Intelligent Identification Method for Low-Altitude Small and Micro UAV Targets Based on Track Features

Abstract

The present invention discloses an intelligent target recognition method for low-altitude small and micro unmanned aerial vehicles based on track features, which specifically includes: designing multi-dimensional features that can effectively characterize target characteristics for the track data of low-altitude small and micro unmanned aerial vehicles; performing data enhancement on the track data through the prime step sampling method combined with the sliding window method; calculating features for the enhanced track data, processing outliers and missing values, and then normalizing and standardizing the data; using the AutoGluon framework for feature importance analysis, screening out key features and using them to construct an AutoGluon unmanned aerial vehicle classifier; then constructing an iTransformer time series classifier based on the enhanced track data and fusing it with the AutoGluon unmanned aerial vehicle classifier, and obtaining the optimal recognition model through training iteration; the present invention can achieve rapid and accurate recognition of low-altitude small and micro unmanned aerial vehicles, has a high recognition rate and strong adaptability, and can be widely applied to the fields of low-altitude unmanned aerial vehicle detection and defense.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent target recognition of unmanned aerial vehicles, and particularly relates to an intelligent target recognition method for low-altitude small and micro unmanned aerial vehicles based on track features. Background Art

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

[0003] Small and micro unmanned aerial vehicles are small in size and flexible in flight, and are easy to be confused with natural objects (such as birds, balloons) in the low-altitude environment or people and vehicles on the ground, bringing huge challenges to the existing monitoring technology and management system. The flight activities of illegal unmanned aerial vehicles may not only violate personal privacy and disrupt normal social order, but may even become tools for terrorist attacks or crimes, seriously threatening public safety and social stability. To effectively address these issues and ensure the safety of the low-altitude area, it is necessary to quickly and accurately identify unmanned aerial vehicle targets and non-unmanned aerial vehicle targets through the point and track data of radar when a suspicious target is detected. Therefore, it is crucial to develop efficient monitoring and recognition technologies.

[0004] Currently, the methods commonly used by researchers in the field of unmanned aerial vehicle target recognition include support vector machines, random forests, and convolutional neural networks, etc. These methods can, to a certain extent, achieve target detection and classification tasks, but they still have many limitations in practical applications. For example, the model optimization process highly depends on the quality and quantity of data, and often shows performance degradation or even misjudgment when facing serious data imbalance problems. Another prominent problem lies in the lack of generalization ability. Many existing models perform well when trained on specific scenarios or datasets, but when applied to different environments or new target categories, the accuracy often drops significantly. This lack of generalization performance makes it difficult for the model to meet the actual needs of low-altitude safety management, especially in the case of a wide variety of unmanned aerial vehicle types and changing environments. Therefore, researchers need to develop efficient and robust data augmentation strategies to increase the diversity and scale of training data and effectively address the problem of data category imbalance. On this basis, it is also necessary to further explore and develop more generalizable model architectures to ensure that the model can still maintain good prediction results when facing unseen data. The combination of data augmentation technology and the optimization of model architecture can effectively promote the development of unmanned aerial vehicle recognition technology. Summary of the Invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention proposes an intelligent target recognition method, device and storage medium for low-altitude small and micro unmanned aerial vehicles (UAVs) based on track features to meet the needs of people for the supervision of low-altitude UAVs and achieve the precise recognition and real-time monitoring of low-altitude small and micro UAVs. At the same time, it has a high recognition rate and strong adaptability and is applicable to the fields of low-altitude UAV monitoring and defense.

[0006] To achieve the above technical objectives, the present invention provides the following technical solutions:

[0007] An intelligent target recognition method for low-altitude small and micro UAVs based on track features specifically includes the following steps:

[0008] S1. For the track data of low-altitude small and micro UAVs, design multi-dimensional features for effectively characterizing the target to be recognized;

[0009] S2. Perform data augmentation on the track data set through the prime step sampling method combined with the sliding window method;

[0010] S3. Calculate the features of the track data set after data augmentation, process the outliers and missing values of the features, and then normalize the processed track features;

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

[0012] S5. Take the screened key feature set as the input, construct an AutoGluon UAV classifier through the AutoGluon framework, 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;

[0013] S6. Based on the augmented track data set, construct an iTransformer time series classifier and combine it with the constructed AutoGluon UAV classifier, and comprehensively consider the recognition ability and accuracy of the two models to achieve the precise recognition of low-altitude targets.

