A target threat assessment model training method and evaluation method

Abstract

The present invention discloses a target threat assessment model training method and evaluation method, which uses an LSTM network to extract the temporal characteristics of threat assessment indicators, and combines the rankNet network structure to directly learn the threat ranking between two targets, making full use of the relative threat information between targets, and effectively improving the accuracy of target threat ranking of multiple targets.

Description

Technical Field

[0001] The present invention relates to the technical field of threat level assessment, and in particular to a target threat level assessment model training method and an assessment method that can fully utilize relative threat level information between two aerial targets. Background Art

[0002] With the application of high-tech and equipment in modern warfare, terminal defense air strike targets have shown characteristics such as a wide variety of targets and all-round multi-batch saturation attacks. At the same time, with the development of sensor technology and collaborative detection and perception technology, battlefield situation awareness capabilities have been significantly improved, and the complexity of situation information that the command and control system needs to process has increased significantly. The situation information contains a large amount of uncertainty and nonlinear information, and presents big data characteristics, which brings new challenges to the terminal defense air target threat assessment technology.

[0003] Traditional deterministic and uncertain multi-attribute target threat assessment methods have shortcomings such as self-learning and insufficient real-time assessment. However, the powerful nonlinear information processing capabilities of machine learning and deep learning can deeply explore the intrinsic relationship between target threat level and evaluation indicators, simplify the threat assessment modeling process, and more easily accept new data for continued training and optimization. They have become important methods in the field of multi-target threat assessment.

[0004] Currently available threat assessment method models based on deep learning and machine learning all establish threat assessment models by learning the relationship between target threat assessment indicators and target threat levels or threat assessment levels. This method is a typical Point Wise ranking method, which achieves target threat level ranking by accurately scoring the target threat level.

[0005] The current threat assessment model based on machine learning and deep learning only establishes a threat assessment model by learning the relationship between target threat assessment indicators and target threat level or threat assessment level, ignoring the relationship between the relative threat level of different targets and the target threat assessment indicators, which in turn affects the accuracy of threat assessment. Summary of the Invention

[0006] In view of the above problems, the present invention provides a target threat assessment model training method and assessment method that overcome the above problems or at least partially solve the above problems.

[0007] The present invention provides the following solutions:

[0008] A target threat assessment model training method, comprising:

[0009] Acquiring a first photoelectric image and first track information of an incoming target, wherein the first photoelectric image and the first track information both include measurement errors;

[0010] Preprocessing the first photoelectric image and the first track information to obtain first image data and a target motion state estimation value;

[0011] Identifying the category of the incoming target using the first image data and the target motion state estimation value to obtain category information;

[0012] Calculating an estimated threat level feature of the incoming target by combining the threat level influencing factors of the incoming target and the threat level membership functions corresponding to the factors; the threat level influencing factors include the category information;

[0013] Acquire a second photoelectric image and second track information of the incoming target; the second photoelectric image and the second track information are both data without adding noise and containing real labels;

[0014] Acquire a true threat level characteristic of the incoming target based on the second photoelectric image and the second track information; and calculate a true threat level using the true threat level characteristic.

[0015] Combining the estimated threat level feature with the actual threat level to obtain a threat level assessment data set;

[0016] Randomly extracting two incoming target samples from the threat assessment dataset, combining the threat level features and labels of the two incoming targets into sample pairs, repeating the sample pair construction process to obtain a threat level assessment sample pair dataset, and inputting the sample pair dataset into a rankNet network in the threat level assessment model so that the threat level assessment model learns the threat level ranking between the two incoming targets;

[0017] Obtaining the true labels learned by the threat level assessment model and the predicted labels of the threat level assessment model, and constructing an objective function for optimizing the threat level assessment model using the true labels and the predicted labels;

[0018] The target threat assessment model is obtained by optimizing the objective function through a gradient descent method.

[0019] Preferably, the objective function is expressed by the following formula:

[0020]

[0021] Where: P ij is the true label, is the predicted label.

