Unmanned aerial vehicle identification method and system based on random forest

Abstract

The invention relates to the technical field of signal processing, in particular to an unmanned aerial vehicle identification method and system based on a random forest, and the method comprises the steps: transmitting a radar signal, receiving an echo signal reflected by an unmanned aerial vehicle, carrying out the signal processing and target detection of the echo signal, and extracting a target trace point set; performing track association and filtering processing based on the target trace point set to generate a target track set; based on a target track state in a preset time window, extracting a multi-dimensional motion feature vector of the target track; and inputting the feature vector into a pre-trained random forest model, and outputting an unmanned aerial vehicle identification result. The objective of the invention is to realize low-cost and high-precision unmanned aerial vehicle identification.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and particularly to an unmanned aerial vehicle (UAV) identification method and system based on random forest. Background Art

[0002] With the rapid development of the low-altitude economy, the application scenarios of UAVs in the urban airspace have shown an explosive growth in fields such as logistics transportation, emergency inspection, and geographical mapping. However, the openness and electromagnetic complexity of the urban low-altitude environment pose severe challenges to UAV supervision technologies: within the radar monitoring area, in addition to UAV targets, there are also densely distributed dynamic interference objects such as bird flocks, kites, moving vehicles, and pedestrians, whose kinematic characteristics are highly coupled with those of low-altitude UAVs, resulting in extremely high misjudgment rates for traditional coarse-grained identification methods based on single motion parameters such as speed and altitude. In addition, the multipath scattering effect, electromagnetic interference, and atmospheric turbulence disturbance caused by urban building complexes further reduce the signal-to-noise ratio (SNR) of radar echoes, significantly decreasing the stability of target feature extraction and further weakening the robustness of the identification system.

[0003] Currently, mainstream UAV identification technologies generally rely on the micro-Doppler effect, that is, classifying by analyzing the radar echo frequency shift characteristics caused by the high-speed rotation of UAV rotors or propellers. This method relies on the radar to capture the subtle vibration signals of the rotors under high sampling rate and high SNR conditions, and combines time-frequency analysis algorithms or deep learning models for feature interpretation. However, in civilian low-cost radar systems, limited by hardware sampling rate, limited computing resources, and power consumption constraints, the micro-Doppler signal resolution of rotor vibration signals is insufficient, resulting in attenuation of feature SNR. At the same time, the coupled interference of complex environmental noise and multipath effects causes spectral aliasing and modal confusion in feature extraction algorithms, ultimately leading to insufficient average identification accuracy of existing technologies in actual scenarios and restricting the reliability of urban low-altitude security systems. Summary of the Invention

[0004] In order to achieve low-cost and high-precision UAV identification, the present invention provides a UAV identification method and system based on random forest, and the specific technical solutions adopted are as follows: The technical solution of the first aspect of the present invention provides a UAV identification method based on random forest, and the method includes: Transmit a radar signal and receive the echo signal reflected by the UAV, perform signal processing and target detection on the echo signal, and extract a set of target traces; Perform track association and filtering processing based on the set of target traces to generate a set of target tracks; Extract the multi-dimensional motion feature vector of the target track based on the target track state within a preset time window; Input the feature vector into a pre-trained random forest model and output the UAV identification result.

[0005] Further, a radar signal is transmitted and the echo signal reflected by the UAV is received, and signal processing and target detection are performed on the echo signal to extract the target plot set, including: The echo signal and the transmitted signal are subjected to mixing and filtering processing to generate an intermediate frequency signal; Moving target detection is performed on the intermediate frequency signal to generate a spectrum containing target distance and speed information; Based on the background noise, the detection threshold is dynamically adjusted, and the target plot set is extracted from the spectrum.

[0006] Further, based on the target plot set, track association and filtering processing are performed to generate a target track set, including: An initial track is generated according to the time and space distribution of the target plots; Based on minimizing the distance cost between the observed value and the predicted state, the target plots in consecutive frames are associated with the track; Iterative processing of state prediction and measurement correction is performed on the associated track, and the position and speed parameters of the track are dynamically smoothed.

[0007] Further, iterative processing of state prediction and measurement correction is performed on the associated track, and the position and speed parameters of the track are dynamically smoothed, including: Based on the historical state of the track, the state parameters at the current moment are predicted; According to the difference between the actual measurement value and the predicted value, the weight is dynamically adjusted to update the track state parameters; Based on the updated track state, a track data set including the target unique identifier, three-dimensional polar coordinates and rectangular coordinates, speed components, and signal-to-noise ratio set is generated.

[0008] Further, based on the target track state within a preset time window, a multi-dimensional motion feature vector of the target track is extracted, including: Based on the preset time window, a continuous state sequence of the target track is intercepted; The statistics of the speed, acceleration, and angle change rate in the continuous state sequence are calculated; The mean value of the signal-to-noise ratio and the number of effective measurement points are extracted to generate a multi-dimensional motion feature vector.

[0009] Further, the statistics of the speed, acceleration, and angle change rate in the continuous state sequence are calculated, including: Calculate the mean value of the speed and the fluctuation range deviating from the mean value within the preset time window; Calculate the mean value of the acceleration and the fluctuation range deviating from the mean value; Calculate the mean value and the fluctuation range of the azimuth angle and the pitch angle change rate.

[0010] Further, input the feature vector into the pre-trained random forest model to output the UAV recognition result, including: Generate training subsets of multiple decision trees through random sampling based on the historical track feature dataset; For each node of each decision tree, randomly select a candidate feature subset and optimize the splitting rule based on the information purity index; Based on the classification prediction results of multiple decision trees, determine the UAV recognition result through the majority voting mechanism.

