Unmanned aerial vehicle identification method fusing course, micro-motion spectrum and transient signal characteristics

Abstract

The invention provides an unmanned aerial vehicle identification method fusing heading, micro-motion spectrum and transient signal features. The method comprises the following steps: extracting unmanned aerial vehicle target heading features; unmanned aerial vehicle target micro-motion spectrum feature extraction; extracting transient features of a target communication signal of the unmanned aerial vehicle; constructing a fusion feature vector; and classifying and identifying the target. According to the method, the heading features, the micro-motion spectrum features and the transient signal features of the light and small unmanned aerial vehicle are extracted, the unmanned aerial vehicle target is classified and identified by using the random forest classification method through feature normalization fusion, and the robustness and accuracy of unmanned aerial vehicle identification are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and relates to radar heading features, micro-motion spectrum feature extraction, communication transient signal feature extraction, feature normalization fusion and classification recognition, and specifically relates to a drone recognition method that integrates heading, micro-motion spectrum and transient signal features. Background Art

[0002] With the rapid development of composite materials, intelligent flight control, super-powered batteries, network communications and other technologies, drones have shown the advantages of miniaturization, low cost and intelligent operation, and are widely used in agriculture, aerial photography, mapping, reconnaissance, strike and other fields. While light and small drones provide new vitality for the rapid development of the low-altitude economy, they also bring severe challenges to the safety of low-altitude airspace. It is urgent to strengthen the development of low-altitude target detection and identification technology to achieve effective control of light and small drones.

[0003] The detection and identification of light and small UAVs mainly include low-altitude radar detection, radio detection, photoelectric detection and acoustic wave monitoring. Low-altitude radar can obtain the target's track, micro-Doppler, radar cross-sectional area and other characteristics, but the target recognition ability is weak when the UAV is hovering or flying slowly, especially when it is blocked by high-rise buildings in the city, low-altitude clutter, bird interference and other environmental influences; electronic reconnaissance equipment can obtain the characteristics of the target wireless remote control signal, satellite navigation signal and image transmission signal, but it is necessary to establish an electromagnetic signal feature database for UAV communication and navigation, and identify it based on the learned information, which puts higher requirements on information collection and cannot detect program-controlled UAVs; photoelectric detection equipment can obtain the image information of the target by infrared imaging or optical imaging means, and it is necessary to extract the image feature information of the UAV target from the image and compare and analyze it with the established UAV image database to achieve target recognition; acoustic wave monitoring equipment uses the voiceprint feature of the target for detection and recognition, but it is easily interfered by the external environment. With the continuous expansion of drone application scenarios, it is difficult to accurately identify light and small drones using a single sensor or detection method in complex electromagnetic, shielding, interference, meteorological and other environments. The fusion of detection signal characteristics between different sensors can enhance the robustness and reliability of drone target recognition.

[0004] At present, some methods have been used to extract high-order cumulative quantities and instantaneous feature statistics of cluster communication signals and integrate radar track features to construct a UAV cluster feature matrix to realize the identification of UAV cluster types. However, there are still some challenges in the close-range identification of UAVs of different configurations such as fixed-wing, rotor-wing, and compound-wing, as well as non-cluster UAV targets. Summary of the invention

[0005] In view of the deficiencies of the above-mentioned existing technologies, the present invention proposes a method for identifying unmanned aerial vehicles (UAVs) that integrates heading, micro-motion spectrum, and transient signal features, including the following steps:

[0006] Step 1: Extract the heading features of the UAV target;

[0007] Use a low-altitude radar to detect the UAV target and obtain the motion track of the UAV target. Each track point is represented by information such as the space, speed, and signal-to-noise ratio of the target at a certain moment. The i-th track point is represented as f(i) = [r(i), α(i), β(i), v(i), SNR(i)] T , where r(i), α(i), and β(i) respectively represent the slant range, elevation angle, and azimuth angle of the target in three-dimensional space, v(i) represents the speed of the target, and SNR(i) represents the signal-to-noise ratio of the target; for a UAV target with N consecutive track points, its motion track is represented as F = [f(1), f(2),..., f(N)];

[0008] 1. Standard deviation of heading angle deflection

[0009] According to the radar's tracking situation of the target, extract a continuous track point F with j points at a certain moment t from the motion track F j = [f(t - j + 1), f(t - j + 2),..., f(t)], where j < N, and the heading angles of any two track points are represented as