[0022] Preferably, after the sample pair passes through the feature extraction layer composed of the LSTM layer and the fully connected layer, two deep features S are obtained. i and S j, the two deep features are passed through the Sigmod layer to obtain the predicted label, which is expressed by the following formula:

[0023]

[0024] Where: β is the model hyperparameter.

[0025] Preferably, preprocessing the first photoelectric image and the first track information separately includes:

[0026] The first photoelectric image information is cropped and data normalized to obtain the first image data; the first track information is coordinate transformed, and a recursive least squares method is used to remove outliers, fill missing values, and then perform Kalman filtering to obtain the target motion state estimation value.

[0027] Preferably, identifying the category of the incoming target using the first image data and the target motion state estimation value includes:

[0028] The target motion state estimation value and the first image data are respectively input into a trajectory feature attribute recognition model based on LSTM and a photoelectric image attribute recognition model based on Res2Net, and the recognition probabilities of the two models for the target category are obtained respectively. The category information is obtained using the DS evidence fusion method.

[0029] Preferably, the probability distribution of the target attribute category after fusion is expressed as follows:

[0030]

[0031] Where: m1(A i ) and m2(A j ) represent the probability that the target attribute is predicted to be of category i and category j by the photoelectric image target attribute recognition model and the trajectory feature target attribute recognition model, respectively; K represents the degree of conflict between the evidences; 1 / (1-K) is the normalization factor to ensure the consistency of fusion quality.

[0032] Preferably, the threat level influencing factors include target category, track shortcut, flight altitude, flight speed, and approach time.

[0033] Preferably, the threat level score of the target category is:

[0034]

[0035] The threat degree membership function of the track shortcut is expressed as follows:

[0036] W p =e -kp

[0037] p=R sin(π-θ)

[0038] Where: R is the target distance, θ is the target entry angle, and the coefficient k is 10 -3 ;

[0039] The threat degree membership function of the flight altitude is expressed by the following formula:

[0040]

[0041] Where: h is the target flight altitude, coefficients k and a are 10 -3 and 50;

[0042] The threat degree membership function of the flight speed is expressed by the following formula:

[0043]

[0044] Where: v is the target flight speed, the coefficients k and a are 10 5 and 5;

[0045] The threat degree membership function of the approach time is expressed by the following formula:

[0046] W T =e -kt

[0047]

[0048] Where: t is the target approach time, and the coefficient k is 10 -2 .

[0049] Preferably, the method of calculating the real threat level using the real threat level feature includes:

[0050] The real threat level is obtained by linearly weighting the real threat level features, and the real threat level is obtained by dividing the real threat level according to a preset threat level division rule.

[0051] A target threat assessment method, comprising:

[0052] Obtain target electro-optical images and target track information of each real incoming target collected on the actual battlefield;

[0053] Obtaining target threat characteristics of each of the real incoming targets according to the target electro-optical image and the target track information;

[0054] Each of the target threat level features is input into the target threat level assessment model to obtain a target threat level ranking.

[0055] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0056] The embodiments of the present application provide a target threat assessment model training method and evaluation method, which uses an LSTM network to extract the time series characteristics of the threat assessment indicator, and combines the rankNet network structure to directly learn the threat ranking between two targets, making full use of the relative threat information between the targets, and effectively improving the accuracy of the target threat ranking of multiple targets.

[0057] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0059] Figure 1 This is a flow chart of a target threat assessment model training method provided by an embodiment of the present invention;

[0060] Figure 2 This is a flow chart of target attribute recognition provided by an embodiment of the present invention;

[0061] Figure 3 1 is a schematic diagram of the rankNet threat assessment network training provided by an embodiment of the present invention;

[0062] Figure 4 Schematic diagram of an aerial target threat assessment method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of the present invention.