[0011] Further, for each node of each decision tree, randomly select a candidate feature subset and optimize the splitting rule based on the information purity index, including: For each candidate feature, calculate the information purity difference of the child nodes under different splitting thresholds; Select the feature and threshold that maximize the information purity difference of the child nodes as the splitting rule.

[0012] Further, the method further includes: According to the UAV recognition result, perform communication interference or navigation deception operations on the target UAV to block the communication connection between the target UAV and the control terminal or induce the target UAV to deviate from the preset flight path; Mark the target UAV that has completed the disposal, record the track data, disposal time, and effect parameters, and update the airspace situation information.

[0013] The technical solution of the second aspect of the present invention provides a UAV recognition system based on a random forest, which adopts the UAV recognition method based on a random forest described in the technical solution of the first aspect of the present invention. The system includes: A signal processing module configured to transmit a radar signal and receive the echo signal reflected by the UAV, perform signal processing and target detection on the echo signal, and extract the target point track set; A data processing module configured to perform track association and filtering processing based on the target point track set to generate a target track set; A feature extraction module configured to extract the multi-dimensional motion feature vector of the target track based on the target track state within a preset time window; A classification module configured to input the feature vector into the pre-trained random forest model and output the UAV recognition result; A countermeasure module configured to perform countermeasures on the UAV according to the UAV recognition result.

[0014] The present invention has the following beneficial effects: The UAV recognition method based on random forest provided by the present invention realizes efficient and accurate recognition of low-altitude UAVs through radar signal processing, trajectory feature extraction, and integrated learning classification. This method abandons the high sampling rate dependence on the micro-Doppler effect, instead, it explores the essential differences in the macroscopic motion patterns between UAVs and interference targets, extracts multi-dimensional motion feature vectors, and combines the robust learning ability of the random forest model for high-dimensional heterogeneous features, effectively overcoming the defects of high signal noise and insufficient feature resolution in low-cost radar systems. This method can achieve high-precision recognition only relying on the historical trajectory data of the target, enhances the interpretability of the model while ensuring the classification accuracy, provides a reliable basis for the precise triggering of subsequent countermeasure strategies, improves the practicability and adaptability of the urban low-altitude security system, and is applicable to low-cost and low-power civilian radar systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0016] Figure 1 It is the method flow chart of the UAV recognition method based on random forest provided by an embodiment of the present invention; Figure 2 It is the structural schematic diagram of the UAV recognition system based on random forest provided by an embodiment of the present invention; Figure 3 It is the visualization diagram of the importance of classification features provided by an embodiment of the present invention; Figure 4 It is the schematic diagram of the measured training data provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the intended invention purpose, the following will, in conjunction with the drawings and preferred embodiments, describe in detail the specific implementation manners, structures, features, and effects of a UAV recognition method and system based on random forest proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0019] The following specifically describes the specific solution of a UAV recognition method and system provided by the present invention in conjunction with the accompanying drawings.

[0020] The UAV recognition method and system based on random forest provided by the present invention are specifically a binary classification problem of UAV recognition. The goal is to determine whether the target is a UAV according to the radar tracking target trajectory information. The following are the specific implementation steps of the present invention: Please refer to Figure 1 , which shows the method flow chart of the UAV recognition method based on random forest provided by an embodiment of the present invention. The method includes: Step S100: Transmit a radar signal and receive the echo signal reflected by the UAV, perform signal processing and target detection on the echo signal, and extract the target point track set; Step S100 specifically includes: Step S110: Mix and filter the echo signal with the transmitted signal to generate an intermediate frequency signal; Specifically, the radar system in this embodiment uses a linear frequency modulation signal, and the transmitted signal of the radar can be expressed as: ; In the formula, is the radar transmitted signal at time; is the signal amplitude, representing the transmission power; is the carrier frequency, determining the radar operating frequency band; is the frequency modulation slope; is the signal bandwidth; is the pulse width; represents the sampling time; represents the cosine operation; represents pi; The radar receives the echo signal of the UAV. After being reflected by the target, the echo signal generates a time delay , is the target distance, is the speed of light, and at the same time, it will be superimposed with the Doppler frequency shift , is the target radial velocity, then the corresponding echo signal can be expressed as:

[0021] In the formula, is the echo signal; is the attenuation coefficient, reflecting the energy loss of the signal during propagation; is the additive noise, including environmental noise and hardware noise; Based on the obtained echo signal Perform signal processing and target detection. The echo signal is mixed with the transmitted signal by a mixer. The mixer multiplies the two signals. According to the sum-to-product principle of trigonometric functions, the high-frequency signal is converted into an intermediate-frequency signal containing target range and velocity information. Then, through a filter, the high-frequency components generated after mixing are filtered out to obtain the intermediate-frequency signal , which can be expressed as:

[0022] In the formula, is the comprehensive attenuation coefficient of the signal after mixing; is the additive noise after filtering. The intermediate-frequency signal after mixing contains two-phase information, which reflects the linear frequency modulation component of the target range , and the Doppler component that reflects the target radial velocity ; represents the frequency modulation slope; represents the speed of light; In this embodiment, through the mixing and filtering process, the high-frequency radar echo signal is converted into a low-frequency intermediate-frequency signal, reducing the complexity of signal processing and hardware costs. The mixing process extracts the joint modulation information of the target range and velocity, providing a baseband signal with a high signal-to-noise ratio for subsequent moving target detection (MTD) and feature extraction. At the same time, it effectively suppresses environmental noise and high-frequency interference, providing a basis for signal preprocessing in low-cost radar systems for UAV identification.

