Prediction Method for Multiple Sea Surface Physical Quantities Based on MoM and Deep Learning Data Driven

Abstract

The present invention discloses a prediction method for multiple sea surface physical quantities based on MoM and deep learning data-driven. The system is as follows: Describe various parameters of the sea surface, extract relevant characteristics of the sea surface under corresponding environmental parameters, generate a static sea surface based on the common geometric sea spectrum model PM spectrum, obtain clutter characteristic data in combination with electromagnetic analysis means according to the sea conditions, assist in calculating the electromagnetic scattering of the sea surface by MoM, discuss the numerical calculation results, and thus construct a sample library of different sea surface environmental characteristics in the form of statistical models and numerical calculation data as samples. Then introduce BP neural network deep learning, use the sea clutter characteristic data as input and various sea surface standardized parameter data as labels, compare different structural neural networks, and obtain the optimal neural network under the minimum error. Based on the optimal neural network structure, iterate and continuously correct the model to obtain the optimal prediction model, and finally realize the online prediction of the geometric characteristics of the sea surface.

Description

Technical Field

[0001] The present invention relates to the technical field of numerical calculation of electromagnetic characteristics of a static sea surface. Specifically, it is a prediction method for multiple sea surface physical quantities based on MoM and deep learning data-driven. Background Technique

[0002] The research on sea surface electromagnetic scattering is closely related to fields such as ocean microwave remote sensing, radar imaging, interception technology, sea target detection, and intelligent recognition. Sea surface electromagnetic scattering has also become an important research direction and development trend with the application of these technologies in various fields.

[0003] In the aspect of remote sensing technology, the concept of synthetic aperture radar (SAR) imaging has emerged for nearly 70 years. The electromagnetic scattering theory is closely related to synthetic aperture radar ocean remote sensing and is also complementary to the radar measurement technology of the sea surface. In the past 40 years, high-resolution imaging radars have been widely used in ocean remote sensing, and the research on sea surface electromagnetic characteristics is deeply needed.

[0004] In the field of military applications, radar has an indispensable position in modern high-tech wars. Without radar, or if the radar does not meet the requirements, surface ships, anti-ship missiles, and shipborne aircraft will be like losing their eyes and lead to the failure of the war. And how to accurately identify targets in the complex sea clutter background, guide aircraft and anti-ship missiles to resist complex electromagnetic interference and accurately attack targets all require in-depth research on sea surface electromagnetic scattering to provide reliable technical support for radar simulation experiments, and then achieve victory in future ocean wars. Summary of the Invention

[0005] The purpose of the present invention is to provide a prediction method for multiple sea surface physical quantities based on MoM and deep learning data-driven. Under single-station conditions, when θ is 0° - 70° and Φ is 0° under single-station conditions, the sea surface RCS echo data is input to predict the sea surface standardized parameters, and the error is guaranteed to be within 5%.

[0006] The technical solution to achieve the purpose of the present invention is as follows: A prediction method for multiple sea surface physical quantities based on MoM and deep learning data-driven, the steps are as follows:

[0007] Step 1: Use the FORTRAN language in the VS2010 software to write a program to generate a sea surface based on the PM spectrum, and obtain a sea surface model.LIS file. The length and width of the sea surface model are 1m × 1m, and the sea surface is divided into 35 × 35 triangles;

[0008] Step 2: For the sea surface model in Step 1, use MoM to calculate the RCS and obtain the RCS echo data of the sea surface at θ = 0° - 70° and Φ = 0° under monostatic conditions. During the calculation process, set the frequency parameter to 1.5 GHz, the dielectric constant of the sea surface corresponding to this frequency to 79.6 - j6.57, the scanning interval to 0.5°, and the polarization mode to VV polarization;

[0009] Step 3: Repeat Steps 1 and 2 6000 times to obtain sufficient required RCS data. Collect the obtained RCS data and the wind speed data in the static sea surface model generated in Step 1 as the neural network sample library for deep learning;

[0010] Step 4: Based on the BP neural network theory, use the RCS to invert the wind speed; use the RCS data as the input and the wind speed data as the neural network sample label; change the hidden layer structure, optimizer, learning rate, and batch of the neural network, and compare the test set error decline curves of different neural networks iterating 1 - 2000 times to obtain the optimal neural network;

[0011] Step 5: Repeat training the optimal neural network in Step 4 and make comparisons to establish the optimal prediction model. The finally obtained model uses the RCS echo data of the sea surface at θ = 0° - 70° and Φ = 0° under monostatic conditions as the input to predict the corresponding sea surface standardized parameter data such as the wind speed, root mean square height, sea surface correlation length, and effective height of the sea surface;

[0012] Step 6: Use PyCharm to package the BP neural network prediction model into an.exe file, use Matlab to write the software interface, and call the executable.exe file of the neural network prediction model to achieve the software encapsulation of the program. At the same time, encapsulate the PM spectrum sea surface code of the software, input the wind speed data, and generate a sea surface model of 200×200 with a dissection interval of 2.