[0064] See also Figure 1 , is a target threat assessment model training method provided by an embodiment of the present invention, such as Figure 1 As shown, the method includes:

[0065] S101: Acquire a first photoelectric image and first track information of an incoming target; the first photoelectric image and the first track information both include measurement errors;

[0066] S102: Preprocessing the first photoelectric image and the first track information to obtain first image data and a target motion state estimation value;

[0067] S103: using the first image data and the target motion state estimation value to identify the category to which the incoming target belongs, and obtain category information;

[0068] S104: Calculating an estimated threat level feature of the incoming target by combining the threat level influencing factors of the incoming target with the threat level membership functions corresponding to the factors; the threat level influencing factors include the category information;

[0069] S105: Acquire a second photoelectric image and second track information of the incoming target; the second photoelectric image and the second track information are both data without adding noise and containing real labels;

[0070] S106: Obtaining a true threat level characteristic of the incoming target based on the second photoelectric image and the second track information; and calculating a true threat level using the true threat level characteristic.

[0071] S107: Combining the estimated threat level feature with the actual threat level to obtain a threat level assessment data set;

[0072] S108: Randomly extract two incoming target samples from the threat assessment dataset, combine the threat level features and labels of the two incoming targets into a sample pair, repeat the sample pair construction process to obtain a threat level assessment sample pair dataset, and input the sample pair dataset into the rankNet network in the threat level assessment model so that the threat level assessment model learns the threat level ranking between the two incoming targets;

[0073] S109: Acquire the true labels learned by the threat level assessment model and the predicted labels of the threat level assessment model, and construct an objective function for optimizing the threat level assessment model using the true labels and the predicted labels;

[0074] S1010: Optimizing the objective function through the gradient descent method to obtain the target threat assessment model.

[0075] The target threat assessment model training method provided in the embodiment of the present application uses a target image and target track database to obtain a multi-target threat assessment dataset through data preprocessing, target attribute recognition, target threat feature extraction, and target real threat extraction steps. The obtained multi-target threat assessment dataset is used to train a rankNet-based threat assessment network. The rankNet-based threat assessment network is deployed to perform data preprocessing, target attribute recognition, and target threat feature extraction steps on the optoelectronic image data and radar track data of each target obtained in the actual battlefield. The obtained target threat feature data is then input into the rankNet-based threat assessment network to obtain a target threat ranking.

[0076] Furthermore, the objective function is expressed as follows:

[0077]

[0078] Where: P ij is the true label, is the predicted label.

[0079] After the sample pair passes through the feature extraction layer composed of the LSTM layer and the fully connected layer, two deep features S are obtained. i and S j , the two deep features are passed through the Sigmod layer to obtain the predicted label, which is expressed by the following formula:

[0080]

[0081] Where: β is the model hyperparameter.

[0082] The preprocessing of the first photoelectric image and the first track information respectively includes:

[0083] The first photoelectric image information is cropped and data normalized to obtain the first image data; the first track information is coordinate transformed, and a recursive least squares method is used to remove outliers and fill missing values ​​before performing Kalman filtering to obtain the target motion state estimation value.

[0084] The step of using the first image data and the target motion state estimation value to identify the category to which the incoming target belongs and obtain category information includes:

[0085] The target motion state estimation value and the first image data are respectively input into a trajectory feature attribute recognition model based on LSTM and a photoelectric image attribute recognition model based on Res2Net, and the recognition probabilities of the two models for the target category are obtained respectively. The category information is obtained using the DS evidence fusion method.

[0086] The probability distribution of the target attribute category after fusion is expressed as follows:

[0087]

[0088] Where: m1(A i ) and m2(A j ) represent the photoelectric image target attribute recognition model and the trajectory feature target attribute recognition model, respectively, and predict the probability of the target attribute being of the i-th category and the j-th category; K represents the degree of conflict between the evidences; 1 / (1-K) is the normalization factor to ensure the consistency of the fusion quality.