[0023] Step S120: Perform moving target detection on the intermediate-frequency signal to generate a spectrum containing target range and velocity information; Specifically, perform a fast Fourier transform (FFT) on the intermediate-frequency signal output in step S110 to convert the time-domain signal into a frequency-domain signal, which can be expressed as:

[0024] In the formula, is the spectrum after the FFT transformation; the peak position of the spectrum reflects the key parameters of the UAV target, including the range frequency and the Doppler frequency ; by detecting the peak position of the spectrum, the range and velocity of the UAV target can be expressed as: , , represents the speed of light, represents the frequency modulation slope; In this embodiment, through moving target detection and joint spectrum analysis, the efficient decoupling of the distance and speed of the UAV target is realized. By mapping the spectrum peak to the actual motion parameters through physical formulas, accurate input data is provided for subsequent trajectory tracking and feature extraction. At the same time, the dependence on high sampling rate in the traditional micro-Doppler method is avoided, adapting to the hardware limitations of low-cost radar systems.

[0025] Step S130: Dynamically adjust the detection threshold based on background noise and extract the target point set from the spectrum. Specifically, in this embodiment, constant false alarm rate detection (CFAR) is adopted. First, protection units and reference units are set around the unit to be detected in the spectrum. The protection units exclude the interference of target signals, and the reference units cover adjacent frequency regions to collect noise samples. For each reference unit, calculate the mean and standard deviation of its background noise, and then dynamically adjust the detection threshold according to the noise statistic, which can be expressed as: ; In the formula, is the detection threshold; is the local noise mean; is the local noise standard deviation; is the false alarm probability control parameter, which is determined by looking up the table according to the preset false alarm rate or empirical value. The larger the value, the stricter the threshold and the lower the false alarm rate; Compare the amplitude value of each unit in the spectrum with . If the amplitude exceeds the threshold, it is determined as a potential target point; otherwise, it is background noise. By traversing and comparing the entire spectrum, all pixel points with signal intensity values greater than the detection threshold are extracted to form a target point set, which contains the signal information that may belong to the target (UAV). In this embodiment, through the adaptive threshold detection technology, the robustness of target point extraction in a complex electromagnetic environment is improved. The dynamically adjusted threshold can adapt to different noise intensities and distribution characteristics, effectively suppressing false signals caused by multipath reflection and industrial interference, while retaining the weak echoes of real targets. By jointly calculating the noise mean and standard deviation, on the premise of ensuring a low false alarm rate, the target detection probability is maximized, providing high-confidence input data for subsequent trajectory generation and feature extraction, thus supporting the reliable operation of low-cost radar systems in urban complex environments.

[0026] Step S200: Perform trajectory association and filtering processing based on the target point set to generate a target trajectory set; Step S200 specifically includes: Step S210: Generate an initial track based on the time and space distribution of target traces; specifically, this embodiment is used to generate an initial track based on discrete target traces, and the initial track can be implemented by methods such as the logical start method, the hypothesis branch start method, the track start method based on the hough transform, etc.; in this embodiment, the logical start method is taken as an example. As a method for determining track start based on logical judgment, the logical start method specifically determines which traces can form an initial track by setting logical rules. For example, spatio-temporal clustering is performed on the target traces detected by the radar in multiple consecutive frames. If a certain trace appears continuously within the time window and the fluctuation range of the spatial position is less than the threshold, it is determined as a potential stable target; for the trace cluster that satisfies spatio-temporal consistency, initialize the track header, record its initial position, timestamp, and motion state, and form a target track header. The track header represents the starting part of a possible target track. In this embodiment, through the logical start method, an initial track is efficiently generated based on spatio-temporal consistency constraints, avoiding the mis-start of noise traces.

[0027] Step S220: Associate the target traces in consecutive frames to the track based on minimizing the distance cost between the observed value and the predicted state; specifically, after the track start is completed, the track association operation can be performed. The main function of track association is to match the target traces detected by the radar at different times with the existing tracks to achieve continuous tracking of the same target; the track association methods include but are not limited to the nearest neighbor data association, the probabilistic data association algorithm, the joint data association algorithm, etc. This embodiment selects but is not limited to the nearest neighbor data association method, and the specific process is as follows: For each generated initial track, according to its historical state information, use the prediction model to predict the state at the current moment, which can be expressed as: ; In the formula, is the predicted state of the th track at the moment based on the information at the moment; is the state transition matrix; represents the state of the th track at the moment; Then, obtain the observed trace detected by the radar at the current moment ; and map the observed trace to the state space to determine the observation matrix . The observation matrix is a pre-defined matrix that converts the true state of the target into a value in the observation space and is used to convert the predicted state to the same space as the observed trace ; For each track, calculate the current observed value With the track prediction state After passing through the observation matrix The converted distance cost can be expressed as: ; In the formula, Represents the distance cost; Represents the minimum operation; the track number that minimizes the distance cost is found through this formula; the current observed target point is associated with the track with the minimum distance cost to achieve continuous tracking of the same target; if the distance costs of all tracks exceed the preset threshold, it is considered that this point may belong to a new target and the track initiation operation needs to be performed again.

