the number of FFT points, χ 2 It is the chi-square distribution in mathematical statistics. The power spectrum signal is divided into k segments, the average power spectrum density of each segment is calculated, and then the smallest average value P is selected from the segment average values. N It is regarded as the noise level value of the entire spectrum, and then the signal-to-noise ratio (SNR) of the power spectrum is calculated; (2) Calculate the radial average velocity using the spectral moment method include: First, for each beam and each range bin, calculate the zero-order moment M0 and first-order moment M1 of the signal power spectrum: N is the number of FFT points, M0 is the zero-order moment, M1 is the first-order moment, S i and f i are the power value and frequency value of the i-th point respectively, and Δf is the frequency resolution; The parameters of the atmospheric wind field are calculated based on the moment difference method. The parameters of the atmospheric wind field include the signal power estimation P S , radial average velocity Signal power estimation P S , radial average velocity The calculation formula is expressed as: P S =M0; Where Δv is the Doppler velocity resolution, λ is the wavelength of the electromagnetic wave emitted by the radar; (3) Use the Gaussian spectrum model to calculate the spectrum width w; the echo power S′ fitted by the Gaussian spectrum is expressed as: Where P r is the echo strength, is the radial average velocity obtained by the least squares fitting method, w is the spectrum width, v r represents the radial velocity of a single spectrum point, and S′ represents the echo power fitted by the Gaussian spectrum; the fitting process is to fit the Gaussian spectrum model to the power spectrum line of each range gate MST radar, and determine the echo intensity P by the least squares fitting method. r , radial average velocity Spectral width w; find a set of P r , The process of minimizing the mean square error between the Gaussian spectrum model spectrum line and the power spectrum line of each range gate MST radar is expressed as: r , Among them, argmin is a function that is used to make the objective function reach the minimum value of the variable, S refers to the actual echo power value, P N is the noise level value.

4. The method for calculating MST radar power spectrum parameters based on deep learning according to claim 1, characterized in that: The Step 3 includes: The symmetry test of five-beam echoes and the principle of space-time continuity are used to determine the selection of spectrum peaks and control the quality of MST radar power spectrum parameters, and inaccurate power spectrum parameters are manually corrected.

5. The method for calculating MST radar power spectrum parameters based on deep learning according to claim 1, characterized in that: In the Step 4, the MST radar power spectrum parameter calculation model based on deep learning includes a one-dimensional convolution layer, a maximum pooling layer, a GRU layer and a fully connected layer.

6. The method for calculating MST radar power spectrum parameters based on deep learning according to claim 5, characterized in that: In the Step 4, two convolution layers with a convolution kernel size of 3 and a step size of 1 and two maximum pooling layers with a pooling kernel size of 5 and a step size of 2 are used to extract the features of the MST radar power spectrum, one GRU layer is used to process the features extracted by the convolution layer, and two fully connected layers are used to output the final power spectrum parameters.

7. The method for calculating MST radar power spectrum parameters based on deep learning according to claim 1, characterized in that: The Step 4 includes: (1) Normalize the original MST radar power spectrum of each height range library, and the normalized output X normalization It is expressed as: When normalizing the data at each height, X normalization is the normalized output, X min and X max It refers to the minimum and maximum values ​​of echo power, and X refers to the power spectrum at each height; (2) The normalized power spectrum is passed through the first convolutional layer and then input into the maximum pooling layer. The output of the maximum pooling layer is used as the input of the second GRU layer after changing the data dimension. Finally, the power spectrum parameter calculation result is obtained through the fully connected layer. The calculation formula for each GRU layer is as follows: Update gate z t :z t =σ(W z x t +U z h t-1 +b z ), W z and U z is the weight matrix of the update gate, b z is the bias term, σ is the Sigmoid activation function, x t Represents the input of the current time step; Reset Gate t :r t =σ(W r x t +U r h t-1 +b r ), W r and U r is the weight matrix of the reset gate, b r is the bias term, σ is the Sigmoid activation function; Candidate hidden states b h is the bias term, h t-1 represents the past hidden state, ⊙ represents element-wise multiplication, tanh is used as a nonlinear activation function, W h , U h They are all weight parameters; Compute the final hidden state: h t represents the final hidden state.

8. The MST radar power spectrum parameter calculation system based on deep learning is characterized by: The system includes: a module for executing the MST radar power spectrum parameter calculation method based on deep learning as described in any one of claims 1 to 7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the MST radar power spectrum parameter calculation method based on deep learning is implemented as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for calculating the power spectrum parameters of an MST radar based on deep learning as described in any one of claims 1 to 7 is implemented.

Images