CN113177356B - Target electromagnetic scattering characteristic rapid prediction method based on deep learning - Google Patents

Target electromagnetic scattering characteristic rapid prediction method based on deep learning Download PDF

Info

Publication number
CN113177356B
CN113177356B CN202110469304.2A CN202110469304A CN113177356B CN 113177356 B CN113177356 B CN 113177356B CN 202110469304 A CN202110469304 A CN 202110469304A CN 113177356 B CN113177356 B CN 113177356B
Authority
CN
China
Prior art keywords
target
neural network
theta
value
network model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110469304.2A
Other languages
Chinese (zh)
Other versions
CN113177356A (en
Inventor
李尧尧
郭俊玲
蔡少雄
胡蓉
苏东林
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beihang University
Original Assignee
Beihang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beihang University filed Critical Beihang University
Priority to CN202110469304.2A priority Critical patent/CN113177356B/en
Publication of CN113177356A publication Critical patent/CN113177356A/en
Application granted granted Critical
Publication of CN113177356B publication Critical patent/CN113177356B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Abstract

The invention discloses a target electromagnetic scattering characteristic rapid prediction method based on deep learning, which comprises the following steps: s1: establishing a target electromagnetic simulation model; s2: determining factors influencing the electromagnetic scattering characteristics of the target; s3: simulating a target model under the condition that the influence factors take different values to obtain a far field RCS of the target model, and establishing a training set and a test set; s4: constructing a BP neural network model by using a BP neural network algorithm; s5: training the BP neural network model by utilizing a training set; s6: testing the BP neural network model obtained by training by using a test set, if the test reaches the standard, entering the step S7, and if the test does not reach the standard, returning to the step S4; s7: and rapidly predicting the electromagnetic scattering property by using the BP neural network model which reaches the standard in the test. The method effectively solves the problems of large calculation amount, low solving efficiency and the like of the existing method, and meets the requirement of fast prediction of the electromagnetic scattering property of the high dynamic target.