[0028] In this embodiment, by associating the target points of consecutive frames to the track based on minimizing the distance cost between the observed value and the predicted state, continuous tracking of the UAV target is achieved. The association method based on distance cost can accurately match the points at different times with the existing tracks, effectively reducing the possibility of incorrect association, and improving the accuracy and stability of target tracking. The system can construct a complete track from discrete target points, providing reliable target motion information for subsequent UAV identification, helping to improve the accuracy and reliability of UAV identification, and enhancing the system's monitoring and identification capabilities for UAV targets in complex environments.

[0029] Step S230: Perform iterative processing of state prediction and measurement correction on the associated track, and perform dynamic smoothing on the position and speed parameters of the track; specifically, after the track association in step S220, a track filtering operation is performed, and its purpose is to filter out track measurement errors; the track filtering methods adopted in this embodiment include but are not limited to extended Kalman filtering, unscented Kalman filtering, particle filtering, etc. The present invention selects but is not limited to extended Kalman filtering, and the specific process is as follows: Step S231: Predict the state parameters at the current moment based on the historical state of the track; based on the historical motion state of the track (such as position, speed), combined with a preset motion model (such as a uniform or uniformly accelerated model), predict the expected position and speed of the UAV at the current moment; Step S232: Dynamically adjust the weights to update the track state parameters according to the difference between the actual measurement value and the predicted value; convert the target position actually detected by the radar into a rectangular coordinate system and compare it with the predicted value; according to the difference between the predicted value and the actual observed value, dynamically adjust the weight ratio of the two. If the credibility of the observed data is high, more reliance is placed on the measurement value; if the prediction model is more reliable, it is biased towards the prediction result; through this dynamic balance, the state jump caused by noise can be effectively suppressed.

[0030] Step S233: Based on the updated track state, generate a track data set containing the target unique identifier, three-dimensional polar coordinates and rectangular coordinates, velocity components, and signal-to-noise ratio set; organize the updated track state, extract key information such as the coordinates of the target in the three-dimensional polar coordinate system and rectangular coordinate system, and velocity components. At the same time, add a unique identifier to each track, record its signal-to-noise ratio set and the marker information indicating whether it is an actual measurement point. Combine this information to form the final track data set, which can comprehensively describe the state of each track in the airspace at the current moment. Scan the airspace Track set at a certain moment Can be expressed as: ; ; In the formula, Represents The track set in the scanned airspace at a certain moment, summarizing all track-related information detected by the radar at that moment; Represents The total number of valid tracks detected by the radar at a certain moment; Is the unique identifier of the th track, which can be generated by the track starting algorithm, and the generation method is based on the time stamp and spatial hash; Is the th track at A set of state estimation values at a certain moment and historical moments, which can be updated through the extended Kalman filtering process; Is the coordinate of the target in the three-dimensional rectangular coordinate system; Is the coordinate of the target in the three-dimensional polar coordinate system; Is the velocity component; Represents the transpose operation; Is the th track at A set of signal-to-noise ratios at a certain moment and before, recording the signal quality of the track at different moments; Is the th track at A set of marker bits indicating whether it is an actual measurement point at a certain moment and historical moments, where 1 represents an actual measurement point, that is, the track information at that moment is actually measured by devices such as radar, and 0 represents that no actual measurement point is obtained, that is, the track information at that moment may be obtained through prediction or other means; Represents the th track at The radial distance parameter of the target in the three-dimensional polar coordinate system at a certain moment, used to describe the distance information of the target relative to observation devices such as radar; Represents the th track at The azimuth angle parameter of the target at a moment in the three-dimensional polar coordinate system; Indicates the th track at The pitch angle parameter of the target at a moment in the three-dimensional polar coordinate system; In this embodiment, through iterative processing of state prediction and measurement correction for the associated tracks, the extended Kalman filter method is used to dynamically smooth the position and velocity parameters of the tracks; the filtering operation effectively filters out the track measurement errors and improves the accuracy of track state estimation. In a complex detection environment, measurement errors may lead to inaccurate track information, while the extended Kalman filter dynamically adjusts the weights according to the historical state and actual measurement values, making the estimated track state closer to the true value. Secondly, the track data set generated in this embodiment contains rich information, providing comprehensive and accurate data support for subsequent UAV identification and tracking, helping to more accurately judge the motion state and characteristics of the target, and improving the reliability and accuracy of UAV identification.

[0031] Step S300: Extract the multi-dimensional motion feature vector of the target track based on the target track state within a preset time window; please refer to Figure 4 As shown, according to Figure 4 It can be understood that the dot traces of different colors in the figure represent tracks with different IDs. It can be seen that there are a large number of different tracks, indicating that there are many targets in this monitoring scenario and their motion trajectories are different, reflecting the diversity of targets and the complexity of motion; the vertical coordinate shows the smoothed value of the target distance, ranging from close to 0 to about 14000, indicating that the distance span between the target and the monitoring device is large, and the distance differences between different targets and the monitoring point at different times are obvious; combined with the UTC (Coordinated Universal Time) on the horizontal coordinate, that is, the coordinated universal time, the distance change trends of each track at different times are different. Some tracks have relatively stable distances, manifested as small fluctuations in the smoothed distance value within a certain time; while some tracks have relatively large distance changes, and the smoothed distance value has large fluctuations in a short time, reflecting the differences in the target motion state in the time dimension. Figure 4 The information of the smoothed distance values of a large number of tracks with different IDs (distinguished by different colors) shown in

[0032] Step S300 specifically includes: Step S310: Intercept the continuous state sequence of the target track based on a preset time window; specifically, based on the track set obtained in step S200 ; preset a time window, and set the sliding window to cover the most recent At a certain moment, the window interval is , and then we extract the track states within the window from the track set to focus on the states of the target track within a specific time range.