[0014] Further, step S1 specifically includes:

[0015] S11. Organize and summarize the radar data used in the UAV recognition method based on track data in the public literature; the radar data includes six basic information: the azimuth angle, slant range, relative height, radial velocity, recording time, and radar cross section (RCS) of the target;

[0016] S12. Design features including acceleration dimension features, speed dimension features, heading angle dimension features, and relative height dimension features based on radar data to finely distinguish the differences in flight trajectories and motion patterns between drones and non-drones; the acceleration dimension features include maximum radial acceleration, maximum radial acceleration value, and acceleration change amount; the speed dimension features include maximum radial speed, minimum radial speed, maximum radial speed value, minimum radial speed value, average speed, speed standard deviation, speed oscillation frequency, maximum speed, and minimum speed; the heading angle dimension features include heading angle mean and heading angle standard deviation; the relative height dimension features include relative height mean and relative height standard deviation;

[0017] S13. Design RCS dimension features based on radar data to coarsely distinguish the differences between specific composite materials, metal alloys used by drones, and natural object materials; the RCS dimension features include average RCS and RCS range.

[0018] Further, step S2 specifically includes:

[0019] S21. Use the sliding window method to generate new sub-trajectories for single-track data; the sizes of the sliding windows are set to 5, 6, 7, 8, 9, 10, and 11 in sequence, and the window slides step by step to intercept sub-trajectories of different scales to comprehensively capture the multi-dimensional features of the trajectory;

[0020] S22. On the basis of generating sub-trajectories by the sliding window, set the sliding window step size to prime numbers that increase in sequence, and sample multiple times to obtain multiple differentiated sub-trajectories;

[0021] S23. Integrate the sampled sub-trajectories, eliminate the duplicate sub-trajectories among them, and obtain a trajectory dataset with enhanced data.

[0022] Further, step S3 specifically includes:

[0023] S31. Calculate the various dimension features in each track data and integrate the various dimension features into a feature vector;

[0024] S32. Eliminate the outliers caused by dirty data in the multi-dimensional features. The outliers include infinite speed standard deviation, infinite heading angle standard deviation, and infinite relative height standard deviation;

[0025] S33. Check whether there are missing values in the obtained feature vector. If missing values are found, delete the corresponding entire data;

[0026] S34. Normalize each dimension feature using maximum-minimum normalization.

[0027] Further, step S4 specifically includes:

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

[0029] S42. Based on the constructed inheritance learning model, use Permutation Importance to calculate the importance scores of each normalized track feature in the track dataset;

[0030] S43. According to the distribution of the feature importance scores, screen the track features; set a feature importance threshold. If the importance score of a feature is less than the threshold, the feature is regarded as unimportant and excluded, and the remaining features form the key feature set.

[0031] Further, step S5 specifically includes:

[0032] S51. From the key feature dataset screened in step S4, load all the remaining key features and their corresponding label data, divide the loaded key feature set into a training set and a validation set, 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 from the following 12 base models: LightGBMXT, RandomForestEntr, XGBoost, ExtraTreesGini, ExtraTreesEntr, RandomForestGini, LightGBM, CatBoost, KNeighborsDist, KNeighborsUnif, NeuralNetFastAI, and NeuralNetTorch; under the configured training time and computing resource limits, select a combination of base models to construct an AutoGluon drone classifier;

[0035] S54. In the Base layer of AutoGluon, first perform hyperparameter optimization based on the selected base model type; through random search and Bayesian optimization methods, find the best hyperparameter combination for each base model and generate a meta-feature matrix containing the predicted values of all base models; then use the meta-feature matrix and target variables output by the Base layer 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 from the training set is selected each time for training and evaluation on the recognition model. Based on multiple trainings, the model combination with the optimal performance is selected as the final AutoGluon UAV classifier. The validation set is input into the final AutoGluon UAV classifier for the recognition of small and micro UAVs. The results of each recognition are recorded, and the average accuracy rate of the AutoGluon UAV classifier is calculated and denoted as w. a 。

[0037] Further, step S6 specifically includes:

[0038] S61. Using the enhanced track data as input, training the iTransformer time - series classifier to learn the time - series features of the UAV flight mode;

[0039] S62. Using the ∈ loss function to guide the learning process of the model; the formula expression of the ∈ loss function is:

[0040]

[0041] where ∈ is the regularization term used to reduce the trust of the model in the training data labels; y is the true label of the target track, is the probability that the iTransformer time - series classifier identifies the target as a UAV;

[0042] S63. Also using the cross - validation method, 20% of the data from the training set is selected each time for training and evaluation. Through multiple trainings, the optimal iTransformer time - series classifier is obtained, and the UAV is recognized using the optimal iTransformer time - series classifier. The calculated average accuracy rate is denoted as w i ;

[0043] S64. Adopting a strategy based on weighted fusion, combining the prediction outputs of the two models to obtain a more accurate UAV classification probability. The calculation formula for the final probability P of identifying the target as a UAV is:

[0044]

[0045] where, is the probability that the AutoGluon UAV classifier identifies the target as a UAV.