[0010]

[0011] Convert the heading angle θ i between two track points into a unit vector p i in the two-dimensional plane, which is represented as

[0012] p i = (cosθ i , sinθ i )

[0013] Using the method of vector synthesis, the heading angle vector of part of the continuous track point F j is represented as

[0014]

[0015] At this time, the calculation of the heading angle of part of the track needs to be converted into a relationship between 0° and 360° according to the quadrant area where the vector p i is located; therefore, the mean value of the heading angles of part of the continuous track point F j is represented as

[0016] According to the definition of the standard deviation, the standard deviation of the heading angle deflection is obtained and represented as

[0017]

[0018] Among them, Setting δ p is the deviation threshold of a heading angle from the mean value of some heading angles. hour, when hour, At this time <j;

[0019] 2. Flight speed standard deviation

[0020] For some continuous track points F j The mean velocity is expressed as The standard deviation of its flight speed is expressed as

[0021]

[0022] 3. Flight altitude standard deviation

[0023] For some continuous track points F j The mean value of its flight height is expressed as The standard deviation of the flight altitude is expressed as

[0024]

[0025] Where h(i) represents the flight altitude of the i-th track point;

[0026] Step 2: Extraction of UAV target micro-motion spectrum features;

[0027] Assume that the echo signal received by the low-altitude radar is represented by x = [x1, x2, ..., x M ], x i represents the target echo signal at the i-th moment, i is 1-M, and the echo signal x is processed by sliding window using a rectangular window function with a length of L. M represents the total number of echo signals, and M-L+1 signals with a length of L are obtained, which are represented as the echo signal s i =[x i ,x i+1 ,…,x i+L-1 ], i = 1, 2, ..., L-M + 1; perform empirical mode decomposition (EMD) on the received signal with a length of L time, and obtain

[0028]

[0029] Among them, m k represents the kth intrinsic mode function IMF, s Δis the remainder after q times decomposition; for the echo signal s i For the target micro-motion time domain signal expression,

[0030]

[0031] The target micro-motion time domain signal s iw Perform FFT processing to obtain the target micro-motion frequency domain signal S iw , obtain the target's micro-motion spectrum characteristics;

[0032] 1. Standard deviation of micro-motion spectrum

[0033] The target micro-motion frequency domain signal S iw The mean value of Then the standard deviation of the target micro-motion spectrum is expressed as

[0034]

[0035] 2. Micro-motion spectrum energy entropy

[0036] The spectrum energy entropy of the target micro-motion signal is expressed as

[0037]

[0038] Among them, the energy mean E k Expressed as

[0039]

[0040] At this time, M k Expressed as the eigenmode function m k Frequency domain expression of ;

[0041] Step 3: Extract transient features of UAV target communication signals;

[0042] Use software radio peripherals to capture radio frequency signals in the monitoring area, pre-process them, and extract the target's wireless signal y(t);

[0043] 1. High-order cumulants of transient signals

[0044] For a stationary random process y(t) with a mean of zero, extract some of its second-order and fourth-order cumulants;

[0045] The second-order cumulant of the target signal is expressed as:

[0046] C 20 = cum(y(t),y(t)) = M 20 =E[y(t)y(t)]

[0047] C 21 = cum(y(t),y* (t)) = M 21 =E[y(t)y * (t)]

[0048] Among them, the second-order central cumulant C 20 The second-order central moment M 20 Indicates that M 20 It is expressed as the mean of the product of the wireless signals y(t) and y(t), the second-order mixed cumulant C 21 The second-order mixing moment M 21 Indicates that M 21 Expressed as wireless signal y(t) and y * (t) is the mean of the product, cum(·) means the cumulative amount of (·), that is, the mean of the function of (·); y * (t) represents the conjugate of y(t);

[0049] The fourth-order cumulant is expressed as:

[0050]

[0051] C 41 =cum(y(t),y(t),y(t),y * (t)) = M 41 -3M 20 M 21

[0052]