[0033] Step S320: Calculate the statistics of the speed, acceleration, and angle change rate in the continuous state sequence; specifically including: Calculate the mean value of the speed and the fluctuation range deviating from the mean value within a preset time window. The speed feature includes the mean speed at each moment within the window and the speed standard deviation , which can be expressed as: ; ; In the formula, , , respectively represent the speed component of the th track in the moment in the three-dimensional rectangular coordinate system in the axis direction; represents the modulus operation, represents the sum from to , represents the square root operation; represents the average motion speed of the th track within the moment; represents the speed standard deviation of the th track, reflecting the degree of fluctuation of the speed deviating from the mean value within the time window; Calculate the mean value and the fluctuation range of the azimuth angle and pitch angle change rate, including the mean azimuth angle turning rate , standard deviation , which can be expressed as: ; ; In the formula, represents the difference between adjacent moment speed vectors; represents the radar scanning period; the mean acceleration reflects the average acceleration situation of the target track within this time window; the acceleration standard deviation reflects the degree of fluctuation of the acceleration deviating from the mean value within the time window; Calculate the mean value and the fluctuation range of the azimuth angle and pitch angle change rate, including the mean azimuth angle turning rate , standard deviation , which can be expressed as: ​ ; ; ; In the formula, represents the azimuth angle of the th track at the moment; represents the absolute value of the difference in azimuth angles calculated for adjacent moments; Mean pitch angle turn rate , standard deviation , can be expressed as: ; ; In the formula, represents the absolute value of the difference in pitch angles calculated for adjacent moments; Step S330: Extract the mean signal-to-noise ratio and the number of effective measurement points, and generate a multi-dimensional motion feature vector; specifically, the mean signal-to-noise ratio is the average of the signal-to-noise ratios within the window can be expressed as: ; Total number of valid points of the track within the window , can be expressed as: ; Set of feature vectors corresponding to each track , can be expressed as: ; ; In the formula, represents the set of feature vectors corresponding to all tracks at the moment; is an element in the set, representing the feature vector of the th track at the moment, represents the 8-dimensional real number space; represents the mean velocity of the th track at the moment; represents the standard deviation of the velocity of the th track at the moment; represents the mean acceleration of the th track at the moment; represents the standard deviation of the acceleration of the th track at the Indicates Tracks in The mean azimuth turning rate at the moment; Indicates Tracks in The standard deviation of the azimuth turning rate at the moment; Indicates Tracks in The mean pitch angle turning rate at the moment; Indicates Tracks in The standard deviation of the pitch angle turning rate at each moment; Indicates Tracks in The mean signal-to-noise ratio at the moment; Indicates Tracks in The number of valid points at the moment; This embodiment extracts the multi-dimensional motion feature vector of the target track based on the target track state within a preset time window, and constructs a feature vector that comprehensively describes the dynamic behavior of the drone. The statistics of speed and acceleration capture the macroscopic motion pattern of the target (such as constant speed cruising and frequent speed changes), the angle change rate quantifies the turning maneuver characteristics (such as circling and sharp turns), and the mean signal-to-noise ratio and the number of effective points reflect the impact of environmental interference on detection stability. The combined effect of these features highlights the essential differences in the motion laws between drones and interference targets such as birds and kites (such as higher turning rates and more violent acceleration fluctuations in drones), providing highly discriminative input data for the random forest model. By integrating time series statistics and signal quality analysis, this method significantly improves the robustness of drone identification in complex low-altitude environments, while avoiding dependence on high-cost radar hardware and micro-Doppler effects.

[0034] Step S400: input the feature vector into the pre-trained random forest model and output the drone recognition result; Step S400 specifically includes: Step S410: Based on the historical track feature data set, generate multiple decision tree training subsets by random sampling; specifically, summarize the feature vectors of all tracks at different times and their corresponding classification results to form a historical track feature set. ,in Indicates The classification result of the track can be the target type or whether it is a drone. The present invention selects but is not limited to setting: , Indicates that the track is a drone. for non-UAV targets; Then, the random forest is determined by It consists of decision trees. In this embodiment, the number of trees is selected, but not limited to . Set the maximum depth of each decision tree to limit the maximum number of layers of branches during the growth of the decision tree and avoid overfitting caused by excessive growth of the decision tree; the minimum number of samples per node . When the number of samples in a node is less than , the node will no longer split, further restricting the growth of the decision tree and ensuring the generalization ability of the model; For each tree, it is independently trained and the results are integrated through a voting mechanism. For the th tree in the random forest , perform the following operations: Draw samples with replacement from the historical track feature set . That is, after each sample is drawn, the sample will be put back into the original set and still has a chance to be drawn again during the next draw, finally forming a subset . The number of samples drawn can usually be set to be the same as the number of samples in the original dataset . In this way, each subset