Description

Target electromagnetic scattering characteristic rapid prediction method based on deep learning
Technical Field
The invention relates to computational electromagnetism, in particular to a target electromagnetic scattering characteristic rapid prediction method based on deep learning.
Background
With the rapid development of intelligent technology and ad hoc network technology, a clustered and intelligent cooperative system has become a necessary trend for electronic information systems to work. For a detection type electronic device, a target is irradiated by emitting an electromagnetic wave, and an echo thereof is received, thereby obtaining information such as a distance, a distance change rate (radial velocity), an azimuth, and an altitude from the target to an electromagnetic wave emission point.
The scattering echo of a target is the basis of the work of detection type electronic equipment, for a complex target with large scale, the traditional numerical method has large calculation amount and slow solving speed, the angle, the posture and the like of the target moving at high speed are constantly changed, the electromagnetic scattering property of the target is changed along with the change, and the requirement of the RCS (radar cross section) quick prediction of the target cannot be met by simply adopting the traditional numerical method frame by frame calculation mode, so the quick prediction of the electromagnetic scattering property of the high-dynamic target increasingly becomes an important factor influencing the normal performance of the electronic equipment.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, provides a target electromagnetic scattering characteristic rapid prediction method based on deep learning, effectively solves the problems of large calculation amount, low solving efficiency and the like of the prior method, and meets the requirement of rapid prediction of high-dynamic target electromagnetic scattering characteristics.
The purpose of the invention is realized by the following technical scheme: a target electromagnetic scattering characteristic rapid prediction method based on deep learning comprises the following steps:
s1: establishing a target electromagnetic simulation model;
s2: determining factors influencing the electromagnetic scattering characteristics of the target;
s3: simulating a target model under the condition that the influence factors take different values to obtain a far field RCS of the target model, and establishing a training set and a test set;
s4: constructing a BP neural network model by using a BP neural network algorithm;
s5: training the BP neural network model by utilizing a training set;
s6: testing the BP neural network model obtained by training by using a test set, if the test reaches the standard, entering the step S7, and if the test does not reach the standard, returning to the step S4;
s7: and rapidly predicting the electromagnetic scattering property by using the BP neural network model which reaches the standard in the test.
The invention has the beneficial effects that: the invention provides a priori knowledge-based rapid prediction idea of 'generation in advance and on-site calling' by introducing deep learning, can effectively solve the problems of large calculation amount, low solving efficiency and the like of the conventional method, and meets the requirement of rapid prediction of the electromagnetic scattering property of a high dynamic target.
Drawings
FIG. 1 is a flow chart of a method of the present invention;
FIG. 2 is a diagram illustrating factors affecting the electromagnetic scattering properties of a target;
FIG. 3 is a schematic diagram of the topology of the BP neural network in the embodiment;
FIG. 4 is a schematic diagram of an electromagnetic simulation model of a single triangular pyramid in an embodiment;
FIG. 5 is a schematic diagram of an electromagnetic simulation model of two triangular pyramids in an embodiment;
fig. 6 shows the predicted RCS and the true RCS when θ equals 105.77 ° and MAE equals 0.6103 in the embodiment
Figure BDA0003044731370000021
A graph of the variation;
fig. 7 shows the predicted RCS and the true RCS when θ equals 115.33 ° and MAE equals 0.7727 in the example
Figure BDA0003044731370000022
A graph of the variation;
fig. 8 shows the predicted RCS and the true RCS when θ equals 140.03 ° and MAE equals 0.6569 in the example
Figure BDA0003044731370000023
A graph of the variation;
fig. 9 shows the predicted RCS and the true RCS when θ equals 167.77 ° and MAE equals 0.3338 in the example
Figure BDA0003044731370000024
The curve of the change is shown schematically.
Detailed Description
The technical solutions of the present invention are further described in detail below with reference to the accompanying drawings, but the scope of the present invention is not limited to the following.
According to the invention, through analysis, the overall electromagnetic scattering characteristics of the target scene have a certain predictable relationship with the constituent elements (namely the structures of all sub-targets) and the relative positions among the targets. Aiming at meeting the demand of quickly predicting the electromagnetic scattering of the high dynamic target and solving the problem that the traditional numerical method can not meet the demand of predicting the electromagnetic scattering property of the high dynamic target in a short time, meanwhile, the characteristics of high dimensionality, more variables and large variation range of electromagnetic scattering characteristic data are considered, a rapid prediction thought based on priori knowledge, which is generated in advance and called on site, is provided by introducing a deep learning method, the isomorphic optimization algorithm fits a model capable of reflecting the data change rule by taking high-quality numerical simulation data samples as a base number, can keep the simulation precision of the traditional numerical method to a certain extent, can meet the requirement of rapid prediction, in practical application, the generalization of the obtained model is benefited, the field simulation calculation is not needed, the demand is directly brought into the model to quickly obtain the prediction result meeting the precision requirement, and specifically:
as shown in fig. 1, a method for rapidly predicting electromagnetic scattering characteristics of a target based on deep learning is characterized in that: the method comprises the following steps:
s1: establishing a target electromagnetic simulation model;
s2: determining factors influencing the electromagnetic scattering characteristics of the target;
s3: simulating a target model under the condition that the influence factors take different values to obtain a far field RCS of the target model, and establishing a training set and a test set;
s4: constructing a BP neural network model by using a BP neural network algorithm;
s5: training the BP neural network model by utilizing a training set;
