Method for identifying unmanned aerial vehicles and bird targets by radar

Abstract

The application provides a UAV and bird target radar identification method, first compares the difference of the two targets in the motion trajectory, carries out feature analysis, proposes a time-related heading oscillation frequency and speed oscillation frequency feature quantity description method, and in an offline state, uses the track data recorded by the actual measurement radar system to extract the effective feature quantities of the two targets, then uses the support vector machine algorithm to train the samples and after obtaining the optimal model parameters, tests the samples, and the test classification result shows that a high accurate identification rate can be achieved, finally in an online state, through a large number of flying experiments, the results show that the method is correct, and the application has light weight, practicality and applicability in the engineering implementation, and has high value.

Description

Technical Field

[0001] This invention relates to radar signal processing, data processing, target recognition, and other problems, and can be widely used in fields such as bird detection radar and anti-drone radar. 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 simple operation. However, the frequent occurrence of incidents such as unauthorized drone flights and interference has become a major hidden danger. In addition, incidents of aircraft colliding with birds during takeoff or flight are also emerging one after another, and the pressure to prevent bird strikes is increasing. Both types of targets pose a serious threat to important places such as airports and security. Generally, depending on the requirements of different combat environments, different countermeasures and measures need to be taken after these two types of targets are detected. For example, airports usually use bird deterrent equipment to drive away birds after they are detected, while for drone targets, early warning and coordination measures are taken. Therefore, the identification of drones and birds is extremely necessary.

[0003] Currently, methods for monitoring and identifying drones and birds mainly fall into two categories. One category focuses on the echo characteristics of the target. Since 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. The French Space Centre (FSC) conducted micro-Doppler measurement experiments on multi-rotor drones using DVB-T signals, while the radar from the Dutch company Robin uses a transmitted frequency-modulated continuous wave waveform, possessing both drone detection and identification capabilities. In subsequent years of development, domestic and international organizations have further refined the extraction of features by designing and optimizing radar operating systems, operating modes, transmitted and received waveforms, and extending dwell time. However, in practical applications in complex and ever-changing environments, this remains challenging, especially when the target is at a distance. The weak echo characteristics of both types of drones lead to difficulties in feature extraction and low time-frequency resolution, limiting their widespread application and development. Another approach focuses on the flight trajectory characteristics of drones and birds. By recording a large amount of historical data, we can become familiar with the activity patterns and flight trajectories of birds and extract the motion trajectory characteristics of drones and birds from the data layer. Currently, there are methods that use multiple motion models and calculate the conversion frequency between models for identification, as well as methods that use flight trajectory information to establish feature quantities and utilize machine learning or deep learning. These methods have further verified the feasibility of this type of approach. Summary of the Invention

[0004] This invention proposes a radar identification method for UAVs and birds. First, based on the analysis and research of a large amount of historical radar system track data, the method studies the characteristic differences between UAVs and birds in their trajectories. It proposes to establish five feature quantities based on track layer information: heading angle standard deviation, heading oscillation frequency, mean velocity, velocity standard deviation, and velocity oscillation frequency. Then, it adopts a support vector machine algorithm to divide the historical samples into training samples and test samples. The optimal hyperparameters are obtained by cross-training the training samples, and the accuracy of identification and classification is verified by using test samples.

[0005] The radar identification method for UAVs and birds proposed in this invention differs from existing identification methods in the following ways: First, it does not employ complex time-frequency analysis methods, but instead uses a lightweight flight track dataset, which greatly reduces the identification cost of the radar system and has strong universality; Second, in order to better reflect the characteristic differences in the flight trajectories of the two birds, a support vector machine algorithm model is adopted, and a time-related heading oscillation frequency and velocity oscillation frequency feature vector description method is proposed. Combined with other feature quantities, online dynamic correction and prediction are achieved, which can better improve the identification accuracy. Attached Figure Description

[0006] Figure 1 This is a block diagram of drone and bird target recognition according to an embodiment of the present invention;

