A formation recognition target selection method based on deep learning

Abstract

The present invention discloses a formation recognition target selection method based on deep learning, comprising the following steps: 1) obtaining surface target formation information and generating a formation formation diagram; 2) constructing a formation formation recognition model based on an improved LeNet convolutional neural network; 3) inputting the formation formation diagram into the formation formation recognition model to obtain a formation recognition result; 4) sorting and numbering formation targets in the target indication phase according to the formation recognition result; 5) selecting a target to be tracked, and transmitting the target number to be tracked to an aircraft; 6) during the flight phase, the aircraft obtains the surface target formation situation through a terminal guidance sensor, sorts and numbers the target formation, then matches the target to be tracked, and tracks the target to be tracked. Based on the improved LeNet network, Hough transform and ISODATAˋ algorithm, the present invention proposes a surface target formation target selection modeling method to improve selection efficiency and enhance engineering operability.

Description

Technical Field

[0001] The present invention relates to the field of formation recognition, and in particular to a formation recognition target selection method based on deep learning. Background Art

[0002] There are usually two types of technologies for surface target selection of guided unmanned aerial vehicles: feature recognition and formation recognition. Feature recognition relies on a single feature that distinguishes the designated target from other members in the formation, or a comprehensive feature of multiple weighted features. However, as the range of the aircraft increases, the self-control time increases, and the change in the terminal formation situation no longer meets the characteristics determined before the launch, thus exceeding the ability of the aircraft's feature recognition. When there are members of the same type in the formation, it is easy to have false hits. The original feature recognition cannot meet the requirements of multi-target recognition. Formation recognition is the process of comparing the formation situation captured by the terminal guidance sensor of the aircraft with the formation situation on the aircraft (real-time detection in the target indication phase or pre-stored in the intelligence database), matching the similarity of the formation characteristics of the two situations, and sorting and identifying the predetermined targets of the formation for precise strikes.

[0003] The formation recognition target selection technology first lies in the formation formation recognition. Formation formation recognition is a type of cluster recognition, an important development direction of terminal guidance technology, and the basis and key to improving the target situation recognition ability. Formation formation recognition technology is to identify and analyze the target formation detected by the sensor through certain methods to understand the target's state and possible action intentions, and to allocate firepower and strike the predetermined target based on this information. In order to ensure the needs of the action and the safety of navigation, the surface formation has stability and regularity, which provides the main basis for formation recognition.

[0004] At present, the main methods for formation recognition include template matching, mathematical analysis, pattern recognition based on domain knowledge, neural networks, etc. With the development of artificial intelligence technology, the method of using deep learning to recognize formations has received attention. These methods have the following defects: 1) When the formation is misplaced and disordered, the formation cannot be accurately recognized. 2) When the V-shaped formation and the circular formation have large noise, the distinction is not high, and the queue line must be initially divided when selecting the target, which requires manual intervention.

[0005] Existing formation recognition technologies often focus on formation recognition rather than target selection, and have defects such as not fully considering the situation changes at the end of the formation and poor real-time performance. In addition, the selection of surface formation targets by guided unmanned aerial vehicles is mostly based on infrared or electromagnetic characteristics, and it is easy to make false hits when facing the same type of targets, making it difficult to achieve the effect of destroying the intended target. Summary of the invention

[0006] The purpose of the present invention is to provide a formation recognition target selection method based on deep learning, comprising the following steps:

[0007] 1) Obtain surface target formation information and generate a formation diagram;

[0008] 2) Construct a formation recognition model based on the improved LeNet convolutional neural network;

[0009] 3) Inputting the formation diagram into the formation recognition model to obtain the formation recognition result;

[0010] 4) Sort and number the formation targets in the target indication phase according to the formation recognition results;

[0011] 5) Select the target to be tracked and transmit the target number to the aircraft;

[0012] 6) In the terminal guidance stage, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, then matches the target to be tracked, and tracks the target to be tracked.

[0013] Furthermore, the formation recognition model includes an input layer, multiple convolutional layers, multiple batch normalization layers, multiple ReLU layers, multiple maximum pooling layers, multiple fully connected layers, a Softmax layer and a classification layer;

[0014] The activation function used by the formation recognition model is as follows:

[0015]

[0016] In the formula, f(x) represents the activation function; x is the input;

[0017] The optimization process of parameters in the formation recognition model is as follows:

[0018]

[0019] Where: β1 is the first-order moment attenuation coefficient, 0<β1<1; β2 is the second-order moment attenuation coefficient, 0<β2<1; g t is the derivative gradient of the objective function; m t is the first-order moment estimate of the gradient in momentum form at time t; v t is the second-order moment estimate of the gradient in momentum form at time t; is the first-order moment estimate after bias correction; is the bias-corrected second-order moment estimate; is β1 to the power of t; is β2 raised to the power of t; θ t+1 is the updated parameter value; η is the learning rate, and ε is a constant. θt is the parameter value before updating.

[0020] Further, the formation recognition result includes the number of queue lines and the formation outline;

[0021] The number of queue lines includes one queue line and two queue lines;

[0022] When the queue line is one, the formation outline includes single horizontal column, single vertical column and azimuth column;

[0023] When there are two queue lines, the formation outlines include double horizontal queue, double vertical queue, V-shaped queue and circular queue.

[0024] Furthermore, when the formation outline is a single horizontal line, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0025]

[0026] Where i = 1, 2, 3, ..., T; T is the number of members; D x , D z is the horizontal spacing and vertical spacing; N(0,σ1 2 )、N(0,σ2 2 ) represents the distribution; σ1 2 ,σ2 2 represents the mean square error of position and heading; γ represents the heading angle of the formation;

[0027] When the formation outline is a double horizontal line, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0028]

[0029] When the formation outline is a single column, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0030]

[0031] When the formation outline is a double column, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0032]

[0033] When the formation outline is a azimuth formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0034]

[0035] When the formation outline is a V-shaped formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0036]

[0037] In the formula, N(γ,σ2 2 ) represents the distribution; β represents the queue angle;

[0038] When the formation outline is a ring formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0039]

[0040] Furthermore, considering the disturbance change of heading around the echo center, the heading vector coordinates of the formation members are as follows:

[0041]

[0042] Where, X i ″、Z i ″ is the heading vector coordinate of the formation member considering the disturbance change of heading around the echo center; X i0 , Z i0 is the intermediate parameter; X i , Z i is the navigation vector coordinate of the formation member before updating; (x i ,z i ) is the coordinate of each member in the measurement space; r i is the heading; γ is the formation heading angle.

[0043] Furthermore, during the autonomous flight phase, if the formation turns around a certain rotation point, the coordinates of the target formation members (X i ′,Z i ′) is as follows:

[0044]

[0045] Where (X0, Z0) is the coordinate of the rotation center point; q is the counterclockwise rotation angle.

[0046] During the autonomous flight phase, if the target formation spacing changes, the coordinates of the target formation members (X n ′,Z n ′) is as follows:

[0047]

[0048] In the formula, (X n+1 ′,Z n+1 ′) is the change coordinate;

[0049] During the autonomous flight phase, if the target formation spacing changes, the target formation spacing (D h ′,Z z ′) is as follows:

[0050]

[0051] Where, when k1, k2∈(0,1), it is the ratio of the team spacing reduction; when k1, k2∈(1,∞), it is the ratio of the team spacing enlargement; (D h , D z ) is the target formation interval before the interval changes.

[0052] Furthermore, when there is one queue line, the step of sorting and numbering the formation targets in the target indication phase includes:

[0053] A straight line is used to fit the measured spatial coordinates of the formation members in the target indication phase, and this straight line is the queue line;

[0054] Then, the reference members are obtained by combining the heading information, and the formation targets are sorted and numbered according to the relative positions of other members other than the reference members and the reference members, with the reference members as reference;

[0055] In the terminal guidance phase, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and matches the target to be tracked in the following steps:

[0056] The aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and then selects the target with the same number as the target to be tracked from the numbered target formation.

[0057] Furthermore, when there are two queue lines, the step of sorting and numbering the formation targets in the target indication phase includes:

[0058] 1) Establish standard template situation formation in real time according to the formation recognition results;

[0059] 2) Mapping the coordinates of the target situation to be tracked and the template situation formation members into the parameter space by Hough transformation;

[0060] 3) In the parameter space, the coordinates are clustered using the ISODATA algorithm to obtain error points;

[0061] 4) In the parameter space, according to the principle of Hough transform, the change of the error point from the template situation to the target indication situation is used to offset the change of the ideal point, and the ideal point in the target indication situation parameter space is found, thereby eliminating the error of the queue line in the target indication situation;

[0062] 5) Inverse Hough transform the ideal point in the target indication situation parameter space to the measurement space to obtain the benchmark members and queue lines of the actual formation, and sort and number them.