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

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

[0048] The present invention proposes a data augmentation method based on prime - step sampling and sliding window. This method makes full use of the prime - step to differentially sample the track data, maximizing the difference between adjacent sampling points, thereby optimizing the data diversity and the learning effect of the model.

[0049] The present invention uses the AutoGluon framework to replace traditional manual hyperparameter tuning. Automating hyperparameter search and model optimization significantly reduces the hyperparameter tuning time and computing resource consumption, improving the development efficiency. At the same time, it introduces the iTransformer time - series classifier to identify the target from the time - series perspective, and proposes the ∈ loss function to reduce the trust degree of the training data labels and reduce the risk of overfitting. Finally, by weighted - fusing the AutoGluon and iTransformer time - series classifiers, it combines the track features and time - series information, improving the model accuracy and enhancing the model robustness. Brief Description of the Drawings

[0050] Figure 1 is the overall step - flow chart of the method proposed by the present invention;

[0051] Figure 2 is the schematic diagram of the process of constructing a low - altitude recognition model by combining the AutoGluon UAV classifier and the iTransformer time - series classifier proposed by the present invention. Detailed Embodiment

[0052] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the 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 numbered, they are not used to limit the sequence of the steps. Unless the sequence of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative sequence of the steps can be adjusted. It can be understood that the term "and / or" used herein involves and covers any and all possible combinations of one or more of the related listed items.

[0054] As Figure 1 shown, the intelligent target recognition method for low - altitude small and micro UAVs based on track features proposed by the present invention is specifically as follows:

[0055] S1. Design multi-dimensional features for effectively characterizing the target to be recognized for the trajectory data of low-altitude small and micro unmanned aerial vehicles (UAVs).

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

[0057] S11. Sort out and summarize the radar data used in the UAV recognition method based on trajectory data in the public literature; the radar data includes six basic information: azimuth angle of the target, slant range, relative height, radial velocity, recording time, and radar cross section (RCS).

[0058] S12. Design features including acceleration dimension features, velocity dimension features, heading angle dimension features, and relative height dimension features based on the radar data to finely distinguish the differences in flight trajectories and motion patterns between UAVs and non-UAVs (such as birds, balloons); the acceleration dimension features include maximum radial acceleration, maximum radial acceleration value, and acceleration change amount; the velocity dimension features include maximum radial velocity, minimum radial velocity, maximum radial velocity value, minimum radial velocity value, average velocity, velocity standard deviation, velocity oscillation frequency, maximum velocity, and minimum velocity; the heading angle dimension features include mean heading angle and heading angle standard deviation; the relative height dimension features include mean relative height and relative height standard deviation.

[0059] In this embodiment, not only acceleration and velocity vectors are considered, but also the corresponding scalar values are considered, that is, only the magnitudes of acceleration and velocity are considered regardless of direction, so as to intuitively judge the flight performance of the target.

[0060] S13. Design RCS dimension features based on the radar data to coarsely distinguish the differences in specific composite materials, metal alloys used in UAVs, and the materials of natural objects (such as birds, balloons); the RCS dimension features include average RCS and RCS range.

[0061] S2. Perform data augmentation on the trajectory data set by combining the prime step sampling method and the sliding window method.

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

[0063] S21. Use the sliding window method for a single trajectory data to generate new sub-trajectories; the sizes of the sliding windows are set to 5, 6, 7, 8, 9, 10, and 11 in sequence, and the window slides step by step to intercept sub-trajectories of different scales to comprehensively capture the multi-dimensional features of the trajectory.

[0064] S22. On the basis of generating sub-trajectories by the sliding window, set the sliding window step size to prime numbers that increase in sequence, and sample multiple times to obtain multiple different sub-trajectories.

[0065] S23. Integrate the sampled sub-trajectories, remove the duplicate sub-trajectories among them, and obtain the trajectory dataset after data augmentation.

