Aerial target track identification method based on LightGBM model

Abstract

The invention discloses an aerial target track identification method based on a LightGBM model. The method mainly solves the problems that a traditional method is low in identification accuracy, poor in efficiency and poor in reliability under the conditions of a complex electromagnetic environment and high target maneuverability. According to the scheme, the method comprises the following steps: 1) constructing a target track data set comprising correct track data and wrong track data; 2) constructing a gradient approximation strategy based on a divide-and-conquer method idea, and performing approximation reduction operation on gradient values in a target track data set; (3) training the LightGBM model by adopting the processed data set; 4) based on the trained model, identifying a correct target track from the to-be-determined target tracks; and 5) converting the original track data of the to-be-measured target into a predicted value matrix, optimizing the predicted value matrix, and obtaining a correct track of the aerial target according to an optimization result. According to the method, the correct track of the high-maneuverability target in the air can be accurately identified in a complex environment, and the probability of misjudgment and missed judgment of the target is effectively reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and further relates to the technology of track recognition. Specifically, it is an air target track recognition method driven by a LightGBM model, which can be used to identify the air target track in a ground analysis system to achieve air traffic control. Background Art

[0002] Track recognition refers to the technology of identifying the correct track based on the real-time or historical track data of an aircraft through a certain recognition method. For modern radars, target tracking is one of its most basic tasks. Air target track recognition is the process of establishing a track before the radar achieves stable target tracking. It is the primary problem and key component of the target tracking task. Due to the increasingly complex radar detection environment, especially the interference of strong clutter environment and the lack of relevant prior knowledge, the uncertainty of radar measurement has become very serious. The problem of radar target track recognition is gradually becoming the main problem in the research of radar data processing algorithms.

[0003] Currently, there are mainly two types of recognition methods based on radar measurement data: sequential processing methods represented by intuitive methods, logical methods and their improved algorithms, and batch processing methods represented by the Hough transform method and its improved algorithms. Among them, although the sequential processing method has the advantages of small computational complexity and fast recognition speed in a weak clutter environment and can achieve good recognition results, the sequential processing method is prone to the problem of "combinatorial explosion" of measurement data and has a serious dependence on prior knowledge; while the Hough transform method itself has good robustness and the transformation ability that is not affected by the dimensional space and the shape of the line. Therefore, the batch processing method can more accurately achieve track recognition in a strong clutter environment, but in a strong clutter environment, the computational complexity of the batch processing method is huge, resulting in low computational efficiency and inability to quickly identify the track. How to achieve fast and accurate track recognition in a strong clutter environment is the main problem faced by traditional track recognition methods.

[0004] In recent years, with the rapid development of machine learning, it has provided a brand-new idea for the problem of track recognition, converting the track recognition problem into the discrimination and differentiation problem between real and false tracks, that is, the binary classification problem of the machine learning model. By constructing a relevant sample data set to train the machine learning model to achieve the classification ability. This method can greatly reduce the dependence of the model on prior knowledge, thereby improving the robustness and adaptability of the model. In current research, the commonly adopted method is to use the heuristic rule method for track recognition. This technology uses kinematic laws for track recognition, judges the true and false tracks through feature information such as the speed and acceleration of the track. At the same time, to reduce the possibility of forming false tracks, a limit on the track yaw angle is added. However, although this method can adjust the threshold setting according to the known motion characteristics of the target, it is extremely dependent on past experience, the accuracy is difficult to guarantee, and the recognition speed will inevitably decrease during the adjustment process. Summary of the Invention

[0005] The purpose of the present invention is to propose an air target track recognition method based on the LightGBM model in view of the above-mentioned deficiencies of the existing technology. It is used to solve the technical problems of low accuracy, poor calculation efficiency and insufficient reliability of the recognition results of traditional air target track recognition methods in complex electromagnetic environments and under the condition of high maneuverability of targets. When training the LightGBM model, the present invention adopts an efficient and fast gradient boosting algorithm. Algorithms such as gradient-based unilateral sampling and exclusive feature bundling effectively reduce the computational complexity while maintaining the performance of the model; in addition, the growth strategy of giving priority to leaf nodes significantly improves the computational efficiency and memory utilization rate. The present invention trains the LightGBM model through the correct track data set and the incorrect track data set, then identifies the correct track through the trained model, and finally optimizes the track prediction value matrix. It has achieved a significant improvement in the accuracy, reliability and recognition efficiency of air target track recognition.