[0063] In the terminal guidance phase, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and matches the target to be tracked in the following steps:

[0064] During the autonomous flight phase, the aircraft undergoes rotation and scaling changes. If the type of formation detected during the terminal guidance phase remains unchanged relative to the target indication phase, the targets in the formation are sorted and numbered, and targets with the same number as those preselected during the target indication phase are selected.

[0065] Further, the steps of Hough transform are as follows:

[0066] 1) Based on the duality of points and lines, the measurement space is established using the Hough transform function Points in the parameter space The corresponding relationship of the curve; (X, Z) is the coordinate;

[0067] For any point in the measurement space The Hough transform function is as follows:

[0068]

[0069] Where, θ is the angle between the perpendicular line from the origin of the measurement space to any straight line passing through the point and the positive direction of the X-axis; |ρ| is the length of the perpendicular line; θ0 and ρ0 are the initial values ​​of the angle and the length of the perpendicular line;

[0070] 2) Record the coordinates of the target formation members in the measurement space In the parameter space, there exists a corresponding sine curve l i ; i = 1, 2, ..., T, T is the number of formation targets, the formation target set in the measurement space There exists a corresponding cluster of sinusoids in parameter space

[0071] Get the whole polyline of the measurement space Calculate the corresponding sine curve intersection point l in parameter space i ∩l j =r k =(θ k ,ρ k ), constitutes a The data set X of samples is {r1, r2, ..., r n}, r k Represents the kth sample in the data sample set X.

[0072] Furthermore, the steps of clustering coordinates by the ISODATA algorithm include:

[0073] 1) Initialization parameters: set the desired number of clusters K and the minimum number of samples in a cluster θ N , the minimum distance between clusters θ C , standard deviation θ S , the maximum number of cluster pairs L for each merger and the number of allowed iterations I;

[0074] 2) Select the initial cluster center: set the initial number of clusters to c and the initial cluster center to m i , assigned to c clusters according to the minimum distance principle; i = 1, 2, ..., c;

[0075] If X satisfies formula (15), then

[0076] ||Xm i ||<||Xm j ||(j=1,2,...,c;i≠j) (15)

[0077] Then X∈w i Divide all samples into corresponding classes.

[0078] 3) Re-clustering: If there is a cluster with a sample size N i <θ N , then discard w i , merge it with the nearest cluster and set c = c-1;

[0079] 4) Update the cluster center and calculate w i The distance between the sample data and its corresponding cluster center is m i The average distance D i , the total distance D from all sample data to their corresponding cluster centers, that is:

[0080]

[0081]

[0082]

[0083] Where N is the total number of samples; N i is the number of samples in the i-th cluster;

[0084] 5) If the current number of iterations is equal to the number of allowed iterations I, let θ N =0, go to step 8);

[0085] If the current iteration is an even number of iterations or c ≥ 2K, go directly to step 8);

[0086] If the current iteration number Then proceed to step 6);

[0087] 6) For each cluster w i , calculate the standard deviation vector σ based on the distance value i =(σ 1i , σ 2i , ..., σ ni ) T , and obtain the component σ with the largest standard deviation in each cluster imax ;

[0088] where the components of the standard deviation vector are as follows:

[0089]

[0090] In the formula, is the jth component of the kth sample; m ij is the jth component of the i-th cluster center; σ ij is the jth component of the standard deviation of the ith cluster; n is the dimension of sample X.

[0091] 7) Split: For any σ imax , if there exists σ imax >θ S , and there is D i >D and N i >2(θ N +1), then cluster w i Split into 2 clusters, the cluster centers are and Cancel cluster center m i , and let c=c+1, go to step 1);

[0092] Cluster Center and As shown below:

[0093]

[0094] In the formula, α, γ i is the coefficient.

[0095] 8) Sorting: For all cluster centers, calculate the distance D between each pair ij

[0096] D ij =||m i -m j ||, (i=1, 2,..., c-1; j=i+1, i+2,...c) (21)

[0097] Comparison D ij and θ c , D ij <θ c The values ​​are arranged in ascending order, namely:

[0098]

[0099] 9) Merge: From the smallest Initially, the distance Two types of cluster centers and Merge and calculate the new cluster center m l , and let c=c-1.

[0100]

[0101] In the formula, is the cluster center and The number of samples in ;

[0102] 10) If the current number of iterations is equal to the allowed number of iterations I or the result converges, then the process ends; otherwise, the number of iterations is increased by 1 and the process goes to step 1).

[0103] It is worth noting that the present invention provides a method for selecting targets for formation identification of guided unmanned aerial vehicles based on deep learning. The method first obtains the situation of the surface target formation through a platform sensor or a data link, establishes a formation echo model based on the surface target formation information to generate a formation formation diagram, and then uses the trained improved LeNet convolutional neural network to perform formation identification on the formation formation diagram, and then establishes a standard template situation in real time based on the formation identification result, sorts and numbers the formation targets in the target indication stage based on the standard template situation, and selects the targets to be hit based on the superior command. During the autonomous flight stage of the aircraft, the formation target undergoes rotation and scaling transformation, the terminal guidance sensor of the aircraft is turned on, the situation of the surface target formation is obtained, and the formation targets are sorted and numbered, and then the target indication target is matched, and finally the aircraft captures and tracks the target.

[0104] The technical effect of the present invention is unquestionable. Based on the improved LeNet network, Hough transform and ISODATA` algorithm, the present invention proposes a surface target formation target selection modeling method, which improves the selection efficiency and enhances the engineering operability. It does not require the initial division of the target formation queue line, is convenient and fast, and has a high accuracy rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 Flowchart for identifying target selection for guiding surface formations of unmanned aerial vehicles; Figure 2 This is a schematic diagram of the simulated platform sensor PPI converted to the formation echo model; Figure 3 It is a schematic diagram of queue elements; Figure 4 It is a schematic diagram of the echo model of single-column and double-column formations; Figure 5 It is a schematic diagram of the echo model of single column and double column formation;

[0106] Figure 6 This is a schematic diagram of the echo model of the azimuth team formation; Figure 7 This is a schematic diagram of the V-shaped formation echo model; Figure 8 This is a schematic diagram of the circular formation echo model; Fig. 9 This is a schematic diagram showing some images of the V-shaped formation; Fig.10 This is a schematic diagram of the formation's heading change;

[0107] Fig.11 It is a zoom diagram; Fig.12 To improve the LeNet network to identify the formation diagram; Fig.13 Schematic diagram of confusion matrix for improving LeNet network to identify formations; Fig.14 It is the measurement space map before Hough transform; Fig.15 is the Hough parameter space diagram;

[0108] Fig.16 It is the measurement space map after inverse Hough transform; Fig.17It is the flow chart of ISODATA algorithm; Fig.18 Select a flow chart for the formation target ranking; Fig.19 It is a standard template to measure the spatial situation map; Fig. 20 This is the situation diagram after the standard template parameter space clustering;

[0109] Fig.21 Hough parameter space diagram for V-shaped team target indication situation; Fig. 22 Indicates the situation parameter space diagram for the V-shaped team target; Fig.23 Measuring spatial situation map for the target indication phase after pre-selection of targets; Fig.24 This is the situation diagram of the V-shaped team terminal guidance sensor; Fig.25 It is the terminal guidance situation parameter space diagram;

[0110] Fig.26 To measure the spatial situation map in the final guidance phase after matching the target; Fig. 27 Provides a situation map for target indication after pre-selection of targets; Fig.28 This is the final guidance situation diagram after confirming the target. DETAILED DESCRIPTION

[0111] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0112] Embodiment 1:

[0113] See also Figures 1 to 28 , a formation recognition target selection method based on deep learning, comprising the following steps:

[0114] 1) Obtain surface target formation information and generate a formation diagram;

[0115] 2) Construct a formation recognition model based on the improved LeNet convolutional neural network;

[0116] 3) Inputting the formation diagram into the formation recognition model to obtain the formation recognition result;

[0117] 4) Sort and number the formation targets in the target indication phase according to the formation recognition results;

[0118] 5) Select the target to be tracked and transmit the target number to the aircraft;

[0119] 6) In the terminal guidance stage, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, then matches the target to be tracked, and tracks the target to be tracked.

