A method, device and medium for detecting and countering unmanned aerial vehicles based on deep learning

Abstract

This invention discloses a method, device, and medium for detecting and countering unmanned aerial vehicles (UAVs) based on deep learning, relating to the field of deep learning technology. The method includes: extracting phased array radar motion features, image features, and acoustic features from candidate electromagnetic feature packets of the UAV to generate a cross-modal UAV state feature sequence; inputting the cross-modal UAV state feature sequence into a cross-modal countermeasure prediction deep learning network for temporal modeling and attention weighting, outputting UAV state prediction results and countermeasure action parameters; selecting countermeasure actions based on the UAV state prediction results and the current available state of the device, and transforming the countermeasure action parameters to generate a UAV detection and countermeasure instruction set. This invention achieves simultaneous generation of state prediction and countermeasure parameters based on deep learning, supporting intelligent decision-making and achieving adaptive and highly timely countermeasure actions.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method, device and medium for detecting and countering unmanned aerial vehicles (UAVs) based on deep learning. Background Technology

[0002] In recent years, with the gradual relaxation of low-altitude airspace management policies and the widespread application of drone technology in civilian, commercial, and even industrial fields, the number of small and micro drones has grown exponentially. While this trend has driven the development of emerging industries such as logistics, inspection, and aerial photography, it has also raised serious concerns about security risks such as illegal intrusion, privacy violations, and even terrorist attacks. Against this backdrop, drone detection and countermeasures technology, as a core component of the low-altitude security system, has received widespread attention from academia and industry. Early systems mostly used single-modal methods such as radar, radio spectrum monitoring, optical imaging, or acoustic sensing to achieve target detection. In recent years, deep learning technology, due to its advantages in high-dimensional feature extraction, temporal modeling, and multimodal fusion, has been gradually introduced into the field of drone perception, driving a paradigm shift from rule-driven to data-driven approaches.

[0003] Current mainstream UAV detection and countermeasure technologies have two main shortcomings: First, multi-source sensing data (such as electromagnetic, optical, and acoustic data) often exhibit asynchronicity and coordinate deviation in time axis and spatial reference frame. Existing methods often employ simple splicing or independent processing strategies, failing to effectively construct spatiotemporally aligned joint representations, resulting in feature redundancy and information loss, which weakens the robustness of subsequent identification and prediction. Second, in the countermeasure decision-making stage, most systems rely on preset rules or static thresholds to trigger countermeasure actions, lacking the ability to dynamically model the evolution trend of UAV states and failing to adaptively generate optimal countermeasure parameters based on the available state of the equipment, thus limiting the intelligence and practical applicability of the system. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based UAV detection and countermeasure method to address the problems of insufficient feature fusion and lack of dynamic adaptation in countermeasure decisions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a deep learning-based method for detecting and countering unmanned aerial vehicles (UAVs), which includes collecting UAV detection data, preprocessing it, and aligning it to the same time axis and spatial reference frame to generate a multi-source sensing time-series dataset. Joint electromagnetic situation reconstruction is performed on multi-source sensing time-series datasets to obtain electromagnetic situation feature tensors. Channel separation and key component enhancement processing are then performed to generate candidate electromagnetic feature packages for UAVs. Extract phased array radar motion features, image features, and acoustic features from candidate electromagnetic feature packages of UAVs to generate a cross-modal UAV state feature sequence; The state feature sequence of the cross-modal UAV is input into the cross-modal countermeasure prediction deep learning network, and time-series modeling and attention weighting are performed to output the UAV state prediction result and countermeasure action parameters. Based on the UAV status prediction results and the current availability of the equipment, countermeasures are selected, and the countermeasure parameters are converted to generate a UAV detection and countermeasure command set.

[0007] As a preferred embodiment of the deep learning-based UAV detection and countermeasure method of the present invention, the UAV detection data includes phased array radar echo data, radio spectrum sampling data, optical or infrared image frame data, and acoustic waveform data. The preprocessing includes noise suppression, outlier removal, missing fragment imputation, and time stamp normalization.

[0008] As a preferred embodiment of the deep learning-based UAV detection and countermeasure method of the present invention, the steps for generating the multi-source sensing time-series dataset are as follows: A standard time scale is constructed according to a unified timeline and aligned with the preprocessed UAV detection data to generate a time-aligned dataset. The location information and observation direction parameters attached to the time-aligned dataset are subjected to coordinate transformation and combined into a multi-source sensing time-series dataset within a unified spatial reference frame.