[0035] has a certain representativeness, and at the same time, due to the characteristics of sampling with replacement, there are differences between different subsets. In this way, a unique training subset is created for each decision tree, so that each decision tree learns based on different data samples during the training process, thereby increasing the diversity between decision trees and improving the overall performance of the random forest model. Step S420: For each node of each decision tree, randomly select a candidate feature subset and optimize the splitting rule based on the information purity index; specifically, for each candidate feature, calculate the information purity difference of the child nodes under different splitting thresholds; select the feature and threshold that maximize the information purity difference of the child nodes as the splitting rule. In this embodiment, in step S410, a training subset has been generated for each decision tree through Bootstrap sampling For each node of each decision tree, randomly select features from all 8 features of all samples in the training subset corresponding to this node as candidate features for subsequent node splitting decisions. represents the floor operation; For each node dataset , which is a part of the training subset and corresponds to the sample set contained in a node in the decision tree, calculate its Gini coefficient, which can be expressed as: In the formula, is the category The number of samples, where the class c takes values of 0 and 1. 1 indicates that the track is a drone, and 0 indicates a non-drone target; the Gini coefficient reflects the impurity of the sample classes in the node dataset S. The smaller the Gini coefficient, the purer the sample classes, that is, most samples in this node belong to the same class; Furthermore, calculate the feature split gain: For each candidate feature and threshold , the left and right subsets after splitting can be expressed as: and ; is the sample feature vector, is the sample value on the candidate feature ; represents the left subset; represents the right subset; Furthermore, the split gain can be expressed as: ; In the formula, represents the split gain; this formula first calculates the Gini coefficient of the node dataset before splitting, and then calculates the Gini coefficients and of the left and right subsets and after splitting respectively, and weights and sums their Gini coefficients according to the proportion of the number of samples in the left and right subsets to the total number of samples. Finally, subtract the result of the weighted sum from the Gini coefficient before splitting to obtain the split gain. The split gain reflects the degree of improvement in the purity of the sample classes after splitting the node according to the feature and threshold . The larger the split gain, the purer the sample classes after splitting. Traverse all candidate features and their corresponding different thresholds, calculate their split gains, and select the feature and threshold that maximize the split gain as the splitting rule for this node, that is, select the feature and threshold that can maximize the purity of the sample classes to split the node, so as to construct a more effective decision tree structure.

[0036] In this embodiment, the stop condition can be configured such that when the number of node samples or the Gini coefficient (such as Stop splitting when reaching the maximum depth or the minimum number of samples in a leaf node

[0037] In this embodiment, by randomly selecting candidate feature subsets for each node of each decision tree and optimizing the splitting rule based on information purity metrics (Gini coefficient and splitting gain), randomly selecting candidate feature subsets increases the differences between decision trees, enabling different decision trees to learn and judge data from different feature perspectives, avoiding all decision trees relying on the same features for classification, and improving the diversity and generalization ability of the random forest model. Secondly, the method of selecting the optimal splitting point based on minimizing the Gini coefficient can effectively divide the samples by category, continuously improving the sample category purity of each node, and thus constructing a more accurate and effective decision tree structure. In this way, the random forest model can better learn the differences in features between drones and non-drone targets, improving the accuracy of drone recognition.

[0038] Step S430: Based on the classification prediction results of multiple decision trees, determine the drone recognition result through a majority voting mechanism; specifically, based on steps S410 and S420, the independent training of decision trees in the random forest is completed. For the input feature vector the classification result can be expressed as:

[0039] In the formula, is the prediction function of the th tree. This function performs classification prediction on the input feature vector according to the structure and training results of the decision tree; is an indicator function used to judge whether the condition holds; represents the value of the independent variable when the function reaches the maximum value; for each category, this formula calculates the number of decision trees that predict this category among all decision trees, that is, obtained by summing the indicator functions, and then selects the category with the largest number as the final classification result; Furthermore, the random forest evaluates the importance by statistically analyzing the splitting contribution degree of features in all trees. For feature its importance calculation formula can be expressed as: ; In the formula, For the collection of nodes split using feature in the th tree, i.e., these nodes selected feature to determine the splitting rule; is the splitting gain of this node; in this way, the importance of each feature for classification decisions in the random forest model can be evaluated; Finally, embed this classification network into the radar product software for real-time recognition and classification tasks. According to the currently acquired data, use the sliding window method to calculate the feature vector of the new track in real time ; then each tree is based on the prediction function to independently classify and generate prediction results; finally, preset the number of votes predicted as a drone in the results. If the number of votes exceeds the threshold

[0040] then determine that the target corresponding to this new track is a drone; otherwise, determine it as a non-drone target. Table 1: Influence of the number of trees and the sliding window length on the recognition performance

[0041] As can be understood from Table 1, the number of trees and the sliding window length have a significant impact on the performance of the drone recognition method based on random forest. Too few trees will lead to underfitting of the model and missed detection of small sample targets; a balance between performance and computational efficiency can be achieved when T = 100; continuing to increase the number of trees to T = 200 results in limited improvement in accuracy and increased latency, with diminishing marginal returns. In terms of the sliding window length, K = 3 has a low accuracy because it is sensitive to short-term fluctuations and is greatly affected by noise; K = 5 can achieve the best balance between time resolution and stability; K = 8 results in a decrease in accuracy because long-term smoothing weakens the recognition of maneuver features. Therefore, appropriate settings of the number of trees and the sliding window length affect the accuracy and overall performance of this drone recognition method. The selection of T = 100 and K = 3 in this embodiment is not a conventional choice.