[0013] Compared with the prior art, the significant advantages of the present invention are as follows: When using FORTRAN to calculate the RCS data in this method, MoM is adopted. The MoM solution process is simple, the solution steps are unified, various types of data can be accurately calculated, and various complex targets can be processed. At the same time, a sufficiently large sample library is established. When using the BP neural network for training, after iterating multiple times, the average error is less than 5%, indicating that a sufficient accuracy is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a schematic diagram of the extraction and inversion of sea surface characteristics based on the deep learning algorithm of the present invention.

[0015] Figure 2 is a sea surface model diagram generated by writing a program based on the PM spectrum using the FORTRAN language in the VS2010 software of the present invention.

[0016] Figure 3 It is the FEKO sea surface coordinate system of the sea surface model when the wind speed is 9.89 m / s.

[0017] Figure 4 It is the comparison diagram of sea surface models with different wind speeds. (a) Sea spectrum model diagram when the wind speed is 9.89 m / s, (b) Sea spectrum model diagram when the wind speed is 9.89 m / s, (c) Sea spectrum model diagram when the wind speed is 6.5 m / s, (d) Sea spectrum model diagram when the wind speed is 4.2 m / s, (e) Sea spectrum model diagram when the wind speed is 2.7 m / s, (f) Sea spectrum model diagram when the wind speed is 0.5 m / s.

[0018] Figure 5 It is the substitution diagram of the single - station RCS varying with the incident angle at different wind speeds. (a) RCS variation diagram when the wind speed is 9.89 m / s, (b) RCS variation diagram when the wind speed is 8.8 m / s, (c) RCS variation diagram when the wind speed is 6.5 m / s, (d) RCS variation diagram when the wind speed is 4.2 m / s, (e) RCS variation diagram when the wind speed is 2.7 m / s, (f) RCS variation diagram when the wind speed is 0.5 m / s.

[0019] Figure 6 It is the sea surface under the 10×10 subdivision surface.

[0020] Figure 7 It is the sea surface diagram under the 15×15 subdivision surface.

[0021] Figure 8 It is the sea surface diagram under the 20×20 subdivision surface.

[0022] Figure 9 It is the sea surface diagram under the 25×25 subdivision surface.

[0023] Figure 10 It is the sea surface diagram under the 30×30 subdivision surface.

[0024] Figure 11 It is the sea surface diagram under the 35×35 subdivision surface.

[0025] Figure 12 It is the sea surface diagram under the 40×40 subdivision surface.

[0026] Figure 13 It is the sea surface model diagram generated by writing a program based on the PM spectrum using the FORTRAN language in the VS2010 software in the example where the wind speed is 9.89 m / s in the present invention.

[0027] Figure 14 It is the waveform diagram of the RCS data change in the example where the wind speed is 9.89 m / s. Specific implementation manner

[0028] The present invention provides a method for modeling and rapidly analyzing the electromagnetic scattering characteristics of the sea surface by constructing a sea surface environment database based on the Method of Moments (MOM) and an inversion method for sea scene parameters driven by deep learning data, so as to realize the extraction of sea surface characteristics and construct a system for predicting sea surface physical quantities. The system is as follows: various parameters describing the sea surface are used to extract the relevant characteristics of the sea surface under the corresponding environmental parameters. A static sea surface is generated based on the common geometric sea spectrum model, the PM spectrum. According to the sea state, combined with electromagnetic analysis means (analytical and numerical simulation), clutter characteristic data are obtained, and the MoM is used to calculate the electromagnetic scattering of the sea surface. The numerical calculation results are discussed, so as to construct a sample library of different sea surface environmental characteristics in the form of statistical models and numerical calculation data. Subsequently, BP neural network deep learning is introduced, with sea clutter characteristic data (such as RCS) as the input and various standardized sea surface parameter data (such as wind speed) as the labels. By comparing neural networks with different structures, the optimal neural network with the minimum error is obtained. Based on the optimal neural network structure, iterative correction is carried out 20,000 times to continuously correct the model, and the optimal prediction model is obtained, finally realizing the online prediction of the geometric characteristics of the sea surface.

