A fuze target matching and burst height adaptive control method

Abstract

This invention discloses a fuze target matching and adaptive detonation height control method, belonging to the field of fuze electronic countermeasures technology. The implementation method of this invention is as follows: The relative velocity is obtained based on detection information, and then the optimal activation delay τ is calculated based on the target's two-dimensional image. opt Optimal startup delay τ opt The goal is to minimize the total detonation deviation of each surface element of the target. Combining the near-field two-dimensional imaging results of the target with the fuze's target, the target centroid is estimated by weighting the grayscale of the near-field extended target's two-dimensional image, and the detonation height is designed based on the target centroid. A convolutional neural network for target matching and recognition is constructed. Based on the probability of target classification output by the convolutional network, a fuze detonation height decision risk factor is designed, and a minimum-risk detonation height decision model is established. The optimal detonation height is adaptively selected according to the minimum-risk principle, thereby achieving adaptive detonation height control for known or unknown targets in complex environments and effectively improving the fuze's accuracy in estimating the detonation height of near-field extended targets.

Description

TECHNICAL FIELD

[0001] The present application relates to a fuze target matching and burst height adaptive control method, belonging to the technical field of fuze electronic countermeasures. BACKGROUND

[0002] The ground attack long-range guided rocket can attack many types of ground targets, and can destroy long-range command and control centers, and can clear armored vehicles, missile sites, communication hubs, radar stations, enemy airfields, port terminals, artillery sites and other targets. However, the best burst height setting is not consistent for different attack objects, the existing burst height control method lacks the use of target information, and the pre-burst height preset mode lacks adaptability to different combat scenes and target types. In the complex and changeable battlefield situation, the burst height control precision is often insufficient and the adaptability is poor, and there is currently no burst height control method suitable for two-dimensional imaging detection fuze. SUMMARY

[0003] The purpose of the present application is to provide a fuze target matching and burst height adaptive control method, which combines the near-field two-dimensional imaging results of the target, estimates the target centroid by weighting the near-field extended target two-dimensional image gray scale, designs the burst height based on the target centroid, designs the fuze burst height decision risk factor according to the target attribution probability of each class output by the convolution network, establishes a fuze minimum risk burst height decision model, and adaptively selects the best burst height according to the minimum risk principle, so as to realize adaptive burst height control of known targets or unknown targets in complex environments, and effectively improve the burst height estimation precision of the fuze on the near-field extended target.

[0004] The purpose of the present application is achieved by the following technical solutions:

[0005] The fuze target matching and burst height adaptive control method disclosed by the present application comprises the following steps:

[0006] Step 1: Precise estimation of burst height based on target two-dimensional image gray scale weighting.

[0007] According to the detection information, the relative speed is obtained, and then the target two-dimensional image is used to calculate the best start delay τ opt , as shown in Figure 1 , the best start delay τ opt should minimize the total burst point deviation of each face element of the target;

[0008] The least square method of gray scale weighting is used to estimate the best start delay τ opt .

[0009]

[0010] Where A is a geometric XY image with M×N pixels; I(x,y) is the gray value of pixel (x,y), where x and y are the row and column index symbols of the pixel in the geometric image, and the gray value is proportional to the target surface area; For the optimal startup delay estimate, τ opt That is, to make equation (1) reach its minimum value The startup delay for pixel (x,y) is Where v r The relative velocity between the fuse and the target; The calculated angle of the fuse is:

[0011]

[0012] From equation (2), we get, And (x0, y0) is the centroid of the image. That is, the explosion delay estimation based on the centroid of a two-dimensional image is equivalent to the least squares start delay estimation based on the gray-level weighting of a two-dimensional image;

[0013] For a fuze with defined parameters, its distance resolution ρ r and Doppler resolution ρ fd It is also certain; the time-frequency image B = [b mn ] M×N Each pixel b mn =((m-1)ρ r ,(n-1)ρ fd ), where m and n are the row and column index symbols of the time-frequency image;

[0014] Based on wavelength λ and relative velocity v between the projectile and the target r The geometric image and the time-frequency image B can be converted, so according to And B, calculate each pixel b mn The startup delay matrix τ = [τ mn ] M×N The startup delay of each pixel

[0015] Before the missile and target meet, the fuze obtains the values ​​of each pixel b within the imaging area based on the guidance information. mn The start-up delay matrix τ, the real-time detection of the fuse to obtain the time-frequency image B, let position b mn The pixel grayscale value is i mn The high delay τ of the explosion is obtained. opt for:

[0016]

[0017] Gray value i below the fuze detection threshold mn It should be set to 0;

[0018] Step two: target matching recognition based on convolutional neural network.

