Track-based anti-radar-free target identification method

Abstract

The invention discloses an unmanned aerial vehicle target identification method based on track characteristics, and the method comprises the steps: 1, obtaining the track data of a target through radar detection, and carrying out the preprocessing of the track data; step 2, sampling the preprocessed track data by using a sliding window to obtain track sample data of different window lengths; step 3, for each piece of track sample data, the abnormal possibility of each track point is judged by using the Doppler principle, and two corresponding backup track samples are acquired and saved; step 4, performing feature extraction on the obtained track sample data and the corresponding backup track sample from the perspective of radar professional knowledge and statistics to obtain a track feature vector and a backup feature vector; 5, performing feature screening on the track feature vector and the backup feature vector; step 6, training a specified first-level model and a specified second-level model; and step 7, for track sample data needing to be predicted, using the completed training model to perform prediction to obtain a prediction identification result.

Description

Technical Field

[0001] This invention relates to radar target recognition technology, and in particular to a method for identifying radar-free targets based on flight paths. Background Technology

[0002] With the rapid development of drone technology, drones have been widely used in both military and civilian fields due to their advantages such as low cost, small size, and ease of operation. Radar, with its advantages of long monitoring range, all-weather operation, and full-airspace scanning, has become the most suitable technology for monitoring drones over a wide area. Radar-based drone identification technology has become an important development direction for modern radar technology, playing a crucial role in improving system performance and ensuring security.

[0003] Radar-based UAV target recognition primarily relies on the echo characteristics or flight path features of UAV targets. Data analysis and mining are performed to extract hidden UAV target features, which are then matched against a known target feature database to achieve UAV target recognition. The main processing flow can be found in [reference needed]. Figure 1 However, radar identification of drones faces numerous challenges. On the one hand, existing drones are typically manufactured using non-metallic composite materials, resulting in small size, low radar echo intensity, weak identifiable features, and significant difficulty in effective detection. On the other hand, small drones usually operate in low-altitude airspace, where birds are another major type of non-cooperative "low, slow, and small" target. Studies have shown that birds and drones share high similarities in radar echo characteristics (such as radar cross-section), making them a major source of interference in drone target surveillance. Furthermore, target tracks may malfunction in dynamic clutter environments, thus affecting drone identification.

[0004] Currently, methods for monitoring and identifying drones and birds mainly fall into two categories. One category focuses on the echo characteristics of the targets. Because the echo signals generated by the multi-rotor rotation of drones and the flapping of birds' wings have different characteristics, time-frequency analysis algorithms are generally used to extract the micro-Doppler features of both. However, this approach remains challenging in complex and variable environments, especially when the target is at a distance. The weak echo characteristics of both types of birds lead to difficulties in feature extraction and low time-frequency resolution, limiting their widespread application and development. The other category focuses on the flight trajectory characteristics of drones and birds. By recording a large amount of historical data and understanding the activity patterns and flight trajectories of birds, the motion trajectory features of drones and birds can be extracted from the data layer. Currently, multiple motion models are used for identification, and the conversion frequency between models is calculated. Some scholars have proposed using trajectory information to establish feature quantities and employing machine learning or deep learning methods for identification. However, existing research has not considered the existence of abnormal tracking points in the target trajectory, which affects the feature extraction effect to some extent, ultimately leading to unsatisfactory identification results. Since the possibility of UAV target maneuvering needs to be considered, these abnormal tracking points are difficult to remove directly from the preprocessing stage. We need to further explore and study a comprehensive and efficient model method architecture to ensure that target trajectory recognition will not lead to recognition errors due to a certain abnormal trajectory point, and ultimately improve the overall recognition stability. Summary of the Invention

[0005] Objective of the Invention: The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a drone identification method based on flight path features. This method enables stable and accurate identification of drone targets even in dynamic clutter environments where there is an abnormal tracking point in the drone target's flight path, thereby meeting the needs of drone surveillance and defense.