[0007] Figure 2 This is a schematic diagram of a typical UAV flight trajectory according to an embodiment of the present invention;

[0008] Figure 3 This is a schematic diagram of a typical bird flight trajectory according to an embodiment of the present invention;

[0009] Figure 4 This is a PPI map for unmanned aerial vehicle target tracking and recognition according to an embodiment of the present invention;

[0010] Figure 5 This is a PPI image for bird target tracking and recognition according to an embodiment of the present invention. Detailed Implementation

[0011] This invention proposes a radar-based method for identifying UAVs and birds. First, under offline conditions, a large amount of historical experimental flight track data recorded by the radar system is analyzed. Then, the differences in the flight trajectories of UAVs and birds are compared, and feature analysis and calculation are performed to determine five feature quantities: standard deviation of heading angle, heading oscillation frequency, mean velocity, standard deviation of velocity, and velocity oscillation frequency. Next, the dataset is divided into training and testing sets. The training set data is used as input, and the SVM algorithm is used to calculate the model and optimal hyperparameters. The classification accuracy is verified using the testing set. The specific process is as follows: Figure 1 .

[0012] 1.1. Feature Analysis

[0013] Describing the flight paths of drones and birds, birds generally have relatively flexible and short flight times, while typical drone operations usually fly according to pre-set routes and tasks. Common routes include continuous straight-line flight, waypoint hovering, turnaround flight, and multi-directional multi-point flight. Therefore, the flight paths of such drone targets are longer and generally more stable. However, if the flight is manually controlled, it has a higher degree of freedom in turning, turning around, and hovering. Continuous manual control of drone maneuvering is called strong maneuvering mode, while operation according to waypoint tasks is called weak maneuvering mode.

[0014] Utilizing a phased array radar system platform, which is capable of detecting and tracking drones and birds, through extensive experimental work and analysis of historical flight track data, typical flight trajectories of drones and birds at the track level are shown as follows: Figure 2 , Figure 3 The difference between the two can be intuitively seen from the distance, azimuth angle, and track length. In order to improve the database, the UAV track data layer contains a large number of UAV tracks of different operation types, such as manual operation flight, straight flight, figure-eight flight, return flight, hovering flight and other different trajectories.

[0015] The differences in characteristic quantities between UAVs and birds are described below from the local and overall characteristics of their motion trajectories. These characteristic quantities are the standard deviation of heading angle, heading oscillation frequency, average speed, speed standard deviation, and speed oscillation frequency.

[0016] Heading angle standard deviation: For the trajectory of a typical bird, its directionality is relatively random, and the standard deviation varies widely. However, when a UAV performs weak maneuvering mode, its standard deviation is smaller. For this type, the difference between the two flight trajectories can be clearly reflected by comparing the difference in heading angle standard deviation. In strong maneuvering mode, the value range varies more widely, making it more difficult to distinguish.

[0017] Heading oscillation frequency: The standard deviation of heading angle mentioned above can only reflect the overall characteristics of both and cannot well describe the local heading change characteristics. Therefore, heading oscillation frequency is introduced to characterize the difference between the two. The heading oscillation frequency of UAVs is relatively small in weak maneuvering mode. However, if manual intervention and control are only performed for part of the flight to keep it in strong maneuvering mode, the heading oscillation frequency will gradually decrease as the number of tracking operations increases. If it is in strong maneuvering mode all the time, the heading oscillation frequency will remain at a high value. Therefore, the heading oscillation frequency of UAVs is always at a large or small value. The flight path of birds is more flexible and relatively smooth. In addition, the tracking time of birds by radar is generally short. Therefore, the heading oscillation frequency of birds is in the range between small and large. Therefore, heading oscillation frequency has the potential to distinguish between UAVs and birds.

[0018] Average speed: It is not very obvious in distinguishing between birds and general drones, but it can still be used as a feature to distinguish some large drones from bird targets, for example, the flight speed of general fixed-wing drones is much higher than that of birds.