[0053] Among them, M pq =E[y(t) p-q ,y * (t) q ], E[·] represents the mathematical expectation, the exponential parameters p and q are positive integers, representing the power; the fourth-order central cumulant C 40 , the first fourth-order mixed cumulant C 41 , the second fourth-order mixed cumulant C 42 is the fourth-order cumulant of the wireless signal y(t), and the fourth-order central moment M 40 , second-order central moment M 20 , the first fourth-order mixing moment M 41 , the second fourth-order mixing moment M 42 , second-order mixing moment M 21 Perform calculations;

[0054] 2. Signal instantaneous amplitude kurtosis

[0055] The signal sequence y is processed by sliding window using a rectangular window function with a length of L, and the segmented signal amplitude is A. p(i) = [A(i), A(i+1), ..., A(i+L-1)], A(i) represents the signal amplitude at the i-th moment, where i = 1, 2, ..., N-L+1. The instantaneous amplitude kurtosis of the segmented signal is expressed as

[0056]

[0057] Step 4: Construct fusion feature vector;

[0058] Normalization is performed at the output of the three feature vectors. The normalization formula is expressed as:

[0059]

[0060] Among them, F'1, F'2, F'3 represent the normalization of the heading feature vector F1, the micro-motion spectrum feature vector F2, and the communication signal transient feature vector F3, respectively. var(·) represents the variance of (·), and "'" represents the feature vector after normalization. The heading feature vector F1 = [p std ,v std ,h std ], micro-motion spectrum feature vector Communication signal transient characteristic vector F3 = [C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,σ],p std ,v std ,h std They represent the target’s heading angle deflection standard deviation, flight speed standard deviation, and flight altitude standard deviation, respectively. γ represents the standard deviation of the target's micro-motion spectrum and the spectrum energy entropy;

[0061] Through normalization processing, the value distribution of the heading feature vector F1, the micro-motion spectrum feature vector F2, and the communication signal transient feature vector F3 is limited to a standard normal distribution with a mean of 0 and a variance of 1, and the fusion feature vector of the target to be identified and classified is constructed as follows:

[0062] F=[F′1,F′2,F′3]

[0063] Step 5: Target classification and identification;

[0064] The random forest classification algorithm is used to classify and identify drone targets. The steps are as follows:

[0065] 1) Bootstrap sampling with replacement;

[0066] In the N1 target training data sets, a random Bootstrap sampling method with replacement is used to extract n sample points as the generation samples of the decision tree;

[0067] 2) Build a decision tree;

[0068] For each Bootstrap data sample, a decision tree is constructed using a random feature subset; m features are randomly extracted from the fused feature vector F, the features are extracted by traversing the training set, and the Gini coefficient of feature partitioning is calculated, which is expressed as where p i It represents the probability of the i-th feature appearing in the training set, and takes the feature with the largest Gini coefficient as the split point of the training set; splits are performed continuously to allow the decision tree to grow completely, and the minimum number of samples included in the node after and during the split is used as the condition for stopping growth;

[0069] 3) Target recognition and classification;

[0070] Repeat steps 1 and 2 K times. The category corresponding to the largest number of samples in the leaf node is the classification result of the decision tree. According to the voting principle, the number of occurrences of each category in all decision trees is counted. The final classification result of the random forest is the category with the largest number of occurrences.

[0071] In step three of an embodiment of the present invention, preprocessing is performed using an adaptive triangular threshold method or a wavelet transform signal processing method.

[0072] The present invention improves the robustness and accuracy of UAV identification by extracting the heading characteristics, micro-motion spectrum characteristics, and transient signal characteristics of light and small UAVs, normalizing and fusing the characteristics, and using the random forest classification method to classify and identify UAV targets.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] (1) Taking into account the differences in the track characteristics of UAVs, flying birds, and raindrop targets, the differences in the micro-motion spectrum characteristics of different rotorcraft UAVs and flying birds, and the differences in the communication modulation methods of different UAVs, representative features are selected, and the target echo signal and radio frequency signal are processed using a rectangular window function with a length of L to slide the micro-motion spectrum and communication signal transient characteristics. The correlation between the heading, micro-motion spectrum, and communication signal transient key features is fully combined to perform feature normalization and fusion, which effectively improves the accuracy of UAV target recognition.

[0075] (2) The random forest classification algorithm is used for feature extraction. The algorithm design is relatively simple, easy to implement in engineering, and has strong practicality.

