Unsupervised identification method and device for drones based on radar data
Through the preprocessing and cluster analysis of radar data, combined with the random forest classification model, unsupervised drone recognition is achieved, which solves the recognition problem of radar systems under unlabeled data and improves the efficiency of drone recognition.
Patent Information
- Application Number
- CN202211377305.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-04
- Publication Date
- 2025-09-30
- Estimated Expiration
- 2042-11-04
AI Technical Summary
Existing radar systems find it difficult to effectively distinguish drones from other low-speed flying targets, such as birds, without labeled data, and the cost of manually labeling data is high, resulting in low efficiency in drone identification.
The target track data is collected by radar, pre-processed and segmented, statistical features are extracted, and the data is aggregated into two categories using a clustering algorithm. The random forest classification model is trained for unsupervised recognition without any data labeling.
It achieves efficient classification and recognition of drones and non-drones, improves recognition efficiency, and avoids the time-consuming and labor-intensive problem of manual labeling.
Smart Images
Figure CN115565090B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) supervised identification, and in particular to a method and device for unsupervised UAV identification based on radar data. Background Art
[0002] Drones have rapidly developed and are now widely used in various fields, including the military. However, radar, the most widely used detection equipment, currently lacks the ability to identify and classify targets. It cannot distinguish drones from "low, small, and slow" targets like birds, limiting its further military benefits. Furthermore, the high cost of manually labeling radar-detected flight data increases the difficulty of identifying target types through radar data.
[0003] Patents and literature such as Chinese patents CN201610896005.6, CN202011419393.1, CN201810851549.X, and CN202111110784.X, as well as "UAV Classification and Identification Based on Dual Radar Micromotion Feature Fusion" by Zhang Pengfei, Li Gang, and Huo Chaoying, and "Drone Classification Using Convolutional Neural Networks With Merged Doppler Images" by BK Kim, HS Kang, and SO Park, all utilize supervised approaches to identify drones. However, the current problem with supervised approaches is that they require a large amount of labeled data. However, labeling radar detection data is time-consuming and labor-intensive. Therefore, it is necessary to consider unsupervised drone identification solutions in the absence of labeled data. Summary of the Invention
[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and device for unsupervised drone identification based on radar data, so as to realize unsupervised drone identification without labeled data.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] Unsupervised drone identification methods based on radar data include:
[0007] Step a. Collect target track data using radar and pre-process and split it into subsequences;
[0008] Step b. extracting subsequence features using statistics;
[0009] Step c. Classify the subsequences of the same track into the same subclass, and use a clustering algorithm to aggregate the subclasses into two categories based on the features extracted in step b: one category is drones and the other category is non-drones;
[0010] Step d. Based on the two major categories obtained in step c, a classification algorithm is trained, and then the input track data to be judged is identified as a drone or a non-drone according to the classification algorithm.
[0011] A further technical solution is: in step a, the collected target track data includes track data of UAV targets and non-UAV targets, and the track data of non-UAV targets in the collected target track data is more than the track data of UAV targets; the target track data includes high-dimensional time series data composed of information such as distance, azimuth, pitch, track number, etc. of each cycle of the target.
[0012] A further technical solution is: in step a, preprocessing is performed to divide the subsequences: missing points on the track are supplemented by interpolation, and each track is divided into subsequences of a specified length L, and data smaller than L is eliminated.
[0013] A further technical solution is:
[0014] In step b, extracting subsequence features using statistics includes:
[0015] For each subsequence, the maximum and minimum moving distances in the x-direction, the mean and variance of the x-direction velocity, the mean and variance of the x-direction acceleration, the maximum and minimum moving distances in the y-direction, the mean and variance of the y-direction velocity, the mean and variance of the y-direction acceleration, the maximum and minimum moving distances in the z-direction, the mean and variance of the z-direction velocity, and the mean and variance of the z-direction acceleration are included.
[0016] A further technical solution is: in step c, a clustering algorithm is used to aggregate the subclasses according to the features extracted in step b into two major categories, specifically including:
[0017] The distances between different subclasses are calculated using the features extracted in step b, and the subclasses with the closest distances are merged; this process is repeated until only two classes remain, and the two classes are then divided into UAV and non-UAV classes based on the number of tracks.
[0018] A further technical solution is: the step d specifically includes:
[0019] d.1) Using the two subsequences of samples from step c, train a classification model based on the features extracted in step b.
[0020] d.2) When the track data to be judged is input, it is divided into sub-tracks of length L; the features of the sub-tracks to be judged obtained by segmentation are extracted according to step b) and then sent to the classification model in step d.1) to obtain the judgment result; the recognition result of the track to be judged is determined by voting on the judgment results of each sub-track to be judged obtained by segmenting the track to be judged.
[0021] A further technical solution is: the classification model is a random forest classification model.
[0022] A further technical solution is: the voting method adopts a hard voting method.