[0019] Speed ​​standard deviation: For birds, their speed is relatively stable, and the standard deviation is generally small. However, drones fly more freely, and maneuvers such as hovering, acceleration, and deceleration affect their flight speed. Therefore, the standard deviation is relatively large during their maneuvers, which can be used as a characteristic to distinguish drones from birds.

[0020] Speed ​​oscillation frequency: For birds, their speed does not change abruptly. However, during the turning flight or hovering process of drones, acceleration or deceleration is required. The speed standard deviation mentioned above can only reflect the overall characteristics of both and cannot well describe the local speed change characteristics. Therefore, speed oscillation frequency is introduced to reflect the local speed change characteristics. Generally, the speed oscillation frequency value of birds is very small. For drones, although the acceleration or deceleration process only takes a short time, the speed oscillation frequency value is relatively large. Therefore, it can be accumulated over a long period of time and can also be used as a flight characteristic to distinguish between birds and drones in strong maneuvering mode.

[0021] 1.2. Feature Calculation

[0022] Define a target set of a track of length N as S = {s(1), s(2), ..., s(N)}, and denote each data sampling point as s, where the i-th track sampling point... These represent the target's distance, azimuth, pitch, and velocity at that point, respectively.

[0023] 1. Standard deviation of heading angle

[0024] To facilitate the calculation of the heading angle, the distance, azimuth, and pitch of each sampling point s(i) in the track data set are used. The transformation formula is as follows: (x(i), y(i), z(i)) into a Cartesian coordinate system.

[0025]

[0026] Calculate the position information x(i-1), y(i-1), x(i), and y(i) in the Cartesian coordinate system based on the previous time s(i-1) and the current time s(i). Calculate the heading angle at the current time i, denoted as h(i), using the following formula:

[0027]

[0028] When calculating the heading angle, the quadrant regions where the values ​​of x(i)-x(i-1) and y(i)-y(i-1) lie need to be considered, thus transforming the value range of -90° to 90° into 0° to 360°. Therefore, the track set S and the heading angle set H are...

[0029] H={h(2),h(3),...,h(N)} (3)

[0030] To calculate the standard deviation of the heading angle, the mean of the heading angle must first be calculated. However, some target trajectories have heading angles that span 0° or 360°, or these spans may occur due to measurement angle errors. Therefore, the conventional mean method cannot be used to calculate the heading angle. This invention uses a vectorization method to describe this type of periodic angle data. The method for calculating the mean and standard deviation of periodic data is to first transform the i-th heading angle h(i) into a unit vector on a two-dimensional plane. The heading angle corresponds to cosh(i) on the x-axis and sinh(i) on the y-axis, as shown in the following formula:

[0031]

[0032] Convert the N-1 heading angles in set H into vectors and add them together to obtain the composite vector. The calculation formula is as follows: mean heading angle h m and standard deviation h std They are defined as follows:

[0033]

[0034]

[0035]

[0036] The mean heading angle needs to be considered. The quadrant in which the heading angle is located will be converted to mean 0° to 360°.

[0037] 2. Heading oscillation frequency

[0038] To reflect the local characteristics of heading changes, a heading oscillation frequency is introduced. First, the difference in adjacent heading angles between the target set S is calculated, denoted as d. h (i), the formula is as follows:

[0039] d h (i)=h(i)-h(i-1),i≥3 (8)

[0040] And determine the angle difference d h (i) and the set threshold γ h The difference in sign between the heading angles is calculated by taking the values ​​between them, and is defined as follows:

[0041]

[0042] Based on the set O of heading angle sign values, two heading oscillation modes are defined:

[0043] O(i-1)+O(i)=0 and O(i-1)≠O(i) (10)

[0044] O(i-1)+O(i+1)=0 and O(i-1)≠O(i+1),O(i)=0 (11)

[0045] When formula (10) or (11) above is satisfied, the system is determined to be in an oscillation mode and the number of oscillations is counted, denoted as U. h ,

[0046]

[0047] Where γ h The threshold for determining the heading angle difference can be set based on the actual radar angle measurement error. After obtaining the number of heading oscillations, the heading oscillation frequency is defined as...