[0076] (3) Before fusing the features, normalization processing is performed, which greatly reduces the computational complexity of feature recognition. Description of the Drawings

[0077] Figure 1 Show the logic block diagram of UAV target feature recognition and classification. Detailed Implementation Manner

[0078] The present invention will be further described below with reference to the accompanying drawings.

[0079] The present invention provides a UAV recognition method that fuses heading, micro-motion spectrum, and transient signal features, including the following steps:

[0080] Step 1: Extract the heading feature of the UAV target;

[0081] Use a low-altitude radar to detect the UAV target, and the motion track of the UAV target can be obtained. Each track point represents information such as the space, speed, and signal-to-noise ratio of the target at a certain moment. The i-th track point can be expressed by the formula f(i) = [r(i), α(i), β(i), v(i), SNR(i)], T where r(i), α(i), and β(i) respectively represent the slant range, elevation angle, and azimuth angle of the target in three-dimensional space, v(i) represents the speed of the target, and SNR(i) represents the signal-to-noise ratio of the target. Therefore, for a UAV target with N consecutive track points, its motion track can be expressed as F = [f(1), f(2),..., f(N)]. By analyzing and processing the target track, different distributions of the target in various information such as heading change and motion fluctuation can be obtained, and then the recognition and classification of the target can be realized.

[0082] Combined with the uncertainty of the UAV in different motion states such as straight line, turning back, hovering, and turning in the air, and associating the direction characteristics and fluctuation characteristics of the target flight track, extract the motion characteristics of the UAV that are different from different substances such as the environment, birds, and raindrops.

[0083] 1. Standard deviation of heading angle deflection

[0084] The heading is the direction vector of adjacent track sampling points, and the standard deviation of heading angle deflection can reflect the fluctuation of the direction deflection of a certain continuous track of the target flight. According to the radar's tracking of the target, at a certain moment t, a continuous track point F with j points is extracted from the motion track F j = [f(t - j + 1), f(t - j + 2),..., f(t)], where j < N, and the heading angles of any two track points are expressed as

[0085]

[0086] In order to normalize the subsequent data, the heading angle θ between the two track points is i Transformed into a unit vector p on a two-dimensional plane i , expressed as

[0087] p i =(cosθ i ,sinθ i )

[0088] Then, using the vector synthesis method, some continuous track points F j The heading angle vector can be expressed as

[0089]

[0090] At this time, the calculation of the heading angle of the partial track needs to be based on the vector p i The quadrant area is converted to the relationship between 0° and 360°. Therefore, some continuous track points F j The mean heading angle is expressed as

[0091] According to the definition of standard deviation, the standard deviation of heading angle deflection can be expressed as

[0092]

[0093] Among them, Setting δ p is the deviation threshold of a heading angle from the mean value of some heading angles. hour, when hour, At this time <j。

[0094] 2. Flight speed standard deviation

[0095] When the UAV target makes a maneuver, the speed values ​​of the track points in different directions are different. The standard deviation of the flight speed can be used to reflect the identification information of different movement patterns of the target. j For , its mean velocity can be expressed as The standard deviation of its flight speed is expressed as

[0096]

[0097] 3. Flight altitude standard deviation

[0098] The standard deviation of flight altitude is mainly used to indicate the change of target flight trajectory, which can reflect the stability or variability of target flight. j For , the mean value of its flight height can be expressed as The standard deviation of the flight altitude is expressed as

[0099]

[0100] Where h(i) represents the flight altitude of the i-th track point.

[0101] Step 2: Extraction of UAV target micro-motion spectrum features;

[0102] The rotation of the drone's rotor and the flapping of the bird's wings will have a modulation effect on the radar echo, causing the Doppler spectrum of the target echo to be broadened accordingly, and have different characteristics such as time-varying and periodicity. Suppose the echo signal of the target received by the low-altitude radar is represented by x = [x1, x2, ..., x M ], x i represents the target echo signal at the i-th moment, i is 1-M, and the echo signal x is processed by sliding window using a rectangular window function with a length of L. M represents the total number of echo signals, and M-L+1 signals with a length of L are obtained, which are represented as the echo signal s i =[x i ,x i+1 ,…,x i+L-1 ], i = 1, 2, ..., L-M + 1. For the received signal with a length of L, empirical mode decomposition (EMD) is performed, and we can get