s6: testing the BP neural network model obtained by training by using a test set, if the test reaches the standard, entering the step S7, and if the test does not reach the standard, returning to the step S4;
s7: and rapidly predicting the electromagnetic scattering property by using the BP neural network model which reaches the standard in the test.
Further, in step S1, the target electromagnetic simulation model refers to a model that can be used for electromagnetic simulation calculation, and includes a target geometry, a structural material, a simulation frequency, a boundary condition, an excitation manner, and a mesh generation manner set in simulation software. The mesh generation mode is a mode of generating a target geometric structure into mesh units in a simulation process, and because the geometric structure of the target cannot be directly subjected to simulation analysis in simulation software; when mesh is divided, the shape and the size of a mesh unit are obtained by mainly considering division; the mesh subdivision mode is based on simulation requirements of simulation software on a target geometric structure, and a tetrahedron, a hexahedron, an octahedron and the like are used as mesh units to divide the target geometric structure into a plurality of small units suitable for electromagnetic simulation calculation;
as shown in fig. 2, in step S2, the factors influencing the electromagnetic scattering property of the target include: factors affecting single target electromagnetic scattering properties and factors affecting group target electromagnetic scattering properties, wherein:
factors that affect the electromagnetic scattering properties of a single target include: material characteristics, working frequency, azimuth angle of incident wave and pitch angle of incident wave;
factors that affect the electromagnetic scattering properties of a population of targets include: material properties, operating frequency, azimuth angle of incident wave, pitch angle of incident wave, distance between cluster targets, and relative position between cluster targets.
The step S3 includes the following sub-steps:
s301, constructing a training set and a test set under the single-target electromagnetic scattering characteristic:
for the electromagnetic scattering property of a single target, the azimuth angle of an incident wave is considered
Figure BDA0003044731370000031
And the pitch angle theta of the incident wave, in
Figure BDA0003044731370000032
A value range of
Figure BDA0003044731370000033
Sampling interval of
Figure BDA0003044731370000034
Under the conditions that the value range of theta is theta 1-theta 2 and the sampling interval is delta theta 1, simulation software is utilized to simulate to obtain the far-field single-station RCS value sigma of a single target1(unit is m)2) Carrying out conversion to obtain an RCS conversion value sigma'1＝10lgσ1(Single sheet)The bit is dBsm),
Figure BDA0003044731370000039
and theta and its corresponding RCS conversion value sigma'1A single target training set is formed;
taking into account the azimuth of the incident wave
Figure BDA0003044731370000037
And the pitch angle theta of the incident wave, in
Figure BDA0003044731370000035
A value range of
Figure BDA0003044731370000036
Sampling interval of
Figure BDA0003044731370000038
Theta is in a range of theta 3-theta 4, and can be simulated by simulation software under the condition of randomly taking values or sampling at equal intervals delta theta 2 within the range to obtain a far-field single-station RCS value sigma of a single target2(unit is m)2) Carrying out conversion to obtain an RCS conversion value sigma'2＝10lgσ2(in units of dBsm),
Figure BDA0003044731370000041
and theta and its corresponding RCS conversion value sigma'2A single target test set is formed;
s302, constructing a training set and a testing set under the electromagnetic scattering characteristic of the group target:
for the electromagnetic scattering characteristics of the group targets, the distance r between the targets and the azimuth angle of an incident wave are considered
Figure BDA0003044731370000042
And the pitch angle theta of incident waves, the value range of r is r 1-r 2, the sampling interval is delta r1,
Figure BDA0003044731370000045
a value range of
Figure BDA0003044731370000043
Sampling interval of
Figure BDA0003044731370000044
Simulating by using simulation software under the condition that the theta value range is theta 1 ' -theta 2 ' and the sampling interval of theta is delta theta 1 ', and obtaining the far-field single-station RCS value sigma of the group target3(unit is m)2) Is converted to give sigma'3＝10lgσ3(in dBsm), r,
Figure BDA0003044731370000046
And theta and RCS conversion value sigma 'under different values and corresponding conditions'3Forming a group target training set;
considering the distance r between the targets, the azimuth of the incident wave
Figure BDA00030447313700000424
And the pitch angle theta of incident waves, the value range of r is r 3-r 4, the sampling interval is delta r2,
Figure BDA0003044731370000049
a value range of
Figure BDA0003044731370000047
Sampling interval of
Figure BDA0003044731370000048
Theta is in a value range of theta 3 ' to theta 4 ', and theta can be simulated by simulation software under the condition of randomly taking values or uniformly spacing delta theta 2 ' in the range to obtain a far-field single-station RCS value sigma of a group target4(unit is m)2) Is converted to give sigma'4＝10lgσ4(in dBsm), r,
Figure BDA00030447313700000410
And theta and RCS conversion value sigma 'under different values and corresponding conditions'4A group target test set is constructed.
In the above process, r,
Figure BDA00030447313700000411
And each value range and sampling interval of theta are parameters which are selected and set according to the actual application requirements, the unit of r is m,
Figure BDA00030447313700000412
and θ in degrees; for example
Figure BDA00030447313700000413
Can be 0-360 degrees, theta 1-theta 2 can be 90-180 degrees,
Figure BDA00030447313700000414
and Δ θ 1 may be 2 °, r 1-r 2 may be 3-38m, Δ r1 may be 1.5 m; wherein
Figure BDA00030447313700000415
In the range of
Figure BDA00030447313700000416
Theta 3 to theta 4 are in the range of theta 1 to theta 2, r3 to r4 are in the range of r1 to r2,
Figure BDA00030447313700000418
in the range of
Figure BDA00030447313700000417
The range of theta 3 'to theta 4' is within the range of theta 1 'to theta 2';
Figure BDA00030447313700000419
and
Figure BDA00030447313700000420
delta theta 2 and Delta theta 1,
Figure BDA00030447313700000421
And
Figure BDA00030447313700000422
Δ θ 2 'and Δ θ 1', and Δ r2 and Δ r1 may be the same or different.
The step S5 includes:
s501, training a single-target neural network model:
in training single targets
Figure BDA00030447313700000423