[0066] In this embodiment, sampling is performed on the historical trajectory data to obtain richer trajectory features. The sampling takes a step size of 2 and 3 as an example, and the sliding window size is 5. Define a trajectory dataset of a certain length of 10 as S = {s(1), s(2), …, s(10)}.

[0067] The sub-trajectories obtained after sampling with a step size of 2 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 size of 3, the new sub-trajectories obtained are: {s(4), s(5), s(6), s(7), s(8)}. Such a sampling strategy can not only retain the local information of the target trajectory due to the characteristics of prime numbers, but also effectively reduce the repeated sampling of sub-trajectories. At the same time, as the sampling step size continues to increase, the redundant calculation is also reduced to a certain extent, significantly improving the calculation efficiency. The model can learn on a richer dataset, thereby enhancing its generalization ability and prediction accuracy.

[0069] S3. Calculate the features of the trajectory dataset after data augmentation, process the outliers and missing values, and then normalize the processed dataset.

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

[0071] S31. Calculate the features of each dimension in each trajectory data, and integrate the features of each dimension into a feature vector.

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

[0073] Define a trajectory dataset of a certain length of N as S = {s(1), s(2), …, s(N)}, where each trajectory data is denoted as s(i), that is, the i-th sampled trajectory s(i) = {θ(i), r(i), h(i), v(i), t(i), rcs(i)}. Each letter in s(i) represents the target azimuth angle, target slant range, relative height, radial velocity, recording time, and RCS of the target at this moment. The calculation processes of each feature are as follows:

[0074] 1) Maximum radial acceleration: First, it is necessary to calculate the acceleration a m between every two adjacent point traces in the trajectory, and its formula is expressed as:

[0075]

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

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

[0078] 2) Maximum radial speed: That is, the maximum value of the radial speed in each track data;

[0079] 3) Minimum radial speed: That is, the minimum value of the radial speed in each track data;

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

[0081] 5) Acceleration change: Among them, a i+1 and a i-1 are the accelerations at two moments;

[0082] 6) Maximum radial speed value: That is, the maximum value of the magnitude of the radial speed in each track data;

[0083] 7) Minimum radial speed value: That is, the minimum value of the magnitude of the radial speed in each track data;

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

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

[0086] 10) Average speed: To calculate the average speed, it is first necessary to calculate the approximate speed of the object through the state change between adjacent echoes. We use to calculate, and d i is the displacement between the i-th echo and the (i - 1)-th echo. The following steps are required:

[0087] Calculate the displacements in each direction:

[0088] The displacement dx in the horizontal direction 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 displacement dy in the vertical direction is determined by the difference in the coordinates 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] Where r i , r i-1 respectively represent the target slant ranges of the two positions, θ i , θ i-1 respectively represent the target azimuth angles of the two positions, H i , H i-1 are the heights of the two positions respectively.

[0095] Calculate the total displacement:

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

[0097]

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

[0099]

[0100] After obtaining the velocity of each point in the track, calculate the average velocity according to the following formula

[0101]

[0102] 11) Standard deviation of velocity: Calculate the standard deviation of velocity through the following formula:

[0103]

[0104] 12) Velocity oscillation frequency: To reflect the local characteristics of velocity changes, the velocity oscillation frequency is introduced. At the current moment i, take the velocity values of the sample points in the historical length segment of L, and calculate the average velocity V mL , defined as

[0105] Then calculate the difference between the velocity V(i) of the current point and the historical average velocity V mL , and denote the sign as τ V(i), which is defined by the formula as:

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

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

[0108]

[0109] where the value of 1 indicates oscillation occurs, and the value of 0 indicates no oscillation occurs;

[0110] Furthermore, the threshold γ v and the length L, and the speed oscillation frequency f v are defined as:

[0111]

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

[0113] 14) Minimum speed: That is, the minimum value of the speed in each track data;

[0114] 15) Calculation of the mean heading angle: First, 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), calculate the heading angle h(i) at the current moment i. The calculation formula for the heading angle h(i) is:

[0115]

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

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

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

[0119]

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

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

[0122]

[0123] 17) Mean relative height: Calculate the mean relative height through the following formula:

[0124]

[0125] 18) Standard deviation of relative height: Calculate the standard deviation of relative height through the following formula:

[0126]

[0127] Thus, the calculation of each dimensional feature is completed;

[0128] S32. Eliminate the outliers caused by the dirty data in the multi-dimensional features. The outliers include infinite standard deviation of speed, infinite standard deviation of heading angle, and infinite standard deviation of relative height;

[0129] S33. Check whether there are missing values in the obtained feature vectors. If missing values are found, delete the corresponding entire data;

[0130] S34. Use the maximum-minimum normalization to normalize each dimensional feature.