[0103]

[0104] Among them, m k represents the kth intrinsic mode function (IMF), s Δ is the remainder after q times decomposition. For the echo signal s i For example, the received signal s can be transformed into i It is decomposed into a series of sub-signals with different frequencies. After simulation analysis, the spectrum distribution of its first intrinsic mode function m1 is similar to that of the echo signal s i The spectrum of the sum of the eigenmode functions except the first eigenmode function component can represent the modulation component of the target micromotion. The target micromotion time domain signal expression is:

[0105]

[0106] The target micro-motion time domain signal s iw Perform FFT processing to obtain the target micro-motion frequency domain signal S iw , the micro-motion spectrum characteristics of the target can be obtained.

[0107] 1. Standard deviation of micro-motion spectrum

[0108] The frequency of the rotor blades of the UAV and the flapping wings of the bird can periodically modulate the frequency of the echo signal. The higher the periodic modulation frequency, the greater the discreteness of the micro-motion spectrum. The standard deviation of the micro-motion spectrum can be used to characterize the size of the periodic modulation frequency, and then classify the target. iw The mean value of Then the standard deviation of the target micro-motion spectrum is expressed as

[0109]

[0110] 2. Micro-motion spectrum energy entropy

[0111] The spectrum energy entropy indicates the complexity of the signal energy distribution in the spectrum. When the frequency composition of the target echo signal is simple, the power spectrum is concentrated on some frequency points, and the corresponding spectrum lines are relatively small. In this case, the spectrum energy entropy value is relatively small. In other words, the larger the energy entropy of the target micro-motion signal spectrum, the more obvious the micro-motion characteristics in the target echo signal, and the more micro-motion spectrum lines. The spectrum energy entropy of the target micro-motion signal is expressed as

[0112]

[0113] Among them, the energy mean E k Expressed as

[0114]

[0115] At this time, M k Expressed as the eigenmode function m k The frequency domain expression of .

[0116] Step 3: Extract transient features of UAV target communication signals;

[0117] Use software radio peripherals to capture radio frequency signals in the monitoring area. Generally speaking, the radio frequency signals of drones include remote control signals that use frequency hopping and image transmission signals that use orthogonal frequency division multiplexing modulation. Considering that the acquired signals contain both drone signals and invalid noise signals, the adaptive triangular threshold method or wavelet transform signal processing method is used for preprocessing to extract the target's wireless signal y(t). The transient part of the captured radio communication signal can reflect the subtle characteristics of the drone signal transmitting equipment. The transient signal will be different if the structural characteristics of the transmitter are different. The transient signal reflects the characteristic information of the signal power jumping from zero to rated power, which can be used for target detection and identification of signal changes such as multi-channel switching and FM communication.

[0118] 1. High-order cumulants of transient signals

[0119] For the target's wireless signal y(t), its high-order cumulants will not be affected by noise. For the stationary random process y(t) with a mean of zero, extracting some of its second-order and fourth-order cumulants can be used to identify the UAV target's transmitting signal.

[0120] The second-order cumulant of the target signal is expressed as:

[0121] C 20 = cum(y(t),y(t)) = M 20 =E[y(t)y(t)]

[0122] C 21 = cum(y(t),y * (t)) = M 21 =E[y(t)y * (T)]

[0123] Among them, the second-order central cumulant C 20 The second-order central moment M 20 Indicates that M 20 It is expressed as the mean of the product of the wireless signals y(t) and y(t), the second-order mixed cumulant C 21 The second-order mixing moment M 21 Indicates that M 21 Expressed as wireless signal y(t) and y * (t) is the mean of the product, cum(·) means the cumulative amount of (·), that is, the mean of the function of (·); y * (t) represents the conjugate of y(t);

[0124] The fourth-order cumulant is expressed as:

[0125]

[0126] C 41 =cum(y(t),y(t),y(t),y * (t)) = M 41 -3M 20 M 21

[0127]

[0128] Among them, M pq =E[y(t) p-q ,y * (t) q ], E[·] represents the mathematical expectation, and the exponential parameters p and q are positive integers, representing the power. The fourth-order central cumulant C 40 , the first fourth-order mixed cumulant C 41 , the second fourth-order mixed cumulant C42 is the fourth-order cumulant of the wireless signal y(t), which can be expressed by the fourth-order central moment M 40 , second-order central moment M 20 , the first fourth-order mixing moment M 41 , the second fourth-order mixing moment M 42 , second-order mixing moment M 21 Perform calculations.