[0006] To address the aforementioned technical problems, this invention discloses a track-based anti-radar-free target identification method, comprising:

[0007] Step 1: The radar detects the target and obtains the target's trajectory data. The trajectory data is preprocessed, including error data processing and the establishment of intermediate variables, to obtain the processed trajectory data.

[0008] Step 2: Use a sliding window to sample the preprocessed track data to obtain track sample data of different window lengths; The sliding window method is used to sample track sample data of different window lengths. Considering that the track data meets the temporal characteristics, track samples of different window lengths ([5,6,7,8,9,10]) need to be taken during the sliding sampling process to meet the prediction requirements of track data of different lengths.

[0009] Step 3: For each track sample data, use the Doppler principle to determine the probability of anomalies for each track point, remove the points with the highest and second highest probability of anomalies respectively, obtain two corresponding backup track samples and save them.

[0010] Step 4: From the perspectives of radar expertise and statistics, extract features from the acquired track sample data and the corresponding backup track samples to obtain track feature vectors and backup feature vectors;

[0011] Step 5: Use the embedding method to filter the track feature vector and backup feature vector to form the filtered track feature vector and backup feature vector; the track feature vector and backup feature vector refer to the feature set formed by combining the signal processing features, data processing features and statistical features of the track and backup track, respectively.

[0012] Step 6: Train the specified first-level and second-level models using the improved temporal stacking ensemble learning framework;

[0013] Step 7: For the trajectory sample data that needs to be predicted, use the improved temporal stacking ensemble learning framework that has been trained to make predictions and obtain the prediction and recognition results of whether the target is a drone.

[0014] The trajectory data mentioned in step 1 includes information in six dimensions: bearing, slant range, relative altitude, Doppler velocity, time, and scattering cross section;

[0015] The error data processing specifically includes: processing error values, outlier values, and duplicate values.

[0016] Error value handling refers to deleting errors caused during data processing and tagging;

[0017] Outlier handling refers to the deletion of values ​​that are significantly higher or lower than the range of values ​​for certain variables, as they do not conform to the norm.

[0018] The duplicate value processing refers to deleting data in frames that are identical to those in preceding frames.

[0019] In step 1, the intermediate variables include: horizontal distance x, vertical distance y, pitch Ele, distance Dis, velocity V, acceleration Acc, radial velocity JxV, and heading angle Dir.

[0020] Horizontal distance:

[0021] Vertical distance:

[0022] Looking up and down:

[0023] distance:

[0024] speed:

[0025] Acceleration:

[0026] Radial velocity:

[0027] Heading angle:

[0028] in, This represents the orientation of the i-th waypoint. Represents the slant distance of the i-th waypoint. Represents the horizontal distance to the i-th waypoint. This represents the vertical distance to the i-th waypoint. Represents the relative altitude of the i-th waypoint. represents the time at the i-th track point, arcsin represents the arcsine function, and atan2 represents the arctangent function in the four quadrants.

[0029] Step 3 describes using the Doppler principle to determine the probability of an anomaly for each track point. Specifically, it involves comparing the difference between the Doppler velocity of the track point and the radial velocity obtained by solving the track point's position. If the radial velocity is within the non-ambiguity threshold range, the larger the difference between the two, the greater the probability that the point is an anomaly.

[0030] The unambiguity threshold range is the maximum unambiguity speed of the radar.

[0031] The track feature vector and backup feature vector extracted from the perspective of radar expertise in step 4 specifically include signal processing features and data processing features.

[0032] Signal processing feature extraction includes: performing Discrete Fourier Transform on horizontal distance, vertical distance, and relative height, and extracting the maximum amplitude from the amplitude and phase spectrum information obtained from the Fourier Transform. and the frequency corresponding to the maximum amplitude and phase information The discrete Fourier transform features with different characteristics specifically include:

[0033] Horizontal distance x: ];

[0034] Vertical distance y: ];

[0035] Relative height h: ];

[0036] Data processing and feature extraction mainly include: Turn Rate and Detection Probability. :

[0037] The formula is as follows:

[0038] Cumulative distance of waypoint i:

[0039] Changes in heading angle at track point i:

[0040] Turn rate of waypoint i: Detection probability:

[0041] Where dotCnt represents the number of waypoints and period represents the number of processing cycles for the waypoint.