[0089] Furthermore, the threat level influencing factors include target category, track shortcut, flight altitude, flight speed, and approach time. The threat level score of the target category is:

[0090]

[0091] The threat degree membership function of the track shortcut is expressed as follows:

[0092] W p =e -kp

[0093] p=R sin(π-θ)

[0094] Where: R is the target distance, θ is the target entry angle, and the coefficient k is 10 -3 ;

[0095] The threat degree membership function of the flight altitude is expressed by the following formula:

[0096]

[0097] Where: h is the target flight altitude, coefficients k and a are 10 -3 and 50;

[0098] The threat degree membership function of the flight speed is expressed by the following formula:

[0099]

[0100] Where: v is the target flight speed, the coefficients k and a are 10 5 and 5;

[0101] The threat degree membership function of the approach time is expressed by the following formula:

[0102] W T =e -kt

[0103]

[0104] Where: t is the target approach time, and the coefficient k is 10 -2 .

[0105] The method of calculating the real threat level using the real threat level feature includes:

[0106] The real threat level is obtained by linearly weighting the real threat level features, and the real threat level is obtained by dividing the real threat level according to a preset threat level division rule.

[0107] The present application may also provide a target threat assessment method, including:

[0108] Obtain target electro-optical images and target track information of each real incoming target collected on the actual battlefield;

[0109] Obtaining target threat characteristics of each of the real incoming targets according to the target electro-optical image and the target track information;

[0110] Each of the target threat level features is input into the target threat level assessment model to obtain a target threat level ranking.

[0111] The following combination Figure 4 , a detailed introduction is given to the target threat assessment model training method and assessment method provided in the embodiments of the present application.

[0112] Step 1: Obtain the first photoelectric image and first track information of the incoming target from the incoming target image database and the incoming target track database.

[0113] There is a random measurement error in the target track measurement of the real radar. The disturbance error can be expressed as a mean of 0 and a variance of σ xr is a normally distributed random variable. For the target observation quantity track observation quantity x, assume that its true value is x u , then when simulating a real radar, the target observation value satisfies:

[0114]

[0115] Step 2: Perform data preprocessing on the first photoelectric image and the first track information.

[0116] The processing process includes coordinate transformation of track information, outlier elimination and missing value filling using recursive least squares method, and finally Kalman filtering to obtain stable target track data information, as well as cropping and data normalization of target photoelectric image information to obtain standard image information.

[0117] Step 2.1: Transform the track information coordinates and use the recursive least squares method to remove outliers and fill in missing values.

[0118] The original data of the target track is usually expressed in spherical coordinates as azimuth angle β(k), elevation angle ε(k) and slant distance d(k). First, the target coordinates in spherical coordinates are converted to rectangular coordinates x(k), y(k) and z(k):

[0119]

[0120] Then, the recursive least squares method as shown in formula (2) is used to eliminate outliers and fill in missing values.

[0121]

[0122] remember and is the one-step forecast value after least squares estimation, if: or or Then the measured value at time k+1 is removed as an outlier and filled with the one-step predicted value. β , σ ε and σ d They are the standard deviations of the tracking equipment's azimuth, elevation, and slant distance measurement errors.

[0123] Step 2.2: Track Kalman filtering.

[0124] According to the target's current track position observation information and the target's estimated state at the previous moment, the target is Kalman filtered according to equations (3) to (7) to obtain a new target motion state estimate.

[0125]

[0126]

[0127]

[0128]

[0129]

[0130] Where: F k is the state transfer matrix, H is the measurement matrix, R k Measure the noise for the target, Q k is the target maneuvering noise, is the posterior state estimate, is the posterior covariance estimate, and K is the filter gain.

[0131] Step 2.3: Crop and normalize the first photoelectric image. Use a segmentation algorithm to preliminarily determine the target location in the photoelectric image, and crop the first image to a uniform size with the target as the center. Then, normalize the first image information matrix.

[0132] Step 3: Target attribute identification. The target category attribute directly affects the accuracy of the target threat assessment. Therefore, it is necessary to use the target estimated motion state and target optoelectronic image data obtained in step 2 to identify the target category.

[0133] like Figure 2 As shown in Figure 2, the estimated motion state and target image data obtained in step 2 are respectively input into the trajectory feature attribute recognition model based on LSTM and the photoelectric image attribute recognition model based on Res2Net, and the recognition probabilities of the two models for the target category are obtained respectively. Then, the DS evidence fusion method is used to obtain the target category.