[0129] 2. Signal instantaneous amplitude kurtosis

[0130] The instantaneous amplitude kurtosis of the signal can reflect the steepness of the instantaneous amplitude distribution of the signal compared with the normal distribution, and can be used to identify the signal modulation mode. Using a rectangular window function with a length of L to perform sliding window processing on the signal sequence y, the segmented signal amplitude can be A p (i) = [A(i), A(i+1), ..., A(i+L-1)], A(i) represents the signal amplitude at the i-th moment, where i = 1, 2, ..., N-L+1. The instantaneous amplitude kurtosis of the segmented signal is expressed as

[0131]

[0132] Step 4: Construct fusion feature vector;

[0133] In order to avoid the large difference in the numerical values ​​of the feature vectors output by the three branches of the target's radar heading feature, micro-motion spectrum feature, and communication signal transient feature, which makes it difficult for the later target recognition algorithm to converge, normalization processing is performed at the output end of the three feature vectors. The formula for normalization processing is expressed as follows:

[0134]

[0135] Among them, F'1, F'2, F'3 represent the normalization of the heading feature vector F1, the micro-motion spectrum feature vector F2, and the communication signal transient feature vector F3, respectively. var(·) represents the variance of (·), and "'" represents that the feature vector has been normalized. Heading feature vector F1 = [p std ,v std ,h std ], micro-motion spectrum feature vector Communication signal transient characteristic vector F3 = [C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,σ],p std ,v std ,h std They represent the target’s heading angle deflection standard deviation, flight speed standard deviation, and flight altitude standard deviation, respectively. γ represents the target's micro-motion spectrum standard deviation and spectrum energy entropy respectively.

[0136] Through normalization processing, the value distribution of the heading feature vector F1, the micro-motion spectrum feature vector F2, and the communication signal transient feature vector F3 can be limited to a standard normal distribution with a mean of 0 and a variance of 1. At this time, the fusion feature vector of the target to be identified and classified is constructed as follows:

[0137] F=[F′1,F′2,F′3]

[0138] Step 5: Target classification and identification;

[0139] The random forest classification algorithm is used to classify and identify drone targets. The random forest classification algorithm mainly uses multiple decision trees as the base classifier for target classification. By combining and constructing the total classifier, it realizes strong learning classification of the target and does not require any prior information. It is suitable for the classification and identification of low-altitude drone targets. The steps taken are:

[0140] 1. Bootstrap sampling with replacement;

[0141] In the N1 target training data sets, a random sampling method with replacement (Bootstrap) is used to extract n sample points as the generation samples of the decision tree.

[0142] 2. Build a decision tree;

[0143] For each Bootstrap data sample, a decision tree is constructed using a random feature subset. m features are randomly extracted from the fused feature vector F, the features are extracted by traversing the training set, and the Gini coefficient of the feature partition is calculated, which is expressed as where p i It represents the probability of the ith feature appearing in the training set, and the feature with the largest Gini coefficient is used as the split point of the training set. The split is continued to allow the decision tree to grow completely, and the minimum number of samples included in the node after and during the split is used as the condition for stopping growth.

[0144] 3. Target recognition and classification;

[0145] Repeat steps 1 and 2 K times. The category corresponding to the largest number of samples in the leaf node is the classification result of the decision tree. According to the voting principle, the number of occurrences of each category in all decision trees is counted. The final classification result of the random forest is the category with the largest number of occurrences. Specific embodiments

[0147] The experiment selected three types of drones, including a helicopter drone, a 4-rotor drone, and a 6-rotor drone, as experimental objects. In terms of wireless signal acquisition, an oscilloscope with a bandwidth of 6GHz and a maximum sampling frequency of 20Gsa / s was used with 2.4GHz directional and omnidirectional antennas for signal reception. The receiving gains of the directional and omnidirectional antennas were 17dBi and 10dBi, respectively. In terms of radar signal acquisition, a K-band radar was used for target detection, with a carrier frequency of 24GHz, a baseband sampling frequency of 128kHz, a transmission power of 16dBm, and an antenna gain of 15dBi. The target flight includes movement modes such as take-off, landing, hovering, fixed-altitude level flight, climbing, descending, and turning. According to the different modulation methods of drone communication signals, the target signal feature samples were constructed. The number of each feature sample is 1000, and a total of 3000 samples are included. All data sets are divided into training sets and test sets in a ratio of 7:3.