[0019] Convolutional neural network (CNN) is a variant of multi-layer perceptron (MLP), which adopts local connection and weight sharing to reduce the number of weights, making the network easy to optimize, and reducing the complexity of the model and the risk of overfitting, and is widely used in image recognition field.

[0020] The convolutional neural network is composed of an input layer, a hidden layer and an output layer;

[0021] First part: input layer, the input layer of convolutional neural network can process multi-dimensional data, in the target matching recognition based on convolutional neural network proposed in the application, the input data is three-dimensional, that is, two-dimensional pixel points on the plane and RGB channels. Since gradient descent method is used for learning, the input features of convolutional neural network need to be standardized, specifically, before the pixel data is input into the convolutional neural network, it needs to be normalized to the interval [0, 1], which is conducive to improving the learning efficiency and performance of the convolutional neural network.

[0022] Second part: hidden layer, the hidden layer of convolutional neural network includes three common structures of convolutional layer, pooling layer and fully connected layer. The function of convolutional layer is to extract features from input data, which contains multiple convolutional kernels inside. Each element of convolutional kernel corresponds to a weight coefficient and a bias. Each neuron in the convolutional layer is connected to multiple neurons in the previous layer with close position, and the size of the region depends on the size of the convolutional kernel, which is also called receptive field; The parameters of convolutional layer include convolutional kernel size, step and padding, which together determine the size of convolutional layer output feature map; In the convolutional layer, there is also an activation function to help express complex features, usually using linear rectifier function ReLU (Rectified Linear Unit) and its variant functions, the expression of ReLU function is as follows:

[0023] f(u)=max(0,u) (4)

[0024] Where u represents the input vector, and ReLU is a non-saturated activation function, which can effectively solve the problems of gradient explosion and gradient disappearance; and compared with other complex activation functions, the expression of ReLU is simple, which simplifies the calculation process and reduces the overall calculation cost of neural network.

[0025] In the convolutional neural network, a pooling layer is usually periodically inserted between successive convolutional layers, which functions to select and filter the features extracted by the convolutional layer, gradually reduce the spatial dimensions of the data volume, reduce the number of parameters in the network, reduce the computing resources, and effectively control overfitting. In addition, the fully connected layer in the convolutional neural network is equivalent to the hidden layer in the traditional feedforward neural network, which is located at the last part of the hidden layer of the convolutional neural network and only transmits signals to other fully connected layers.

[0026] The third part is the output layer, and the upstream of the output layer in the convolutional neural network is usually a fully connected layer. For image classification problems, the output layer uses a logistic function or a normalized exponential function (softmax function) to output a classification label.

[0027] The target matching recognition network based on the convolutional neural network comprises a plurality of convolutional layers and a plurality of fully connected layers, and the convolutional layers, nonlinear activation functions and pooling layers are used in combination. The sampling window of the convolutional layer and the pooling layer is 3*3 or 2*2, and the pooling mode is maximum pooling. Each depth slice is sampled with a step of 2, more image texture information is retained, and the influence of useless information is reduced.

[0028] The flowchart of using the convolutional neural network to perform target matching recognition on the fuse target near-field two-dimensional imaging graph is shown in Figure 2 The specific steps are as follows:

[0029] Step 2.1: The fuse target near-field two-dimensional imaging data is cropped, scaled and rotated, the image data is divided into a training set and a test set, and the distance coordinates, azimuth coordinates, true class of the target and the size of the bounding box are taken as labels;

[0030] Step 2.2: The convolutional neural network extracts features in the image data and classifies the image;

[0031] Step 2.3: The weights of the convolutional neural network are trained using the training set data;

[0032] Step 2.4: The test set data is input into the trained neural network, and the class of the target and the corresponding probability are output.