[0120] Embodiment 2:

[0121] A formation recognition target selection method based on deep learning, the technical content is the same as that of Example 1, further, the formation formation recognition model includes an input layer, multiple convolutional layers, multiple batch normalization layers, multiple ReLU layers, multiple maximum pooling layers, multiple fully connected layers, a Softmax layer and a classification layer;

[0122] The activation function used by the formation recognition model is as follows:

[0123]

[0124] In the formula, f(x) represents the activation function; x is the input;

[0125] The optimization process of parameters in the formation recognition model is as follows:

[0126]

[0127] Where: β1 is the first-order moment attenuation coefficient, 0<β1<1; β2 is the second-order moment attenuation coefficient, 0<β2<1; g t is the derivative gradient of the objective function; m t is the first-order moment estimate of the gradient in momentum form at time t; v t is the second-order moment estimate of the gradient in momentum form at time t; is the first-order moment estimate after bias correction; is the bias-corrected second-order moment estimate; is β1 to the power of t; is β2 raised to the power of t; θ t+1 is the updated parameter value; η is the learning rate, and ε is a constant. θ t is the parameter value before updating.

[0128] Embodiment 3:

[0129] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-2, further, the formation recognition result includes the number of queue lines and the formation outline;

[0130] The number of queue lines includes one queue line and two queue lines;

[0131] When the queue line is one, the formation outline includes single horizontal column, single vertical column and azimuth column;

[0132] When there are two queue lines, the formation outlines include double horizontal queue, double vertical queue, V-shaped queue and circular queue.

[0133] Embodiment 4:

[0134] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-3, further, when the formation profile is a single horizontal row, the coordinates (x i ,z i ), heading r i As shown below:

[0135]

[0136] Where i = 1, 2, 3, ..., T; T is the number of members; D x , D z is the horizontal spacing and vertical spacing; N(0,σ1 2 )、N(0,σ2 2 ) represents the distribution; σ1 2 ,σ2 2 represents the mean square error of position and heading; γ represents the heading angle of the formation;

[0137] When the formation outline is a double horizontal line, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0138]

[0139] When the formation outline is a single column, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0140]

[0141] When the formation outline is a double column, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0142]

[0143] When the formation outline is a azimuth formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0144]

[0145] When the formation outline is a V-shaped formation, the coordinates of each member in the measurement space (x i ,z i ), heading ri As shown below:

[0146]

[0147] In the formula, N(γ,σ2 2 ) represents the distribution; β represents the queue angle;

[0148] When the formation outline is a ring formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

[0149]

[0150] Embodiment 5:

[0151] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-4, further, considering the disturbance change of heading around the echo center, the heading vector coordinates of the formation members are as follows:

[0152]

[0153] Where, X i ″、Z i ″ is the heading vector coordinate of the formation member considering the disturbance change of heading around the echo center; X i0 , Z i0 is the intermediate parameter; X i , Z i is the navigation vector coordinate of the formation member before updating; (x i ,z i ) is the coordinate of each member in the measurement space; r i is the heading; γ is the formation heading angle.

[0154] Embodiment 6:

[0155] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-5, further, during the autonomous flight phase, if the formation turns around a certain rotation point, the coordinates (X i ′,Z i ′) is as follows:

[0156]

[0157] Where (X0, Z0) is the coordinate of the rotation center point; q is the counterclockwise rotation angle.

[0158] During the autonomous flight phase, if the target formation spacing changes, the coordinates of the target formation members (X n ′,Z n ′) is as follows:

[0159]

[0160] In the formula, (X n+1 ′,Z n+1 ′) is the change coordinate;

[0161] During the autonomous flight phase, if the target formation spacing changes, the target formation spacing (D h ′,Z z ′) is as follows:

[0162]

[0163] Where, when k1, k2∈(0,1), it is the ratio of the team spacing reduction; when k1, k2∈(1,∞), it is the ratio of the team spacing enlargement; (D h , D z ) is the target formation interval before the interval changes.

[0164] Embodiment 7:

[0165] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-6, further, when the queue line is one, the step of sorting and numbering the formation targets in the target indication stage includes:

[0166] A straight line is used to fit the measured spatial coordinates of the formation members in the target indication phase, and this straight line is the queue line;

[0167] Then, the reference members are obtained by combining the heading information, and the formation targets are sorted and numbered according to the relative positions of other members other than the reference members and the reference members, with the reference members as reference;

[0168] Embodiment 8:

[0169] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-7, further,

[0170] In the terminal guidance phase, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and matches the target to be tracked in the following steps:

[0171] The aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and then selects the target with the same number as the target to be tracked from the numbered target formation.

[0172] Embodiment 9:

[0173] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-8, further, when there are two queue lines, the step of sorting and numbering the formation targets in the target indication stage includes:

[0174] 1) Establish standard template situation formation in real time according to the formation recognition results;

[0175] 2) Mapping the coordinates of the target situation to be tracked and the template situation formation members into the parameter space by Hough transformation;

[0176] 3) In the parameter space, the coordinates are clustered using the ISODATA algorithm to obtain error points;

[0177] 4) In the parameter space, according to the principle of Hough transform, the change of the error point from the template situation to the target indication situation is used to offset the change of the ideal point, and the ideal point in the target indication situation parameter space is found, thereby eliminating the error of the queue line in the target indication situation;

[0178] 5) Inverse Hough transform the ideal point in the target indication situation parameter space to the measurement space to obtain the benchmark members and queue lines of the actual formation, and sort and number them.

[0179] Embodiment 10:

[0180] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-9, further, in the terminal guidance stage, the aircraft obtains the surface target formation situation through the terminal guidance sensor, and sorts and numbers the target formation, and the steps of matching the target to be tracked are:

[0181] During the autonomous flight phase, the aircraft undergoes rotation and scaling changes. If the type of formation detected during the terminal guidance phase remains unchanged relative to the target indication phase, the targets in the formation are sorted and numbered, and targets with the same number as those preselected during the target indication phase are selected.

[0182] Embodiment 11:

[0183] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-10, and further, the steps of Hough transform are as follows:

[0184] 1) Based on the duality of points and lines, the measurement space is established using the Hough transform function Points in the parameter space The corresponding relationship of the curve; (X, Z) is the coordinate;

[0185] For any point in the measurement space The Hough transform function is as follows:

[0186]

[0187] Where, θ is the angle between the perpendicular line from the origin of the measurement space to any straight line passing through the point and the positive direction of the X-axis; |ρ| is the length of the perpendicular line; θ0 and ρ0 are the initial values ​​of the angle and the length of the perpendicular line;

[0188] 2) Record the coordinates of the target formation members in the measurement space In the parameter space, there exists a corresponding sine curve l i ; i = 1, 2, ..., T, T is the number of formation targets, the formation target set in the measurement space There exists a corresponding cluster of sinusoids in parameter space

[0189] Get the whole polyline of the measurement space Calculate the corresponding sine curve intersection point l in parameter space i ∩l j =r k =(θ k ,ρ k ), constitutes a The data set X of samples is {r1, r2, ..., r n}, r k Represents the kth sample in the data sample set X.

[0190] Embodiment 12:

[0191] A formation recognition target selection method based on deep learning, the technical content is the same as any one of embodiments 1-11, and further, the step of clustering coordinates by ISODATA algorithm includes:

[0192] 1) Initialization parameters: set the desired number of clusters K and the minimum number of samples in a cluster θ N , the minimum distance between clusters θ C , standard deviation θ S , the maximum number of cluster pairs L for each merger and the number of allowed iterations I;

[0193] 2) Select the initial cluster center: set the initial number of clusters to c and the initial cluster center to m i , assigned to c clusters according to the minimum distance principle; i = 1, 2, ..., c;

[0194] If X satisfies formula (15), then

[0195] ||Xm i ||<||Xm j ||(j=1,2,...,c;i≠j) (15)

[0196] Then X∈w i Divide all samples into corresponding classes.

[0197] 3) Re-clustering: If there is a cluster with a sample size N i <θ N , then discard w i , merge it with the nearest cluster and set c = c-1;

[0198] 4) Update the cluster center and calculate w i The distance between the sample data and its corresponding cluster center is m i The average distance D i , the total distance D from all sample data to their corresponding cluster centers, that is:

[0199]

[0200]

[0201]

[0202] Where N is the total number of samples; N i is the number of samples in the i-th cluster;

[0203] 5) If the current number of iterations is equal to the number of allowed iterations I, let θ N =0, go to step 8);

[0204] If the current iteration is an even number of iterations or c ≥ 2K, go directly to step 8);

[0205] If the current iteration number Then proceed to step 6);

[0206] 6) For each cluster w i , calculate the standard deviation vector σ based on the distance value i =(σ 1i , σ 2i , ..., σ ni ) T , and obtain the component σ with the largest standard deviation in each cluster imax ;

[0207] where the components of the standard deviation vector are as follows:

[0208]

[0209] In the formula, is the jth component of the kth sample; m ij is the jth component of the i-th cluster center; σ ij is the jth component of the standard deviation of the ith cluster; n is the dimension of sample X.