[0042] Step 4 describes the extraction of statistical features from the calculation object of track point i from a statistical perspective, including statistical features such as the mean. Standard deviation Range Maximum value Minimum value , median Maximum fluctuation value Maximum volatility Average fluctuation value Average volatility Coefficient of variation Autocorrelation coefficient ;

[0043] The formula is as follows:

[0044] Mean: ;

[0045] Standard deviation: ;

[0046] Maximum value: ;

[0047] Minimum value: ;

[0048] Range: ;

[0049] Median: ;

[0050] Maximum fluctuation value: ;

[0051] Maximum volatility: ;

[0052] Average volatility: ;

[0053] Average volatility: ;

[0054] Coefficient of variation: ;

[0055] Autocorrelation coefficient: ;

[0056] In the formula, This represents a feature of a specific dimension in the track feature vector or backup feature vector. , The number of track points representing a single track sample or a backup track sample. Represents the maximum value. Represents the minimum value. Represents the median. Represents covariance, Represents variance. Represents the root mean square.

[0057] The calculation object for the waypoint i includes: distance ,position Looking up and down ,speed acceleration radial velocity Doppler velocity Turning rate , scattering cross-section Rcs.

[0058] The specific steps for using the embedding method to filter the track feature vector and backup feature vector in step 5 are as follows: First, a random forest model is used to train the track feature vector. The model is used to obtain the weight coefficients of each feature. Features are selected from large to small according to the weight coefficients. Feature variables with weight coefficients greater than a certain threshold are selected as the filtered track feature vector. Feature variables with the same meaning are also selected as the filtered backup feature vector.

[0059] Step 6 describes using the improved temporal stacking ensemble learning framework to fundamentally restructure the ensemble learning paradigm, which specifically includes the following steps:

[0060] Step 6-1: Use the filtered track feature vector as input and track type label as output to train the first-level model, and obtain the prediction probability output of each first-level model as the original prediction result.

[0061] Step 6-2: Use the two corresponding filtered backup feature vectors as input to the first-level model, and use the corresponding first-level model output as the backup prediction result.

[0062] Step 6-3: Calculate the mean and variance of the three predictions (original prediction, backup prediction 1, and backup prediction 2) obtained in steps 6-1 and 6-2, and use them as enhancement features;

[0063] Step 6-4: When training the secondary model, combine the original prediction results, two backup prediction results, enhanced features (mean and variance), and the filtered track feature vector as the input of the secondary model, and use the track type label as the output to train the secondary model.

[0064] Preferably, the first-level model in step 6 uses the three most popular and high-performance classic tree model algorithms: XGboost, LightGBm, and CatBoost. The second-level model uses the LightGBm model.

[0065] Beneficial effects:

[0066] (1) This invention fully analyzes and mines radar UAV target trajectory data, extracts key effective features, and establishes a suitable model, which can achieve high recognition accuracy and low false alarm rate of UAVs.

[0067] (2) This invention proposes an outlier analysis method based on the Doppler principle, which creates two “denoised” versions (backup track 1 and backup track 2) for each track sample, providing a cleaner and more reliable data foundation for subsequent model training, and effectively avoiding the impact of track anomaly tracking points;

[0068] (3) This invention proposes an improved temporal stacking ensemble learning framework, which considers the predictions and original features from three perspectives (original track, backup track 1, and backup track 2) during the training and prediction of the secondary model. This is equivalent to the secondary model seeing three "expert opinions" (probabilities) and a set of "original data" (original features), and also considers the mean and variance of the three predictions, thereby improving the robustness of the entire system's prediction. Attached Figure Description

[0069] Figure 1 This is a schematic block diagram of the radar target identification process;

[0070] Figure 2 This is a flowchart of the improved Stacking ensemble learning framework;

[0071] Figure 3 This is a flowchart illustrating the drone identification method based on improved Stacking ensemble learning provided in an embodiment of the present invention.