[0023] At the same time, in order to achieve the above purpose, the present invention also provides the following technical solutions:
[0024] An unsupervised drone identification device based on radar data consists of a memory, a processor, a bus, and a radar. The radar detects aerial targets and generates target track data, and sends the track data to the memory and processor via a bus. The bus connects the processor, memory, and radar. The memory stores a computer program that enables the processor to execute the unsupervised drone identification method based on radar data as described above.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] The present invention realizes the classification and identification of UAV targets and non-UAV targets based on the target track data collected by radar. Compared with the methods in existing related patent literature, the present invention does not require any data labeling work, which solves the problem that radar data labeling is time-consuming and labor-intensive, and sometimes even impossible to label, thereby improving the efficiency of UAV identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 This is a flow chart of the unsupervised identification method of drones based on radar data of the present invention.
[0028] Figure 2 The track diagrams of several targets on a certain radar are given.
[0029] Figure 3 A schematic diagram of the specific process of feature extraction is given.
[0030] Figure 4 This is a structural diagram of the unsupervised identification device for drones based on radar data of the present invention. DETAILED DESCRIPTION
[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0032] Example 1;
[0033] Refer to the attached Figure 1 , unsupervised drone recognition method based on radar data, including:
[0034] Step a. Collect target track data using radar and pre-process and split it into subsequences;
[0035] Among them, using radar to collect target track data refers to using radar to collect UAV targets and various non-UAV targets such as flying birds; since non-UAV targets such as flying birds are easy to collect, their number will far exceed UAV targets.
[0036] In this embodiment, a total of 325 target track data are collected, most of which are bird tracks.
[0037] Specifically, target track data refers to high-dimensional time series data consisting of information such as the distance, azimuth, pitch, and track number of each target cycle. Figure 2 Given a track diagram of several targets on a radar, the target's track is depicted using track data. A track is a time series. In this example, the normal interval of the time series, i.e., the radar cycle, is 2 seconds, and the dimensions are 4, namely, range, azimuth, pitch, and track number. The track number remains unchanged for the same track.
[0038] Preprocessing to split the subsequences means interpolating to supplement the missing points on the track, and then splitting each track into subsequences of a specified length L, and removing data smaller than L. It is worth noting that all data do not need to be labeled.
[0039] In this embodiment, when the time interval between two adjacent points on a track exceeds one radar cycle, linear interpolation is used to add interpolation points for supplementation; the subsequence length is set to 7, and track data with a length less than 7 is eliminated; thus, a total of 1824 subsequences are obtained.
[0040] Step b. extracting subsequence features using statistics;
[0041] Specifically, using statistics to extract subsequence features means extracting the maximum and minimum moving distances in the x-direction, the mean and variance of the x-direction velocity, the mean and variance of the x-direction acceleration, the maximum and minimum moving distances in the y-direction, the mean and variance of the y-direction velocity, the mean and variance of the y-direction acceleration, the maximum and minimum moving distances in the z-direction, the mean and variance of the z-direction velocity, and the mean and variance of the z-direction acceleration for each subsequence. Figure 3 The specific process of feature extraction is shown.
[0042] In this embodiment, the 18 features are extracted from each subsequence. The speed in a certain direction is obtained by dividing the first-order difference of the displacement in that direction by two seconds, and the acceleration in a certain direction is obtained by dividing the first-order difference of the speed in that direction by two seconds.
[0043] The maximum and minimum movement distances in a certain direction, such as the x direction, are calculated as follows:
[0044] Let the position of the subsequence in the x direction be [d(t_1), d(t_2), …d(t_L)], where L is the length of the subsequence. Then the maximum moving distance d(max) = max(|d(t_2)-d(t_1)|, |d(t_3)-d(t_2)|, …, |d(t_L)-d(t_(L-1))|); the minimum moving distance d(min) = min(|d(t_2)-d(t_1)|, |d(t_3)-d(t_2)|, …, |d(t_L)-d(t_(L-1))|). ))|); In this embodiment, specifically: the position of the subsequence in the x direction is recorded as [d(t_1), d(t_2), … d(t_7)], the maximum moving distance d(max) = max(|d(t_2) - d(t_1)|, |d(t_3) - d(t_2)|, …, |d(t_7) - d(t_6|); the minimum moving distance d(min) = min(|d(t_2) - d(t_1)|, |d(t_3) - d(t_2)|, …, |d(t_7) - d(t_6|).
[0045] Step c. Classify the subsequences of the same track into the same subclass, and use a clustering algorithm to aggregate these subclasses into two major categories based on the features extracted in step b: one for drones and the other for non-drones;
[0046] Furthermore, the clustering algorithm can be various hierarchical clustering methods;
[0047] Aggregating these subclasses into two major categories based on the features extracted in step b means calculating the distances between different subclasses using the features extracted in step b, and then merging the subclasses with the closest distances; repeating this process until only two categories remain, and then setting the two categories as drone and non-drone. Among them, drone refers to the category with fewer tracks in the two major categories; non-drone refers to the category with more tracks.
[0048] In this embodiment, the centroid clustering method is used to merge the 325 subclasses into two major categories; one category contains 41 tracks and is a drone category, and the other category contains 284 tracks and is a non-drone category.