[0006] The specific steps for the present invention to achieve the above purpose are as follows:

[0007] (1) Construct a target track data set including correct track data P c and incorrect track data P e ;

[0008] (2) Based on the idea of the divide-and-conquer method, construct a gradient approximation strategy for approximately reducing the gradient value;

[0009] (3) Use the gradient approximation strategy to process the target track data set, and use the processed data to train the LightGBM model to obtain the final recognition model;

[0010] (4) Input the original track data of the target to be measured into the final recognition model, identify the correct target track through the model, and output the recognition result;

[0011] (5) Convert the original track data of the target to be measured into a prediction value matrix, and optimize the prediction value matrix by comparing the model recognition result with the original track data and replacing the incorrect track data in the prediction value matrix.

[0012] (6) Obtain the true track of the airborne target according to the optimized prediction value matrix.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] First, the present invention adopts the method driven by the LightGBM model, effectively integrates multi-source data such as radar cross-section area, and accurately characterizes its distribution by using the Gaussian function. Through a series of gradient calculations and divide-and-conquer optimization operations, the accuracy of airborne target track recognition is significantly improved. In a complex environment, compared with traditional methods, it can more accurately identify the correct track, greatly reducing the probability of target misjudgment and missed judgment, and solving the problem of low recognition accuracy of traditional methods in a complex electromagnetic environment and high target maneuverability.

[0015] Second, since the present invention performs gradient approximation through the idea of divide-and-conquer, the computational complexity is greatly reduced, and the running efficiency of the algorithm is improved. When processing large-scale track data, it can quickly complete the training and recognition process, saving a large amount of computational time and resources compared with traditional recognition methods, further ensuring that the airborne target track can be processed in a timely and efficient manner in practical applications, enhancing the real-time performance and practicality of the system, and enabling the system to better handle the track recognition task in a complex environment.

[0016] Third, the present invention further improves the reliability of the recognition result by optimizing the prediction value matrix. By comparing and adjusting with the maximum values of rows and columns, the final recognition result is more stable and accurate, avoiding the influence of local extreme values or outliers on the recognition result, improving the quality of track recognition again, and further consolidating the advantages of the present invention in solving the defects of traditional methods, providing more reliable technical support for air traffic management and military defense, etc. Description of the Drawings

[0017] Figure 1 It is a flowchart for implementing the method of the present invention. Detailed Embodiments

[0018] The present invention will be further described below with reference to the drawings.

[0019] Embodiment 1. Referring to the attached Figure 1 A method for identifying the track of an airborne target based on the LightGBM model proposed by the present invention includes the following steps:

[0020] Step 1: Construct a target track data set including correct track data P c and incorrect track data P e where:

[0021] P c ={(x c , y c ) | x c = f1(x1, y1, x 1t , y 1t ), y c = f2(x1, y1, x 1t , y 1t )};

[0022] In the formula, x c and y c respectively represent the abscissa and ordinate values of the predicted track points corresponding to the true track of the correct track in the Cartesian coordinate system. (x1, y1) represents the coordinates in the radar coordinate system of the correct track, (x 1t , y 1t ) represents the coordinates in the radar coordinate system of the correct track corresponding to the true track. f1 and f2 are functions for calculating the coordinates of the predicted track points of the correct track in the Cartesian coordinate system based on the coordinates (x1, y1) and (x 1t , y 1t ) respectively;

[0023] P e ={(x e , y e ) | x e = g1(x2, y2, x 2t , y 2t ), y e = g2(x2, y2, x 2t , y 2t )};

[0024] In the formula, x e and y e respectively represent the abscissa and ordinate values of the predicted track points corresponding to the true track of the incorrect track in the Cartesian coordinate system. (x2, y2) represents the coordinates in the radar coordinate system of the incorrect track, (x 2t , y 2t ) represents the coordinates in the radar coordinate system of the incorrect track corresponding to the true track. g1 and g2 are functions for calculating the coordinates of the predicted track points of the incorrect track in the Cartesian coordinate system based on the coordinates (x2, y2) and (x 2t , y 2t ) respectively.