And theta is used as a characteristic vector to be input into the constructed BP neural network model, a predicted value of the neural network model is obtained through a forward propagation algorithm, loss value (loss) calculation is carried out on the predicted value and an RCS (remote control system) conversion value in a single target training set, then network weight is updated through backward propagation, and then forward and backward propagation processes are carried out in a circulating mode until the loss value achieves an expected effect, namely the single target training set is used for completing the training of the constructed neural network model to obtain the single target neural network model; wherein the loss value is calculated by mean square error or mean absolute error;
in the embodiment of the application, the training set has multiple groups due to the single target
Figure BDA0003044731370000051
And theta, each group
Figure BDA0003044731370000052
And theta is used as a sample, each sample has a corresponding RCS conversion value (real value) in a single target training set, the real values form a real vector, and after the characteristic vector formed by each sample is input into the neural network model, the predicted vector output by the model comprises a predicted value corresponding to each sample; comparing the predicted value contained in the predicted vector with the real value contained in the real vector to obtain a loss value;
s502, training a group target neural network model:
r in the group target training set,
Figure BDA0003044731370000053
And theta as the characteristic directionInputting the quantity into a BP neural network model, obtaining a predicted value of the neural network model through a forward propagation algorithm, calculating a loss value (loss) with an RCS (Radar Cross section) conversion value in a group target training set, updating network weight through backward propagation, and then recycling the forward and backward propagation processes until the loss value reaches an expected effect, namely completing the training of the established neural network model by utilizing a group target training set to obtain the group target neural network model; where the loss value is calculated by mean square error or mean absolute error.
In the embodiment of the present application, since there are multiple sets r in the group target training set,
Figure BDA0003044731370000054
And theta, each of r,
Figure BDA0003044731370000055
And theta is taken as a sample, in the group target training set, each sample has a corresponding RCS conversion value (real value), the real values form a real vector, after the characteristic vector formed by each sample is input into the neural network model, the predicted vector output by the model comprises a predicted value corresponding to each sample; comparing the predicted value contained in the predicted vector with the real value contained in the real vector to obtain a loss value;
in steps S501 to S502, if the loss value is lower than a set threshold, it is determined that an expected effect is achieved, wherein the set threshold may be set by self-definition according to specific requirements;
the step S6 includes:
s601. centralizing single target test
Figure BDA0003044731370000056
And the theta value is input into the single-target neural network model trained in the step S5 as a feature vector, a predicted value of the single-target neural network model is obtained through a forward propagation algorithm, the predicted value and an expected output value, namely an RCS (Rich coupled computing) conversion value in a single-target test set, are compared and analyzed, and whether the single-target neural network model is in a single-target test set or not is judged according to the prediction accuracyTesting to reach the standard; if the test is up to the standard, the step S7 is entered, and if the test is not up to the standard, the step S4 is returned;
the judgment mode for whether the test single-target neural network model reaches the standard is as follows: if the error between the predicted value of the single-target neural network model and the RCS conversion value in the single-target test set is smaller than the standard-reaching threshold, the test is up to the standard, and if the error between the predicted value of the single-target neural network model and the RCS conversion value in the single-target test set is not smaller than the standard-reaching threshold, the test is not up to the standard; the error comprises a mean square error or a mean absolute error; the standard reaching threshold value can be set in a self-defined mode according to specific requirements;
in the embodiment of the application, the single target test set has multiple groups
Figure BDA0003044731370000061
And theta, each group
Figure BDA0003044731370000062
And theta is used as a sample, in the single target test set, each sample has a corresponding RCS conversion value (real value), the real values form a real vector, after the characteristic vector formed by each sample is input into the single neural network model obtained by training, the prediction vector output by the model comprises a prediction value corresponding to each sample; comparing the predicted value contained in the predicted vector with the real value contained in the real vector to obtain an error;
s602, concentrating the group target test,
Figure BDA0003044731370000063
And the theta value is used as a feature vector and is input into the swarm target neural network model obtained by training in the step S5, a predicted value of the swarm target neural network model is obtained through a forward propagation algorithm, the predicted value and an expected output value, namely an RCS conversion value in the swarm target test set, are compared and analyzed, and whether the swarm target neural network model reaches the standard or not is judged according to the prediction precision; if the test is up to the standard, the step S7 is entered, and if the test is not up to the standard, the step S4 is returned;
the judgment mode of whether the test group target neural network model reaches the standard is as follows: if the error between the predicted value of the group target neural network model and the RCS conversion value in the group target test set is smaller than a standard threshold, the test is standard, and if the error between the predicted value of the group target neural network model and the RCS conversion value in the group target test set is not smaller than the standard threshold, the test is not standard; the error comprises a mean square error or a mean absolute error; wherein, the standard reaching threshold value can be self-defined according to specific requirements.
The step S7 includes the following substeps
S701, judging whether the single target electromagnetic scattering characteristics or the group target electromagnetic scattering characteristics need to be predicted according to the prediction requirements;
if the single target electromagnetic scattering characteristic needs to be predicted, the step S702 is carried out;
if the electromagnetic scattering characteristics of the group targets need to be predicted, the step S703 is executed;
s702, acquiring azimuth angle of incident wave
Figure BDA0003044731370000064