[0072] Figure 4 This is a screenshot showing the fully automatic tracking and identification of clutter areas in a low-altitude anti-radar application of the present invention. Detailed Implementation

[0073] Example 1: The scenario in this example is: anti-drone radar plays an important role in the anti-drone system. In order to ensure the effectiveness of real-time monitoring of drones, it is necessary to accurately and stably identify the drone's flight path.

[0074] This invention proposes a track-based anti-radar-free target identification method, characterized by comprising:

[0075] Step 1: The radar detects the target and obtains the target's trajectory data. The trajectory data contains information in six dimensions: azimuth, slant range, relative altitude, Doppler velocity, time, and cross section. The trajectory data is preprocessed to obtain the processed trajectory data.

[0076] Step 2: Use a sliding window to sample the preprocessed track data to obtain track sample data of different window lengths, that is, sub-tracks of different lengths. The window lengths are set to 5, 6, 7, 8, 9, and 10 respectively. The window length corresponds to the number of points of the sub-tracks obtained.

[0077] Step 3: For each track sample data, use the Doppler principle to determine the probability of anomalies for each track point, remove the points with the highest and second highest probability of anomalies respectively, obtain two corresponding backup track samples and save them.

[0078] Step 4: From the perspective of radar expertise and statistics, perform feature extraction on the acquired track sample data and the corresponding backup track samples to obtain track feature vectors and backup feature vectors; the dimension of the track feature vectors and backup feature vectors obtained by feature extraction is equal to the six dimensions contained in the original track data plus the types of intermediate variables.

[0079] Step 5: Use the embedding method to filter the track feature vector and backup feature vector to form the filtered track feature vector and backup feature vector;

[0080] Step 6: Train the specified first-level and second-level models using the improved temporal stacking ensemble learning framework;

[0081] Step 7: For the trajectory sample data that needs to be predicted, use the improved temporal stacking ensemble learning framework that has been trained to make predictions and obtain the prediction and recognition results of whether the target is a drone.

[0082] Step 1, which involves preprocessing the track data, includes handling error values, outliers, and duplicate values. Specifically, this includes:

[0083] The error value handling mainly refers to deleting errors caused during data processing and tagging;

[0084] Outlier handling mainly refers to the fact that the values ​​of certain variables are much higher or lower than the range of values ​​for that variable. These abnormal values ​​can be treated as outliers and deleted.

[0085] The processing of duplicate values ​​mainly refers to the need to delete duplicate data when the data of some frames is the same as that of the preceding frames.

[0086] On the other hand, it refers to establishing intermediate variables to facilitate subsequent feature extraction. In this example, the intermediate variables include horizontal distance x, vertical distance y, and pitch. ,distance ,speed acceleration radial velocity heading angle .

[0087] The calculation formula is as follows:

[0088] Horizontal distance:

[0089] Vertical distance:

[0090] Looking up and down:

[0091] distance:

[0092] speed:

[0093] Acceleration:

[0094] Radial velocity:

[0095] Heading angle:

[0096] in, This represents the orientation of the i-th waypoint. Represents the slant distance of the i-th waypoint. Represents the horizontal distance to the i-th waypoint. This represents the vertical distance to the i-th waypoint. Represents the relative altitude of the i-th waypoint. represents the time at the i-th track point, arcsin represents the arcsine function, and atan2 represents the arctangent function in the four quadrants.

[0097] Step 2 describes using the sliding window method to sample the preprocessed track data to obtain track sample data of different window lengths. This means that, considering the temporal characteristics of the track data, different window lengths ([5,6,7,8,9,10]) are used during the sliding sampling process to meet the prediction requirements of track data of different lengths. For example, a window length of 8 is used, defining a track data of length 10. ,in Feature information representing the i-th point of the track sample ,in, Represents direction, Represents slant distance, Represents relative altitude, DropV represents Doppler velocity, Time represents time information, and Rcs represents scattering cross-section. The track sample data obtained after sampling for a length of 8 is... .