[0210] 7) Split: For any σ imax , if there exists σ imax >θ S , and there is D i >D and N i >2(θ N +1), then cluster w i Split into 2 clusters, the cluster centers are and Cancel cluster center m i , and let c=c+1, go to step 1);

[0211] Cluster Center and As shown below:

[0212]

[0213] In the formula, α, γ i is the coefficient.

[0214] 8) Sorting: For all cluster centers, calculate the distance D between each pair ij

[0215] D ij =||m i -m j ||, (i=1, 2,..., c-1; j=i+1, i+2,...c) (21)

[0216] Comparison D ij and θ c , D ij <θ c The values ​​are arranged in ascending order, namely:

[0217]

[0218] 9) Merge: From the smallest Initially, the distance Two types of cluster centers and Merge and calculate the new cluster center m l , and let c=c-1.

[0219]

[0220] In the formula, is the cluster center and The number of samples in ;

[0221] 10) If the current number of iterations is equal to the allowed number of iterations I or the result converges, then the process ends; otherwise, the number of iterations is increased by 1 and the process goes to step 1).

[0222] Embodiment 13:

[0223] A formation recognition target selection method based on deep learning, with the same technical content as any one of embodiments 1-12. Furthermore, the surface target formation model uses the formation heading in the target indication stage as the positive direction of the Z axis and the right-hand direction perpendicular to the Z axis as the positive direction of the X axis to establish a plane rectangular coordinate system, and a formation queue is marked in the plane rectangular coordinate system; the formation queue includes a reference member J, a queue line, a formation heading γ, a formation spacing D, an alignment angle α, and a queue angle β; the formation model is a formation queue marked in the plane rectangular coordinate system, and the positive direction of the Z axis of the rectangular coordinate system is the formation heading in the target indication stage.

[0224] The queue line refers to the line connecting the points of each member in the queue starting from the base member;

[0225] The formation heading γ refers to the mean angle between the bow and stern lines of the reference members and the true north direction;

[0226] The alignment angle α refers to the angle between the bow and stern lines of other members and the queue line;

[0227] The queue angle β = 180° - α refers to the angle between the bow and stern lines of the reference member and the queue line.

[0228] Embodiment 14:

[0229] A formation recognition target selection method based on deep learning, the steps are as follows:

[0230] First, the surface target formation situation is obtained through platform sensors or data links. A formation echo model is established based on the surface target formation information to generate a formation formation diagram. Then, the trained improved LeNet convolutional neural network is used to perform formation recognition on the formation formation diagram. A standard template situation is then established in real time based on the formation recognition results. The formation targets in the target indication phase are sorted and numbered according to the standard template situation, and the targets to be hit are selected according to superior commands. During the autonomous flight phase of the aircraft, the formation targets rotate and scale, and the terminal guidance sensor of the aircraft is turned on to obtain the surface target formation situation, sort and number the formation targets, and then match the target indication targets. Finally, the aircraft captures and tracks the target. The relevant formation recognition target selection flow chart is shown below. Figure 1 shown.

[0231] The steps of identifying the target formation and selecting the target based on deep learning include:

[0232] 1) According to the surface target formation information, a formation echo model is established to generate a formation diagram. Figure 2 As shown, firstly, the PPI recording of the platform sensor is simulated during the target indication phase, and the polar coordinate system is converted to the rectangular coordinate system with the origin as the center to unify the line of sight and obtain a formation echo model with only distance and azimuth information.

[0233] In order to construct the formation echo model, the following reasonable assumptions are put forward:

[0234] (1) The number of members in the formation is known, and the distances and headings of different members can be detected and distinguished by sensors. However, each member has position noise and heading noise, and both position noise and heading noise obey a normal distribution with a mean of 0.

[0235] (2) Ignoring the environmental influences such as sea clutter, weather, and interference from false targets, the established formation echo model only contains surface targets and ocean background;

[0236] (3) A black background is used to represent the ocean. The combination of a white equilateral triangle and a rectangle formed by the five feature points at the lower left, left, upper, right, and lower right represents the member. The size of the rectangle represents the echo size, which is uniformly set to 0.22×0.22, in units of nmile. The target coordinate position is at the center of the rectangle, and the direction of the equilateral triangle represents the member's heading.

[0237] (4) The type and degree of change in the target formation are unrelated to the autonomous flight time of the aircraft;

[0238] (5) The terminal guidance phase does not change the formation type of the target indication phase, or changes from one typical linear formation to another typical linear formation, and the formation targets do not exchange positions, that is, the order based on the reference members does not change;

[0239] (6) Ignoring the impact of sea conditions, weather, and network electromagnetic space environment conditions on aircraft, the formation does not take any hard countermeasures;

[0240] (7) Δθ = 0.1 rad;

[0241] (8) If the target confirmed in the terminal guidance phase is consistent with the pre-selected target in the target indication phase, the identification and selection of the formation target is correct.

[0242] A plane rectangular coordinate system is established in the measurement space with the true north direction as the positive direction of the Z axis and the due east direction as the positive direction of the X axis. Then, the heading noise and position noise that obey N(0,σ) are applied to the single member to establish the formation echo model. In order to standardize and unify the method of describing the formation information, Figure 3 As shown in the figure, taking the V-shaped formation as an example, the definition of the formation parameter elements is given: the reference member J is a member that provides a reference for other members in the formation to maintain their formation position, usually located in a special position. The formation line refers to a straight line that passes through the theoretical location of the member and can characterize the formation characteristics. The formation heading γ refers to the azimuth of the reference member relative to the true north direction. The formation spacing D f It can be decomposed into the lateral spacing D along the vertical formation heading x and the longitudinal spacing D along the formation heading z The alignment angle α refers to the angle between the queue line and the bow and stern lines of other members, and the queue angle β refers to the angle between the bow and stern lines of the reference member and the queue line. Obviously, β = 180°-α. Further, the formation echo model includes single horizontal column and double horizontal column, single vertical column and double vertical column, azimuth column, V-shaped column, and circular column.

[0243] 1.1) Single-horizontal column and double-horizontal column: Both single-horizontal column and double-horizontal column belong to horizontal column. The formation heading of horizontal column is perpendicular to the queue line, that is, the queue angle β=90°. The two queue lines of double-horizontal column are parallel to each other, and the formation members move forward in parallel.

[0244] Assume that the formation heading of the horizontal team is γ = 0°, the total number of members is T, and the horizontal spacing is D x , the vertical spacing is D z , the mean square error of position is σ1, and the mean square error of heading is σ2. Then the spatial coordinates (x i ,z i ), heading r i (i=1,2,3,...,T) is as follows:

[0245]

[0246] Each member of the double team measures the spatial coordinates (x i ,z i ), heading r i (i=1,2,3,...,T) is as follows:

[0247]

[0248] Let single row T = 6, double row T = 12, σ1 = 0.1nmile, D x =1nmile, D Z =1nmile, then the formation echo model of single-column and double-column is as follows Figure 4 shown.

[0249] 1.2) Single column and double column: Both single column and double column belong to column. The formation heading of the column is parallel to the queue line, that is, the queue angle β=180°. The two queue lines of the double column are parallel to each other, and the formation members move forward in parallel.

[0250] Assume that the column's formation heading γ = 0°, the total number of members is T, and the lateral spacing is D x , the vertical spacing is D z , the mean square error of position is σ1, and the mean square error of heading is σ2. Then the spatial coordinates (x i ,z i ), heading r i (i=1,2,3,...,T) is as follows:

[0251]

[0252] Each member of the double column measures the spatial coordinates (x i ,z i ), heading r i (i=1,2,3,...,T) is as follows:

[0253]

[0254] Let single column T = 6, double column T = 12, σ1 = 0.1nmile, D x =1nmile, D Z = 1nmile, then the formation echo model of single column and double column is as follows Figure 5 shown.

[0255] 1.3) Azimuth formation: Azimuth formation is a formation in which all members are located on a certain azimuth line of the reference member. Unlike the echelon formation, the coordinate distance of the azimuth formation remains unchanged, but the alignment angle changes. In the present invention, in order to verify the ability of the convolutional neural network to identify the heading, the azimuth formation is listed separately and distinguished from the single column and single horizontal formation.