[0148] Target recognition is performed using a drone identification method that integrates heading, micro-motion spectrum, and transient signal features. When only heading features are used for target recognition, the target recognition rate is 96.33%; when only micro-motion spectrum features are used for target recognition, the target recognition rate is 95.21%; when only communication signal transient features are used for target recognition, the target recognition rate is 97.62%; when the heading, micro-motion spectrum, and transient signal features are combined for target recognition, the target recognition accuracy is 99.35%.

Claims

1. A method for identifying unmanned aerial vehicles by integrating heading, micro-motion spectrum, and transient signal features, characterized in that: The following steps are involved: Step 1: Extract the heading features of the UAV target; The UAV target is detected by low-altitude radar to obtain the movement track of the UAV target. Each track point represents the space, speed, signal-to-noise ratio and other information of the target at a certain moment. The i-th track point is expressed as f(i) = [r(i), α(i), β(i), v(i), SNR(i)] T , where r(i), α(i), β(i) represent the slant range, pitch angle, and azimuth of the target in three-dimensional space, v(i) represents the speed of the target, and SNR(i) represents the signal-to-noise ratio of the target. For a UAV target with N consecutive track points, its motion track is expressed as F = [f(1), f(2), ..., f(N)]; 1. Standard deviation of heading angle deflection According to the radar's tracking of the target, a certain moment t is extracted from the motion track F, and the continuous track points F with j points are obtained j = [f(t - j + 1), f(t - j + 2),..., f(t)], where j < N, and the course angles of any two track points are expressed as The heading angle θ between the two track points i Transformed into a unit vector p on a two-dimensional plane i , expressed as p i =(cosθ i ,sinθ i ) Using vector synthesis, some continuous track points F j The heading angle vector is expressed as At this time, the calculation of the heading angle of the partial track needs to be based on the vector p i The quadrant area is converted to the relationship between 0o and 360o; therefore, some continuous track points F j The mean heading angle is expressed as According to the definition of standard deviation, the standard deviation of heading angle deflection is expressed as Among them, Setting δ p is the deviation threshold of a heading angle from the mean value of some heading angles. hour, when hour, At this time <j; 2. Flight speed standard deviation For some continuous track points F j The mean velocity is expressed as The standard deviation of its flight speed is expressed as 3. Flight altitude standard deviation For some continuous track points F j The mean value of its flight height is expressed as The standard deviation of the flight altitude is expressed as Where h(i) represents the flight altitude of the i-th track point; Step 2: Extraction of UAV target micro-motion spectrum features; Assume that the echo signal received by the low-altitude radar is represented by x = [x1, x2, ..., x M ], x i represents the target echo signal at the i-th moment, i is 1-M, and the echo signal x is processed by sliding window using a rectangular window function with a length of L. M represents the total number of echo signals, and M-L+1 signals with a length of L are obtained, which are represented as the echo signal s i =[x i ,x i+1 ,…,x i+L-1 ], i = 1, 2, ..., L-M + 1; perform empirical mode decomposition (EMD) on the received signal with a length of L time, and obtain Among them, m k represents the kth intrinsic mode function IMF, s Δ is the remainder after q times decomposition; for the echo signal s i For the target micro-motion time domain signal expression, The target micro-motion time domain signal s iw Perform FFT processing to obtain the target micro-motion frequency domain signal S iw , obtain the target's micro-motion spectrum characteristics; 1. Standard deviation of micro-motion spectrum The target micro-motion frequency domain signal S iw The mean value of Then the standard deviation of the target micro-motion spectrum is expressed as 2. Micro-motion spectrum energy entropy The spectrum energy entropy of the target micro-motion signal is expressed as Among them, the energy mean E k Expressed as At this time, M k Expressed as the eigenmode function m k Frequency domain expression of ; Step 3: Extract transient features of UAV target communication signals; Use software radio peripherals to capture radio frequency signals in the monitoring area, pre-process them, and extract the target's wireless signal y(t); 1. High-order cumulants of transient signals For a stationary random process y(t) with a mean of zero, extract some of