[0098] Step 3, which describes using the Doppler principle to determine the anomaly probability of each waypoint, specifically involves comparing the difference between the Doppler velocity of the waypoint and the radial velocity obtained through the waypoint's position calculation. If the radial velocity is within the maximum unambiguous velocity range, the larger the difference between the two, the greater the probability that the point is an anomaly. In the track data defined above, taking the 5th waypoint as an example, its Doppler velocity is... The radial velocity calculated using track points 5 and 6 is Based on the design principles of radar signal processing, the maximum unambiguous velocity of this radar is obtained as follows: In this embodiment The possibility of an anomaly at the 5th waypoint The calculation formula is as follows:

[0099]

[0100] The feature extraction described in step 4 from the perspective of radar expertise includes signal processing features and data processing features.

[0101] Signal processing feature extraction includes: performing Discrete Fourier Transform on horizontal distance, vertical distance, and relative height, and extracting the maximum amplitude from the amplitude and phase spectrum information obtained from the Fourier Transform. and the frequency corresponding to the maximum amplitude and phase information The discrete Fourier transform features with different characteristics specifically include:

[0102] Horizontal distance x: ];

[0103] Horizontal distance y: ];

[0104] Height H: ];

[0105] Data processing feature extraction mainly includes: turning rate and detection probability .

[0106] The formula is as follows:

[0107] Cumulative distance:

[0108] Changes in heading angle:

[0109] Turn rate: Detection probability:

[0110] in, Represents the distance of the intermediate variable that has been created. The intermediate variable heading has been created; dotCnt represents the number of waypoints, and period represents the number of processing cycles for the waypoint.

[0111] The statistical features extracted from a statistical perspective in step 4 include statistical features such as the mean. Standard deviation Range Maximum value Minimum value , median Maximum fluctuation value Maximum volatility Average fluctuation value Average volatility Coefficient of variation Autocorrelation coefficient ;

[0112] The formula is as follows:

[0113] Mean: ;

[0114] Standard deviation: ;

[0115] Maximum value: ;

[0116] Minimum value: ;

[0117] Range: ;

[0118] Median: ;

[0119] Maximum fluctuation value: ;

[0120] Maximum volatility: ;

[0121] Average volatility: ;

[0122] Average volatility: ;

[0123] Coefficient of variation: ;

[0124] Autocorrelation coefficient: ;

[0125] In the formula, This represents a feature of a specific dimension in the track feature vector or backup feature vector. , This represents the number of points in a single track sample or a backup track sample. Represents the maximum value. Represents the minimum value. Represents the median. Represents covariance, Represents variance.