[0256] Assume that the formation heading of the azimuth team is γ = 45, the total number of members is T, the alignment angle is α, and the lateral spacing is D x , the vertical spacing is D z , the position mean square error is σ1, and the heading mean square error is σ2. Then the spatial coordinates (x i ,z i ), heading r i (i=1,2,3,...,T) is as follows:

[0257]

[0258] Let the position team T = 6, σ1=0.1nmile, D x =1nmile, D Z =1nmile, then the formation echo model is as follows Figure 6shown.

[0259] 1.4) V-shaped formation: V-shaped formation is a common formation in the process of formation action. It has a wide coverage and strong early warning and detection capabilities. It can be rotated 180° according to the mission requirements, that is, a reverse V-shaped formation. The V-shaped formation usually takes the member located at the connection of two queue lines as the reference member. The queue angle and alignment angle on both sides of the reference member are the same.

[0260] Assume that the total number of members in the V-shaped team is T, the queue angle is β, the queue navigation is γ = 0°, and the lateral spacing is D x , the vertical spacing is D z , the mean square error of position is σ1, the mean square error of heading is σ2, then each member measures the spatial coordinates (x i , z i ), heading r i (i=1,2,3,...,T) is as follows:

[0261]

[0262] Let V-shaped team T = 9, σ1 = 0.1nmile, D x =0.5nmile, D Z =1nmile, then the formation echo model is as follows Figure 7 shown.

[0263] 1.5) Ring formation: The ring formation is conducive to giving full play to the early warning and detection capabilities of the peripheral members. Usually the middle member is taken as the benchmark member, and the horizontal queue line and the vertical queue line are perpendicular to each other.

[0264] Assume that the circular formation heading γ = 0°, the total number of members in the formation is T, and the lateral spacing is D x , the vertical spacing is D z , the mean square error of position is σ1, the mean square error of heading is σ2, and the coordinates of each member in the measurement space (x i ,z i ), heading r i (i=1,2,3,...,T) is as follows:

[0265]

[0266] Let the ring team T = 9, σ1 = 0.1nmile, D x =1nmile, D Z =1nmile, then the formation echo model is as follows Figure 8 shown.

[0267] 1.6) The 5-dimensional row vector composed of the coordinates of the five feature points of the target echo: lower left, left, upper, right, and lower right:

[0268]

[0269] The coordinates of the heading vector of the echo feature point considering the disturbance change of heading around the echo center are:

[0270]

[0271] 1.7) After the formation echo model is established, MATLAB simulation is first used to obtain 200 pictures of each formation, and then each formation picture is expanded by random rotation to obtain a formation data set. The formation data set contains 7 types of formations, namely single horizontal team, single vertical team, azimuth team, double horizontal team, double vertical team, V-shaped team, and circular team, with a total of 7,000 formation images. When the pictures in the formation data set are prepared, they are first pre-processed, all pictures are formatted, the picture format is uniformly adjusted to .jpg format, and pictures of the same type are named in order, such as V-shaped team 1_1, which is the first picture of the V-shaped formation. The processed images are stored in a folder, such as Fig. 9 shown.

[0272] 1.8) During the autonomous flight phase, if the formation turns around a certain rotation point, the coordinates of the target formation member (X n ′,Z n ′) is as follows:

[0273]

[0274] Where (X0, Z0) is the coordinate of the rotation center point; q is the rotation angle.

[0275] 1.9) During the autonomous flight phase, if the distance between the target formations changes, the coordinates (X n ′,Z n ′) is as follows:

[0276]

[0277] Alternatively, the target formation interval (D h ′,Z z ′) is as follows:

[0278]

[0279] In the formula, when k1, k2∈(0,1), the distance between the formations is reduced; when k1, k2∈(1,∞), the distance between the formations is enlarged. When k1=k2, the target formation is scaled proportionally.

[0280] 1.10) The indicators used to evaluate the quality of the target formation recognition algorithm are detection speed and detection accuracy. The detection speed refers to the test time for a formation diagram, and the detection accuracy mainly refers to the accuracy rate. The accuracy rate P is expressed as the probability that the formation is a correct sample. The expression is shown in formula (13):

[0281]

[0282] Where T r Indicates the number of samples detected correctly, T e Indicates the number of samples in the test set.

[0283] 2) Use the trained improved LeNet convolutional neural network to recognize the formation diagram. The improved LeNet network structure is shown in Table 2:

[0284] Table 2 Improved LeNet convolutional neural network structure

[0285]

[0286]

[0287] Improved LeNet has made some improvements and optimizations to LeNet-5. The specific improvements are as follows:

[0288] 2.1) Added batch normalization layer. Batch normalization layer can accelerate the convergence speed of model training, reduce the gradient explosion and vanishing problems in the optimization process, improve the generalization ability of the model, make the neural network more robust to small changes in the input data, improve the distribution of activation functions, make the distribution of activation functions more stable, reduce the degree to which the input data distribution of the later layer of the model will change with the change of parameters of the previous layer, and improve the stability of the model.

[0289] 2.2) Use ReLU activation function instead of sigmoid function.

[0290] The expression of the sigmoid function is as follows:

[0291]

[0292] The expression of the ReLU function is as follows:

[0293]

[0294] Compared with the sigmoid function, the ReLU function as an activation function can reduce the number of model parameters, has the advantages of faster calculation, avoiding gradient disappearance, improving sparsity and robustness, and can enhance the nonlinear expression ability of the model.

[0295] 2.3) The pooling method is changed from average pooling to maximum pooling. Maximum pooling only considers the maximum value, which can better retain important information related to the feature and better extract edge information. Since maximum pooling only considers the maximum value, it can reduce redundant information in the feature map, thereby reducing the risk of overfitting of the model. In addition, the nonlinear transformation of maximum pooling can increase the nonlinear expression ability of the model and further improve the generalization performance of the model.

[0296] 2.4) Adam (Adaptive Moment Estimation) is an optimization algorithm based on gradient descent. Compared with the SGDM momentum stochastic gradient descent algorithm, it uses the first-order moment and second-order moment estimation of the gradient to adjust the learning rate of each parameter. Therefore, it is more stable when dealing with non-stationary objective functions and noisy data. At the same time, it can adaptively adjust the learning rate of each weight and converge faster. It has the advantages of being suitable for large-scale data and high-dimensional features and fast convergence speed. The relevant principle is as shown in formula (16):

[0297]

[0298] Where: β1 is the first-order moment attenuation coefficient, 0<β1<1; β2 is the second-order moment attenuation coefficient, 0<β2<1; g t is the derivative gradient of the objective function; m t is the first-order moment estimate of the gradient in momentum form at time t; v t is the second-order moment estimate of the gradient in momentum form at time t; is the first-order moment estimate after bias correction; is the bias-corrected second-order moment estimate; is β1 to the power of t; It is β2 to the power of t.

[0299] Since the elements in v0 and m0 are initialized to 0, when t is small, the sum of the random gradient weights of the small batch in each time step in the past will be small, so error correction is required. The last item is the update formula, θ t+1 is the updated parameter value, where η is the learning rate and ε is a constant added to maintain numerical stability.

[0300] 3) After using the improved LeNet network to perform formation recognition, the steps of formation target selection include:

[0301] 3.1) After the formation is identified according to the improved LeNet network, the number of formation queue lines can be obtained, that is, the number of cluster centers clustered in the parameter space using the ISODATA algorithm. In the present invention, the formation can be divided into two categories. One is that there is only one queue line, including a single horizontal queue, a single vertical queue and an orientation queue; the other is that there are two queue lines, that is, there are two cluster centers, including double horizontal queues, double vertical queues, V-shaped queues and circular queues.

[0302] When there is only one queue line, the method for sorting and selecting formation targets is relatively simple. It only needs to use a straight line to fit the measured spatial coordinates of the formation members in the target indication phase. This straight line is the queue line. Then, the reference member is obtained in combination with the heading information. Then, the formation targets can be sorted and numbered based on the relative positions of other members and the reference member with the reference member as a reference. Similarly, the formation targets can be sorted and numbered in the terminal guidance phase, and the targets with the same pre-selected numbers in the target indication phase can be selected and attacked.

[0303] When there are two queue lines, the sorting and selection of formation targets is more complicated, and this is the focus of the following research. The target sorting and selection method based on Hough transform and ISODATA algorithm adopts the idea of ​​template method and error compensation method. After the formation is recognized in the target indication stage, the standard template situation of the corresponding formation is established in real time. The standard template formation is obtained by using the improved LeNet network for formation recognition. The number and position of members of the standard template formation, the formation heading, the formation spacing, the queue line and the reference members are known, and the target has no position noise and heading noise, which is in the theoretically optimal state.