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

[0132] As a preferred implementation method, 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, and then automatically construct an ensemble learning model through the TabularPredictor function and perform hyperparameter optimization;

[0134] S42. Based on the constructed model, use Permutation Importance to calculate the importance score of each feature;

[0135] S43. Set the threshold to 0.002 according to the distribution of the feature importance scores. If the importance score of a feature is less than 0.002, then the feature is regarded as unimportant and removed. After calculation, the RCS range difference and the acceleration change amount are removed, and the remaining features are the key feature set.

[0136] S5. Use the filtered key feature set to construct an efficient small and micro AutoGluon UAV classifier to achieve accurate identification of low-altitude targets;

[0137] In this embodiment, step S5 constructs a UAV recognition model by using the AutoGluon framework. First, manually limit the integration among 12 basic models, and then input to the Base layer. Generate the prediction results of each model through independent training and cross-validation, and output the meta-feature matrix (including the predicted values of all basic models). These meta-features will be used as the input of the Stack layer. Utilize the meta-feature matrix output by the Base layer and the target variable 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, as Figure 2 shown, step S5 specifically includes:

[0139] S51. From the key feature dataset filtered in step S4, load all the remaining 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, and select to search for the best combination among the following 12 basic models: LightGBMXT, RandomForestEntr, XGBoost, ExtraTreesGini, ExtraTreesEntr, RandomForestGini, LightGBM, CatBoost, KNeighborsDist, KNeighborsUnif, NeuralNetFastAI, and NeuralNetTorch; under the configured training time and computing resource limits, select the basic model combination to construct the AutoGluon UAV classifier;

[0142] S54. In the Base layer of AutoGluon, first, hyperparameter optimization is performed based on the selected basic model type; through random search and Bayesian optimization methods, the best hyperparameter combination for each basic model is found, and a meta-feature matrix containing the predicted values of all basic models is generated; then, using the meta-feature matrix output by the Base layer and the target variable, a meta-model is trained, and the combined weights of the basic models are optimized to improve the overall performance;

[0143] S55. Model training and result verification; the cross-validation method is adopted, and 20% of the data in the training set is selected each time for training and evaluation on the recognition model. According to multiple trainings, the model combination with the optimal performance is selected as the final AutoGluon drone classifier. The validation set is input into the final AutoGluon drone classifier for the recognition of small and micro drones, 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. Next, more refined recognition will be combined with iTransformer.

[0145] S6. Based on the enhanced track dataset, an iTransformer time series classifier is constructed and combined with the constructed AutoGluon drone classifier. Considering the recognition ability and accuracy of the two models, the accurate recognition of low-altitude targets is realized; in this embodiment, in step S6, the prediction results of the iTransformer and AutoGluon models are fused, and the training is optimized through a customized loss function. Finally, weighted averaging is used to improve the classification accuracy and robustness;

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

[0147] S61. Using the enhanced track data as input, training the iTransformer time series classifier to learn the time series features of the drone flight mode; the time series features 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] where ∈ is a regularization term used to reduce the trust of the model in the training data labels; y is the true label of the target track, is the probability that the iTransformer time series classifier identifies the target as a drone;

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

[0152] S63. Similarly, the cross-validation method is adopted. Each time, 20% of the data in the training set is selected for training and evaluation. Through multiple trainings, the optimal iTransformer time series classifier is obtained, and the optimal iTransformer time series classifier is used for drone identification. The calculated average accuracy is denoted as w i ;

[0153] S64. Adopt a strategy based on weighted fusion, combine the prediction outputs of the two models to obtain a more accurate drone classification probability. The calculation formula for the final probability P of identifying the target as a drone is:

[0154]

[0155] where, is the probability that the AutoGluon drone classifier identifies the target as a drone.

[0156] In this embodiment, first, the information in the track feature dimension extracted by the AutoGluon model is combined with the information based on the time series dimension extracted by the iTransformer time series classifier, and the output probabilities of the two are weighted and fused, which can improve the overall accuracy of the model, reduce the overfitting phenomenon, and enhance the robustness of the model to noise and uncertainty in the data, and finally obtain a more accurate recognition result.