[0126] The statistical features mentioned in step 4 are those features that require statistical calculation, including: distance. ,position Looking up and down ,speed acceleration radial velocity Doppler velocity Turning rate The scattering cross-section Rcs is used to obtain the final statistical eigenvector StatVector = [R avg , R std , R max , R min ,R ptp , R med , R fluctmax , R fluctRatioMax , R fluctAvg , R fluctRatio , R cv , R autocorr Azi avg Azi std Azi max Azi min Azi ptp Azimed , Azi fluctmax , Azi fluctRatioMax , Azi fluctAvg , Azi fluctRatio ,Azi cv , Azi autocorr , Ele avg , Ele std , Ele max , Ele min , Ele ptp , Ele med , Ele fluctmax ,Ele fluctRatioMax , Ele fluctAvg , Ele fluctRatio , Ele cv , Ele autocorr ,V avg , V std , V max , V min ,V ptp ,V med , V fluctmax , V fluctRatioMax , V fluctAvg , V fluctRatio , V cv , V autocorr , Acc avg , Acc std , Acc max ,Acc min , Acc ptp , Acc med , Acc fluctmax , Acc fluctRatioMax , Acc fluctAvg , Acc fluctRatio , Acc cv ,Acc autocorr , JxV avg , JxV std , JxV max , JxV min , JxV ptp , JxV med , JxV fluctmax , JxV fluctRatioMax ,JxV fluctAvg , JxV fluctRatio , JxV cv , JxV autocorr , DropV avg , DropV std , DropVmax DropV min DropV ptp DropV med DropV fluctmax DropV fluctRatioMax DropV fluctAvg DropV fluctRatio DropV cv DropV autocorr TurnRateV avg TurnRateV std TurnRateV max TurnRateV min TurnRateV ptp TurnRateV med TurnRateV fluctmax TurnRateV fluctRatioMax TurnRateV fluctAvg TurnRateV fluctRatio TurnRateV cv TurnRateV autocorr , Rcs max , Rcs min , Rcs ptp , Rcs med ,Rcs fluctmax , Rcs fluctRatioMax , Rcs fluctAvg , Rcs fluctRatio , Rcs cv , Rcs autocorr ]

[0127] The meaning of each subscript corresponds to the statistical characteristics mentioned above.

[0128] The track feature vector and backup feature vector mentioned in step 4 refer to the feature set formed by combining the signal processing features, data processing features, and statistical features of the track and backup track, respectively. The feature set is AllVector=[

[0129] TopAmp x TopFreq x TopPhase x TopAmp y TopFreq y TopPhase y TopAmp y TopFreq hTopPhase h ,TurnRate,DetectProba,StatVector].

[0130] Step 5 describes using the embedding method to filter the track feature vector and backup feature vector. Specifically, the track feature vector is first trained using a random forest model. The model is then used to obtain the weight coefficients of each feature. Features are selected from largest to smallest based on their weight coefficients. Feature variables with weight coefficients greater than a certain threshold are selected as the filtered track feature vector. Similarly, feature variables with the same meaning are selected as the filtered backup feature vector. In this embodiment, the threshold is set to 0.02.

[0131] Step 6 describes using the improved temporal stacking ensemble learning framework to fundamentally restructure the ensemble learning paradigm, which specifically includes the following steps:

[0132] Step 6-1: Use the filtered track feature vector as input and track type label as output to train the first-level model, and obtain the prediction probability output of each first-level model as the original prediction result.

[0133] Step 6-2: Use the two corresponding filtered backup feature vectors as input to the first-level model, and use the corresponding first-level model output as the backup prediction result.

[0134] Step 6-3: Calculate the mean and variance of the three predictions (original prediction, backup prediction 1, and backup prediction 2) obtained in steps 6-1 and 6-2, and use them as enhancement features;

[0135] Step 6-4: When training the secondary model, combine the original prediction results, two backup prediction results, enhanced features (mean and variance), and the filtered track feature vector as the input of the secondary model, and use the track type label as the output to train the secondary model.

[0136] The first-level model described in step 6 uses the three most popular and high-performance classic tree model algorithms: XGboost, LightGBm, and CatBoost.

[0137] The secondary model uses the LightGbm model.

[0138] In practical implementation, the first-level model can include the three most popular and high-performing classic tree model algorithms: XGboost, LightGBm, and CatBoost.

[0139] The parameter settings during the training process of the Level 1 XGBoost model are as follows:

[0140] Table 1. Parameters to be optimized

[0141]

[0142] The parameter settings during the training process of the Level 1 LightGBm model are as follows:

[0143] Table 2. Parameter Settings for Level 1 LightGbm Model

[0144]

[0145] The parameter settings during the training process of the Level 1 CatBoost model are as follows:

[0146] Table 3. Parameter Settings for Level 1 CatBoost Model

[0147]

[0148] In practice, the secondary model may include the LightGbm model.