[0304] 3.2) First, the coordinates of the target indication situation and template situation formation members are mapped to the parameter space by Hough transformation, and then clustered in the parameter space by the ISODATA algorithm. In the parameter space, according to the principle of Hough transformation, the change of the error point from the template situation to the target indication situation is used to offset the change of the ideal point, and the ideal point of the target indication situation parameter space is found, thereby eliminating the error of the queue line in the target indication situation. The ideal point is the point where the ideal queue line is mapped in the parameter space, which is the common intersection of at least [T / 2] sine curves in the template situation parameter space. The error point is the two clustering centers of the ISODATA algorithm in the parameter space, which is the point where the error queue line is mapped to the parameter space. Then, the ideal point of the target indication situation parameter space is inversely Hough transformed to the measurement space to obtain information such as the benchmark members and queue lines of the actual formation, and sort and number them.

[0305] 3.3) When the aircraft flies to the terminal guidance sensor power-on stage, the type of formation detected remains unchanged compared to the target indication stage, but changes such as rotation and scaling occur. Similarly, the targets in the formation in the terminal guidance stage can be sorted and numbered, and the targets with the same number as the pre-selected targets in the target indication stage can be selected and attacked.

[0306] 4) The steps of Hough transform are as follows:

[0307] 4.1) Based on the duality of points and lines, the measurement space is established using the Hough transform function Points in the parameter space The corresponding relationship of the curve; (X, Z) is the coordinate;

[0308] For any point in the measurement space The Hough transform function is as follows:

[0309]

[0310] Where θ is the angle between the perpendicular line from the origin of the measurement space to any straight line passing through the point and the positive direction of the X-axis; |ρ| is the length of the perpendicular line;

[0311] 4.2) Record the coordinates of the target formation members in the measurement space In the parameter space, there exists a corresponding sine curve l i ; i = 1, 2, ..., T, T is the number of formation targets, the formation target set in the measurement space There exists a corresponding cluster of sinusoids in parameter space Get the whole polyline of the measurement space Find the corresponding sine curve intersection point l in parameter space i ∩l j =r k =(θ k ,ρ k ), constitutes a The data set X of samples is {r1, r2, ..., r n}, r k Represents the kth sample in the data sample set X.

[0312] 5) The steps of ISODATA algorithm clustering include:

[0313] 5.1) Initialization parameters: set the desired number of clusters K and the minimum number of samples in a cluster θ N , the minimum distance between clusters θ C , standard deviation θ S , the maximum number of cluster pairs L for each merge and the number of allowed iterations I, etc.

[0314] 5.2) Select the initial cluster center: Set the initial number of clusters to c and the initial cluster center to m i (i=1,2,...,c), are assigned to c clusters according to the minimum distance principle.

[0315] ||Xm i ||<||Xm j ||(j=1,2,...,c;i≠j) (18)

[0316] Then X∈w i Divide all samples into corresponding classes.

[0317] 5.3) Re-clustering: If there is a cluster with a sample size N i <θ N , then discard w i , merge it with the cluster with the nearest distance, and let c=c-1.

[0318] 5.4) Update the cluster center and calculate the parameters:

[0319]

[0320] Compute updated cluster centers.

[0321] Calculate w i The distance between the sample data and its corresponding cluster center is m i The average distance D i for

[0322]

[0323] Calculate the total distance D from all sample data to their corresponding cluster centers:

[0324]

[0325] 5.5) Iteration operation:

[0326] If the number of iterations is equal to the number of allowed iterations I, then let θ N =0, go to step 5.8);

[0327] like That is, the number of clusters is too small, go to step 5.6);

[0328] If the iteration is an even number or c ≥ 2K, go to step 5.8), otherwise proceed as follows.

[0329] 5.6) Calculate the standard deviation: For each cluster w i , calculate the standard deviation vector σ based on the distance value i σ i =(σ1i , σ 2i , ..., σ ni ) T , where the vector components are

[0330]

[0331] In the formula is the jth component of the kth sample; m ij is the jth component of the i-th cluster center; σ ij is the jth component of the standard deviation of the i-th cluster; n is the dimension of sample X. And the component σ with the largest standard deviation in each cluster is obtained imax (i=1, 2, ..., c).

[0332] 5.7) Splitting: If for any σ imax (i=1,2,...,c), there exists σ imax >θ S , and there are

[0333] D i >D and N i >2(θ N +1), that is, the number of samples in the cluster is twice as high as the threshold or When i Split into 2 clusters, whose centers are and The original m i Cancel, and set c=c+1, go to step 5.1). and The calculation of is as follows:

[0334]

[0335] In the formula, the value of α should be selected so that w i The samples in and The distance is different, but w i The samples from are still in the two new sets.

[0336] 5.8) Sorting: For all cluster centers, calculate the distance D between each pair ij

[0337] D ij =||m i -m j ||, (i=1, 2,..., c-1; j=i+1, i+2,...c) (24)

[0338] Comparison D ij and θc , D ij <θ c The values ​​are arranged in ascending order as

[0339]

[0340] 5.9) Merge: From the smallest Initially, the distance Two types of cluster centers and Merge and calculate the new cluster center m l , and let c=c-1.

[0341]

[0342] Embodiment 15:

[0343] A simulation of a formation target selection method based on deep learning is as follows:

[0344] First, the formation data set with a position error of 0.1 nmile and a heading error of 10° is divided into a training set and a test set in a ratio of 8:2, and the improved LeNet network is used for training and testing. The specific experimental environment is shown in Table 3. The test results are shown in Fig.12 and Fig.13 shown.

[0345] Table 3 Experimental environment

[0346]

[0347] (1) Comparison of different convolutional neural networks

[0348] In this experiment, a formation data set of 7000 samples with a position error of 0.1 and a heading error of 10° is first divided into a training set and a test set in a ratio of 8:2. The training images and test images are adjusted to 160*120 3-channel images with the same pixel size as the input layer. Three network models, LeNet-5, AlexNet and improved LeNet, are used for training and testing. The simulation results are shown in Table 4.

[0349] Table 4 Recognition accuracy and duration of three network models: LeNet-5, AlexNet and improved LeNet

[0350]

[0351]

[0352] Training time refers to the time used for training the network model, which is one of the indicators used to judge the quality of the model. Test time refers to the time used to detect the test sample set after the network model is trained. As can be seen from Table 4, as the number of layers of the improved LeNet network increases and the structure becomes more complex, the time required to train the network increases slightly, but there is no significant difference in the detection speed of the three models, all less than 5ms. At the same time, the features extracted by the improved LeNet network model are richer, more sensitive to heading information, and better able to distinguish between single horizontal, single vertical and azimuth teams. The overall accuracy of formation recognition is greatly improved, meeting the requirements of detection speed and detection accuracy. It is worth noting that the formation heading of the single vertical team is consistent with the queue line, which can enhance recognition; the formation heading of the single horizontal team is perpendicular to the queue line, which is easy to distinguish; the formation heading of the azimuth team is at a certain angle to the queue line, which is difficult to distinguish; the double-row team can be compared with each other, which is helpful for recognition.

[0353] (2) Comparison between Convolutional Neural Networks and Artificial Neural Networks

[0354] In this example, the conditions of each formation sample in the formation data set are changed, 3 are randomly selected as training and 1 is selected as test, the images are grayed out, and converted into a 160*120 digital matrix with the same pixel size as the input layer of the convolutional neural network for experiment. The GRNN with the best performance in the literature is selected for simulation test. The correct result is [1 2 3 4 5 67], and the test results are shown in Table 5.

[0355] Table 5 Recognition accuracy of artificial neural network

[0356] Experimental conditions Test Results Accuracy Only rotate, remove noise [3 3 2 2 5 3 2] 14.3％ No rotation, no noise [1 2 3 4 5 6 7] 100％ No rotation, only position noise 0.1 [1 2 3 4 5 6 7] 28.6％ No rotation, only position noise 0.2 [5 2 3 3 2 6 7] 57.2％ No rotation, only position noise 0.5 [3 2 3 6 2 6 3] 42.9％

[0357] It can be found that GRNN cannot cope with the rotation situation. It can cope with certain position noise without rotation and cannot realize the formation recognition of the formation data set.

[0358] (3) Position noise impact

[0359] The heading noise is removed, and position noise with a normal distribution and a mean of 0 and a mean square error ranging from 0.1 to 0.5 is added to all formation targets. The improved LeNet network is used to train and test the formation data set. The recognition accuracy is shown in Table 6.