[0042] In summary, in terms of determining the recognition result in this embodiment, the majority voting mechanism combines the classification prediction results of multiple decision trees, making full use of the diversity of decision trees in the random forest. Since each decision tree learns based on different training subsets and feature selections during the training process, their prediction results may vary. Through the majority voting mechanism, the influence of misjudgment of a single decision tree can be reduced, and the accuracy and reliability of the recognition result can be improved. Secondly, the calculation of feature importance can help us understand the contribution degree of each feature to the classification decision in the random forest model. This is of great significance for further optimizing the model, selecting more effective features, and understanding the feature differences between drone and non-drone targets. For example, if the importance of a certain feature is low, we can consider whether this feature can be removed in the subsequent model optimization to simplify the model structure and improve the calculation efficiency; According to Figure 3 It can be understood that if the importance score of the signal-to-noise ratio is the highest and the importance score of the direction change rate is the lowest, the direction change rate feature can be removed to simplify the model structure and improve the calculation efficiency; In the classification result inference stage, the random forest classification network is embedded in the radar product software to perform real-time recognition and classification tasks, enabling this method to be applied to the actual drone monitoring scenario. Real-time sliding window calculation of the new track feature vector and fast classification can timely identify newly emerging targets, meeting the requirements for real-time performance in practical applications.

[0043] Preferably, the method further includes: according to the drone recognition result, performing communication interference or navigation deception operations on the target drone to block the communication connection between the target drone and the control terminal or induce the target drone to deviate from the preset flight path; Specifically, based on steps S100 - S400, the recognition of the drone is completed, and the target classification result and the voting rate of the random forest are obtained; when the target classification result is 1 and the random forest voting rate , it is considered that the current target is a drone and the recognition result has a high confidence level, and at this time, the countermeasure device is triggered; Specifically, for communication interference, first determine the frequency band of the interference signal. In this solution, the 2.4GHz / 5.8GHz frequency band is selected, which is a common communication frequency band for drones. Transmitting interference signals in a specific frequency band can block the communication link between the drone and the remote controller or satellite. Start the directional electromagnetic interference device, which emits interference signals in the direction of the target drone according to preset parameters (such as transmission power, interference waveform, etc.). The transmission power of the interference signal needs to be adjusted according to the actual situation, ensuring both effective interference with the drone's communication and avoiding unnecessary impacts on other surrounding devices. Continuously monitor the communication status between the drone and the control terminal, and judge whether the interference has successfully blocked the communication link by analyzing parameters such as the intensity and frequency of the communication signal; Navigation deception specifically means that, through false GPS signals, it simulates the characteristics of real GPS satellite signals, including signal frequency, coding method, timestamp, etc. It emits false GPS signals to the target UAV. After the GPS receiver of the UAV receives the false signals, it will calculate its own position based on the position information in the signals, thereby inducing the UAV to deviate from its original flight path. It can also adjust the parameters of the false GPS signals according to the actual flight trajectory of the UAV and the preset deception target to ensure that the UAV can deviate from the flight path in the expected manner and finally reach a safe area; Mark the target UAV that has completed the disposal, record the track data, disposal time and effect parameters, and update the airspace situation information; specifically, mark the target UAV that has taken countermeasure measures as "disposed" so that it can be quickly identified as having been processed in subsequent monitoring and management to prevent repeated interference; record the track data of the target UAV, including its position, speed, acceleration, etc. before and after being countered. At the same time, record the countermeasure time, that is, the specific time when the countermeasure device is triggered, for subsequent time series analysis. Record the disposal effect parameters, such as the time length of the communication link interruption in the case of directional electromagnetic interference, the distance and angle of the UAV deviating from the flight path in the case of navigation deception, etc. These parameters can be obtained by monitoring the flight state and communication signals of the UAV. Finally, update the airspace situation map in real time according to the disposal situation of the target UAV. In the airspace situation map, mark the status of the disposed target UAV accordingly, such as changing its icon color or shape, etc. At the same time, update the information of other targets in the airspace to ensure that the airspace situation map can accurately reflect the target distribution and status in the current airspace and provide accurate information support for subsequent monitoring and management.

[0044] In summary, the UAV recognition method based on random forest provided by the present invention has the following beneficial effects: Break through the traditional hardware limitations and reduce the deployment cost: This method abandons the high-cost hardware solution that traditionally relies on the micro-Doppler effect. By extracting the essential differences in the macroscopic motion trajectory characteristics (such as speed volatility, acceleration dynamic range, turning rate) between the UAV and the interference target, it significantly reduces the demand for a high-sampling-rate radar system. The hardware cost is reduced, adapting to civilian low-cost radar equipment and supporting the economical deployment of large-scale urban low-altitude monitoring networks.

[0045] High noise resistance and robustness: This method uses sliding window statistical feature extraction and random forest ensemble learning to effectively suppress multipath reflection, electromagnetic noise and clutter interference. By integrating the classification results of multiple decision trees, the model has strong robustness to local noise, and the measured recognition accuracy is above 93%, which is better than the traditional coarse-grained method based on a single motion parameter.

[0046] Real-time processing and low resource consumption: Through parallel decision tree inference, the random forest model achieves an inference time of less than 10 ms at the scale of hundreds of trees, meeting the real-time processing requirements of embedded devices and supporting the second-level anti-aircraft response in urban low-altitude areas. Compared with recurrent neural networks (such as LSTM), the computational resource occupancy is reduced, adapting to the power consumption and computing power limitations of edge computing scenarios.

[0047] Multi-dimensional interpretability to support system optimization: This method quantifies the weights of different motion features on the classification results through feature importance analysis, which can guide the optimization of radar parameters (such as scanning frequency, signal bandwidth) and the adjustment of track tracking strategies.