Inputting a tested standard single-target neural network model according to the pitch angle theta of the incident wave, and outputting a prediction result by the single-target neural network model;
s703, acquiring the distance r between targets and the azimuth angle of incident waves
Figure BDA0003044731370000065
And the pitch angle theta of the incident wave is input into the group target neural network model, and a prediction result is output by the group target neural network prediction model.
In embodiments of the present application, since there are multiple sets r,
Figure BDA0003044731370000066
And theta, each of r,
Figure BDA0003044731370000067
And θ as a sample, each sample having a corresponding RCS conversion value (true value) in the cluster target test setForming a real vector by the real values, inputting the feature vector formed by each sample into the trained group target neural network model, wherein the prediction vector output by the model comprises a prediction value corresponding to each sample; comparing the predicted value contained in the prediction vector with the real value contained in the real vector to obtain an error value;
as shown in fig. 3, in the embodiment of the present application, the BP neural network model is adopted to be composed of an input layer, an output layer and a plurality of hidden layers. The factors affecting RCS are taken as a feature vector [ x ]1,x2,…,xM]Input to a neural network model using an activation function f1,f2,fm,…,fnAnd the nonlinear mapping relation between the RCS influence factors and the values thereof is realized, and the weight value and the threshold value of the network are continuously adjusted through the processes of information forward propagation and error backward propagation, so that the RCS output value y predicted by the network is as close to the real RCS value as possible, and the training of the model is completed.
In the embodiment of the application, the simulation software FEKO is used to establish an electromagnetic simulation model of a single triangular pyramid and two triangular pyramids, the simulation model is as shown in fig. 4 and fig. 5 below, wherein the size of the single triangular pyramid is 3.9 mx 2.7 mx 1.2m, the material is set to be all-metal, the two triangular pyramids are arranged side by side, and the distance between the two triangular pyramids is adjustable. The present embodiment takes a single triangular pyramid and two triangular pyramids as examples, respectively, to perform feasibility verification of the proposed method.
Single triangular pyramid method validation:
under the conditions that the frequency of an incident wave is 500MHz, the incident angle theta is 90-180 degrees,
Figure BDA0003044731370000074
under the conditions that the angle is 0-360 degrees and the value intervals are all 2 degrees, the far-field single-station RCS value of a single pyramid is obtained through simulation. Then, a BP neural network comprising seven hidden layers with the neuron number of 64 is established and trained by simulation data. The model was then tested at angles of incidence theta of 93.33 deg., 105.77 deg., 115.33 deg., 123.33 deg., 140.03 deg., 153.33 deg., 167.77 deg.,
Figure BDA0003044731370000075
the simulation is carried out under the conditions that the temperature is 1.33-359.33 degrees and the value interval is 2 degrees, and a test set is established to evaluate the prediction capability and the generalization of the model. We selected three evaluation indices, which are defined as follows:
determination of coefficient R2(R-square)
Figure BDA0003044731370000071
Mean square error MSE (mean Squared error)
Figure BDA0003044731370000072
Mean Absolute error MAE (mean Absolute error)
Figure BDA0003044731370000073
Wherein, yiIs the true value of the ith sample, yiFor the prediction value of the i-th sample,
Figure BDA0003044731370000077
the mean of the true values.
The decision coefficient R2 is used for representing the quality of a fitting through the change of data, and the normal value range of the decision coefficient is [0, 1] according to the above expression, and the closer to 1, the better the fitting of the model to the data is. Mean square error MSE and mean absolute error MAE indicate the difference between the true and predicted values, the closer to 0 the better the model fits the data.
The R2, MSE and MAE values for the test set as a whole are shown in the following table:
R2 MSE MAE predicting time
0.9778 1.3438 0.7000 ＜0.6s
The test is concentrated under a certain fixed theta angle, and the predicted RCS and the real RCS are followed
Figure BDA0003044731370000076
The change curves are shown in FIGS. 6 to 9 below.
As can be seen from FIGS. 6 to 9, the difference between the predicted RCS and the true RCS is not large, and the true RCS can be well fitted
Figure BDA0003044731370000081
A trend of change.
Two triangular pyramid methods verify:
taking the distance between the two triangular pyramids to be 3, 4.5, 6, 7.5, … …, 39m, respectively, taking into account the distance and angle factors, the frequency of the incident wave is 1GHz, theta is 90-180 DEG,
Figure BDA0003044731370000082
and (3) carrying out simulation under the conditions that the angle is 0-360 degrees and the value intervals are 2 degrees, and establishing a training set of two triangular pyramid models. After that we can see that when r is 3m, 5.5m, 8m, 10.5m, …, 38m, theta is 95.5 deg., 108.2 deg., 121.3 deg., 134.6 deg., 147.4 deg., 159.1 deg., 173.7 deg.,
Figure BDA0003044731370000083
at 0 deg., 2 deg., 4 deg., … deg., and 360 degAnd (3) simulating, establishing a test set, establishing a BP (back propagation) neural network comprising 15 hidden layers, and predicting RCS (Radar Cross section) values of multiple targets under different conditions after training of a neural network model is completed. The MAE and the prediction time of the RCS of the two triangular pyramid models are shown in the table below under the condition of different distances, the MAE does not exceed 6dB, the prediction time is better than 0.6s, and the feasibility of the method is verified. The predicted results are evaluated as shown in the following table:
r(m) 3 5.5 8 10.5 13 15.5 18
MAE(dB) 5.1609 5.4314 5.4662 5.6168 5.8039 5.4677 5.7825
predicting time(s) 0.5195 0.5808 0.5443 0.5531 0.5495 0.5499 0.5536
r(m) 20.5 23 25.5 28 30.5 33 35.5
MAE(dB) 5.8564 5.2480 5.3825 5.9863 5.4970 5.8431 5.3439
Predicting time(s) 0.5395 0.5533 0.5360 0.5584 0.5441 0.5191 0.5149
While the foregoing description shows and describes a preferred embodiment of the invention, it is to be understood, as noted above, that the invention is not limited to the form disclosed herein, but is not intended to be exhaustive or to exclude other embodiments and may be used in various other combinations, modifications, and environments and may be modified within the scope of the inventive concept described herein by the above teachings or the skill or knowledge of the relevant art. And that modifications and variations may be effected by those skilled in the art without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (5)