[0048]

[0049] 3. Average speed

[0050] The formula for calculating the average velocity of the target set S is as follows:

[0051]

[0052] 4. Speed ​​Standard Deviation

[0053] The formula for calculating the standard deviation of the velocity of the set of targets on the track is as follows:

[0054]

[0055] 5. Velocity oscillation frequency

[0056] To reflect the local characteristics of velocity changes, a velocity oscillation frequency is introduced. At the current time i, the velocity values ​​of historical sample points over a length L are taken, and the mean velocity v over that length is calculated. mL The definition is as follows:

[0057]

[0058] Calculate the current velocity v(i) and the historical average velocity v mL The difference between them is denoted by τ. v (i) is defined as follows:

[0059] τ v (i)=v(i)-v mL ,i≥L+1 (17)

[0060] Set the speed change threshold γ v When the difference is greater than γ v Then this point is defined as a velocity oscillation, denoted by U. v (i) is defined as follows:

[0061]

[0062] The threshold γ can be adjusted based on the actual radar tracking data rate and tracking accuracy. v And with length L, the velocity oscillation frequency is defined as

[0063]

[0064] Finally, the characteristic quantities for track calculation are determined to be the standard deviation of heading angle, the heading oscillation frequency, the mean velocity, the standard deviation of velocity, and the velocity oscillation frequency, denoted by G = [h]. std ,f h ,v m ,v std ,f v Using these five features may lead to misjudgment of the target type at the beginning of tracking. However, as the target is tracked stably for longer periods and more times, the time-related feature values ​​can be dynamically corrected online and the target type can be re-predicted, thus gradually improving the accuracy of the identification.

[0065] 1.3. Offline Training and Testing

[0066] 1. Model Selection

[0067] After obtaining the five features mentioned above, considering that the sample features are non-linear and to prevent overfitting, a soft-margin support vector machine algorithm is adopted, and its model formula is as follows:

[0068]

[0069] Where w is the hyperplane normal vector, C is a constant greater than 0, ξ is a slack variable, φ(x) is the mapping function, b is the hyperplane intercept, and x... i For the sample vector, the calculation in the above formula introduces the kernel function κ(x). i ,x j In this invention, a Gaussian kernel function is used, and the formula is as follows:

[0070]

[0071] In the formula, σ is the bandwidth of the Gaussian kernel function, and the SMO (Sequential Minimal Optimization) algorithm is subsequently used to solve the model.

[0072] 2. Model Training and Validation

[0073] After determining the support vector machine model, the hyperparameters of different sizes have a significant impact on the recognition accuracy. In the Gaussian kernel function, the hyperparameters that need to be adjusted are the penalty factor C and the kernel bandwidth σ. First, the track dataset is divided into training and test sets using random sampling. Then, the training dataset is trained using a 5-fold cross-training method to find the optimal hyperparameters. After obtaining the training model and the optimal hyperparameters, the model is validated and evaluated using the test set. The test set contains 114 drone tracks and 76 bird tracks, respectively. The confusion matrix is ​​shown in Table 1. The accuracy rate reached 87%, verifying the effectiveness of the proposed method.

[0074] Table 1

[0075]

[0076] 1.4. Online Experiment Results

[0077] The experiment used a phased array radar system as a platform, deployed in an environment with high bird activity. The drone used was a DJI Phantom 4. To ensure the completeness of the experiment, the drone simulated multiple missions involving different operations and manual flight scenarios. In one follow-up flight experiment, the radar target tracking and identification display (PPI) showed the following: Figure 4 , Figure 5As shown in the figure, the UAV target, batch number 223, has an azimuth angle of 356°. Its trajectory involves flying away from the radar direction for a period of time, then turning and changing direction upon reaching the designated heading point, and flying in the direction of entering the radar. The bird target, batch number 337, has an azimuth angle of 33° and its flight trajectory is relatively disordered. Under stable and continuous tracking, the radar correctly identified both the UAV and the bird target. The experimental results verify the lightweight nature and feasibility of this method from an engineering perspective.