[0025] Step 2: Construct a gradient approximation strategy based on the divide-and-conquer idea. This strategy is used to approximately reduce the gradient value. Specifically: First, use the Gaussian function to characterize the distribution of radar cross-sectional area data in the target track dataset, then calculate its gradient value, divide the gradient value based on the divide-and-conquer idea and generate a feature histogram, and finally calculate the gradient value of the histogram. Use the histogram gradient value to replace the sample gradient value, thereby reducing the computational complexity and improving the running efficiency of the algorithm.

[0026] Step 3: Use the gradient approximation strategy to process the target track dataset, and train the LightGBM model with the processed data to obtain the final recognition model;

[0027] Step 4: Input the original track data of the target to be measured into the final recognition model, identify the correct target track through the model, and output the recognition result;

[0028] Step 5: Convert the original track data of the target to be measured into a prediction value matrix. By comparing the model recognition result with the original track data and replacing the incorrect track data in the prediction value matrix, the prediction value matrix is optimized;

[0029] The optimization operation performed in this step of this embodiment is specifically to judge the size relationship between the predicted value in the model output matrix and the maximum value in the row or column where the predicted value matrix is located. If the predicted value is less than the maximum value in the row or column, replace the corresponding predicted value in the predicted value matrix with the predicted value in the model output matrix; otherwise, no replacement is performed; it is expressed as follows:

[0030]

[0031] In the formula, P opt [i][j] represents the optimized predicted value matrix, P[i][j] represents the predicted value in the i-th row and j-th column of the predicted value matrix, i represents the row index, j represents the column index, and k represents the traversed index.

[0032] Step 6: Obtain the true track of the airborne target according to the optimized predicted value matrix.

[0033] Embodiment 2. The overall implementation steps of the recognition method proposed in this embodiment are the same as those in Embodiment 1. Now, a further detailed description is made on the implementation of the gradient value approximation strategy in Step 2:

[0034] Step 2.1) Use the Gaussian function to characterize the distribution of radar cross-sectional area data in the target track dataset, and take the derivative of the radar cross-sectional area data to obtain the cross-sectional area data gradient value.

[0035] The above use of the Gaussian function to characterize the distribution of radar cross-sectional area data is specifically as follows:

[0036]

[0037] Among them, x is the radar cross-sectional area data, which follows a Gaussian distribution with a mean of μ and a variance of σ 2 .

[0038] The gradient value of the cross-sectional area data is obtained by differentiating the above Gaussian function f(x) and is expressed as follows:

[0039]

[0040] Step 2.2) Based on the divide-and-conquer idea, divide the gradient value into n equal parts, use it to generate a histogram, and obtain the gradient approximation G in the upper and lower limits l and u of each equal part according to the following formula approx (l, u):

[0041]

[0042] In the formula, F is the gradient value-related function;

[0043] Step 2.3) Use the gradient approximation to obtain the error ∈ of the histogram hist :

[0044]

[0045] In the formula, G approx (l z , u z ) represents the approximation of the gradient value of the current feature in the upper and lower limits of the Z-th equal-part histogram, and G true (l z , u z ) represents the true value of the gradient value of the current feature in the upper and lower limits of the Z-th equal-part histogram, where Z = 1, 2,..., n;

[0046] Step 2.4) Calculate the maximum error ΔG of the gradient change according to the following formula max :

[0047] ΔG max = max Z |G approx (l Z , u Z ) - G true (l Z , u Z )|;

[0048] Step 2.5) Compare ΔG max with the number of samples m in the interval where the current feature is located. If , the new splitting point of the divide-and-conquer method for this feature is Otherwise, proceed to the next histogram to compare ΔG max with comparisons until all histogram traversals are completed; where l Z represents the lower limit of the Z-th equal part, and u Z represents the upper limit of the Z-th equal part;

[0049] Step 2.6) Calculate the histogram gradient values in sequence, and replace the G of each sample in its corresponding interval true (l Z , u Z ) with the histogram G approx (l Z , u Z ), and input it into the LightGBM model for training.