its second-order and fourth-order cumulants; The second-order cumulant of the target signal is expressed as: C 20 =cum(y(t),y(t))=M 20 =E[y(t)y(t)] C 21 =cum(y(t),y * (t))=M 21 =E[y(t)y * (t)] Among them, the second-order central cumulant C 20 The second-order central moment M 20 Indicates that M 20 It is expressed as the mean of the product of the wireless signals y(t) and y(t), the second-order mixed cumulant C 21 The second-order mixing moment M 21 Indicates that M 21 Expressed as wireless signal y(t) and y * (t) is the mean of the product, cum(·) means the cumulative amount of (·), that is, the mean of the function of (·); y * (t) represents the conjugate of y(t); The fourth-order cumulant is expressed as: Among them, M pq =E[y(t) p-q ,y * (t) q ], E[·] represents the mathematical expectation, the exponential parameters p and q are positive integers, representing the power; the fourth-order central cumulant C 40 , the first fourth-order mixed cumulant C 41 , the second fourth-order mixed cumulant C 42 is the fourth-order cumulant of the wireless signal y(t), and the fourth-order central moment M 40 , second-order central moment M 20 , the first fourth-order mixing moment M 41 , the second fourth-order mixing moment M 42 , second-order mixing moment M 21 Perform calculations; 2. Signal instantaneous amplitude kurtosis The signal sequence y is processed by sliding window using a rectangular window function with a length of L, and the segmented signal amplitude is A. p (i) = [A(i), A(i+1), ..., A(i+L-1)], A(i) represents the signal amplitude at the i-th moment, where i = 1, 2, ..., N-L+1. The instantaneous amplitude kurtosis of the segmented signal is expressed as Step 4: Construct fusion feature vector; Normalization is performed at the output of the three feature vectors. The normalization formula is expressed as: Among them, F'1, F'2, F'3 represent the normalization of the heading feature vector F1, the micro-motion spectrum feature vector F2, and the communication signal transient feature vector F3, respectively. var(·) represents the variance of (·), and "'" represents the feature vector after normalization. The heading feature vector F1 = [p std ,v std ,h std ], micro-motion spectrum feature vector Communication signal transient characteristic vector F3 = [C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,σ],p std ,v std ,h std They represent the target’s heading angle deflection standard deviation, flight speed standard deviation, and flight altitude standard deviation, respectively. γ represents the standard deviation of the target's micro-motion spectrum and the spectrum energy entropy; Through normalization processing, the value distribution of the heading feature vector F1, the micro-motion spectrum feature vector F2, and the communication signal transient feature vector F3 is limited to a standard normal distribution with a mean of 0 and a variance of 1, and the fusion feature vector of the target to be identified and classified is constructed as follows: F=[F′1,F′2,F′3] Step 5: Target classification and identification; The random forest classification algorithm is used to classify and identify drone targets. The steps are as follows: 1) Bootstrap sampling with replacement; In the N1 target training data sets, a random Bootstrap sampling method with replacement is used to extract n sample points as the generation samples of the decision tree; 2) Build a decision tree; For each Bootstrap data sample, a decision tree is constructed using a random feature subset; m features are randomly extracted from the fused feature vector F, the features are extracted from the training set, and the Gini coefficient of the feature partition is calculated, which is expressed as where p i It represents the probability of the i-th feature appearing in the training set, and the feature with the largest Gini coefficient is used as the split point of the training set; the split is performed continuously to allow the decision tree to grow completely, and the minimum number of samples included in the node after and during the split is used as the condition for stopping growth; 3) Target recognition and classification; Repeat steps 1 and 2 K times. The category corresponding to the largest number of samples in the leaf node is the classification result of the decision tree. According to the voting principle, the number of occurrences of each category in all decision trees is counted. The final classification result of the random forest is the category with the largest number of occurrences.

2. The method for identifying unmanned aerial vehicles by integrating heading, micro-motion spectrum, and transient signal characteristics as claimed in claim 1, characterized in that: In step three, preprocessing is performed using an adaptive triangular threshold method or a wavelet transform signal processing method.

Images