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Combined with Figures 1 - 3 , a prediction method for multiple sea surface physical quantities driven by MoM and deep learning data includes the following steps:

[0031] 1. Generation of the sea surface model

[0032] Using the FORTRAN language in the VS2010 software, a program is written based on the PM spectrum to generate the sea surface, and a sea surface model.LIS file is obtained. The length and width of the sea surface model are 1m×1m, and the sea surface is divided into 35×35 triangles.

[0033] 2. Calculation process of the dielectric constant of seawater

[0034] In the study of the interaction between electromagnetic waves and the sea surface, the dielectric constant of seawater is a very important parameter. The dielectric constant of seawater is usually a complex function of the electromagnetic wave frequency, the temperature of seawater, and the salinity of seawater. It is usually assumed that this functional relationship follows the so-called Debye function, that is

[0035]

[0036] Table 1 Dielectric constant diagram of each microwave band

[0037]

[0038] From the table in Table 1, the real part of the relative dielectric constant corresponding to the frequency of 1.5 GHz is 79.6, and the imaginary part is -6.57.

[0039] 3. Calculation of sea surface scattering characteristics in the far-field case

[0040] The essence of calculating the electromagnetic scattering problem of the sea surface is to solve the Maxwell equations under certain boundary conditions. When solving practical problems, in most cases, the integral equations derived from the Maxwell equations combined with the boundary conditions are directly solved.

[0041] In the present invention, the VV polarization mode is adopted in calculating the electromagnetic scattering of the rough sea surface, and the electromagnetic in three-dimensional problems needs to be calculated. According to the vector Green's theorem

[0042]

[0043] Let P = E and Q = aG, where a is a vector in an arbitrary direction and G is the aforementioned Green's function. Then, by continuous calculation, the following equation in the rectangular coordinate system is finally obtained:

[0044]

[0045] This equation can solve the electromagnetic scattering problem in three dimensions, and the scattering field of the two-dimensional scattering problem can be easily obtained from it.

[0046] 1) Far-field case

[0047] In three-dimensional problems, considering the far field:

[0048]

[0049] It can be obtained that

[0050]

[0051] 4. MoM of dielectric rough surface

[0052] The pulse function expansion is actually a constant approximation. It divides the boundary into several segments {ΔC 1 , ΔC 2 , …, ΔC N}, and assumes that the values of the surface current and surface magnetic current are constants {J z (1), J z (2), …, J z (N)} and {K l (1), K l (2), … K l (N)} on each segment. In this way, the surface current and surface magnetic current on the boundary can be respectively expressed as and where P q is the pulse function.

[0053]

[0054] According to the square matrix

[0055]

[0056] where A - , D + and D - have similar forms.

[0057] Any element in + A

[0058] A - is

[0059] Any element in + D

[0060] Any element in - D

[0061] P q is the distance from the q-th observation point to the integration point. It should be noted that since the pulse function is used as the expansion function, that is, the surface current and surface magnetic current are constant in each segment, to find + , A - , D + and D - The key is to find the integral therein. It should be noted that when q = p, that is, the observation point and the integration point coincide, the integral is singular and needs to be processed by analytical methods.

[0062] Using MoM to calculate the RCS, the monostatic RCS echo data of the sea surface with θ from 0° to 70° and Φ = 0° are obtained. During the calculation process, the frequency parameter is set to 1.5 GHz, the sea surface dielectric constant at the corresponding frequency is 79.6 - j6.57, the scanning interval is 0.5°, and the polarization mode is VV polarization.

[0063] Here, six sea surface models are taken as examples first. The meshing surfaces of these six models are all 35×35, and the wind speeds are different. Taking the wind speed of 9.89 m / s as an example, the Beaufort scale is 5, strong breeze. The sea surface model is as Figure 4 shown. Using the method of moments to calculate the sea surface echo data. The size of the sea surface is 1 m×1 m, the sea surface is meshed into 35×35, the incident wave frequency is 1.5 GHz, the corresponding sea surface dielectric constant is 79.6 - j6.57, the polarization mode is VV polarization, and the monostatic RCS echo data of the sea surface with θ from 0° to 70° and Φ = 0° are calculated. The parameters of the other five groups are similar to the above and are presented in the following table.

[0064] Table 2 Model Calculation Parameter Table

[0065]

[0066] From the comparison chart Figure 5 , as the wind speed increases, the sea surface waves become larger. When the wind speed is 9.98 m / s, there are already significant undulations on the sea surface. Considering the actual application scenario of this project and combining with the wind scale division in the real scenario, the inversion wind speed is selected as 0.3 - 10 m / s;

[0067] Taking RCS / dBm 2 as the unit to generate a variation chart. It can be seen from the chart that at the same wind speed, as the incident angle changes, the RCS echo of the sea surface fluctuates greatly. And it can be seen that the smaller the wind speed, the larger the RCS data during the entire change process of the incident angle. This is consistent with the actual situation, so the correctness is proved.