[0134] Assume that the basic probability distribution function provided by the photoelectric image target attribute recognition model is the probability m1(A i ), the basic probability distribution function provided by the trajectory feature target attribute recognition model is the probability m2(A j ).

[0135] Then the probability distribution of the target attribute category after fusion is:

[0136]

[0137] Where m1(A i ) and m2(A j ) represent the photoelectric image target attribute recognition model and the trajectory feature target attribute recognition model, respectively, and predict the probability of the target attribute being of class i and class j. K represents the degree of conflict between the evidence and is called the conflict factor. 1 / (1-K) is a normalization factor to ensure the consistency of fusion quality. K is calculated as:

[0138]

[0139] Step 4: Extract target threat feature.

[0140] The target threat level is related to factors such as the target category of the incoming target, target track shortcut, target flight altitude, target flight speed, and target approach time.

[0141] The target category belongs to qualitative threat factors and uses expert prior scoring:

[0142]

[0143] The target track shortcut combines the influence of target distance and target entry angle. The closer the target distance, the greater the target threat. The target entry angle is the angle between the target flight speed direction and the target position direction. The smaller the entry angle, the stronger the target attack intention and the greater the target threat. The threat degree membership of the track shortcut can be expressed as:

[0144]

[0145] Where: R is the target distance, θ is the target entry angle, and the coefficient k is 10 -3 .

[0146] The lower the target's flight altitude, the higher the probability of a surprise attack and the success rate of the attack, and therefore the greater the corresponding threat. The threat degree membership function of the target's flight altitude is expressed as:

[0147]

[0148] Where: h is the target flight altitude, coefficients k and a are 10 -3 and 50.

[0149] The greater the target flight speed, the higher the target threat level. The membership function of the target flight speed is expressed as:

[0150]

[0151] Where: v is the target flight speed, the coefficients k and a are 10 5 and 5.

[0152] The target approach time reflects the urgency of intercepting the target. The shorter the target approach time, the greater the target threat. The threat degree membership function of the target approach time is expressed as:

[0153]

[0154] Where: t is the target approach time, and the coefficient k is 10 -2 .

[0155] Substituting the target motion state parameters obtained in step 2 and the target category obtained in step 3 into equations (10) to (14) can obtain the estimated threat level features.

[0156] Step 5: Obtain the target's true threat level.

[0157] The second track information without adding noise obtained from the database in step 1 and the target's true label are processed through steps 2 and 4 to obtain the target's true threat level feature. The true threat level feature is linearly weighted to obtain the target's true threat level. The target's true threat level W is:

[0158]

[0159] Wherein: weight coefficients w1, w2, w3, w4, and w5 are 0.2, 0.2, 0.2, 0.4, and 0.3 respectively. When allocating firepower and sorting interception of targets, the target threat level is usually divided into threat level levels. In the embodiment of the present application, the target threat level is divided into 10 levels:

[0160] Table 1 Threat level classification

[0161] Target threat level Target Threat Level Target threat level Target Threat Level 9＜W≤10 10 4＜W≤5 5 8＜W≤9 9 3＜W≤4 4 7＜W≤8 8 2＜W≤3 3 6＜W≤7 7 1＜W≤2 2 5＜W≤6 6 0＜W≤1 1

[0162] The target's real threat level is divided according to Table 1 to obtain the target's real threat level.

[0163] Step 6: Construct a target threat assessment dataset. The target threat assessment dataset is obtained by combining the estimated threat features obtained in step 4 with the target's true threat level obtained in step 5.

[0164] Step 7: Use the target threat assessment dataset to train the rankNet threat assessment network.