[0009] As a preferred embodiment of the deep learning-based UAV detection and countermeasure method of the present invention, the steps for jointly reconstructing the electromagnetic situation from the multi-source sensing time-series dataset to obtain the electromagnetic situation feature tensor are as follows: Frequency sub-interval power distribution and spatial node echo features are extracted from multi-source sensing time-series datasets according to a unified time scale and combined into joint electromagnetic situation reconstruction input nodes. The input nodes of the joint electromagnetic situation reconstruction are mapped to an electromagnetic situation grid in the time, frequency and spatial dimensions, and then normalized and scaled to generate a multidimensional electromagnetic situation feature tensor.

[0010] As a preferred embodiment of the deep learning-based UAV detection and countermeasure method of the present invention, the steps of performing channel separation and key component enhancement processing to generate candidate electromagnetic feature packets for UAVs are as follows: The multidimensional electromagnetic situation feature tensor is input into the electromagnetic situation channel separation deep learning network. Channel separation is performed through multi-layer convolutional structure and attention structure, and the electromagnetic channel separation feature set is output. Based on the electromagnetic channel separation feature set, the response amplitude in the multidimensional electromagnetic situation feature tensor is enhanced, and the feature set is combined according to the time scale and spatial reference frame to generate candidate electromagnetic feature packages for UAVs.

[0011] As a preferred embodiment of the deep learning-based UAV detection and countermeasure method of the present invention, the steps of extracting phased array radar motion features, image features, and acoustic features from the candidate electromagnetic feature package of the UAV to generate a cross-modal UAV state feature sequence are as follows. Based on the candidate electromagnetic feature packages of UAVs, image regions are cropped from the multi-source sensing time-series dataset, phased array radar echo segments are extracted, and acoustic segments are truncated to generate a multimodal UAV candidate dataset. Image regions from the multimodal drone candidate dataset are input into an AI visual recognition deep learning network to extract drone contour shape, size, attitude, and image confidence features, generating image features. The phased array radar echo segments are used to extract the UAV's position coordinates, radial velocity, motion trend, and radar signal confidence to generate phased array radar motion characteristics. Rotor frequency, harmonic structure, and energy change characteristics are extracted from acoustic segments and combined with image features and phased array radar motion characteristics to generate a cross-mode UAV state feature sequence.

[0012] As a preferred embodiment of the deep learning-based UAV detection and countermeasure method of the present invention, the steps of inputting the cross-modal UAV state feature sequence into the cross-modal countermeasure prediction deep learning network, performing temporal modeling and attention weighting, and outputting the UAV state prediction result and countermeasure action parameters are as follows: The state feature sequences of cross-modal UAVs are spliced ​​and normalized in chronological order, and then divided into cross-modal UAV state time segments according to time windows. Input the time segments of the cross-modal UAV state into the temporal construction structure of the cross-modal countermeasure prediction deep learning network, extract the continuous change pattern of the UAV state, and generate temporal intermediate features. The intermediate temporal features and the cross-modal UAV state feature sequence are input into the attention weighting structure, and time weights and modal weights are assigned to generate attention weighted state feature vectors. The attention-weighted state feature vector is input into the output mapping structure. Through multi-level feature recombination, the UAV state prediction result is output, and the components are decomposed and combined to output the countermeasure action parameters.

[0013] As a preferred embodiment of the deep learning-based UAV detection and countermeasure method of the present invention, the steps of selecting countermeasure actions based on UAV state prediction results and the current availability of the device, converting the countermeasure action parameters, and generating a UAV detection and countermeasure instruction set are as follows. Read the current available status of the drone's equipment, bind the drone status prediction results with the countermeasure action parameters, and generate a countermeasure action filtering input set; The countermeasure action parameters in the input set are compared with the current availability status of the equipment, and countermeasure actions that do not meet the equipment capability conditions are eliminated to generate a target countermeasure action record set. The frequency band range, time window, spatial pointing and power parameters in the target countermeasure action record set are disassembled and recombined into executable control fields to generate a UAV detection and countermeasure instruction set.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the deep learning-based drone detection and countermeasure method described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the deep learning-based drone detection and countermeasure method described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by combining electromagnetic situation reconstruction and channel separation enhancement, high-dimensional structured extraction of weak signals from UAVs based on deep learning is achieved, candidate targets are accurately screened, and the signal-to-noise ratio of electromagnetic domain detection is improved; by cross-modal temporal modeling and attention weighting, state prediction and countermeasure parameter generation based on deep learning are achieved simultaneously, supporting intelligent decision-making and achieving adaptive and timely countermeasure actions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a deep learning-based drone detection and countermeasure method.