[0050] In this embodiment, the gradient value of the feature histogram is calculated according to the following steps:

[0051] (2.6.1) Divide the feature space into m' intervals, let the width of each interval be Δk, and the central value of the i-th interval be denoted as k i , and the frequency within the interval be n i , then the feature histogram function h(k) is expressed as follows:

[0052]

[0053] In the formula, I is the indicator function, I = 1 when x falls within the interval , otherwise I = 0;

[0054] (2.6.2) Use the numerical difference method to calculate the gradient of h(k), that is, approximately calculate the gradient h′(k) of the feature histogram:

[0055]

[0056] (2.6.3) Replace the gradient value x of each sample within the corresponding histogram interval with the gradient value h′(k) of the histogram.

[0057] The present invention can identify the flight track in a timely and accurate manner, quickly understand the target information and make corresponding deployment arrangements, so as to enhance the military's rapid response and air traffic control capabilities. Aiming at the problem that after the radar obtains the flight track information of the air target, the traditional machine learning method is difficult to accurately and efficiently identify the target flight track in today's complex environment, and thus the key information of the air target cannot be obtained in a timely manner according to the flight track information, the present invention adopts the method driven by the LightGBM model, effectively integrates multi-source data such as radar cross section, and uses the Gaussian function to accurately characterize its distribution. Through a series of gradient calculations and divide-and-conquer optimization operations, the accuracy and efficiency of the air target flight track identification are significantly improved, and it has broad application prospects in the fields of air traffic control and national defense. The identification model adopted by the present invention is extremely fast in processing data features, effectively ensuring the identification speed of this method. Since the present invention has a fast flight track identification speed, it can be applied to the flight track identification field with high real-time requirements in the future. At the same time, this method has a small memory footprint and can also save valuable hardware resources for the device. In addition, flight track identification is an important part of ground intelligence analysis. This method can also be combined with radar sensors and cameras to form a more comprehensive air target analysis system when conducting research on flight track tracking, prediction, etc. in the future.

[0058] The parts not described in detail in the present invention belong to the common general knowledge of those skilled in the art.

[0059] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Obviously, for those skilled in the art, after understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention. However, these modifications and changes based on the idea of the present invention are still within the protection scope of the claims of the present invention.

Claims

1. A method for aerial target track recognition based on LightGBM model, characterized in that: Includes the following (1) Construct the correct track data P c and the error track data P e Target track dataset; (2) Based on the divide-and-conquer approach, a gradient approximation strategy is constructed to perform approximate reduction operations on gradient values; (3) The target track data set is processed using the gradient approximation strategy, and the LightGBM model is trained using the processed data to obtain the final recognition model; (4) Inputting the original track data of the target to be measured into the final recognition model, identifying the correct target track through the model, and outputting the recognition result; (5) Converting the original track data of the target to be measured into a prediction value matrix, optimizing the prediction value matrix by comparing the model recognition result with the original track data and replacing the erroneous track data in the prediction value matrix; (6) Based on the optimized prediction value matrix, the actual track of the aerial target is obtained.

2. The method according to claim 1, characterized in that: The correct track data P in the target track data set in step (1) c and the error track data P e , respectively, according to the following formula: P c ={(x c ,and c )∣x c =f1(x1,y1,x 1t ,and 1t ),and c =f2(x1,y1,x 1t ,and 1t )}; In the formula, x c and c They represent the horizontal and vertical coordinates of the predicted track point of the correct track corresponding to the true track in the Cartesian coordinate system, (x1, y1) represents the coordinates of the correct track in the radar coordinate system, (x 1t ,y 1t ) represents the coordinates of the correct track in the radar coordinate system corresponding to the true track, and f1 and f2 are respectively based on the coordinates (x1, y1) and (x 1t ,y 1t ) calculates the function of the coordinates of the correct track prediction track points in the Cartesian coordinate system; P e ={(x e ,y e )∣x e =g1(x2,y2,x 2t ,y 2t ),y e =g2(x2,y2,x 2t ,y 2t )}; In the formula, x e and e They represent the horizontal and vertical coordinates of the predicted track point of the error track corresponding to the true track in the Cartesian coordinate system, (x2, y2) represents the coordinates of the error track in the radar coordinate system, (x 2t ,y 2t ) represents the coordinates of the error track in the radar coordinate system corresponding to the true track, g1 and g2 are respectively based on the coordinates (x2, y2) and (x 2t ,y 2t ) is a function that calculates the coordinates of the predicted track points of the error track in the Cartesian coordinate system.

3. The method according to claim 1, characterized in that: The gradient value approximation strategy in step (2) is to first use a Gaussian function to characterize the distribution of radar cross-sectional area data in the target track data set, then calculate its gradient value, divide the gradient value based on the divide-and-conquer method and generate a feature histogram, and finally calculate the gradient value of the histogram, and use the histogram gradient value to replace the sample gradient value; the implementation steps are as follows: (2.1) Use the Gaussian function to characterize the distribution of radar cross-sectional area data in the target track data set, take the derivative of the radar cross-sectional area data, and obtain the gradient value of the cross-sectional area data; (2.2) Based on the divide-and-conquer method, the gradient value is divided into n equal parts, and a histogram is generated using it. The gradient approximation G in the upper and lower limits l and u of the gradient value in each equal part is obtained according to the following formula approx (l,u): Where F is the gradient value related function; (2.3) Using gradient approximation to obtain the histogram error ∈ hist : In the formula, G approx (l z ,u z ) represents the approximate value of the gradient value of the current feature in the upper and lower limits of the Z-th equal-division histogram, G true (l z ,u z ) represents the true value of the gradient value of the current feature in the upper and lower limits of the Z-th equal-division histogram, Z = 1, 2, ..., n; (2.4) The maximum error ΔG of the gradient change is calculated according to the following formula max : ΔG max =max Z |G approx (l Z ,u Z )-G true (l Z ,u Z )|; (2.5) max Compared with the number of samples m in the interval where the current feature is located, if When , the new partition point of the feature is On the contrary, proceed to the next histogram ΔG max and until all histogram traversals are completed; where l Z represents the lower limit of the Zth equal part, u Z represents the upper limit of the Zth equal part; (2.6) Calculate the histogram gradient values ​​one by one and assign them to the G of each sample in the corresponding interval true (l Z ,u Z ) is replaced by histogram G approx (l Z ,u Z ), and input it into the LightGBM model for training.

4. The method according to claim 3, characterized in that: In step (2.1), a Gaussian function is used to characterize the distribution of radar cross-sectional area data, as follows: Among them, x is the radar cross-sectional area data, which has a mean of μ and a variance of σ. 2 Gaussian distribution.

5. The method according to claim 4, characterized in that: The gradient value of the cross-sectional area data in step (2.1) is obtained by taking the derivative of the Gaussian function f(x), and is expressed as follows:

6. The method according to claim 3, characterized in that: The gradient value of the feature histogram calculated in step (2.6) is implemented as follows: (2.6.1) Divide the feature space into m' intervals, let the width of each interval be Δk, and the center value of the i-th interval be k i , the frequency in the interval is n i , then the characteristic histogram function h(k) is expressed as follows: Where I is the indicator function. When x falls within the interval When it is inside, I=1, otherwise I=0; (2.6.2) The numerical difference method is used to calculate the gradient of h(k), that is, the gradient n′(k) of the feature histogram is approximately calculated: (2.6.3) Replace the gradient value x of each sample in the corresponding histogram interval with the gradient value h′(k) of the histogram.

7. The method according to claim 1, characterized in that: In step (4), the correct target track is identified through the model. Specifically, based on the trained LightGBM model, the model output matrix corresponding to the target track data set is obtained, and the correct target track is found based on the model output track matrix.

8. The method according to claim 1, characterized in that: The prediction value matrix optimization in step (5) is specifically to determine the size relationship between the prediction value in the model output matrix and the maximum value in the row or column where the prediction value matrix is ​​located. If the prediction value is less than the maximum value in the row or column, the prediction value in the model output matrix is ​​used to replace the corresponding prediction value in the prediction value matrix; otherwise, no replacement is performed; it is expressed as follows: Where P opt [i][j] represents the optimized prediction value matrix, P[i][j] represents the prediction value of the i-th row and j-th column in the prediction value matrix, i represents the row index, j represents the column index, and k represents the traversed index.

Images