1. A target electromagnetic scattering characteristic rapid prediction method based on deep learning is characterized in that: the method comprises the following steps:
s1: establishing a target electromagnetic simulation model;
s2: determining factors influencing the electromagnetic scattering characteristics of the target;
s3: simulating a target model under the condition that the influence factors take different values to obtain a far field RCS of the target model, and establishing a training set and a test set;
the step S3 includes the following sub-steps:
s301, constructing a training set and a test set under the single-target electromagnetic scattering characteristic:
for the electromagnetic scattering property of a single target, the azimuth angle of an incident wave is considered
Figure FDA0003257272300000011
And the pitch angle theta of the incident wave, in
Figure FDA0003257272300000012
A value range of
Figure FDA0003257272300000013
Sampling interval of
Figure FDA0003257272300000014
The value range of theta is theta 1-theta 2, and the sampling interval is delta theta 1, simulation software is used for simulating to obtain the single-target far distanceField single station RCS value σ1Carrying out conversion to obtain an RCS conversion value sigma'1＝10lgσ1，
Figure FDA0003257272300000015
And theta and its corresponding RCS conversion value sigma'1A single target training set is formed;
taking into account the azimuth of the incident wave
Figure FDA0003257272300000016
And the pitch angle theta of the incident wave, in
Figure FDA0003257272300000017
A value range of
Figure FDA0003257272300000018
Sampling interval of
Figure FDA0003257272300000019
Theta is in a range of theta 3-theta 4, and can be simulated by simulation software under the condition of randomly taking values or sampling at equal intervals delta theta 2 within the range to obtain a far-field single-station RCS value sigma of a single target2Carrying out conversion to obtain an RCS conversion value sigma'2＝10lgσ2，
Figure FDA00032572723000000110
And theta and its corresponding RCS conversion value sigma'2A single target test set is formed;
s302, constructing a training set and a testing set under the electromagnetic scattering characteristic of the group target:
for the electromagnetic scattering characteristics of the group targets, the distance r between the targets and the azimuth angle of an incident wave are considered
Figure FDA00032572723000000111
And the pitch angle theta of incident waves, the value range of r is r 1-r 2, the sampling interval is delta r1,
Figure FDA00032572723000000112
a value range of
Figure FDA00032572723000000113
Sampling interval of
Figure FDA00032572723000000114
Simulating by using simulation software under the condition that the theta value range is theta 1 ' -theta 2 ' and the sampling interval of theta is delta theta 1 ', and obtaining the far-field single-station RCS value sigma of the group target3Is converted to obtain sigma3′＝10lgσ3，r、
Figure FDA00032572723000000115
And theta and RCS conversion value sigma under different values and corresponding conditions3' forming a group target training set;
considering the distance r between the targets, the azimuth of the incident wave
Figure FDA00032572723000000116
And the pitch angle theta of incident waves, the value range of r is r 3-r 4, the sampling interval is delta r2,
Figure FDA00032572723000000117
a value range of
Figure FDA00032572723000000118
Sampling interval of
Figure FDA00032572723000000119
Theta is in a value range of theta 3 ' to theta 4 ', and theta can be simulated by simulation software under the condition of randomly taking values or uniformly spacing delta theta 2 ' in the range to obtain a far-field single-station RCS value sigma of a group target4Is converted to obtain sigma4′＝10lgσ4，r、
Figure FDA00032572723000000120
And theta are different values andRCS conversion value sigma under the corresponding conditions4' forming a group target test set;
s4: constructing a BP neural network model by using a BP neural network algorithm;
s5: training the BP neural network model by utilizing a training set;
s6: testing the BP neural network model obtained by training by using a test set, if the test reaches the standard, entering the step S7, and if the test does not reach the standard, returning to the step S4;
s7: and rapidly predicting the electromagnetic scattering property by using the BP neural network model which reaches the standard in the test.
2. The method for rapidly predicting the electromagnetic scattering property of the target based on the deep learning as claimed in claim 1, wherein: in step S1, the target electromagnetic simulation model is a model that can be used for electromagnetic simulation calculation, and includes a target geometric structure, a structural material, a simulation frequency, a boundary condition, an excitation mode, and a mesh generation mode set in simulation software; the mesh generation mode is a mode of generating a target geometric structure into mesh units in a simulation process.
3. The method for rapidly predicting the electromagnetic scattering property of the target based on the deep learning as claimed in claim 1, wherein: in step S2, the factors influencing the electromagnetic scattering property of the target include: factors affecting single target electromagnetic scattering properties and factors affecting group target electromagnetic scattering properties, wherein:
factors that affect the electromagnetic scattering properties of a single target include: material characteristics, working frequency, azimuth angle of incident wave and pitch angle of incident wave;