[0019] Figure 2 This is a flowchart for electromagnetic situation reconstruction and channel separation.

[0020] Figure 3 This is a flowchart for cross-modal feature extraction.

[0021] Figure 4 A flowchart for cross-modal countermeasure prediction and command generation. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a deep learning-based method for detecting and countering unmanned aerial vehicles (UAVs), comprising the following steps: S1: Collect UAV detection data, preprocess it, and align it to the same time axis and spatial reference frame to generate a multi-source sensing time-series dataset; S1.1: UAV detection data includes phased array radar echo data, radio spectrum sampling data, optical or infrared image frame data, and acoustic waveform data; Furthermore, within the UAV detection area, phased array radar, radio spectrum acquisition equipment, optical or infrared imaging equipment, and acoustic acquisition equipment are activated respectively. The phased array radar continuously transmits detection signals according to beam direction, range grid, and scanning period, and receives echo pulse sequences. The radio spectrum acquisition equipment performs continuous spectrum sampling within the frequency coverage area to obtain the spectrum distribution affected by the UAV control link, image transmission link, and navigation link. The optical or infrared imaging equipment records image frame sequences covering the UAV's motion space at a fixed frame rate. The acoustic acquisition equipment records the acoustic waveforms of UAV rotor noise and background sound field at a specified sampling accuracy. By uniformly processing the time stamps, spatial positions, and sampling parameter sets output by different acquisition devices, UAV detection data is generated.

[0026] S1.2: Preprocessing includes noise suppression, outlier removal, missing fragment imputation, and time stamp normalization; Furthermore, phased array radar echo data, radio spectrum sampling data, optical or infrared image frame data, and acoustic waveform data are filtered out according to their respective sampling characteristics by reading the noise reference parameter set of the acquisition equipment. Data records with amplitudes below the noise threshold are filtered out, and segments with abnormal frequency distribution or discontinuous spatial energy distribution are smoothed. After noise suppression, outlier records are removed based on whether the amplitude jump exceeds the abnormal threshold or whether the time interval exceeds the allowable range of the sampling period. After removing outliers, short-time missing segments in the four types of data sequences are interpolated according to the changing trend of adjacent time slices to maintain continuity. After interpolation, the time stamps of the four types of data records are converted into a unified time format.

[0027] It should be noted that the noise threshold (example range: 3% to 8% of the background noise level) is set based on the lowest resolvable noise level in the noise reference parameter set of the acquisition device, and the upper limit is set based on the maximum acceptable noise suppression amplitude without accidentally deleting valid signals.

[0028] The abnormal threshold (15% to 30% of the range of variation in the example) is set based on the maximum allowable fluctuation range of drone detection data under normal conditions, and the upper limit is set based on the range boundary that ensures abnormal records can be removed without accidentally affecting normal transitional changes.

[0029] S1.3: Construct a standard time scale according to a unified time axis and align it with the preprocessed UAV detection data to generate a time-aligned dataset; Furthermore, a standard timescale sequence covering the entire sampling duration is constructed based on the time scale interval of a unified time axis (e.g., 10 milliseconds in the example). By reading the time stamps in the preprocessed phased array radar echo data, radio spectrum sampling data, optical or infrared image frame data, and acoustic waveform data records, and comparing them one by one with the standard timescale sequence, linear and spline interpolation is performed on the observation values ​​of data falling between adjacent scales, and data with higher sampling rates are aggregated and resampled by window, so that each standard time scale corresponds to a set of synchronously output phased array radar echo data records, radio spectrum sampling data records, optical or infrared image frame data records, and acoustic waveform data records. A time-aligned dataset is generated by sequentially arranging the synchronous records of all standard time scales.