[0165] Randomly extract two samples from the threat assessment dataset to form a sample pair, then the input sample form of the rankNet network is similar to {(x i ,y i ),(x j ,y j )}, the model learns the threat ranking between the two targets, and the true label P ij The calculation method is:

[0166]

[0167] like Figure 3 As shown, after the sample pair passes through the feature extraction layer composed of LSTM layer and fully connected layer, two deep features s are obtained. i and s j , the two deep features pass through the Sigmod layer to obtain the labels predicted by the model:

[0168]

[0169] Where: is the β model hyperparameter, generally set to 1.

[0170] The objective function of model optimization can be expressed as:

[0171]

[0172] By optimizing the objective function through the gradient descent method, the rankNet threat assessment network can be obtained.

[0173] Step 8: Deploy a RankNet-based threat assessment network to implement target threat assessment. In practice, target electro-optical image information and target track information are obtained from the electro-optical system and radar, respectively. These information undergoes data preprocessing, target attribute identification, and target threat feature extraction. The obtained target threat feature data is then fed into the RankNet-based threat assessment network, and the final target threat ranking is obtained based on the RankNet network's prediction results.

[0174] In order to evaluate the performance of the threat assessment model based on rankNet proposed in the embodiment of this application, the comparison model adopts a typical Point Wise ranking model based on a single LSTM. The number of neurons in each layer of this network is set to be consistent with the LSTM network in the rankNet network, the threat assessment data set used for training is consistent, the loss function adopts the cross-entropy loss function, and the gradient descent method is used for training.

[0175] The model's performance was evaluated through simulations using 1,000 sets of target tracks and electro-optical images randomly selected from a database not previously used in model training. Each set of incoming targets included six target categories: cruise missiles, guided bombs, air-to-ground missiles, armed helicopters, small and medium-sized unmanned aerial vehicles, and fixed-wing aircraft. During simulation, noise was added to the target tracks, with 8 mils of noise in azimuth and elevation and 12 meters of noise in slant range. The information was then subjected to data preprocessing, target attribute identification, and target threat feature extraction. The extracted features were then fed into a Point Wise ranking model and a RankNet-based threat ranking model. The evaluation rankings from both models were collected and compared with the actual target threat rankings to calculate the threat ranking accuracy. The threat ranking accuracy is expressed as the ratio of the number of correctly ranked samples to the total number of simulated samples.

[0176] The simulation results show that when the total threat level calculation method in step 5 is directly used without adopting the threat level assessment model, the ranking accuracy is 63%. This is because the existence of track measurement noise leads to incorrect target attribute recognition of some targets and inaccurate target motion state estimation, which leads to errors in threat level factors such as target track shortcut, target flight altitude, target flight speed, and target approach time, ultimately resulting in low target threat level ranking accuracy.

[0177] When using the Point Wise sorting model based on a single LSTM, the sorting accuracy is 75%, which shows that the deep learning model has deeply mined the relationship between the target threat index and the target threat level to a certain extent, and has a certain degree of robustness.

[0178] The evaluation ranking accuracy of the rankNet-based threat assessment model proposed in the embodiment of the present application is 90%, which shows that the relative threat levels between different targets have a direct impact on the threat ranking accuracy. The rankNet-based threat assessment method provided in the embodiment of the present application can improve the accuracy of multi-target threat assessment.

[0179] In summary, the target threat level assessment method provided in the embodiment of the present application uses an LSTM network to extract the time series characteristics of the threat level assessment indicator, and at the same time combines the rankNet network structure to directly learn the threat level ranking between two targets, making full use of the relative threat level information between the targets, and effectively improving the accuracy of the target threat level ranking of multiple targets.

[0180] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0181] Through the description of the above embodiments, it can be seen that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present application.