[0068] Combined with Figures 6 - 11 , a comparison is shown with different numbers of subdivisions at the same wind speed (6.5 m / s):

[0069] It can be seen from the above six groups of sea surface maps with different subdivisions that if the subdivision size is too large, the sea surface details will be lost, so it is not adopted. When choosing a 40×40 subdivision, it takes a lot of time to establish the database. During continuous experiments, a 35×35 subdivision surface is finally selected, which retains most of the sea surface information and has almost no impact on subsequent operations. Therefore, the 35×35 subdivision surface is finally selected.

[0070] 5. BP Neural Network Training and Prediction Model

[0071] In the present invention, the RCS data obtained by repeatedly performing Steps 1 and 2 is used as the input of the neural network, and the corresponding wind speed data is used as the output to continuously train the neural network. During the training process, a prediction model with an error within 5% is obtained after 20,000 iterations. The parameters during the operation are as follows:

[0072] Table 3 BP Neural Network Parameter Table

[0073]

[0074] Finally, a system for predicting the current static sea surface wind speed and other sea surface standardization parameters based on the RCS echo data of the sea surface with θ from 0° to 70° and Φ being 0° under single - station conditions is obtained.

[0075] Example 1

[0076] A sea surface model is given, such as Figure 1As shown in the figure, the RCS is calculated using MoM. The size of the sea surface is 1m×1m, which is divided into 35×35 grids. The incident wave frequency is 1.5GHz, and the polarization mode is VV polarization. Under the condition of monostatic, θ ranges from 0° to 70°, Φ is 0°, and the scanning angle interval is 0.5°. The relative permittivity of the sea surface is 79.6 - j6.57. The wind speed of the test case is 9.89m / s, and the Beaufort scale is 5 (fresh breeze).

[0077] Table 4 Model Calculation Parameter Table

[0078] Parameter name Value / Method Polarization method VV polarization Incident frequency 1.5 GHz Angle of incidence θ is 0° - 70°, Φ is 0° Scanning interval 0.5° Beaufort scale Strong breeze of Force 5 Sea surface dielectric constant 79.6-j6.57

[0079] Taking RCS / dBm 2 as the unit to generate a waveform diagram. As can be seen from Figure 12 、 13 、14, at the same wind speed, with the change of the incident angle, the RCS echo of the sea surface fluctuates greatly.

Claims

1. A prediction method for multiple sea surface physical quantities based on MoM and deep learning data-driven, It is characterized in that Here are the steps: Step 1: Use the FORTRAN language in VS2010 software to write a program based on the PM spectrum to generate the sea surface and obtain a sea surface model .LIS file; divide the sea surface into triangles; Step 2: For the sea surface model in step 1, use MoM to calculate RCS and obtain the sea surface RCS echo data with θ of 0°-70° and Φ of 0° under single-station conditions; during the calculation process, set the frequency parameter to 1.5 GHz, the sea surface dielectric constant at the corresponding frequency to 79.6-j6.57, the scanning interval to 0.5°, and the polarization mode to VV polarization; Step 3: Repeat steps 1 and 2 N times to obtain sufficient required RCS data, where N is greater than 10,000; collect the obtained RCS data and the wind speed data in the static sea surface model generated in step 1 as a neural network sample library for deep learning; Step 4: Based on BP neural network theory, use RCS to invert wind speed; use RCS data as input and wind speed data as neural network sample labels; Change the hidden layer structure, optimizer, learning rate, and batch of the neural network, compare the test set error reduction curves of different neural network iterations 1-2000 times, and obtain the optimal neural network; Step 5: Repeat the training of the optimal neural network in step 4 and compare them to establish the optimal prediction model; the final model uses the sea surface RCS echo data with θ of 0°-70° and Φ of 0° under single-station conditions as input to predict the sea surface corresponding wind speed, root mean square height, sea surface correlation length, effective height and other sea surface standardized parameter data; Step 6: Use PyCharm to encapsulate the BP neural network prediction model into an .exe file, use Matlab to write the software interface, call the neural network prediction model executable exe file, and realize the software packaging of the program; at the same time, the software encapsulates the PM spectrum sea surface code, inputs the wind speed data, and generates a 200×200, subdivision interval 2 sea surface model.

Images