factors that affect the electromagnetic scattering properties of a population of targets include: material properties, operating frequency, azimuth angle of incident wave, pitch angle of incident wave, distance between cluster targets, and relative position between cluster targets.
4. The method for rapidly predicting the electromagnetic scattering property of the target based on the deep learning as claimed in claim 1, wherein: the step S5 includes:
s501, training a single-target neural network model:
in training single targets
Figure FDA0003257272300000022
And theta is used as a characteristic vector and is input into the constructed BP neural network model, a predicted value of the neural network model is obtained through a forward propagation algorithm, loss value calculation is carried out on the predicted value and an RCS (Radar Cross section) conversion value in a single target training set, then network weight is updated through backward propagation, and then forward and backward propagation processes are carried out in a circulating mode until the loss value reaches an expected effect, namely, the single target training set is used for completing the training of the constructed neural network model, and the single target neural network model is obtained; wherein the loss value is calculated by mean square error or mean absolute error;
s502, training a group target neural network model:
r in the group target training set,
Figure FDA0003257272300000021
And theta is used as a characteristic vector and is input into the BP neural network model, a predicted value of the neural network model is obtained through a forward propagation algorithm, loss value calculation is carried out on the predicted value and an RCS conversion value in a group target training set, then network weight is updated through backward propagation, and then the forward and backward propagation processes are carried out in a circulating mode until the loss value achieves the expected effect, namely training on the established neural network model is completed through the group target training set, and the group target neural network model is obtained; where the loss value is calculated by mean square error or mean absolute error.
5. The method for rapidly predicting the electromagnetic scattering property of the target based on the deep learning as claimed in claim 4, wherein: the step S6 includes:
s601. centralizing single target test
Figure FDA0003257272300000031
And the theta value is used as a feature vector and is input into the single-target neural network model obtained by training in the step S5, a predicted value of the single-target neural network model is obtained through a forward propagation algorithm, the predicted value and an expected output value, namely an RCS (Radar Cross section) conversion value in a single-target test set, are compared and analyzed, and whether the single-target neural network model reaches the standard or not is judged according to the prediction precision; if the test is up to the standard, the step S7 is entered, and if the test is not up to the standard, the step S4 is returned;
the judgment mode for whether the test single-target neural network model reaches the standard is as follows: if the error between the predicted value of the single-target neural network model and the RCS conversion value in the single-target test set is smaller than a set threshold, the test reaches the standard, and if the error between the predicted value of the single-target neural network model and the RCS conversion value in the single-target test set is not smaller than the set threshold, the test does not reach the standard; the error comprises a mean square error or a mean absolute error;
s602, concentrating the group target test,
Figure FDA0003257272300000032
And the theta value is used as a feature vector and is input into the swarm target neural network model obtained by training in the step S5, a predicted value of the swarm target neural network model is obtained through a forward propagation algorithm, the predicted value and an expected output value, namely an RCS conversion value in the swarm target test set, are compared and analyzed, and whether the swarm target neural network model reaches the standard or not is judged according to the prediction precision; if the test is up to the standard, the step S7 is entered, and if the test is not up to the standard, the step S4 is returned;
the judgment mode of whether the test group target neural network model reaches the standard is as follows: if the error between the predicted value of the group target neural network model and the RCS conversion value in the group target test set is smaller than a set threshold, the test is up to standard, and if the error between the predicted value of the group target neural network model and the RCS conversion value in the group target test set is not smaller than the set threshold, the test is not up to standard; the error comprises a mean square error or a mean absolute error.
CN202110469304.2A 2021-04-28 2021-04-28 Target electromagnetic scattering characteristic rapid prediction method based on deep learning Active CN113177356B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110469304.2A CN113177356B (en) 2021-04-28 2021-04-28 Target electromagnetic scattering characteristic rapid prediction method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110469304.2A CN113177356B (en) 2021-04-28 2021-04-28 Target electromagnetic scattering characteristic rapid prediction method based on deep learning