[0182] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0183] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A target threat assessment model training method, characterized in that: include: Acquire a first photoelectric image and first track information of an incoming target; The first photoelectric image and the first track information both include measurement errors; Preprocessing the first photoelectric image and the first track information to obtain first image data and a target motion state estimation value; Identifying the category of the incoming target using the first image data and the target motion state estimation value to obtain category information; Calculating an estimated threat level feature of the incoming target by combining the threat level influencing factors of the incoming target and the threat level membership functions corresponding to the factors; the threat level influencing factors include the category information; Acquire a second photoelectric image and second track information of the incoming target; The second photoelectric image and the second track information are both data without adding noise and containing real labels; Acquire a true threat level characteristic of the incoming target based on the second photoelectric image and the second track information; and calculate a true threat level using the true threat level characteristic. Combining the estimated threat level feature with the actual threat level to obtain a threat level assessment data set; Randomly extracting two incoming target samples from the threat level assessment dataset, combining the threat level features and labels of the two incoming targets into a sample pair, repeating the sample pair construction process to obtain a threat level assessment sample pair dataset, and inputting the sample pair dataset into a rankNet network in the threat level assessment model so that the threat level assessment model learns the threat level ranking between the two incoming targets; Obtaining the true labels learned by the threat level assessment model and the predicted labels of the threat level assessment model, and constructing an objective function for optimizing the threat level assessment model using the true labels and the predicted labels; The target threat assessment model is obtained by optimizing the objective function through a gradient descent method.

2. The target threat assessment model training method according to claim 1, characterized in that: The objective function is expressed as follows: Where: P ij is the true label, is the predicted label.

3. The target threat assessment model training method according to claim 2, characterized in that: After the sample pair passes through the feature extraction layer composed of LSTM layer and fully connected layer, two deep features are obtained. and , the two deep features are passed through the Sigmod layer to obtain the predicted label, which is expressed by the following formula: Where: β is the model hyperparameter.

4. The target threat assessment model training method according to claim 1, characterized in that: Preprocessing the first photoelectric image and the first track information respectively includes: The first photoelectric image information is cropped and data normalized to obtain the first image data; the first track information is coordinate transformed, and a recursive least squares method is used to remove outliers, fill missing values, and then perform Kalman filtering to obtain the target motion state estimation value.

5. The target threat assessment model training method according to claim 1, characterized in that: Identifying the category of the incoming target using the first image data and the target motion state estimation value includes: The target motion state estimation value and the first image data are respectively input into a trajectory feature attribute recognition model based on LSTM and a photoelectric image attribute recognition model based on Res2Net, and the recognition probabilities of the two models for the target category are obtained respectively. The category information is obtained using the DS evidence fusion method.

6. The target threat assessment model training method according to claim 5, characterized in that: The probability distribution of the target attribute category after fusion is expressed as follows: Where: and They represent the target attributes predicted by the photoelectric image target attribute recognition model and the trajectory feature target attribute recognition model as the first i Class and j class probability; K Represents the degree of conflict between evidences; 1 / (1− K ) is a normalization factor to ensure the consistency of fusion quality.

7. The target threat assessment model training method according to claim 1, characterized in that: The threat level influencing factors include target category, track shortcut, flight altitude, flight speed, and approach time.

8. The target threat assessment model training method according to claim 7, characterized in that: The threat level score of the target category is: The threat degree membership function of the track shortcut is expressed as follows: Where: R is the target distance, is the target entry angle, and the coefficient k is 10 -3 ; The threat degree membership function of the flight altitude is expressed by the following formula: Where: h is the target flight altitude, coefficients k and a are 10 -3 and 50; The threat degree membership function of the flight speed is expressed by the following formula: Where: v is the target flight speed, the coefficients k and a are 10 5 and 5; The threat degree membership function of the approach time is expressed by the following formula: Where: t is the target approach time, and the coefficient k is 10 -2 .

9. The target threat assessment model training method according to claim 1, characterized in that: Calculating the real threat level using the real threat level feature includes: The real threat level is obtained by linearly weighting the real threat level features, and the real threat level is obtained by dividing the real threat level according to a preset threat level division rule.

10. A target threat assessment method, characterized in that: include: Obtain target electro-optical images and target track information of each real incoming target collected on the actual battlefield; Obtaining target threat characteristics of each of the real incoming targets according to the target electro-optical image and the target track information; Input each of the target threat level features into the evaluation model obtained by the target threat level evaluation model training method according to any one of claims 1 to 9 to obtain the target threat level ranking.

Images