[0030] S1.4: Perform coordinate transformation on the location information and observation direction parameters attached to the time-aligned dataset, and combine them into a multi-source sensing time-series dataset within a unified spatial reference frame; Furthermore, the location information and observation direction parameters attached to the time-aligned dataset are read, and coordinate transformation is performed on the acquisition location coordinates and observation direction according to the unified spatial reference system of the UAV detection area. The coordinate expression methods used by different devices are unified into a consistent spatial coordinate form. Based on the transformed coordinates, the phased array radar beam coverage, radio spectrum acquisition direction, and optical or infrared imaging field of view are repositioned. At the same time, the spatial coordinates of the corresponding acquisition location are supplemented for the acoustic waveform data records. After the coordinate transformation is completed, the phased array radar echo data records, radio spectrum sampling data records, optical or infrared image frame data records, and acoustic waveform data records are recombined in time-aligned order, so that each standard time scale forms a spatially consistent set of sensing information. The sets of all time scales are arranged in sequence to form a multi-source sensing time-series dataset.

[0031] S2: Perform joint electromagnetic situation reconstruction on the multi-source sensing time-series dataset, obtain the electromagnetic situation feature tensor, and perform channel separation and key component enhancement processing to generate candidate electromagnetic feature packages for UAVs. S2.1: Extract frequency sub-interval power distribution and spatial node echo features from the multi-source sensing time-series dataset according to a unified time scale, and combine them into joint electromagnetic situation reconstruction input nodes; Furthermore, radio spectrum sampling data records and phased array radar echo data records corresponding to each time scale are read sequentially according to a unified time scale. The continuous frequency range is divided into multiple frequency sub-intervals in the radio spectrum sampling data records according to frequency division rules, and the power distribution of each frequency sub-interval within the corresponding time scale is statistically analyzed to form the power distribution of the frequency sub-interval. The spatial node positions are determined in the phased array radar echo data records according to the range grid and azimuth division rules, and the echo intensity and Doppler variation characteristics of each spatial node within the corresponding time scale are extracted to form the spatial node echo characteristics. The time scale markers, frequency sub-interval power distributions, and spatial node echo characteristics are combined in a unified format to form the joint electromagnetic situation reconstruction input node.

[0032] It should be noted that the frequency division rule is based on the frequency range of the commonly used control link, image transmission link and navigation link of UAV. The continuous spectrum is divided into equal-width intervals or critical points of energy distribution change, so that each frequency sub-interval can independently reflect the local power characteristics.

[0033] The range grid and azimuth division rules are based on the measurement accuracy and beam coverage of the phased array radar. The observation space is discretized according to fixed range intervals and fixed angle intervals, so that each spatial node can independently express the echo characteristics in the corresponding direction and range.

[0034] S2.2: Map the input nodes of the joint electromagnetic situation reconstruction to an electromagnetic situation grid in the time, frequency and spatial dimensions, and normalize and unify the scale to generate a multidimensional electromagnetic situation feature tensor. Furthermore, based on the correspondence between the time, frequency, and spatial dimensions, the power distribution of frequency sub-intervals and the echo characteristics of spatial nodes in each time scale are filled into the electromagnetic situation grid structure. By establishing a continuous time scale index in the time dimension, arranging frequency sub-intervals according to frequency division rules in the frequency dimension, and locating spatial nodes according to distance grids and azimuth division rules in the spatial dimension, the power distribution is mapped to the frequency axis and the echo characteristics are mapped to the spatial axis. After grid filling, normalization and scale unification processing are performed on the electromagnetic situation grid to adjust the power characteristics and echo characteristics of different orders of magnitude to a comparable range. The feature records of all grid positions are extracted from the three-dimensional structure and arranged in a unified format to generate a multidimensional electromagnetic situation feature tensor.

[0035] S2.3: Input the multidimensional electromagnetic situation feature tensor into the electromagnetic situation channel separation deep learning network, perform channel separation through multi-layer convolutional structure and attention structure, and output electromagnetic channel separation feature set; Furthermore, the multidimensional electromagnetic situation feature tensor is input into the electromagnetic situation channel separation deep learning network in chronological order. The multi-layer convolutional structure in the electromagnetic situation channel separation deep learning network extracts local energy patterns and echo structure changes in the frequency and spatial directions, and uses cross-layer connections to maintain the continuity of features at different scales. After the convolution processing is completed, the output is sent to the attention structure. The attention structure assigns weights to the frequency and spatial dimensions according to feature correlation, strengthens the electromagnetic features related to the UAV link and weakens the background communication services, electromagnetic noise and multipath interference components. The electromagnetic features output by the attention structure are reorganized and combined according to channel category to form an electromagnetic channel separation feature set.