[0157] In summary, the method proposed by the present invention is based on a multi-dimensional feature set that systematically characterizes the differences in motion performance and shape structure between drones and non-drones. The method of data augmentation is adopted by using a sliding window and prime step sampling. The AutoGluon framework is used to calculate the feature importance, and unimportant features are removed. The AutoGluon drone classifier based on important features is combined with the iTransformer time series classifier to achieve efficient discrimination between drones and non-drones in multiple feature dimensions. The final recognition model proposed by the present invention has a high recognition rate and strong adaptability, and can be widely applied to the fields of low-altitude drone detection and defense.

[0158] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, 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, it should be considered as the scope described in this specification.

[0159] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, 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 appended claims.

Claims

1. An intelligent recognition method for low-altitude small and micro UAV targets based on track features, 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 intelligent target recognition method for low-altitude micro UAVs based on track features according to claim 1, 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 intelligent target recognition method for low-altitude micro UAVs based on track features according to claim 1, 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 intelligent target recognition method for low-altitude small and micro unmanned aerial vehicles based on track features according to claim 2, characterized in that Step S3 specifically includes: S31. Calculate the feature of each dimension in each track data, and integrate the features of each dimension into a feature vector; S32. Remove the outliers caused by dirty data in the multi-dimensional features. The outliers include infinite standard deviation of speed, infinite standard deviation of heading angle, and infinite standard deviation of relative height; S33. Check whether there are missing values in the obtained feature vector. If missing values are found, delete the corresponding whole data; S34. Use min-max normalization to normalize the features of each dimension.

5. The intelligent target recognition method for low-altitude small and micro UAVs based on track features according to claim 1, characterized in that, Step S4 specifically includes: S41. Use the TabularDataset function in the AutoGluon library to load the track dataset, and then automatically construct an ensemble learning model through the TabularPredictor function and perform hyperparameter optimization; S42. Based on the constructed inheritance learning model, use Permutation Importance to calculate the importance scores of each normalized track feature in the track dataset; S43. Screen the track features according to the distribution of the feature importance scores; set the feature importance threshold. If the importance score of a feature is less than the threshold, the feature is regarded as unimportant and removed. The remaining features form the key feature set.

6. The intelligent target recognition method for low-altitude micro UAVs based on track features according to claim 1, characterized in that, Step S5 specifically includes: S51. From the key feature dataset screened in step S4, load all the remaining key features and their corresponding label data. Divide the loaded key feature set into a training set and a validation set. 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 from the following 12 base models: LightGBMXT, RandomForestEntr, XGBoost, ExtraTreesGini, ExtraTreesEntr, RandomForestGini, LightGBM, CatBoost, KNeighborsDist, KNeighborsUnif, NeuralNetFastAI, and NeuralNetTorch; under the configured training time and computing resource limits, select the base model combination to construct the AutoGluon drone classifier; S54. In the Base layer of AutoGluon, first perform hyperparameter optimization based on the selected base model type; through random search and Bayesian optimization methods, find the best hyperparameter combination for each base model and generate a meta-feature matrix containing the predicted values of all base models; these meta-features will be used as the input of the Stack layer, and then use the meta-feature matrix and target variables output by the Base layer 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 and evaluation each time. Based on multiple trainings, the model combination with the best performance is selected as the final AutoGluon UAV classifier. The validation set is input into the final AutoGluon UAV classifier for the identification of small and micro UAVs. The results of each identification are recorded, and the average accuracy of the AutoGluon UAV classifier is calculated, denoted as w a 。 7. The intelligent target recognition method for low-altitude micro UAVs based on track features according to claim 6, wherein Step S6 specifically includes: S61. Use the enhanced track data as input to train the iTransformer time series classifier to learn the time series features of the UAV flight mode; S62. Use the ∈ loss function to guide the learning process of the model; the formula expression of the ∈ loss function is: where ∈ is the regularization term used to reduce the model's confidence in the training data labels; y is the true label of the target track, is the probability that the iTransformer time series classifier identifies the target as a drone; S63. Similarly, the cross-validation method is adopted. Each time, 20% of the data in the training set is selected for training and evaluation. The optimal iTransformer time series classifier is obtained through multiple trainings, and the drone is identified using the optimal iTransformer time series classifier. The calculated average accuracy is denoted as w i ; S64. Adopt a weighted fusion-based strategy to combine the prediction outputs of the two models to obtain a more accurate UAV classification probability. The calculation formula for the final probability P of identifying the target as a UAV is: Among them, is the probability of the drone identified by the AutoGluon drone classifier being the target drone.

Images