[0033] Step three: fuse height adaptive control based on optimal risk decision.

[0034] The fuse height adaptive control needs to be based on the target matching recognition result, adopts the fuse height adaptive control method based on the optimal risk decision, establishes a fuse minimum risk height decision model according to the target attribution probabilities output by the convolutional network, and adaptively selects the best height according to the minimum risk principle, so as to realize the adaptive height control of the known target or unknown target in the complex environment.

[0035] p(c i |y) is the probability of the current target belonging to each target type. Assuming that the fuze has three grades of burst height a, b, c, and the identification result of the target belonging to each type is: p(c1|y), p(c2|y), p(c N |y), then the burst height calculation result is:

[0036]

[0037] In the formula, f(a, b, c, y) represents the burst height calculated according to the current target identification result, represents the probability sum of belonging to the a-type burst height, represents the probability sum of belonging to the b-type burst height, represents the probability sum of belonging to the c-type burst height.

[0038] In order to obtain the current best burst height selection of the fuze, it is necessary to evaluate the burst height selection of the fuze, and through the design of the burst height decision risk and the risk factor, the optimal burst height selection that minimizes the burst height decision risk is obtained. The burst height decision risk of the fuze can be represented as:

[0039]

[0040] In the formula, R a (y), R b (y), and R c (y) are the risks of judging the current test target signal burst height as a, b, and c-type burst height respectively, is the burst height decision risk factor, wherein, represents the loss of misjudging the burst height belonging to b and c as a, is the loss of misjudging the burst height belonging to a and c as b, represents the loss of misjudging the burst height belonging to a and b as c.

[0041] Then the adaptive burst height category matched by the current target identification result is:

[0042]

[0043] In order to obtain the risk factor, the relationship between the kill probability and the burst height needs to be fully considered. The target conditional coordinate kill probability G0(y, z) describes the relationship between the guidance error (y, z) and the warhead kill probability. Under the given warhead, fuze and target characteristics, the target conditional coordinate kill law G0(y, z) is the burst height error Also often related to the azimuth angle η function. When the calculation of G0(y, z) encounter difficulties, can be obtained by theoretical analysis plus experimental data semi-empirical formula G0(y, z) approximation to determine. With the increase of r, the fragment distribution of the surface density and the impact of the target speed will drop, so the probability of damage of ammunition weapon system is also small. Obviously, the semi-empirical formula for determining G0(y, z) can be expressed as:

[0044]

[0045] In the formula, δ0(η) is a comprehensive parameter related to η.

[0046] Target condition coordinate kill probability G0(r, η) mainly depends on the size of the height error r, and the relationship with the azimuth angle η is not obvious. Therefore, in the calculation of the kill probability of warhead, the target of the circular condition coordinate kill law is replaced by the target of the two-dimensional condition coordinate kill law. The circular condition coordinate kill probability of the target can be expressed as:

[0047]

[0048] The warhead kill probability is inversely proportional to the height error, the greater the height error, the worse the damage effect, so the height decision risk factor can be expressed as:

[0049]

[0050] In the formula, K is a constant related to G0(r, η), and the fuse height decision risk is:

[0051]

[0052] Then according to the optimal risk decision principle, the fuse self-adaptive height estimation result is

[0053]

[0054] In order to improve the fuse height control precision and damage efficiency, the fuse needs to measure the height information converted into the relative height of the warhead centroid and the ground surface target centroid, and the measured data is converted to the warhead centroid position, so as to eliminate the height error, therefore, the inertial coordinate system o-xyz, o-xy plane is the ground plane, z axis is vertical to the ground vertically upward, O wh is the centroid of warhead, H wh is the relative height of warhead centroid and ground, H d is the relative height of fuse and ground, θ m is the angle of fall, |O d O wh | represents the relative distance between the fuse and the warhead centroid, and H wh and Hd The relationship is shown in the following formula:

[0055]

[0056] The target burst height matching control accuracy is only affected by the fuze target identification result and the risk function, and by optimizing the target identification result and the burst height decision risk, the target burst height matching control accuracy is not less than 90%.

[0057] According to the best start delay τ of step 1 opt The fuze will delay τ wh Time detonation when the centroid of the warhead reaches the estimated height H opt , so that the warhead is closest to the target centroid position, the total burst point deviation of each surface element of the target is minimized, and the near-burst damage effect is optimized, that is, the fuze burst height adaptive control based on the optimal risk decision.

[0058] Beneficial effects:

[0059] 1. The fuze target matching and burst height adaptive control method disclosed by the application estimates the target centroid by two-dimensional image gray weight of the near-field extended target, compared with the method of spectrum estimation and the burst height estimation principle based on Doppler, the resolution of the distance dimension is increased, and the measurement error is reduced, and the method is more superior in the case that the target data frame is discontinuous in actual combat.

[0060] 2. The fuze target matching and burst height adaptive control method disclosed by the application combines the probability of the target belonging to each type output by the convolution network, and adaptively selects the best burst height according to the minimum risk principle, thereby improving the burst height estimation precision of the fuze on the near-field extended target.

[0061] 3. The fuze target matching and burst height adaptive control method disclosed by the application, based on the beneficial effects 1 and 2, the fuze burst height adaptive control method based on target matching and minimum risk decision, combines the target identification probability result output by the convolution network, designs a fuze burst height decision risk factor, constructs a fuze minimum risk burst height decision model, obtains the optimal burst height selection of the current target, thereby completing the adaptive burst height control of the fuze on different targets, realizes the three-grade burst height matching of 20m / 25m / 30m on typical targets such as armored vehicles, radar stations and artillery positions, and the control precision is not less than 0.5m, and the target adaptability and damage control capability of the medium and large caliber rocket bomb radio fuze can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 It is a schematic diagram of the burst height estimation based on two-dimensional image gray weight;

[0063] Figure 2 It is a target matching and identification flowchart disclosed by the application;

[0064] Figure 3 This is a flowchart of the adaptive control of fuse detonation height based on the publicly disclosed optimal risk decision of the present invention;

[0065] Figure 4 A vector model for the deviation between the fuze detonation height and the warhead's center of mass;

[0066] Figure 5 The simulation results are for bomb height delay estimation based on two-dimensional image grayscale weighting; among them, Figure 5 (a) is the spectrum diagram of a Doppler fuze; Figure 5 (b) is the time-frequency image of the imaging fuze;

[0067] Figure 6 The figure shows the statistical results of the estimation error of the fuse detonation height by different algorithms. Detailed Implementation

[0068] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0069] Example 1:

[0070] like Figure 3 As shown in the figure, this example discloses a fuse target matching and detonation height adaptive control method, and the specific implementation steps are as follows:

[0071] Step 1: Accurate estimation of the bomb height based on grayscale weighting of the target's two-dimensional image;

[0072] The fuze requires high real-time performance, and real-time identification of vulnerable target areas is currently challenging. Damage based on target geometry is more suitable for practical applications. The relative velocity v can be obtained from the detection information. r Given this premise, the optimal delay time τ of the target can be calculated based on the two-dimensional image of the target. opt ,like Figure 1 As shown, the optimal startup delay τ opt The goal should be to minimize the total deviation of the explosion points of all surface sources on the target.

[0073] Two-dimensional imaging is the circular projection of a three-dimensional target onto the imaging plane. For a target with uniform surface reflectivity, the grayscale value of the target image represents the concentration of the target's surface area. Therefore, bright spots in the target image should have high weights, and vice versa. A grayscale-weighted least squares method is used to estimate the startup delay τ. opt :

[0074]

[0075] where A is a geometric X-Y image with MxN pixels; I(x,y) is the gray value of pixel (x,y), x and y are the row index and column index of pixel in geometric image, and the gray value is proportional to the target surface area; For the estimation of the optimal initiation delay, τ opt is the gray value of pixel b fd , and x0 and y0 are the row index and column index of the image centroid. The initiation delay of pixel (x,y) is where v r is the relative velocity between fuze and target. The impact angle of fuze is calculated as

[0076]

[0077] From equation (15), we have where (x0,y0) is the centroid of image, It is proved that the least square initiation delay estimation based on the gray weighting of two-dimensional image is equivalent to the impact delay estimation based on the centroid of two-dimensional image.

[0078] For fuze with fixed parameters, the range resolution ρ r and the Doppler resolution ρ fd are fixed. The time-frequency image matrix can be expressed as B=[b mn ] M×N where each pixel b mn = ((m-1)ρ r ,(n-1)ρ fd ), where m and n are the row index and column index of time-frequency image.

[0079] According to the wavelength λ and the relative velocity v r between projectile and target, the time-frequency image B can be converted to geometric image, so according to B and , the initiation delay matrix τ=[τ mn ] M×N of each pixel b mn is calculated, where

[0080] Before the encounter between projectile and target, the initiation delay matrix τ of each pixel b mn in the imaging area can be obtained by fuze according to the guidance information. The two-dimensional image B of target is obtained by fuze in real time. Assuming that the gray value of pixel b mn is i mn , the impact delay τ opt can be obtained as:

[0081]

[0082] The gray value i mn lower than the detection threshold of fuze should be set to 0.

[0083] Figure 5 For the simulation results of the proposed two-dimensional image gray-weighted burst height delay estimation, 256 intersection simulations were conducted. In each intersection process, the fuze obtained 20 frames of signal processing results. Figure 5 For the 23rd intersection, the fuze 11th frame processing result. Finally, the mean square deviation of the burst height estimation deviation from the true value in the 256 intersection processes was statistically analyzed. Figure 6 For the error statistical results of the fuze burst height estimation using Doppler two-point method, Doppler least square method, and two-dimensional image centroid method, as shown in the figure, the error of the fuze “Doppler two-point method” start delay is large, and the least square method using Doppler centroid can only reduce the error by relying on sufficient data frame number. Even so, its error is still significantly larger than that of the two-dimensional imaging fuze. When the fuze uses spectral estimation method with measurement error, the Doppler-based burst height estimation principle will “amplify” the error under short measurement baseline. The two-dimensional imaging of the fuze increases the resolution in the distance dimension, and only one frame of data can achieve a small burst height estimation error. This is more superior in the case of discontinuous target data frames in actual application. Through optimization of the centroid estimation result, the final burst height control accuracy of not less than 0.5m is realized.

[0084] Step two: target matching recognition based on convolutional neural network.

[0085] Step 2.1: crop, scale and rotate the fuze target near-field two-dimensional imaging data, divide the image data into training set and test set two parts, and take the distance coordinate, azimuth coordinate, true class of the target and the size of the bounding box as the label;

[0086] Step 2.2: convolutional neural network extracts features from image data and classifies images;

[0087] Step 2.3: use the training set data to train the weights of the convolutional neural network;

[0088] Step 2.4: input the test set data into the trained neural network, and output the class and corresponding probability of the target.

[0089] Step three: fuze burst height self-adaptive control based on optimal risk decision.

[0090] The output of step two is the classification of the target (armored vehicle, radar station, artillery position) and the probability of belonging to this category. For typical targets such as armored vehicles, radar stations, and artillery positions, the fuze can set three burst heights a, b, and c, where a = 20, b = 25, and c = 30. The identification results of the target belonging to each category are: p(c1|y), p(c2|y), and p(c3|y). The burst height calculation result is: N |y).

[0091]

[0092] where f(a, b, c, y) denotes the current target recognition result, P(a) denotes the probability sum of a class, P(b) denotes the probability sum of b class, P(c) denotes the probability sum of c class.

[0093] As shown in Figure 3 To get the optimal fuse height selection, the fuse height selection needs to be evaluated, and the optimal height selection that minimizes the height decision risk is obtained by designing the height decision risk and risk factor. The fuse height decision risk can be expressed as:

[0094]

[0095] where R a (y), R b (y), R c (y) are the risks of judging the current test target signal height as a, b, and c class height respectively, is the height decision risk factor, where, L is the loss of misjudging the height belonging to b and c class as a class, is the loss of misjudging the height belonging to a and c class as b class,

[0096] The adaptive height class matched by the current target recognition result is:

[0097]

[0098] To obtain the risk factor, the relationship between the kill probability and the height needs to be fully considered. The target conditional coordinate kill probability G0(y, z) describes the relationship between the guidance error (y, z) and the warhead kill probability. Under the given warhead, fuse and target characteristics, the target conditional coordinate kill law G0(y, z) is the height error It is also often related to the azimuth angle η. When calculating G0(y, z) is difficult, G0(y, z) can be determined by semi-empirical formula obtained from theoretical analysis plus experimental data. As r increases, the surface density of the fragment distribution and the speed of hitting the target will decrease, and thus the kill probability of the ammunition weapon system will also decrease. Obviously, the semi-empirical formula for determining G0(y, z) can be expressed as:

[0099]

[0100] wherein δ0(η) is a comprehensive parameter related to η.

[0101] Generally, the target condition coordinate kill probability G0(r,η) mainly depends on the size of the burst height error r, and has no obvious relationship with the azimuth angle η. Therefore, when calculating the kill probability of the warhead, the target two-dimensional condition coordinate kill law is often replaced by the target circular condition coordinate kill law. The target circular condition coordinate kill probability can be expressed as:

[0102]

[0103] The warhead kill probability is inversely proportional to the burst height error. The greater the burst height error, the worse the damage effect, and therefore the burst height decision risk factor can be expressed as:

[0104]

[0105] wherein K is a constant related to G0(r,η). From the above formula, the fuze burst height decision risk is

[0106]

[0107] Therefore, the fuze self-adaptive burst height estimation result obtained according to the optimal risk decision principle is

[0108]

[0109] It can be seen that the target burst height matching control accuracy is only affected by the fuze target recognition result and the risk function. Through optimization of the target recognition result and the burst height decision risk, the target burst height matching control accuracy can be realized to be not less than 90%:

[0110] The burst height information measured by the fuze needs to be converted into the relative height of the warhead centroid and the ground target centroid. The measured data is converted to the warhead centroid position, so as to eliminate the burst height error, as shown in Figure 4 The inertial coordinate system o-xyz, o-xy plane is the ground plane, and the z-axis is perpendicular to the ground and vertically upward. Figure 4 wherein O wh is the warhead centroid, H wh is the relative height of the warhead centroid and the ground, H d is the relative height of the fuze and the ground, θ m is the projectile fall angle, |O d O wh | represents the relative distance between the fuze and the warhead centroid, and then the relationship between H wh and H d is shown in the following formula:

[0111] H wh = H d + |O d Owh |sinθ m (25)

[0112] According to the best start delay τ opt , the fuze will be delayed τ wh time when the warhead centroid reaches the estimated height H opt , and the warhead is closest to the target centroid position, the total point deviation of each surface element of the target reaches the minimum, and the near-blast damage effect reaches the best.

[0113] The above specific description further details the purpose, technical solutions and beneficial effects of the application. It should be understood that the above description is only a specific embodiment of the application and is not intended to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A fuze target-matching and burst height adaptive control method, characterized in that: The method comprises the following steps, Step one: accurate estimation of burst height based on target two-dimensional image gray weight; According to the detection information, the relative speed is obtained, and then according to the target two-dimensional image, the optimal starting delay τ is calculated opt , the optimal starting delay τ opt , so that the total burst deviation of each surface element of the target is minimized A gray-weighted least squares method to estimate the optimal start-up delay τ opt ; where A is a geometric X-Y image with MxN pixels; I(x, y) is the gray value of pixel (x, y), x, y are the row index and column index symbols of the pixel in the geometric image, and the gray value is proportional to the target surface area; for the estimation of the optimal start delay, τ opt that is, to make formula (1) reach the minimum time The start delay of pixel (x, y) is where v r is the relative motion speed between the fuze and the target; is the impact angle of the fuze, which is calculated as from equation (2), where (x0, y0) is the centroid of the image, The centroid-based burst height delay estimation is equivalent to the least square start delay estimation based on the gray-weighted two-dimensional image. For a fuze determining parameters, its range resolution p r and Doppler resolution p fd are also determined; the time-frequency image B = [b mn ] M×N , each pixel b mn = ((m-1) p r , (n-1) p fd ), where m, n are the row index and column index symbols of the time-frequency image; According to the wavelength λ and the relative motion speed v r , the conversion between the geometric image and the time-frequency image B can be realized, so according to and B, the starting delay matrix τ = [τ mn ] of each pixel b mn is calculated M×N , wherein the starting delay τ Before the missile-target encounter, the fuze gets the start delay matrix τ of each pixel b in the imaging area according to the guidance information, and the fuze detects the time-frequency image B in real time, and supposes that the pixel gray value of position b is i, and the obtained burst height delay τ is: mn mn mn opt ​​​​ Gray value i below the fuze detection threshold mn Set to 0; Step two: target matching identification based on convolutional neural network; The convolutional neural network is composed of a plurality of convolutional layers and a plurality of fully connected layers, and the convolutional layers, nonlinear activation functions and pooling layers are combined; wherein the sampling window of the convolutional layer and the pooling layer is 3*3 or 2*2, the pooling mode is maximum pooling, and each depth slice is down-sampled; Based on the convolutional neural network, the target matching identification of the fuze target near-field two-dimensional imaging graph is carried out, and the probability of each type of target attribution is obtained; Step three: fuze burst height adaptive control based on optimal risk decision; The fuze burst height adaptive control needs to be based on the target matching identification result, adopt the burst height adaptive control method based on the optimal risk decision, establish a fuze minimum risk burst height decision model according to the target attribution probability output by the convolutional network, and adaptively select the best burst height according to the minimum risk principle, so as to realize the adaptive burst height control of the known target or unknown target in the complex environment.

2. The fuze target matching and burst height adaptive control method according to claim 1, characterized in that: The convolutional neural network of step two is composed of an input layer, a hidden layer and an output layer; The first part: the input layer of the convolutional neural network can process multi-dimensional data; the input data is three-dimensional, that is, two-dimensional pixel points on the plane and RGB channels; since the gradient descent method is used for learning, the input features of the convolutional neural network need to be standardized, specifically, before the pixel data is input into the convolutional neural network, it needs to be normalized to the interval [0, 1]; The second part: the hidden layer of the convolutional neural network includes convolutional layers, pooling layers and fully connected layers; the function of the convolutional layer is to extract features from the input data, which contains a plurality of convolutional kernels inside, each element of the convolutional kernel corresponds to a weight coefficient and a bias; each neuron in the convolutional layer is connected with a plurality of neurons in the region close to the position in the previous layer, and the size of the region depends on the size of the convolutional kernel, and the region is also called a receptive field; the parameters of the convolutional layer include the size of the convolutional kernel, the step and the padding, which jointly determine the size of the convolutional layer output feature map; in the convolutional layer, an activation function is also included, a linear rectifier function and its variant function are used, and the expression of the ReLU function is as follows: f(u)=max(0,u) (4) Wherein u represents an input vector, and the ReLU function is a non-saturated activation function; A pooling layer is periodically inserted between consecutive convolutional layers, which functions to select and filter the features extracted by the convolutional layer, gradually reduce the spatial size of the data body, and reduce the number of parameters in the network; in addition, the fully connected layer in the convolutional neural network is equivalent to the hidden layer in the feedforward neural network, which is located at the last part of the hidden layer of the convolutional neural network and only transmits signals to other fully connected layers; The third part: the output layer of the convolutional neural network, the upstream of the output layer is a fully connected layer, and for image classification problems, the output layer uses a logistic function or a normalized exponential function to output a classification label.

3. The fuze target matching and burst height adaptive control method according to claim 2, characterized in that: The specific steps of using the convolutional neural network of step two to perform target matching identification on the fuze target near-field two-dimensional imaging graph are as follows: Step 2.1: The fuze target near-field two-dimensional image data is cropped, scaled and rotated, the image data is divided into two parts of training set and test set, and the distance coordinate, azimuth coordinate, real class of the target and the size of the bounding box are taken as labels; Step 2.2: The convolutional neural network extracts features in the image data and classifies the image; Step 2.3: The weights of the convolutional neural network are trained using the training set data; Step 2.4: The test set data is input into the trained neural network, and the class and corresponding probability of the target are output.

4. The method of fuze target matching and burst height adaptive control according to claim 3, characterized in that: In step three, p(c i |y) is the probability of the current target belonging to each target type; assuming that the fuze has three grades of burst height a, b, c, and the identification result of the target belonging to each type is: p(c1|y), p(c2|y), p(c N |y), then the burst height calculation result is: In the formula, f(a, b, c, y) represents the burst height calculated according to the current target recognition result, represents the probability sum of the a-type burst height, represents the probability sum of the b-type burst height, represents the probability sum of the c-type burst height; The optimal burst height selection that minimizes the burst height decision risk is obtained through the burst height decision risk and risk factor design; the fuze burst height decision risk is represented as: wherein R a (y), R b (y), R c (y) are the risks of judging the current test target signal height as a, b, c class signal height respectively, is the signal height decision risk factor, wherein, represents the loss of misjudging the signal height belonging to b and c class as a class signal height, is the loss of misjudging the signal height belonging to a and c class as b class signal height, represents the loss of misjudging the signal height belonging to a and b class as c class signal height; The adaptive burst height class matched by the current target recognition result is: To obtain the risk factor, the relationship between kill probability and burst height must be considered; the conditional target coordinate kill probability G0(y,z) describes the relationship between guidance error (y,z) and warhead kill probability; under given warhead, fuze and target characteristics, the conditional target coordinate kill probability G0(y,z) is a function of burst height error and the azimuth angle η; when it is difficult to calculate G0(y,z), a semi-empirical formula obtained from theoretical analysis and experimental data is used to approximate G0(y,z); as r increases, the area density of the fragment distribution and the speed of the fragments impacting the target decrease, so the kill probability of the munition weapon system decreases; the semi-empirical formula for determining G0(y,z) is given by: In the formula, δ0(η) is a comprehensive parameter related to η; Therefore, when calculating the damage probability of the warhead, the two-dimensional conditional coordinate damage rule of the target is replaced by the circular conditional coordinate damage rule of the target; the damage probability of the target circular conditional coordinate is represented as: The warhead damage probability and burst height error are inversely proportional, the greater the burst height error, the worse the damage effect, and the burst height decision risk factor is represented as: In the formula, K is a constant related to G0(r,η), and the fuze burst height decision risk is obtained from the above formula: Therefore, the fuze adaptive burst height estimation result obtained according to the optimal risk decision principle is The target burst height matching control accuracy is only affected by the fuze target recognition result and the risk function, and the target burst height matching control accuracy is not less than 90% by optimizing the target recognition result and the burst height decision risk; The measured burst height information of the fuze is converted into the relative height of the center of mass of the warhead and the center of mass of the ground target, the measured data is converted to the position of the center of mass of the warhead, so as to eliminate the burst height error; an inertial coordinate system o-xyz is established, the o-xy plane is the ground plane, and the z axis is vertical upward and perpendicular to the ground. wh H is the center of mass of the warhead. wh H is the relative height of the center of mass of the warhead and the ground. d H is the relative height of the fuze and the ground, after obtaining the estimation result of the burst height, there are |O d O wh | indicates the relative distance of the fuze and the center of mass of the warhead, then H wh and H d The relationship is shown in the following formula: According to the best start delay τ of step 1 opt , the fuze will delay τ opt time again when the centroid of the warhead reaches the estimated height H wh , and the warhead is closest to the centroid position of the target, the total deviation of the warhead is minimized, that is, the fuze height adaptive control based on the optimal risk decision.

5. The fuze target matching and burst height adaptive control method according to claim 3 or 4, characterized in that: In step two, each depth slice is down-sampled with a step of 2.

Images