[0360] Table 6 Impact of position noise on recognition accuracy

[0361]

[0362] From the simulation results, it can be seen that as the position mean square error σ1 increases from 0.1 to 0.3, the recognition accuracy rate decreases, but it is still very high, maintained at more than 96%, and the distinction between all formations is good. However, as the position mean square error σ1 increases from 0.3 to 0.5, an abnormal phenomenon occurs, and the recognition accuracy rate increases instead of decreasing. This is because when the position noise increases to a certain extent, the position information as one of the features cannot effectively distinguish the formation. During the training process of the convolutional neural network, the weight of the position information as a feature extracted by the convolutional neural network decreases, and the weight of the number of members and the heading of the members as feature information increases.

[0363] (4) Impact of heading noise

[0364] After removing the position noise, a heading noise with a mean of 0 and a mean square error of 10° to 40° that follows a normal distribution is added to the headings of all formation targets. The formation dataset is trained and tested using the improved LeNet network. The recognition accuracy is shown in Table 7.

[0365] Table 7 Impact of heading noise on recognition accuracy

[0366]

[0367] From the simulation results, it can be seen that as the heading mean square error σ2 increases, the recognition accuracy decreases, especially when σ2=20° to σ2=30°, the recognition accuracy of the single horizontal team, single vertical team and azimuth team drops sharply. When σ2=40°, the recognition accuracy of the azimuth team drops below 70%, and it cannot be effectively distinguished, but the overall recognition accuracy is high, maintained above 92%. It is found that the improved LeNet network can effectively combat heading noise interference.

[0368] The formation recognition experiment conducted by the improved LeNet convolutional neural network added position noise and heading noise, expanded the sample set, made it more referenceable, and produced more accurate results. The simulation experiment showed that the formation recognition method based on the improved LeNet network can quickly and accurately identify the formation, and has good recognition effect even when there is heading noise and position noise in the target information, and has good robustness. The improved LeNet convolutional neural network model can be trained in normal times and used directly during operations, with good real-time performance. It can effectively assist commanders in quickly analyzing and recognizing information collection to meet operational needs.

[0369] Embodiment 16:

[0370] Taking the V-shaped team as an example, the simulation of the target sorting selection method based on Hough transform and ISODATA algorithm is as follows:

[0371] like Fig.18The figure shows the template situation of the V-shaped team. In the template situation, the coordinates of each member in the formation are (X n ,Y n )(n=1,2,3,…,9), the reference member is (X5, Y5). Perform Hough transform on the member position coordinates in the standard template and map them to Fig.19 The parameter space shown in Figure 1 yields 9 curves with a total of There are intersection points, where two ideal queue lines correspond to two ideal points 0 in the parameter space, which are the common intersection points of the two 5 curves and are marked with blue circles.

[0372] Then, the ISODATA algorithm is used to cluster the intersection points in the parameter space, such as Fig.19 As shown in the figure, the two green hollow dots are the clustering centers of the ISODATA algorithm clustering, called error point 0.

[0373] like Fig.21 and Fig. 22 As shown in the figure, the formation measurement space coordinates under the target indication situation are mapped to parameter space 1 by Hough transform and clustered using ISODATA algorithm to obtain error point 1. Error point 1 is Fig. 22 Then, based on the change from error point 0 to the corresponding error point 1, the same change is applied to the ideal point 0 to obtain the ideal point 1. Fig. 22 The blue circles are shown in the figure. Fig. 22 Take as an example, and keep the coordinates of the point to two decimal places. The specific operations are as follows.

[0374]

[0375] Then perform inverse Hough transform on the ideal point 1 and the error point 1 to obtain the corresponding ideal queue line and error queue line, as shown in Fig.23 As shown, the ideal queue line and the error queue line are represented by solid lines and dotted lines respectively. After obtaining the ideal queue line, the formation targets are sorted and numbered, and target No. 3 is pre-selected for attack.

[0376] Similarly, if Fig.24 , Fig.25 Fig.26 As shown, the formation measurement space coordinates in the terminal guidance situation are sorted and numbered to match the pre-selected targets in the target indication phase.

[0377] Since double horizontal columns, double vertical columns, V-shaped columns and circular columns are symmetrical, after finding the benchmark members according to the ideal queue line, the formation targets can be sorted and numbered in combination with the heading information.

[0378] In the V-shaped team and the ring team, the echo signal closest to the intersection of the ideal queue line is selected as the reference member.

[0379] In a V-shaped formation, if the total number of members in the formation is T, the coordinates of each member are (X n ,Y n )(n=1,2,3,…,T), the intersection of the two ideal queue lines is (X,Y), then the coordinates of the reference member (J X ,J Z )for:

[0380]

[0381] In the ring formation, if the total number of members in the formation is T, the coordinates of each member are (X n ,Y n )(n=1,2,3,…,T), the intersection of the two ideal queue lines is (X,Y), then the coordinates of the reference member (J X ,J Z )for:

[0382]

[0383] In the double horizontal and double vertical queues, the error queue lines are two nearly parallel straight lines. However, due to the existence of position noise, the two error queue lines must have an intersection in the measurement space. The echo signal farthest from the intersection is used as the reference member. If the intersection of the two error queue lines is (X, Y), the total number of members in the formation is T, and the coordinates of each member are (X n ,Y n )(n=1,2,3,…,T), then the coordinates of the base member (J X ,J Z )for:

[0384] (J X ,J Z )=(X1,Z1)(30)

[0385] Embodiment 17:

[0386] A comparative simulation of a formation target selection method based on deep learning is as follows:

[0387] V-shaped team γ=0°, D x =0.5nmile, D z =1nmile, α=26.6°, γ of the ring team=0°, D x =1nmile, D z =1nmile, γ of the double horizontal (vertical) team = 0°, D x =1nmile, D z=1nmile, and use random numbers with a position mean square deviation of 0.2 to simulate the position noise of four typical formation target members. Suppose that during the autonomous flight phase of the aircraft, the formation heading becomes 330°, and the formation spacing is scaled by k1=k2=2, and the formation targets are sorted and selected.

[0388] like Fig. 27 As shown, during the target indication phase, the formation is identified and the formation targets are sorted and numbered. The benchmark members of the V-shaped team, circular team, and double horizontal (vertical) team are No. 5, No. 3, and No. 1, respectively. The pre-selected target is No. 3, which is marked with a red box.

[0389] like Fig.28 As shown, according to the selected target No. 3, the anti-surface target guided unmanned aerial vehicle is launched. When the aircraft flies to the end of the self-control section and switches to the terminal guidance sensor to be turned on, the target is sorted and selected, and the selected members are represented by blue boxes for capture and tracking.

[0390] By changing the position mean square error σ1 from 0.1 to 0.4, under the assumed conditions, the algorithm of the present invention is compared with the existing method, and Monte Carlo simulation is performed 100 times. The accuracy of formation target selection is shown in Table 8.

[0391] Table 8 Formation target selection accuracy

[0392]

[0393] After testing, the proposed formation target selection method based on Hough transform and ISODATA algorithm has an average running time of about 1.25s, indicating that the algorithm has good real-time performance. Inheriting the existing ideas for formation target selection, the error source of the queue line is analyzed, and the error cancellation method is used when the parameter space is fully connected, avoiding the problem of manually performing initial division of the queue line. Although the detection speed has decreased, the accuracy of formation target selection has been improved, and automatic target selection can be achieved, which has good engineering applicability.

Claims

1. A formation recognition target selection method based on deep learning, characterized in that: The following steps are involved: 1) Obtain surface target formation information and generate a formation diagram; 2) Construct a formation recognition model based on the improved LeNet convolutional neural network; 3) Inputting the formation diagram into the formation recognition model to obtain the formation recognition result; 4) Sort and number the formation targets in the target indication phase according to the formation recognition results; 5) Select the target to be tracked and transmit the target number to the aircraft; 6) In the terminal guidance stage, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, then matches the target to be tracked, and tracks the target to be tracked; The formation recognition model includes an input layer, multiple convolutional layers, multiple batch normalization layers, multiple ReLU layers, multiple maximum pooling layers, multiple fully connected layers, a Softmax layer and a classification layer; The activation function used by the formation recognition model is as follows: In the formula, f(x) represents the activation function; x is the input; The optimization process of parameters in the formation recognition model is as follows: Where: β1 is the first-order moment attenuation coefficient, 0<β1<1; β2 is the second-order moment attenuation coefficient, 0<β2<1; g t is the derivative gradient of the objective function; m t is the first-order moment estimate of the gradient in momentum form at time t; v t is the second-order moment estimate of the gradient in momentum form at time t; is the first-order moment estimate after bias correction; is the bias-corrected second-order moment estimate; is β1 to the power of t; is β2 raised to the power of t; θ t+1 is the updated parameter value; η is the learning rate, ε is a constant, and θ t is the parameter value before updating; The formation recognition result includes the number of queue lines and the formation outline; The number of queue lines includes one queue line and two queue lines; When the queue line is one, the formation outline includes single horizontal column, single vertical column and azimuth column; When there are two queue lines, the formation outlines include double horizontal queue, double vertical queue, V-shaped queue and circular queue.