[0036] It should be noted that the electromagnetic situation channel separation deep learning network consists of an input organization structure, a multi-layer convolutional structure, a cross-layer connection structure, an attention structure, and a channel reconstruction structure. The input organization structure is used to receive multi-dimensional electromagnetic situation feature tensors arranged in chronological order and maintain temporal continuity. The multi-layer convolutional structure extracts energy distribution patterns and echo structure changes at different scales along the frequency and spatial directions. The cross-layer connection structure is used to fuse shallow and deep features to maintain multi-scale feature continuity. After convolution and cross-layer connection are completed, the attention structure assigns weights to the frequency and spatial dimensions based on feature correlation to highlight UAV link-related features and suppress background communication services, electromagnetic noise, and multipath interference. Finally, the channel reconstruction structure organizes and combines the weighted features according to channel categories to form an electromagnetic channel separation feature set.

[0037] The training of the electromagnetic situational awareness channel separation deep learning network involves collecting multidimensional electromagnetic situational awareness feature tensors covering different electromagnetic environments. Training sample sets are constructed by classifying UAV link features, background communication service features, electromagnetic noise features, and multipath interference features. Supervision signals are set according to the target channel category, and the parameters of the convolutional and attention structures are updated. During training, enhanced interference band samples, low signal-to-noise ratio samples, and mixed link samples are added to improve the network's adaptability to complex environments, enabling the trained network to stably output a clear and separable electromagnetic channel separation feature set.

[0038] S2.4: Based on the electromagnetic channel separation feature set, enhance the response amplitude in the multidimensional electromagnetic situation feature tensor, and combine them according to the time scale and spatial reference frame to generate candidate electromagnetic feature packages for UAVs; Furthermore, by reading the electromagnetic feature records corresponding to the UAV control link, UAV image transmission link, and UAV navigation link in the electromagnetic channel separation feature set, the relevant positions are located in the multidimensional electromagnetic situation feature tensor based on the electromagnetic feature records, and the response amplitude of the relevant positions is enhanced according to the enhancement coefficient threshold, so that the UAV link features have higher contrast in the multidimensional electromagnetic situation feature tensor. After the amplitude enhancement is completed, the enhanced multidimensional electromagnetic situation feature tensor is sliced ​​and recombined according to the time scale marks and spatial reference coordinates in the electromagnetic channel separation feature set, so that each time scale forms a spatial energy distribution structure consistent with the UAV link features. The enhanced segments corresponding to all time scales are arranged in order and packaged in a unified format to generate UAV candidate electromagnetic feature packages.

[0039] It should be noted that the enhancement factor threshold (example range 1.2 to 1.5) is set based on the minimum feasible enhancement ratio that prevents background noise from being excessively amplified when enhancing the drone link features, and the upper limit is set based on the maximum allowable enhancement magnitude that ensures the drone link features are not distorted.

[0040] S3: Extract phased array radar motion features, image features, and acoustic features from the candidate electromagnetic feature package of UAVs to generate a cross-modal UAV state feature sequence; S3.1: Based on the candidate electromagnetic feature package of the UAV, crop the image region from the multi-source sensing time series dataset, extract the phased array radar echo segment and extract the acoustic segment to generate a multimodal UAV candidate dataset; Furthermore, the time scale markers and spatial reference system coordinates in the candidate electromagnetic feature package of the UAV are read. The corresponding optical or infrared image frame data records, phased array radar echo data records, and acoustic waveform data records are located in the multi-source sensing time-series dataset. The image region containing the UAV candidate target is cropped according to the spatial reference system coordinates and the image field of view. The phased array radar echo segments in the direction of the UAV candidate target are extracted according to the spatial reference system coordinates and the phased array radar beam direction. The acoustic segments corresponding to the activity period are extracted from the acoustic waveform data records according to the time scale markers. The image region, phased array radar echo segments, and acoustic segments are combined according to the time scale to generate a multimodal UAV candidate dataset.