[0149] The parameter settings during the training of the Level 2 LightGbm model are as follows:

[0150] Table 4. Parameter Settings for the Second-Level LightGbm Model

[0151]

[0152] The parameters in the table above are explained as follows:

[0153] n_estimators: The number of base learners. Generally, the larger the number of base learners, the stronger the model's fitting ability, but the greater the model's complexity and the easier it is to overfit.

[0154] The ratio of positive to negative sample weights is generally used to deal with class imbalance. The ratio of positive to negative sample weights can be adjusted, and it is usually set to the reciprocal of the ratio of the number of positive to negative samples.

[0155] Class weights are generally used to adjust class weights to address the problem of class imbalance in samples. They are usually set to 'balanced' or set according to the proportion of the number of samples.

[0156] : Maximum depth of decision tree: Specifies the maximum depth to which the decision tree can grow. Generally, the larger the maximum depth, the stronger the model's fitting ability, but the greater the model complexity and the easier it is to overfit.

[0157] The maximum number of leaf nodes in a decision tree directly affects the complexity of the decision tree and the level of detail in the data that the model can capture.

[0158] During specific training, ten-fold cross-validation is used to avoid overfitting of the model.

[0159] The following is combined Figure 4 The invention will be further illustrated through its application in engineering and the evaluation of its effects. Figure 4 This is a screenshot of a radar display used by a certain anti-drone radar for real-time drone identification. TV / W represents the track header, where TV indicates a non-drone track, and W indicates a track identified as a drone. The "-I" in the track header indicates a priority target. For example, TV0825 indicates that the track with batch number 0825 was identified as a non-drone, while W0490-I indicates that the track with batch number 0490 was identified as a drone and is a priority target. Through manual comparison with real data, all targets identified as drones in the image are confirmed to be genuine drones, while all targets not identified as drones are non-drone targets. This demonstrates that this method can accurately identify drones in complex environments, ensuring the accuracy and stability of target intelligence and significantly improving the system's automated intelligence capabilities.

[0160] As described above, this invention is a UAV identification method based on flight track features. It extracts multi-dimensional features from radar expertise and statistical perspectives, and uses the embedding method to automatically select more important features as filtered features. It performs anomaly analysis on flight track points using the Doppler principle, establishes two backup flight track samples for each flight track to be predicted to avoid the impact of flight track tracking errors, and establishes an improved temporal stacking ensemble learning model. It comprehensively utilizes the advantages of multi-view data features (original flight track, backup flight track 1, backup flight track 2, and enhanced features) and multi-model fusion to improve the accuracy of UAV identification, and finally achieves accurate and stable identification of UAV targets in complex environments.

[0161] Compared to traditional Stacking frameworks, the innovation of the improved Stacking integration framework is specifically reflected in the following aspects:

[0162] 1) Innovation in information fusion mechanism: It breaks the one-way information flow of the traditional cascade model and realizes the parallel fusion of multi-source information;

[0163] 2) Improved hierarchical error correction capability: The second-level model not only receives the results of the first-level model, but also receives the original information and historical information as references, and has the ability to correct the first-level model.

[0164] It enables cross-level feature interaction: it allows interaction between the original feature space, the historical prediction information space, and the prediction probability space, thus achieving cross-level feature engineering.

[0165] This invention provides a track-based anti-radar-free target identification method. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A track-based anti-radar-free target identification method, characterized in that, include: Step 1: The radar detects the target and obtains the target's trajectory data. The trajectory data is preprocessed, including error data processing and the establishment of intermediate variables, to obtain the processed trajectory data. Step 2: Use a sliding window to sample the preprocessed track data to obtain track sample data with different window lengths; Step 3: For each track sample data, use the Doppler principle to determine the probability of anomalies for each track point, remove the points with the highest and second highest probability of anomalies respectively, obtain two corresponding backup track samples and save them. Step 4: From the perspectives of radar expertise and statistics, extract features from the acquired track sample data and the corresponding backup track samples to obtain track feature vectors and backup feature vectors; Step 5: Use the embedding method to filter the track feature vector and backup feature vector to form the filtered track feature vector and backup feature vector; Step 6: Train the specified first-level and second-level models using the improved temporal stacking ensemble learning framework; Step 7: For the trajectory sample data that needs to be predicted, use the improved temporal stacking ensemble learning framework that has been trained to make predictions and obtain the prediction and recognition results of whether the target is a drone.