Publications (2)

Publication Number Publication Date
CN113177356A CN113177356A (en) 2021-07-27
CN113177356B true CN113177356B (en) 2021-10-15

Family

ID=76925220

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110469304.2A Active CN113177356B (en) 2021-04-28 2021-04-28 Target electromagnetic scattering characteristic rapid prediction method based on deep learning

Country Status (1)

Country Link
CN (1) CN113177356B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113722951A (en) * 2021-08-17 2021-11-30 西安中电科西电科大雷达技术协同创新研究院有限公司 Scatterer three-dimensional finite element grid optimization method based on neural network
CN114722690B (en) * 2022-06-09 2022-09-02 华中科技大学 Acoustic super-surface sound field rapid prediction method based on variable reliability neural network
CN115146531A (en) * 2022-06-22 2022-10-04 中国电力科学研究院有限公司 Electromagnetic structure optimization design method based on deep learning
CN115983138A (en) * 2023-01-28 2023-04-18 华工未来通信(江苏)有限公司 Method, system, equipment and medium for predicting electromagnetic response of reflective array antenna unit

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008109859A1 (en) * 2007-03-07 2008-09-12 The Macaleese Companies, Inc. D/B/A Safe Zone Systems Object detection method and apparatus
CN107607942A (en) * 2017-08-31 2018-01-19 北京大学 Based on the large scale electromagnetic scattering of deep learning model and the Forecasting Methodology of back scattering
CN108229404A (en) * 2018-01-09 2018-06-29 东南大学 A kind of radar echo signal target identification method based on deep learning
CN111127246A (en) * 2019-11-21 2020-05-08 国网福建省电力有限公司经济技术研究院 Intelligent prediction method for transmission line engineering cost
CN111428407A (en) * 2020-03-23 2020-07-17 杭州电子科技大学 Electromagnetic scattering calculation method based on deep learning
CN111444601A (en) * 2020-03-23 2020-07-24 杭州电子科技大学 AI learning type electromagnetic scattering calculation method suitable for any incident field

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10732277B2 (en) * 2016-04-29 2020-08-04 The Boeing Company Methods and systems for model based automatic target recognition in SAR data
CN109901159A (en) * 2019-04-04 2019-06-18 哈尔滨理工大学 A kind of metre wave radar target identification method based on deep learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008109859A1 (en) * 2007-03-07 2008-09-12 The Macaleese Companies, Inc. D/B/A Safe Zone Systems Object detection method and apparatus
CN107607942A (en) * 2017-08-31 2018-01-19 北京大学 Based on the large scale electromagnetic scattering of deep learning model and the Forecasting Methodology of back scattering
CN108229404A (en) * 2018-01-09 2018-06-29 东南大学 A kind of radar echo signal target identification method based on deep learning
CN111127246A (en) * 2019-11-21 2020-05-08 国网福建省电力有限公司经济技术研究院 Intelligent prediction method for transmission line engineering cost
CN111428407A (en) * 2020-03-23 2020-07-17 杭州电子科技大学 Electromagnetic scattering calculation method based on deep learning
CN111444601A (en) * 2020-03-23 2020-07-24 杭州电子科技大学 AI learning type electromagnetic scattering calculation method suitable for any incident field

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于深度学习的复杂目标RCS估计;郑星权;《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》;20200615;I136-613 *

Also Published As

Publication number Publication date
CN113177356A (en) 2021-07-27

Similar Documents

Publication Publication Date Title
CN113177356B (en) Target electromagnetic scattering characteristic rapid prediction method based on deep learning
CN107798383B (en) Improved positioning method of nuclear extreme learning machine
CN106709216B (en) Microphone array optimization design method considering acoustic propagation correlation loss
CN110213003B (en) Wireless channel large-scale fading modeling method and device
CN107832789B (en) Feature weighting K nearest neighbor fault diagnosis method based on average influence value data transformation
CN112367130B (en) Wireless network signal transmission intensity calculation method and computer storage medium
CN111983619A (en) Underwater acoustic target forward scattering acoustic disturbance positioning method based on transfer learning
CN114996658B (en) Projection-based hypersonic aircraft aerodynamic heat prediction method
CN112149883A (en) Photovoltaic power prediction method based on FWA-BP neural network
CN109783960B (en) Direction-of-arrival estimation method based on grid part refinement
CN111967184B (en) Multi-target antenna design method based on sequence model
Xi et al. Air combat maneuver trajectory prediction model of target based on chaotic theory and IGA-VNN
CN113993205A (en) UWB positioning system and method based on digital twinning
Choi et al. Information-maximizing adaptive design of experiments for wind tunnel testing
CN111859241A (en) Unsupervised sound source orientation method based on sound transfer function learning
CN114814776B (en) PD radar target detection method based on graph attention network and transfer learning
Khurjekar et al. Uncertainty aware deep neural network for multistatic localization with application to ultrasonic structural health monitoring
CN116306199A (en) Efficient optimization method based on multivariate vector control cross eye interference technology
CN115567131A (en) 6G wireless channel characteristic extraction method based on dimensionality reduction complex convolution network
CN111610514B (en) Inversion method and device for propagation characteristics of evaporation waveguide
CN113722951A (en) Scatterer three-dimensional finite element grid optimization method based on neural network
CN113919221A (en) Fan load prediction and analysis method and device based on BP neural network and storage medium
CN111965601A (en) Underwater sound source passive positioning method based on nuclear extreme learning machine
CN116595883B (en) Real-time online system state correction method for numerical reactor
CN113820700B (en) Distributed radar target positioning method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant