Track feature-based low-altitude small and micro unmanned aerial vehicle target intelligent identification method

Abstract

The invention discloses a low-altitude small and micro unmanned aerial vehicle target intelligent identification method based on track characteristics. The method specifically comprises the steps that multi-dimensional characteristics capable of effectively representing target characteristics are designed for track data of a low-altitude small and micro unmanned aerial vehicle; performing data enhancement on the track data by combining a prime number step sampling method with a sliding window method; carrying out feature calculation on the enhanced track data, processing abnormal values and missing values, and then normalizing and standardizing the data; performing feature importance analysis by using an AutoGluon framework, screening out key features, and constructing an AutoGluon unmanned aerial vehicle classifier by using the key features; an iTransform time sequence classifier is constructed on the basis of the enhanced flight path data, the iTransform time sequence classifier is fused with an AutoGluon unmanned aerial vehicle classifier, and an optimal recognition model is obtained through training iteration; the method can realize rapid and accurate identification of low-altitude small and micro unmanned aerial vehicles, has high identification rate and strong adaptability, and can be widely applied to the field of low-altitude unmanned aerial vehicle detection and defense.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent recognition of unmanned aerial vehicle targets, and in particular relates to an intelligent recognition method for low-altitude small and micro unmanned aerial vehicle targets based on track characteristics. Background Art

[0002] With the rapid development of drone technology, it has been widely used in military and civilian applications due to its low cost, small size and easy operation. From agricultural monitoring, disaster relief to commercial photography, the scope of drone use continues to expand, greatly improving work efficiency and convenience. However, this technological advancement has also brought many new challenges and safety issues, especially in important places such as government agencies, commercial centers, airports and transportation hubs. Low-altitude safety management has become a difficult problem waiting to be solved.

[0003] Small and micro drones are easily confused with natural objects (such as birds and balloons) in low-altitude environments or people and vehicles on the ground due to their small size and flexible flight, which poses a huge challenge to existing monitoring technology and management systems. Illegal drone flight activities may not only infringe on personal privacy and disrupt normal social order, but may even become a tool for terrorist attacks or crimes, seriously threatening public safety and social stability. In order to effectively deal with these problems and ensure the safety of low-altitude areas, it is necessary to quickly and accurately identify drone targets and non-drone targets through radar point traces and track data when suspicious targets are discovered. To this end, it is crucial to develop efficient monitoring and identification technologies.

[0004] At present, the methods commonly used by researchers in the field of drone target recognition include support vector machines, random forests, and convolutional neural networks. These methods can achieve target detection and classification tasks to a certain extent, but they still have many limitations in practical applications. For example, the model optimization process is highly dependent on data quality and quantity, and often shows performance degradation or even misjudgment when faced with serious data imbalance problems. Another prominent problem is the lack of generalization ability. Many existing models perform well when trained on specific scenarios or data sets, but when applied to different environments or new target categories, they often show a significant decrease in accuracy. This lack of generalization performance makes it difficult for the model to meet the needs of actual low-altitude safety management, especially when there are many types of drones and changing environments. Therefore, researchers need to develop efficient and robust data enhancement strategies to increase the diversity and scale of training data and effectively deal with the problem of data category imbalance. On this basis, it is necessary to further explore and develop a more generalized model architecture to ensure that the model can still maintain good prediction results when facing unseen data. The combination of data enhancement technology and model architecture optimization can effectively promote the development of drone recognition technology. Summary of the invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention proposes a method, device and storage medium for intelligent identification of low-altitude small and micro UAV targets based on track characteristics, so as to meet people's needs for low-altitude UAV supervision and realize accurate identification and real-time monitoring of low-altitude small and micro UAVs; at the same time, it has high recognition rate and strong adaptability, and is suitable for the field of low-altitude UAV monitoring and defense.

[0006] In order to achieve the above technical objectives, the present invention provides the following technical solutions:

[0007] The intelligent identification method of low-altitude small and micro UAV targets based on track characteristics specifically includes the following steps:

[0008] S1. Design a multi-dimensional feature map to effectively characterize the target to be identified based on the track data of low-altitude micro-UAVs;

[0009] S2, data enhancement of the track data set is performed by combining the prime number step sampling method with the sliding window method;

[0010] S3, calculating the features of the track data set after data enhancement, processing the outliers and missing values ​​of the features, and then normalizing the processed track features;

[0011] S4. Use the AutoGluon framework to perform importance analysis on the normalized track features to select key feature sets to optimize model performance;

[0012] S5. Using the selected key feature set as input, the AutoGluon UAV classifier is constructed through the AutoGluon framework, the key feature set is divided into a training set and a validation set, the optimal model is trained, and the average accuracy of the AutoGluon UAV classifier is calculated;

[0013] S6. Based on the enhanced track data set, the iTransformer time series classifier is reconstructed and combined with the constructed AutoGluon UAV classifier, comprehensively considering the recognition ability and accuracy of the two models to achieve accurate recognition of low-altitude targets.

[0014] Furthermore, step S1 specifically includes:

[0015] S11. Arrange and summarize the radar data used in the UAV identification method based on track data in the public literature; the radar data includes six basic information: the target's azimuth, slant range, relative altitude, radial velocity, recording time and radar cross section RCS;

[0016] S12. Based on radar data, design features including acceleration dimension features, velocity dimension features, heading angle dimension features and relative height dimension features to distinguish the differences between UAVs and non-UAVs in flight trajectories and motion modes in a fine-grained manner; the acceleration dimension features include the maximum radial acceleration, the maximum radial acceleration value, and the acceleration change; the velocity dimension features include the maximum radial velocity, the minimum radial velocity, the maximum radial velocity value, the minimum radial velocity value, the average speed, the speed standard deviation, the speed oscillation frequency, the maximum speed, and the minimum speed; the heading angle dimension features include the heading angle mean value and the heading angle standard deviation; the relative height dimension features include the relative height mean value and the relative height standard deviation;

[0017] S13. Design RCS dimensional features based on radar data to distinguish the differences between specific composite materials, metal alloys and natural object materials used by drones in a coarse-grained manner; the RCS dimensional features include average RCS and RCS range.

[0018] Furthermore, step S2 specifically includes:

[0019] S21, using the sliding window method to generate new sub-tracks for single track data; the size of the sliding window is set to 5, 6, 7, 8, 9, 10, and 11 in sequence, and the window is gradually slid to intercept sub-tracks of different scales to fully capture the multi-dimensional characteristics of the track;

[0020] S22, on the basis of the sub-tracks generated by the sliding window, setting the sliding window step size to a prime number that increases sequentially, and obtaining multiple differentiated sub-tracks by multiple sampling;

[0021] S23, integrating the sampled sub-tracks, removing the repeated sub-tracks, and obtaining a track data set after data enhancement.

[0022] Furthermore, step S3 specifically includes:

[0023] S31, calculating each dimensional feature in each track data, and integrating each dimensional feature into a feature vector;

[0024] S32, remove outliers caused by dirty data in multi-dimensional features, including speed standard deviation infinite, heading angle standard deviation infinite, relative height standard deviation infinite;

[0025] S33, checking whether the obtained feature vector has any missing values, and if any missing values ​​are found, deleting the corresponding entire data;

[0026] S34. Use maximum and minimum normalization to normalize the features of each dimension.

[0027] Furthermore, step S4 specifically includes:

[0028] S41. Use the TabularDataset function in the AutoGluon library to load the track dataset, then automatically build an integrated learning model through the TabularPredictor function and perform hyperparameter optimization;

[0029] S42, based on the construction of inheritance learning model, the importance score of each normalized track feature in the track data set is calculated using Permutation Importance;

[0030] S43. According to the distribution of feature importance scores, the track features are screened; a feature importance threshold is set. If the feature importance score is less than the threshold, the feature is considered unimportant and is removed, and the remaining features constitute a key feature set.

[0031] Furthermore, step S5 specifically includes:

[0032] S51, from the key feature data set screened in step S4, load all the retained key features and their corresponding label data, divide the loaded key feature set into a training set and a validation set, and use 80% of the data as the training set and 20% of the data as the validation set;

[0033] S52, set the training time and computing resource limits, limit the training time to 300 seconds, and specify the use of 24 CPU cores and 1 GPU; at the same time, check whether the program can correctly access these computing resources;

[0034] S53, start the training process, select the best combination among the following 12 basic models: LightGBMXT, RandomForestEntr, XGBoost, ExtraTreesGini, ExtraTreesEntr, RandomForestGini, LightGBM, CatBoost, KNeighborsDist, KNeighborsUnif, NeuralNetFastAI and NeuralNetTorch; select the basic model combination to build the AutoGluon drone classifier under the configured training time and computing resource constraints;

[0035] S54. In the Base layer of AutoGluon, firstly, hyperparameter optimization is performed based on the selected base model type. The best hyperparameter combination of each base model is found through random search and Bayesian optimization methods, and a meta-feature matrix containing the predicted values ​​of all base models is generated. Then, the meta-feature matrix and target variable output by the Base layer are used to train the meta-model and optimize the combined weights of the base models to improve the overall performance.

[0036] S55, model training and result verification; using the cross-validation method, 20% of the data is selected from the training set each time to train and evaluate the recognition model. After multiple trainings, the model combination with the best performance is selected as the final AutoGluon drone classifier. The validation set is input into the final AutoGluon drone classifier for small and micro drone recognition. The results of each recognition are recorded, and the average accuracy of the AutoGluon drone classifier is calculated, denoted as w a .

[0037] Furthermore, step S6 specifically includes:

[0038] S61. Using the enhanced track data as input, train the iTransformer time series classifier to learn the time series characteristics of the UAV flight mode.

[0039] S62, using the ∈ loss function to guide the learning process of the model; the formula of the ∈ loss function is expressed as:

[0040]

[0041] Among them, ∈ is a regularization term, which is used to reduce the model's trust in the training data label; y is the true label of the target track, The probability of the target being a drone identified by the iTransformer time series classifier;

[0042] S63, the cross-validation method is also used, and 20% of the data is selected from the training set for training evaluation each time. The optimal iTransformer time series classifier is obtained through multiple trainings, and the optimal iTransformer time series classifier is used for drone recognition. The calculated average accuracy is recorded as w i ;

[0043] S64. A weighted fusion-based strategy is adopted to combine the prediction outputs of the two models to obtain a more accurate classification probability of drones. The calculation formula for the probability P of the target being a drone is finally as follows:

[0044]

[0045] in, The probability of the target being a drone identified by the AutoGluon drone classifier.

[0046] Based on the above technical solution, the present invention has at least the following beneficial effects:

[0047] The present invention combines the prime number step sampling method with the AutoGluon framework to significantly improve the generalization ability of the model. While improving the performance of the model, it reduces the input of manual intervention and computing resources, thereby significantly reducing the development cost.

[0048] The present invention proposes a data enhancement method based on prime number step sampling and sliding window. The method makes full use of prime number step to perform differential sampling on track data, maximizes the difference between adjacent sampling points, and thus optimizes the diversity of data and the learning effect of the model.

[0049] This invention uses the AutoGluon framework to replace traditional manual parameter adjustment. Automated hyperparameter search and model optimization significantly reduce parameter adjustment time and computing resource consumption, and improve development efficiency. At the same time, the iTransformer time series classifier is introduced to identify targets from the perspective of time series, and an ∈ loss function is proposed to reduce the trust of training data labels and reduce the risk of overfitting. Finally, by weighted fusion of AutoGluon and iTransformer time series classifiers, track features and time series information are combined to improve model accuracy and enhance model robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The overall step flow chart of the method proposed in the present invention;

[0051] Figure 2 This is a schematic diagram of the process of constructing a low-altitude recognition model proposed in the present invention by combining the AutoGluon UAV classifier with the iTransformer time series classifier. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0053] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0054] like Figure 1 As shown, the low-altitude small and micro UAV target intelligent identification method based on track characteristics proposed by the present invention is shown, which specifically includes:

[0055] S1. Design a multi-dimensional feature map to effectively characterize the target to be identified based on the track data of low-altitude micro-UAVs;

[0056] As a preferred implementation, step S1 specifically includes:

[0057] S11. Arrange and summarize the radar data used in the UAV identification method based on track data in the public literature; the radar data includes six basic information: the target's azimuth, slant range, relative altitude, radial velocity, recording time and radar cross section RCS;

[0058] S12. Based on radar data, design features including acceleration dimension features, velocity dimension features, heading angle dimension features and relative height dimension features to distinguish the differences in flight trajectories and motion patterns between drones and non-drones (such as birds and balloons) in a fine-grained manner; the acceleration dimension features include the maximum radial acceleration, the maximum radial acceleration value, and the acceleration change; the velocity dimension features include the maximum radial velocity, the minimum radial velocity, the maximum radial velocity value, the minimum radial velocity value, the average speed, the speed standard deviation, the speed oscillation frequency, the maximum speed, and the minimum speed; the heading angle dimension features include the heading angle mean and the heading angle standard deviation; the relative height dimension features include the relative height mean and the relative height standard deviation;

[0059] In this embodiment, not only the acceleration and velocity vectors are considered, but also the corresponding scalar values ​​are considered, that is, simply considering the magnitude of the acceleration and velocity, regardless of the direction, the flight performance of the target can be intuitively judged;

[0060] S13. Design RCS dimensional features based on radar data to distinguish the differences between specific composite materials and metal alloys used by drones and natural objects (such as birds and balloons) in a coarse-grained manner; the RCS dimensional features include average RCS and RCS range.

[0061] S2, data enhancement of the track data set is performed by combining the prime number step sampling method with the sliding window method;

[0062] As a preferred implementation, step S2 specifically includes:

[0063] S21, using the sliding window method to generate new sub-tracks for single track data; the size of the sliding window is set to 5, 6, 7, 8, 9, 10, and 11 in sequence, and the window is gradually slid to intercept sub-tracks of different scales to fully capture the multi-dimensional characteristics of the track;

[0064] S22, on the basis of the sub-tracks generated by the sliding window, setting the sliding window step size to a prime number that increases sequentially, and obtaining multiple differentiated sub-tracks by multiple sampling;

[0065] S23, integrating the sampled sub-tracks, removing the repeated sub-tracks, and obtaining a data-enhanced track data set.

[0066] In this embodiment, the historical track data is sampled to obtain richer track features; the sampling is performed with a step size of 2, 3, and a sliding window size of 5 as an example, and a track data set with a length of 10 is defined as S = {s(1), s(2), ..., s(10)};

[0067] After sampling with a step size of 2, the sub-tracks that can be obtained are: {s(1), s(2), s(3), s(4), s(5)}, {s(3), s(4), s(5), s(6), s(7)}, {s(5), s(6), s(7), s(8), s(9)};

[0068] After sampling with a step length of 3, the new sub-tracks are: {s(4), s(5), s(6), s(7), s(8)}. This sampling strategy can not only retain the local information of the target track due to the characteristics of prime numbers, but also effectively reduce the repeated sampling of sub-tracks. At the same time, as the sampling step length continues to increase, it also reduces redundant calculations to a certain extent, significantly improving the calculation efficiency. The model can learn on a richer data set, thereby improving its generalization ability and prediction accuracy.

[0069] S3, performing feature calculation on the track data set after data enhancement, processing outliers and missing values, and then normalizing the processed data set;

[0070] As a preferred implementation, step S3 specifically includes:

[0071] S31, calculating each dimensional feature in each track data, and integrating each dimensional feature into a feature vector;

[0072] In this embodiment, the specific calculation process of each dimensional feature is as follows:

[0073] Define a track data set of length N as S = {s(1), s(2), ..., s(N)}, where each track data is denoted as s(i), that is, the i-th sample track s(i) = {θ(i), r(i), h(i), v(i), t(i), rcs(i)}, and each letter in s(i) represents the target azimuth, target slant range, relative altitude, radial velocity, recording time, and RCS of the target at that moment; the feature calculation processes are as follows:

[0074] 1) Maximum radial acceleration: First, we need to calculate the acceleration a between every two points in the track. m , its formula is expressed as:

[0075]

[0076] Among them, v i+1 、v i-1 are the speeds at two moments, t i+1、 t i-1 for the moment;

[0077] Then the maximum value of radial acceleration in each track data is calculated;

[0078] 2) Maximum radial velocity: the maximum value of radial velocity in each track data;

[0079] 3) Minimum radial velocity: the minimum value of radial velocity in each track data;

[0080] 4) Maximum radial acceleration value: the maximum value of radial acceleration in each track data;

[0081] 5) Acceleration change: where a i+1 、a i-1 is the acceleration at two moments;

[0082] 6) Maximum radial velocity value: the maximum value of the radial velocity in each track data;

[0083] 7) Minimum radial velocity value: the minimum value of the radial velocity in each track data;

[0084] 8) Average RCS: The average value for each point in the track.

[0085] 9) RCS extreme: rcs j =max(rcs m )-min(rcs m ), m∈[1,N]

[0086] 10) Average speed: To calculate the average speed, we first need to calculate the approximate speed of the object by the state changes between adjacent points. We use To calculate, d i is the displacement between the i-th point and the i-1-th point, which requires the following steps:

[0087] Calculate the displacement in each direction:

[0088] The horizontal displacement dx is determined by the coordinate difference of the two positions on the x-axis

[0089] dx=r i ×sinθ i -r i-1 ×sinθ i-1 ;

[0090] The vertical displacement dy is determined by the coordinate difference of the two positions on the y-axis

[0091] dx=r i ×cosθ i -r i-1 ×cosθ i-1 ;

[0092] The displacement dz in the height direction is directly equal to the height difference between the two positions

[0093] dh=H i -H i-1 ;

[0094] Among them, r i 、r i-1 Respectively represent the target slant distances at the two positions, θ i ,θ i-1 Represent the target azimuths at the two positions, H i , H i-1 The heights of the two positions are

[0095] Calculate the total displacement:

[0096] The total displacement d is calculated using the distance formula in three-dimensional space:

[0097]

[0098] Calculate speed: Calculate speed according to the following formula:

[0099]

[0100] After obtaining the speed of each point in the track, the average speed is calculated according to the following formula:

[0101]

[0102] 11) Speed ​​standard deviation: The speed standard deviation is calculated by the following formula:

[0103]

[0104] 12) Speed ​​oscillation frequency: In order to reflect the local characteristics of speed change, speed oscillation frequency is introduced. At the current time i, the speed value of the sample point of the historical length L is taken, and the speed mean value V under this length is calculated mL , defined as

[0105] Then calculate the current point speed V(i) and the historical average speed V mL The difference between V(i), defined by the formula:

[0106] τ V (i) = V(i) - v mL , i ≥ L + 1;

[0107] Set speed change threshold γ v , when the difference is greater than γ v , it is considered that the point produces velocity oscillation; through U V (i) indicates whether the point generates velocity oscillation, and the formula is:

[0108]

[0109] Among them, the value of 1 indicates that oscillation occurs, and the value of 0 indicates that oscillation does not occur;

[0110] The threshold γ can be adjusted according to the actual radar tracking data rate and tracking accuracy. v and L length, speed oscillation frequency f v Defined as:

[0111]

[0112] 13) Maximum speed: the maximum value of the speed in each track data;

[0113] 14) Minimum speed: the minimum speed in each track data;

[0114] 15) Calculation of heading angle mean: First, calculate the heading angle h(i) at the current moment i based on the position information x(i-1), y(i-1), x(i), y(i) in the rectangular coordinate system at the previous moment s(i-1) and the current moment s(i). The calculation formula for the heading angle h(i) is:

[0115]

[0116] 16) Calculation of heading angle standard deviation: In order to calculate the standard deviation of the heading angle, the mean of the heading angle needs to be calculated first. Since the heading angle of some target trajectories may cross 0 or 360°, the conventional mean calculation method cannot be used directly; in this embodiment, a vectorization method is used to convert each heading angle h(i) into a unit vector r(i) on a two-dimensional plane:

[0117] r(i)=(cos(h(i)),sin(h(i)));

[0118] Then, add the vectors corresponding to all N-1 heading angles to get the resultant vector r s :

[0119]

[0120] Finally, calculate the mean value h of the heading angle m and standard deviation h std :

[0121] h m =tan -1 (r s );

[0122]

[0123] 17) Relative height mean: The relative height mean is calculated by the following formula:

[0124]

[0125] 18) Relative height standard deviation: Calculate the relative height standard deviation using the following formula:

[0126]

[0127] At this point, the calculation of each dimension feature is completed;

[0128] S32, remove outliers caused by dirty data in multi-dimensional features, including speed standard deviation infinite, heading angle standard deviation infinite, relative height standard deviation infinite;

[0129] S33, checking whether the obtained feature vector has any missing values, and if any missing values ​​are found, deleting the corresponding entire data;

[0130] S34. Use maximum and minimum normalization to normalize the features of each dimension.

[0131] S4. Use the AutoGluon framework to perform feature importance analysis on the normalized data to select key feature sets to optimize model performance;

[0132] As a preferred implementation, step S4 specifically includes:

[0133] S41, based on the normalized track data set in step S3, use the TabularDataset function in the AutoGluon library to load the track data set, then automatically build an integrated learning model through the TabularPredictor function, and perform hyperparameter optimization;

[0134] S42. Based on the constructed model, the importance score of each feature is calculated using Permutation Importance.

[0135] S43. According to the distribution of feature importance scores, the threshold is set to 0.002. If the feature importance score is less than 0.002, the feature is considered unimportant and is eliminated. After calculation, the RCS range and acceleration change are eliminated, and the remaining features are the key feature set.

[0136] S5. Use the selected key feature set to build an efficient small and micro AutoGluon drone classifier to achieve accurate recognition of low-altitude targets;

[0137] In this embodiment, step S5 uses the AutoGluon framework to build a drone recognition model. First, it manually limits the integration to 12 basic models, then inputs the Base layer, generates the prediction results of each model through independent training and cross-validation, and outputs the meta-feature matrix (containing the prediction values ​​of all basic models). These meta-features will be used as inputs to the Stack layer, and the meta-feature matrix and target variables output by the Base layer will be used to train the meta-model, further optimize the combined weights of the basic models, and improve the overall performance;

[0138] As a preferred implementation method, Figure 2 As shown, step S5 specifically includes:

[0139] S51, from the key feature data set screened in step S4, load all the retained key features and their corresponding label data, divide the loaded key feature set into a training set and a validation set, and use 80% of the data as the training set and 20% of the data as the validation set;

[0140] S52, set the training time and computing resource limits, limit the training time to 300 seconds, and specify the use of 24 CPU cores and 1 GPU; at the same time, check whether the program can correctly access these computing resources;

[0141] S53, start the training process, select the best combination among the following 12 basic models: LightGBMXT, RandomForestEntr, XGBoost, ExtraTreesGini, ExtraTreesEntr, RandomForestGini, LightGBM, CatBoost, KNeighborsDist, KNeighborsUnif, NeuralNetFastAI and NeuralNetTorch; select the basic model combination to build the AutoGluon drone classifier under the configured training time and computing resource constraints;

[0142] S54. In the Base layer of AutoGluon, firstly, hyperparameter optimization is performed based on the selected base model type. The best hyperparameter combination of each base model is found through random search and Bayesian optimization methods, and a meta-feature matrix containing the predicted values ​​of all base models is generated. Then, the meta-feature matrix and target variable output by the Base layer are used to train the meta-model and optimize the combined weights of the base models to improve the overall performance.

[0143] S55, model training and result verification; using the cross-validation method, 20% of the data is selected from the training set each time to train and evaluate the recognition model. After multiple trainings, the model combination with the best performance is selected as the final AutoGluon drone classifier. The validation set is input into the final AutoGluon drone classifier for small and micro drone recognition. The results of each recognition are recorded, and the average accuracy of the AutoGluon drone classifier is calculated, denoted as w a .

[0144] After step S5, the AutoGluon drone classifier is constructed, and it will be combined with iTransformer for more precise recognition.

[0145] S6. Reconstruct the iTransformer time series classifier based on the enhanced track data set, and combine it with the constructed AutoGluon UAV classifier, comprehensively consider the recognition ability and accuracy of the two models, and achieve accurate recognition of low-altitude targets; In this embodiment, step S6 integrates the prediction results of the iTransformer and AutoGluon models, optimizes the training through a customized loss function, and finally uses weighted averaging to improve classification accuracy and robustness;

[0146] As a preferred implementation, step S6 specifically includes:

[0147] S61. Use the enhanced track data as input to train the iTransformer time series classifier to learn the time series characteristics of the UAV flight mode; the time series characteristics are each track data obtained by radar sampling, that is, s(i) = {θ(i), r(i), h(i), v(i), t(i), rcs(i)};

[0148] S62, using the ∈ loss function to guide the learning process of the model; the formula of the ∈ loss function is expressed as:

[0149]

[0150] Among them, ∈ is a regularization term, which is used to reduce the model's trust in the training data label; y is the true label of the target track, The probability of the target being a drone identified by the iTransformer time series classifier;

[0151] In this embodiment, the label is a binary classification label, which indicates whether the target is a drone. Therefore, when calculating the ∈ loss function, the label takes two values: 0 and 1. If the true label is 1 to indicate that the target is a drone, the closer the probability of the time series classifier identifying the target as a drone is to 1, the smaller the loss function.

[0152] S63, the cross-validation method is also used, and 20% of the data is selected from the training set for training evaluation each time. The optimal iTransformer time series classifier is obtained through multiple trainings, and the optimal iTransformer time series classifier is used for drone recognition. The calculated average accuracy is recorded as w i ;

[0153] S64. A weighted fusion-based strategy is adopted to combine the prediction outputs of the two models to obtain a more accurate classification probability of drones. The calculation formula for the probability P of the target being a drone is finally as follows:

[0154]

[0155] in, The probability of the target being a drone identified by the AutoGluon drone classifier.

[0156] In this embodiment, by first combining the information in the track feature dimension extracted by the AutoGluon model with the information based on the time series dimension extracted by the iTransformer time series classifier, and performing weighted fusion on the output probabilities of the two, the overall accuracy of the model can be improved, overfitting can be reduced, and the model's robustness to noise and uncertainty in the data can be enhanced, ultimately obtaining more accurate recognition results.

[0157] In summary, the method proposed in the present invention uses sliding window and prime number step sampling methods to perform data enhancement based on a multi-dimensional feature set that systematically characterizes the differences in motion performance and appearance structure between drones and non-drones, and uses the AutoGluon framework to calculate feature importance and eliminate unimportant features; the AutoGluon drone classifier built based on important features is combined with the iTransformer time series classifier to achieve efficient distinction between drones and non-drones in multiple feature dimensions. The final recognition model proposed in the present invention has a high recognition rate and strong adaptability, and can be widely used in the field of low-altitude drone detection and defense.

[0158] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0159] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A low-altitude small and micro UAV target intelligent recognition method based on track characteristics, characterized in that: The specific steps include: S1. Design a multi-dimensional feature map to effectively characterize the target to be identified based on the track data of low-altitude micro-UAVs; S2, data enhancement of the track data set is performed by combining the prime number step sampling method with the sliding window method; S3, calculating the features of the track data set after data enhancement, processing the outliers and missing values ​​of the features, and then normalizing the processed track features; S4. Use the AutoGluon framework to perform importance analysis on the normalized track features to select key feature sets to optimize model performance; S5. Using the selected key feature set as input, construct the AutoGluon UAV classifier, divide the key feature set into a training set and a validation set, train the optimal model, and calculate the average accuracy of the AutoGluon UAV classifier in identifying the target object; S6. Based on the enhanced track data set, the iTransformer time series classifier is reconstructed and combined with the constructed AutoGluon UAV classifier. The recognition ability and average accuracy of the two models are comprehensively considered to achieve accurate recognition of low-altitude targets.

2. The low-altitude small and micro-UAV target intelligent identification method based on track characteristics according to claim 1 is characterized in that: Step S1 specifically includes: S11. Arrange and summarize the radar data used in the UAV identification method based on track data in the public literature; the radar data includes six basic information: the target's azimuth, slant range, relative altitude, radial velocity, recording time and radar cross section RCS; S12. Based on radar data, design features including acceleration dimension features, velocity dimension features, heading angle dimension features and relative height dimension features to distinguish the differences between UAVs and non-UAVs in flight trajectories and motion modes in a fine-grained manner; the acceleration dimension features include the maximum radial acceleration, the maximum radial acceleration value, and the acceleration change; the velocity dimension features include the maximum radial velocity, the minimum radial velocity, the maximum radial velocity value, the minimum radial velocity value, the average speed, the speed standard deviation, the speed oscillation frequency, the maximum speed, and the minimum speed; the heading angle dimension features include the heading angle mean value and the heading angle standard deviation; the relative height dimension features include the relative height mean value and the relative height standard deviation; S13. Design RCS dimensional features based on radar data to distinguish the differences between specific composite materials, metal alloys and natural object materials used by drones in a coarse-grained manner; the RCS dimensional features include average RCS and RCS range.

3. The low-altitude small and micro-UAV target intelligent identification method based on track characteristics according to claim 1 is characterized in that: Step S2 specifically includes: S21, using the sliding window method to generate new sub-tracks for single track data; the size of the sliding window is set to 5, 6, 7, 8, 9, 10, and 11 in sequence, and the window is gradually slid to intercept sub-tracks of different scales to fully capture the multi-dimensional characteristics of the track; S22, on the basis of the sub-tracks generated by the sliding window, setting the sliding window step size to a prime number that increases sequentially, and obtaining multiple differentiated sub-tracks by multiple sampling; S23, integrating the sampled sub-tracks, removing the repeated sub-tracks, and obtaining a track data set after data enhancement.

4. The low-altitude small and micro-UAV target intelligent identification method based on track characteristics according to claim 2 is characterized in that: Step S3 specifically includes: S31, calculating each dimensional feature in each track data, and integrating each dimensional feature into a feature vector; S32, remove outliers caused by dirty data in multi-dimensional features, including speed standard deviation infinite, heading angle standard deviation infinite, relative height standard deviation infinite; S33, checking whether the obtained feature vector has any missing values, and if any missing values ​​are found, deleting the corresponding entire data; S34. Use maximum and minimum normalization to normalize the features of each dimension.

5. The low-altitude small and micro-UAV target intelligent identification method based on track characteristics according to claim 1 is characterized in that: Step S4 specifically includes: S41. Use the TabularDataset function in the AutoGluon library to load the track dataset, then automatically build an integrated learning model through the TabularPredictor function and perform hyperparameter optimization; S42, based on the construction of inheritance learning model, the importance score of each normalized track feature in the track data set is calculated using Permutation Importance; S43. According to the distribution of feature importance scores, the track features are screened; a feature importance threshold is set. If the feature importance score is less than the threshold, the feature is considered unimportant and is removed, and the remaining features constitute a key feature set.

6. The low-altitude small and micro-UAV target intelligent identification method based on track characteristics according to claim 1 is characterized in that: Step S5 specifically includes: S51, from the key feature data set screened in step S4, load all the retained key features and their corresponding label data, divide the loaded key feature set into a training set and a validation set, and use 80% of the data as the training set and 20% of the data as the validation set; S52, set the training time and computing resource limits, limit the training time to 300 seconds, and specify the use of 24 CPU cores and 1 GPU; at the same time, check whether the program can correctly access these computing resources; S53, start the training process, select the best combination among the following 12 basic models: LightGBMXT, RandomForestEntr, XGBoost, ExtraTreesGini, ExtraTreesEntr, RandomForestGini, LightGBM, CatBoost, KNeighborsDist, KNeighborsUnif, NeuralNetFastAI and NeuralNetTorch; select the basic model combination to build the AutoGluon drone classifier under the configured training time and computing resource constraints; S54. In the Base layer of AutoGluon, hyperparameter optimization is first performed based on the selected base model type. The best hyperparameter combination for each base model is found through random search and Bayesian optimization methods, and a meta-feature matrix containing the predicted values ​​of all base models is generated. These meta-features will be used as inputs to the Stack layer, and then the meta-feature matrix and target variable output by the Base layer are used to train the meta-model and optimize the combined weights of the base models to improve the overall performance. S55, model training and result verification; using the cross-validation method, 20% of the data is selected from the training set for training evaluation each time. After multiple trainings, the model combination with the best performance is selected as the final AutoGluon drone classifier. The validation set is input into the final AutoGluon drone classifier for small and micro drone identification. The results of each identification are recorded, and the average accuracy of the AutoGluon drone classifier is calculated, denoted as w a .

7. The low-altitude small and micro-UAV target intelligent identification method based on track characteristics according to claim 6 is characterized in that: Step S6 specifically includes: S61. Using the enhanced track data as input, train the iTransformer time series classifier to learn the time series characteristics of the UAV flight mode. S62, using the ∈ loss function to guide the learning process of the model; the formula of the ∈ loss function is expressed as: Among them, ∈ is a regularization term, which is used to reduce the model's trust in the training data label; y is the true label of the target track, The probability of the target being a drone identified by the iTransformer time series classifier; S63, the cross-validation method is also used, and 20% of the data is selected from the training set for training evaluation each time. The optimal iTransformer time series classifier is obtained through multiple trainings, and the optimal iTransformer time series classifier is used for drone recognition. The calculated average accuracy is recorded as w i ; S64. A weighted fusion-based strategy is adopted to combine the prediction outputs of the two models to obtain a more accurate probability of drone classification. The calculation formula for the probability P of the target being a drone is finally as follows: in, The probability of the target being a drone identified by the AutoGluon drone classifier.

Images