[0049] Step d. Based on the two major categories obtained in step c, a classification algorithm is trained, and then the input track data to be judged is identified as a drone or a non-drone according to the classification algorithm.
[0050] The two categories obtained in step c are used to train a classification algorithm, and then the classification algorithm is used to identify whether the input track data to be judged is a drone or a non-drone, including:
[0051] d.1) Using the two sample subsequences obtained in step c, train a classification model based on the features extracted in step b.
[0052] d.2). When the track data to be judged is input, it is divided into sub-tracks of length L; the features of the sub-tracks to be judged obtained by segmentation are extracted according to step b, and then sent to the classification model in step d.1 to obtain the judgment result; the recognition result of the track to be judged is determined by voting on the judgment results of each sub-track obtained by segmentation.
[0053] In this embodiment, the drone class contains 577 subsequences, and the non-drone class contains 1247 subsequences. These data are used to train a random forest classification model. For the input track data to be judged, every 7 points are cut off as a sub-track to extract the features and input them into the random forest classification model for judgment. The recognition result of the track to be judged is obtained by voting on the recognition results of all the sub-tracks cut out from the track up to the current moment. This embodiment adopts the hard voting method.
[0054] It should be noted that the radar period, subsequence length, sliding time window size and step size, clustering algorithm used, classification algorithm, etc. can all be determined according to actual conditions and are not limited to the specific values listed in this embodiment.
[0055] Example 2;
[0056] Refer to the attached Figure 4The unsupervised identification device for drones based on radar data is composed of a memory, a processor, a bus, and a radar. The radar detects aerial targets and forms corresponding tracks, and sends the track data to the memory and processor through a bus; the bus connects the processor, memory, and radar; the memory stores a computer program, and the computer program enables the processor to execute the unsupervised identification method for drones based on radar data.
[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. Unsupervised identification method of drones based on radar data, characterized by: include: Step a. Collect target track data using radar and pre-process and split it into subsequences; Step b. extracting subsequence features using statistics; In step b, extracting the features of the subsequence using statistics includes: The maximum and minimum moving distances in the x-direction, the mean and variance of the x-direction velocity, the mean and variance of the x-direction acceleration, the maximum and minimum moving distances in the y-direction, the mean and variance of the y-direction velocity, the mean and variance of the y-direction acceleration, the maximum and minimum moving distances in the z-direction, the mean and variance of the z-direction velocity, and the mean and variance of the z-direction acceleration for each subsequence; Step c. Classify the subsequences of the same track into the same subclass, and use a clustering algorithm to aggregate the subclasses into two categories based on the features extracted in step b: one category is drones and the other category is non-drones; Step d. Based on the two major categories obtained in step c, a classification algorithm is trained, and then the input track data to be judged is identified as a drone or a non-drone according to the classification algorithm.
2. The unsupervised identification method of drones based on radar data according to claim 1 is characterized in that: In step a, the collected target track data includes track data of UAV targets and non-UAV targets, and the track data of non-UAV targets in the collected target track data is more than the track data of UAV targets; the target track data includes high-dimensional time series data composed of information such as distance, azimuth, pitch, track number, etc. of each cycle of the target.
3. The unsupervised identification method of drones based on radar data according to claim 2 is characterized in that: In step a, preprocessing is performed to divide the subsequences: missing points on the track are supplemented by interpolation, and each track is divided into subsequences of a specified length L, and data smaller than L is eliminated.
4. The unsupervised identification method of drones based on radar data according to any one of claims 1 to 3, characterized in that: In step c, a clustering algorithm is used to aggregate the subclasses into two categories according to the features extracted in step b, specifically including: The distances between different subclasses are calculated using the features extracted in step b, and the subclasses with the closest distances are merged; this process is repeated until only two classes remain, and the two classes are then divided into UAV and non-UAV classes based on the number of tracks.
5. The unsupervised identification method of drones based on radar data according to claim 4 is characterized in that: The step d specifically includes: d.1) Using the two subsequences of samples from step c, train a classification model based on the features extracted in step b. d.2) When the track data to be judged is input, it is divided into sub-tracks of length L; the features of the sub-tracks to be judged obtained by segmentation are extracted according to step b) and then sent to the classification model in step d.1) to obtain the judgment result; the recognition result of the track to be judged is determined by voting on the judgment results of each sub-track to be judged obtained by segmenting the track to be judged.
6. The unsupervised identification method of drones based on radar data according to claim 5, characterized in that: The classification model is a random forest classification model.
7. The unsupervised identification method of drones based on radar data according to claim 5, characterized in that: The voting method adopts a hard voting method.
8. Unsupervised identification device for drones based on radar data, characterized in that: It is composed of a memory, a processor, a bus and a radar. The radar detects aerial targets and generates target track data, and transmits the track data to the memory and processor via the bus. bus connecting the processor, memory and radar; The memory stores a computer program, which enables the processor to execute the unsupervised drone identification method based on radar data as described in any one of claims 1 to 7.