[0048] Closed-loop optimization and multi-modal scalability. The system supports the automatic transmission of anti-aircraft logs and track data, and combines an online iterative training mechanism to continuously optimize the adaptability of the model to new types of drones and environmental interferences. In addition, it can expand and fuse multi-modal sensor data such as infrared thermal imaging and voiceprint features, break through the perception limitations of a single radar, and improve the all-weather and full-scenario monitoring capabilities.

[0049] Please refer to Figure 2 , which shows a schematic structural diagram of a drone recognition system based on a random forest provided by an embodiment of the present invention. The system includes: A signal processing module configured to transmit radar signals and receive echo signals reflected by drones, perform signal processing and target detection on the echo signals, and extract a set of target point traces; A data processing module configured to perform track association and filtering processing based on the set of target point traces to generate a set of target tracks; A feature extraction module configured to extract a multi-dimensional motion feature vector of the target track based on the target track state within a preset time window; A classification module configured to input the feature vector into a pre-trained random forest model and output a drone recognition result; An anti-aircraft module configured to perform anti-aircraft operations on the drones according to the drone recognition results.

[0050] It should be noted that the above sequence of embodiments of the present invention is only for description and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or may be beneficial.

[0051] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments.

Claims

1. The drone identification method based on random forest is characterized by: The method comprises: Transmit radar signals and receive echo signals reflected by the drone, perform signal processing and target detection on the echo signals, and extract the target point trace set; Perform track association and filtering based on the target point track set to generate a target track set; Based on the target track state within a preset time window, a multi-dimensional motion feature vector of the target track is extracted; The feature vector is input into the pre-trained random forest model and the drone recognition result is output.

2. The method for identifying drones based on random forests according to claim 1, characterized in that: Transmit radar signals and receive echo signals reflected by drones, perform signal processing and target detection on the echo signals, and extract target point trace sets, including: Perform mixing and filtering on the echo signal and the transmission signal to generate an intermediate frequency signal; Perform moving target detection on the intermediate frequency signal to generate a spectrum containing target distance and speed information; The detection threshold is dynamically adjusted based on the background noise, and a set of target points is extracted from the spectrum.

3. The method for identifying drones based on random forests according to claim 1, characterized in that: Based on the target point track set, track association and filtering are performed to generate a target track set, including: Generate initial track based on time and space distribution of target point track; Associating the target points of consecutive frames to the track based on minimizing the distance cost between the observed value and the predicted state; The associated track is iteratively processed with state prediction and measurement correction, and the position and speed parameters of the track are dynamically smoothed.

4. The method for identifying drones based on random forests according to claim 3, characterized in that: The associated track is iterated through state prediction and measurement correction, and the position and speed parameters of the track are dynamically smoothed, including: Predict the current state parameters based on the track history state; According to the difference between the actual measured value and the predicted value, the weight is dynamically adjusted to update the track state parameters; Based on the updated track status, a track data set containing a target unique identifier, three-dimensional polar and rectangular coordinates, velocity components, and a signal-to-noise ratio set is generated.

5. The method for identifying drones based on random forests according to claim 1, characterized in that: Based on the target track state within the preset time window, the multi-dimensional motion feature vector of the target track is extracted, including: Intercepting a continuous state sequence of the target track based on a preset time window; Calculate the statistics of velocity, acceleration and rate of change of angle in a continuous state sequence; The mean value of signal-to-noise ratio and the number of effective measurement points are extracted to generate a multi-dimensional motion feature vector.

6. The method for identifying drones based on random forests according to claim 5, characterized in that: Calculate the statistics of velocity, acceleration, and angle change rate in a continuous state sequence, including: Calculate the mean value of the speed within a preset time window and the fluctuation range that deviates from the mean value; Calculate the mean value of acceleration and the fluctuation range from the mean value; Calculate the mean and fluctuation range of the azimuth and elevation angle change rates.

7. The method for identifying drones based on random forests according to any one of claims 1 to 6, characterized in that: The feature vector is input into the pre-trained random forest model, and the drone recognition results are output, including: Based on the historical track feature data set, multiple decision tree training subsets are generated by random sampling; For each decision tree node, a subset of candidate features is randomly selected, and the splitting rule is optimized based on the information purity index; Based on the classification prediction results of multiple decision trees, the drone identification results are determined through a majority voting mechanism.

8. The method for identifying drones based on random forests according to claim 7, characterized in that: For each decision tree node, a subset of candidate features is randomly selected, and the splitting rules are optimized based on the information purity index, including: For each candidate feature, calculate the difference in information purity of child nodes under different split thresholds; The features and thresholds that maximize the difference in the purity of sub-node information are selected as the splitting rules.

9. The method for identifying drones based on random forests according to claim 7, characterized in that: The method further comprises: According to the drone identification result, communication interference or navigation deception operation is applied to the target drone to block the communication connection between the target drone and the control terminal or induce the target drone to deviate from the preset flight path; The target drones that have been disposed of are marked, and the track data, disposal time and effect parameters are recorded, and the airspace situation information is updated.

10. The drone identification system based on random forest is characterized by: The method for identifying drones based on random forests according to any one of claims 1 to 9 is adopted, wherein the system comprises: A signal processing module is configured to transmit radar signals and receive echo signals reflected by the UAV, perform signal processing and target detection on the echo signals, and extract a target point trace set; A data processing module, configured to perform track association and filtering processing based on the target point track set to generate a target track set; a feature extraction module configured to extract a multi-dimensional motion feature vector of a target track based on a target track state within a preset time window; A classification module, configured to input the feature vector into a pre-trained random forest model and output the drone identification result; The countermeasure module is configured to counter the drone according to the drone identification result.

Images