Claims

1. A radar identification method for unmanned aerial vehicles (UAVs) and flying birds, characterized in that, include: Step S1: Analyze the data of UAVs and birds collected in the offline state. Based on the characteristics of UAV mission operations, describe strong maneuvering mode and weak maneuvering mode. Based on the differences in the characteristics of UAVs and birds in these two modes, propose a time-related heading oscillation frequency and velocity oscillation frequency feature description method. Combine the three feature quantities of mean velocity, standard deviation of velocity, and standard deviation of heading angle to improve the ability to distinguish UAVs from birds by multi-dimensional features. The standard deviation of the heading angle is calculated using a vectorized approach, transforming the i-th heading angle h(i) into a unit vector on a two-dimensional plane. The heading angle corresponds to cosh(i) on the x-axis and sinh(i) on the y-axis. Convert the N-1 heading angles in the heading angle set H into vectors and add them together to obtain the composite vector. Then the standard deviation of the heading angle h std Defined as This avoids the impact of measurement errors causing the heading angle to span 0° or 360°.

2. The radar identification method for UAVs and birds as described in claim 1, characterized in that, Step S1 includes: In step S11, in offline mode, based on five features—heading angle standard deviation, heading oscillation frequency, average speed, speed standard deviation, and speed oscillation frequency—a support vector machine algorithm is used. Historical data is divided into training and testing sets. The optimal hyperparameters and model are obtained by cross-training on the training set, and the model is validated using the testing set. In online mode, the five relevant features of the stably tracked target are calculated in real time using the optimal model and hyperparameters. As the number of tracking times increases, the feature values ​​are repeatedly corrected to gradually improve the recognition accuracy.

3. The radar identification method for UAVs and birds as described in claim 1, characterized in that, Step S11 includes: The difference in heading angles between adjacent heading angles of the target set S is calculated and denoted by the symbol d. h (i), the formula is as follows: d h (i)=h(i)-h(i-1), i≥3 (8) And determine the angle difference d h (i) and the set threshold γ h The difference in sign between the heading angles is calculated by taking the values ​​between them, and is defined as follows: Based on the set O of heading angle sign values, two heading oscillation modes are defined: O(i-1)+O(i)=0 and O(i-1)≠O(i) (10) O(i-1)+O(i+1)=0 and O(i-1)≠O(i+1),O(i)=0 (11) When formula (10) or (11) above is satisfied, the system is determined to be in an oscillation mode and the number of oscillations is counted, denoted as U. h , Where, γ h The threshold for determining heading angle difference is set based on the actual radar angle measurement error. After obtaining the number of heading oscillations, the heading oscillation frequency is defined as...

4. The radar identification method for UAVs and birds as described in claim 3, characterized in that, Step S11 includes: The formula for calculating the average velocity of the target set S is as follows:

5. The radar identification method for UAVs and birds as described in claim 4, characterized in that, Step S11 includes: The formula for calculating the standard deviation of the velocity of the set of targets on the track is as follows:

6. The radar identification method for UAVs and birds as described in claim 5, characterized in that, Step S11 includes: At the current time i, take the velocity values ​​of sample points along a historical length L, and calculate the average velocity v over that length. mL The definition is as follows: Calculate the current velocity v(i) and the historical average velocity v mL The difference between them is denoted by τ. v (i) is defined as follows: τ v (i)=v(i)-v mL ,i≥L+1(17) Set the speed change threshold γ v When the difference is greater than γ v Then this point is defined as a velocity oscillation, denoted by U. v (i) is defined as follows: Among them, the threshold γ is adjusted according to the actual radar tracking data rate and tracking accuracy. v And with length L, the velocity oscillation frequency is defined as

Images