2. The anti-radar-free target identification method based on flight path according to claim 1, characterized in that, The trajectory data mentioned in step 1 includes information in six dimensions: bearing, slant range, relative altitude, Doppler velocity, time, and scattering cross section; The error data processing specifically includes: processing error values, outlier values, and duplicate values.

3. The anti-radar-free target identification method based on flight path according to claim 2, characterized in that, In step 1, the intermediate variables include: horizontal distance x, vertical distance y, pitch Ele, distance Dis, velocity V, acceleration Acc, radial velocity JxV, and heading angle Dir.

4. The anti-radar-free target identification method based on flight path according to claim 2, characterized in that, Step 3 describes using the Doppler principle to determine the probability of an anomaly for each track point. Specifically, it involves comparing the difference between the Doppler velocity of the track point and the radial velocity obtained by solving the track point's position. If the radial velocity is within the non-ambiguity threshold range, the larger the difference between the two, the greater the probability that the point is an anomaly.

5. The anti-radar-free target identification method based on flight path according to claim 4, characterized in that, The unambiguity threshold range is the maximum unambiguity speed of the radar.

6. The anti-radar-free target identification method based on flight path according to claim 5, characterized in that, The track feature vector and backup feature vector extracted from the perspective of radar expertise in step 4 specifically include signal processing features and data processing features; Signal processing feature extraction includes: performing Discrete Fourier Transform on horizontal distance, vertical distance, and relative height, and extracting the maximum amplitude from the amplitude and phase spectrum information obtained from the Fourier Transform. and the frequency corresponding to the maximum amplitude and phase information ; Data processing and feature extraction mainly include: Turn Rate and Detection Probability. .

7. The anti-radar-free target identification method based on flight path according to claim 5, characterized in that, Step 4 describes the extraction of statistical features from the calculation object of track point i from a statistical perspective, including statistical features such as the mean. Standard deviation Range Maximum value Minimum value , median Maximum fluctuation value Maximum volatility Average fluctuation value Average volatility Coefficient of variation Autocorrelation coefficient .

8. The anti-radar-free target identification method based on flight path according to claim 7, characterized in that, The calculation object for the waypoint i includes: distance ,position Looking up and down ,speed acceleration radial velocity Doppler velocity Turning rate , scattering cross-section Rcs.

9. The anti-radar-free target identification method based on flight path according to claim 7, characterized in that, The specific steps for using the embedding method to filter the track feature vector and backup feature vector in step 5 are as follows: First, a random forest model is used to train the track feature vector. The model is used to obtain the weight coefficients of each feature. Features are selected from large to small according to the weight coefficients. Feature variables with weight coefficients greater than a certain threshold are selected as the filtered track feature vector. Feature variables with the same meaning are also selected as the filtered backup feature vector.

10. A method for anti-radar-free target identification based on flight paths according to claim 8, characterized in that, Step 6 describes using the improved temporal stacking ensemble learning framework to fundamentally restructure the ensemble learning paradigm, which specifically includes the following steps: Step 6-1: Use the filtered track feature vector as input and track type label as output to train the first-level model, and obtain the prediction probability output of each first-level model as the original prediction result. Step 6-2: Use the two corresponding filtered backup feature vectors as input to the first-level model, and use the corresponding first-level model output as the backup prediction result; Step 6-3: Calculate the mean and variance of the three prediction results obtained in steps 6-1 and 6-2, and use them as enhancement features; Step 6-4: When training the secondary model, the original prediction results, two backup prediction results, enhanced features, and filtered track feature vectors are combined as the input of the secondary model, and the track type label is used as the output to train the secondary model.

Images