2. The method for selecting a formation recognition target based on deep learning according to claim 1, characterized in that: When the formation outline is a single horizontal line, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below: Where i = 1, 2, 3, ..., T; T is the number of members; D x , D z is the horizontal spacing and vertical spacing; N(0,σ1 2 )、N(0,σ2 2 ) represents the distribution; σ1 2 ,σ2 2 represents the mean square error of position and heading; γ represents the heading angle of the formation; When the formation outline is a double horizontal line, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below: When the formation outline is a single column, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below: When the formation outline is a double column, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below: When the formation outline is a azimuth formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below: When the formation outline is a V-shaped formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below: In the formula, N(γ,σ2 2 ) represents the distribution; β represents the queue angle; When the formation outline is a ring formation, the coordinates of each member in the measurement space (x i ,z i ), heading r i As shown below:

3. The method for selecting a formation recognition target based on deep learning according to claim 1, characterized in that: Considering the disturbance change of heading around the echo center, the heading vector coordinates of the formation members are as follows: Where, X i ″、Z i ″ is the heading vector coordinate of the formation member considering the disturbance change of heading around the echo center; X i0 , Z i0 is the intermediate parameter; X i , Z i is the navigation vector coordinate of the formation member before updating; (x i ,z i ) is the coordinate of each member in the measurement space; r i is the heading; γ is the formation heading angle.

4. The method for selecting a formation recognition target based on deep learning according to claim 1, characterized in that: During the autonomous flight phase, if the formation turns around a certain rotation point, the coordinates of the target formation members (X i ′,Z i ′) is as follows: Where, (X0, Z0) is the coordinate of the rotation center point; q is the counterclockwise rotation angle; During the autonomous flight phase, if the target formation spacing changes, the coordinates of the target formation members (X n ′,Z n ′) is as follows: In the formula, when k1, k2∈(0,1), it is the ratio of the team spacing reduction; when k1, k2∈(1,∞), it is the ratio of the team spacing enlargement; (X n+1 ′,Z n+1 ′) is the change coordinate; During the autonomous flight phase, if the target formation spacing changes, the target formation spacing (D h ′,Z z ′) is as follows: Where, when k1, k2∈(0,1), it is the ratio of the team spacing reduction; when k1, k2∈(1,∞), it is the ratio of the team spacing enlargement; (D h , D z ) is the target formation interval before the interval changes.

5. The method for selecting a formation recognition target based on deep learning according to claim 1, characterized in that: When there is only one queue line, the steps for sorting and numbering the formation targets in the target indication phase include: A straight line is used to fit the measured spatial coordinates of the formation members in the target indication phase, and this straight line is the queue line; Then, the reference members are obtained by combining the heading information, and the formation targets are sorted and numbered according to the relative positions of other members other than the reference members and the reference members, with the reference members as reference; In the terminal guidance phase, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and matches the target to be tracked in the following steps: The aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and then selects the target with the same number as the target to be tracked from the numbered target formation.

6. The method for selecting a formation recognition target based on deep learning according to claim 1, characterized in that: When there are two queue lines, the steps for sorting and numbering the formation targets in the target indication phase include: 1) Establish standard template situation formation in real time according to the formation recognition results; 2) Mapping the coordinates of the target situation to be tracked and the template situation formation members into the parameter space by Hough transformation; 3) In the parameter space, the coordinates are clustered using the ISODATA algorithm to obtain error points; 4) In the parameter space, according to the principle of Hough transform, the change of the error point from the template situation to the target indication situation is used to offset the change of the ideal point, and the ideal point in the target indication situation parameter space is found, thereby eliminating the error of the queue line in the target indication situation; 5) Inverse Hough transform the ideal point of the target indication situation parameter space into the measurement space to obtain the benchmark members and queue lines of the actual formation, and sort and number them; In the terminal guidance phase, the aircraft obtains the situation of the surface target formation through the terminal guidance sensor, sorts and numbers the target formation, and matches the target to be tracked in the following steps: During the autonomous flight phase, the aircraft undergoes rotation and scaling changes. If the type of formation detected during the terminal guidance phase remains unchanged relative to the target indication phase, the targets in the formation are sorted and numbered, and targets with the same number as those preselected during the target indication phase are selected.

7. A formation recognition target selection method based on deep learning according to claim 6, characterized in that: The steps of Hough transform are as follows: 1) Based on the duality of points and lines, the measurement space is established using the Hough transform function Points in the parameter space The corresponding relationship of the curve; (X, Z) is the coordinate; For any point in the measurement space The Hough transform function is as follows: Where, θ is the angle between the perpendicular line from the origin of the measurement space to any straight line passing through the point and the positive direction of the X-axis; |ρ| is the length of the perpendicular line; θ0 and ρ0 are the initial values ​​of the angle and the length of the perpendicular line; 2) Record the coordinates of the target formation members in the measurement space In the parameter space, there exists a corresponding sine curve l i ; i = 1, 2, ..., T, T is the number of formation targets, the formation target set in the measurement space There exists a corresponding cluster of sinusoids in parameter space Get the whole polyline of the measurement space Calculate the corresponding sine curve intersection point l in parameter space i ∩l j =(θ k ,ρ k ), forming a data set X with n samples = {r1, r2, ..., r n }, r k Represents the kth sample in the data sample set X.

8. The method for selecting a formation recognition target based on deep learning according to claim 6, characterized in that: The steps to cluster coordinates using the ISODATA algorithm include: 1) Initialization parameters: set the desired number of clusters K and the minimum number of samples in a cluster θ N , the minimum distance between clusters θ C , standard deviation θ S , the maximum number of cluster pairs L for each merger and the number of allowed iterations I; 2) Select the initial cluster center: set the initial number of clusters to c and the initial cluster center to m i , assigned to c clusters according to the minimum distance principle; i = 1, 2, ..., c; If the data sample set X satisfies formula (15), all samples are divided into corresponding classes; ||X-m i ||<||X-m j ||,j=1,2,...,c;i≠j (15) In the formula, m i 、m j is the cluster center; 3) Re-clustering: If there is a cluster with a sample size N i <θ N , then discard cluster w i , cluster w i Merge with the closest cluster and set c = c-1; 4) Update the cluster center and calculate the cluster w i The distance between the sample data and its corresponding cluster center is m i The average distance D i , the total distance D from all sample data to their corresponding cluster centers, that is: Where N is the total number of samples; N i is the number of samples in the i-th cluster; 5) If the current number of iterations is equal to the number of allowed iterations I, let θ N =0, go to step 8); If the current iteration is an even number of iterations or c ≥ 2K, go directly to step 8); If the current iteration number Then proceed to step 6); 6) For each cluster w i , calculate the standard deviation vector σ based on the distance value i =(σ 1i , σ 2i , ..., σ ni ) T , and obtain the component σ with the largest standard deviation in each cluster imax ; where the components of the standard deviation vector are as follows: In the formula, is the jth component of the kth sample; m ij is the jth component of the i-th cluster center; σ ij is the jth component of the standard deviation of the i-th cluster; n is the dimension of sample X; 7) Split: For any σ imax , if there exists σ imax >θ S , and there is D i >D and N i >2(θ N +1), then cluster w i Split into 2 clusters, the cluster centers are and Cancel cluster center m i , and let c=c+1, go to step 1); Cluster Center and As shown below: In the formula, α, γ i is the coefficient; 8) Sorting: For all cluster centers, calculate the distance D between each pair ij D ij =||m i -m j ||,i=1,2,...,c-1;j=i+1,i+2,...c (21) Comparison D ij and θ c , D ij <θ c The values ​​are arranged in ascending order, namely: 9) Merge: From the smallest Initially, the distance Two types of cluster centers and Merge and calculate the new cluster center m l , and let c = c-1; In the formula, is the cluster center and The number of samples in ; 10) If the current number of iterations is equal to the allowed number of iterations I or the result converges, then the process ends; otherwise, the number of iterations is increased by 1 and the process goes to step 1).

Images