[0041] It should be noted that spatial reference system coordinates refer to a unified spatial coordinate expression established within the UAV's detection area. This expression is used to describe the spatial positional relationship of UAV candidate targets in terms of geographical location, distance direction, and azimuth angle, enabling the observation results of different sensing devices to correspond and match under the same spatial reference.

[0042] A phased array radar beam refers to the directional radiation and reception direction formed by controlling the phase of the array elements. It is used to limit the spatial pointing range of the radar for energy illumination and echo reception of candidate targets of UAVs within a specific azimuth and elevation angle range.

[0043] S3.2: Input the image regions in the multimodal drone candidate dataset into the AI ​​visual recognition deep learning network to extract the drone's outline shape, size, attitude, and image confidence features, and generate image features; Furthermore, image regions from the multimodal drone candidate dataset are input into an AI visual recognition deep learning network in chronological order. The convolutional structure in the AI ​​visual recognition deep learning network extracts edge information and texture change patterns of the image regions. The region localization structure identifies the position of the drone outline in the image region and generates outline shape and size features. The attitude analysis structure analyzes the pixel distribution of the drone outline in different directions to deduce the drone's orientation, pitch change, and rotation trend. The confidence evaluation structure calculates image confidence features based on image sharpness, contrast, and texture integrity. The outline shape features, size features, attitude features, and image confidence features are combined into image features in a unified format.

[0044] It should be noted that the AI ​​visual recognition deep learning network is used to perform structured feature extraction on image regions in the multimodal drone candidate dataset. It obtains the drone's contour edges, surface texture and local detail changes through convolutional structure, determines the spatial position of the drone in the image region and extracts size features through region localization structure, derives pitch, roll and orientation changes based on pixel distribution differences through attitude parsing structure, and generates image confidence based on image sharpness, brightness consistency and texture stability through confidence evaluation structure. This enables the image region to have the ability to express contour shape features, size features, attitude features and image confidence features.

[0045] The training of the AI ​​visual recognition deep learning network involves constructing an image training set covering different lighting conditions, background complexity, drone models, and observation angles. Supervisory labels are created for each image sample, specifying its contour shape, pose state, and size. The parameters of the convolutional structure, region localization structure, and pose resolution structure are updated using the supervisory signals. Blurred images, occluded images, and low-contrast images are added as augmentation samples during training to improve the network's robustness to scenes with missing textures, offsets, and partial occlusion. This ensures that the trained AI visual recognition deep learning network can output stable image features under complex conditions.

[0046] The formula for calculating image confidence features is: ; in, Represents image confidence features. Represents the effective edge energy in the image region. Represents texture noise energy. This represents the weighting coefficient for edge noise. This represents the actual contrast value of the image region. This indicates the maximum achievable contrast ratio. Represents the contrast weighting coefficient. Represents the texture integrity benchmark factor. This represents the texture integrity index.

[0047] It should be noted that the edge noise weighting coefficient (example range: 0.3 to 0.7) is determined based on a balance condition where the impact of edge sharpness on overall confidence is neither weakened nor overemphasized. The contrast weighting coefficient (example range: 0.2 to 0.6) is determined based on the contribution of actual contrast changes to recognition reliability under different lighting conditions; The texture integrity index (example range: 0 to 1) is determined based on the actual performance range of the texture structure from completely destroyed to completely continuous.

[0048] When calculating image confidence features, the effective edge energy and texture noise energy of the image region are read, and their energy proportions are weighted according to the noise weighting coefficient to measure structural clarity. Then, the actual contrast and maximum contrast of the image region are read, and the contrast ratio is processed according to the contrast weighting coefficient to reflect usability under different lighting and texture conditions. On this basis, the texture continuity of the image region is corrected according to the texture integrity benchmark factor and texture integrity index to incorporate texture stability into the confidence expression. The above content is used to calculate the image confidence features that can reflect whether the image region is suitable for UAV contour and attitude recognition.

[0049] S3.3: Extract the UAV's position coordinates, radial velocity, motion trend, and radar signal confidence from the phased array radar echo segments to generate phased array radar motion characteristics; Furthermore, the range grid index and azimuth index in the phased array radar echo segments are read, and the range and angle information are converted into the position coordinates of the UAV in a unified spatial reference system based on the spatial reference system of the phased array radar. The radial velocity is estimated based on the Doppler frequency shift or phase difference, and the trend of radial velocity change is combined with the trend of echo phase change to deduce the motion trend of the UAV on the time scale. At the same time, the radar signal confidence is obtained by reading the signal energy distribution and noise level in the phased array radar echo segments to evaluate the credibility of the motion characteristics in the echo segments. The position coordinates, radial velocity, motion trend and radar signal confidence are combined in a unified format to form the motion characteristics of the phased array radar.

[0050] S3.4: Extract rotor frequency, harmonic structure and energy change features from acoustic segments, and combine them with image features and phased array radar motion features to generate a cross-mode UAV state feature sequence; Furthermore, based on the energy distribution of acoustic segments in the time dimension, time-frequency transformation is performed on the acoustic segments to obtain energy structures in different frequency ranges. Then, based on the rotor frequency characteristic threshold, the harmonic structure formed by the rotor fundamental frequency and its integer multiples is located within the energy structure (first, a spectral peak search is performed to obtain a set of candidate fundamental frequencies, and then the optimal fundamental frequency is selected by combining the harmonic consistency score, rather than hard-cutting with a fixed threshold). Energy change features are extracted through the energy envelope changes of adjacent time segments, and the rotor frequency, harmonic structure, and energy change features are organized according to the acoustic feature format. After the acoustic features are generated, they are aligned with the image features and phased array radar motion features corresponding to the same time scale based on the time scale markings. The three types of features are combined in a fixed splicing order to form a cross-modal UAV state feature vector. A cross-modal UAV state feature sequence is generated by sequentially arranging the cross-modal UAV state feature vectors of all time scales.

[0051] It should be noted that the rotor frequency characteristic threshold (example range: 80Hz to 200Hz) is set with the lower limit based on the acoustic frequency corresponding to the lowest rotor speed of a small drone when hovering stably, and the upper limit based on the range of the highest detectable rotor fundamental frequency of a medium-sized rotor drone during high-speed maneuvering.

[0052] S4: Input the cross-modal UAV state feature sequence into the cross-modal countermeasure prediction deep learning network, perform temporal modeling and attention weighting, and output the UAV state prediction result and countermeasure action parameters; S4.1: The state feature sequence of the cross-modal UAV is spliced ​​and normalized in chronological order, and then divided into cross-modal UAV state time segments according to the time window; Furthermore, each cross-modal UAV state feature vector in the cross-modal UAV state feature sequence is read in chronological order. The image features, phased array radar motion features, and acoustic features contained within the cross-modal UAV state feature vector are normalized according to feature normalization rules to adjust the numerical range of different modal features to a uniform scale. The processed image features, phased array radar motion features, and acoustic features are then spliced ​​together in a fixed splicing order to form a cross-modal UAV state feature record with a consistent format. After the cross-modal UAV state feature record is formed, the cross-modal UAV state feature record corresponding to the continuous time scale is divided into multiple cross-modal UAV state time segments according to the time window length threshold.

[0053] It should be noted that the time window length threshold (example range 0.5 seconds to 1 second) is set with the lower limit based on the minimum detectable state change cycle of the drone during short-term maneuvers, and the upper limit based on the longest accommodating time range that maintains temporal continuity and does not lose key action features.

[0054] S4.2: Input the time segment of the cross-modal UAV state into the temporal construction structure of the cross-modal countermeasure prediction deep learning network, extract the continuous change pattern of the UAV state, and generate temporal intermediate features; Furthermore, each cross-modal UAV state time segment is input into the temporal construction structure of the cross-modal countermeasure prediction deep learning network in chronological order. The multi-layer temporal feature extraction structure in the temporal construction structure is used to process the cross-modal UAV state feature records corresponding to the continuous time scales within the cross-modal UAV state time segment. By the change amplitude of image features between continuous time scales, the velocity change trend of phased array radar motion features, and the energy change pattern of acoustic features, the structural motion pattern of the UAV within the time segment is identified. The long-term and short-term temporal dependencies of cross-modal UAV state features between continuous time scales are captured through the temporal association structure. The time axis information in the cross-modal UAV state time segment is encoded into temporal features that reflect motion trends, attitude trends, and link change trends. The processed temporal features are organized into temporal intermediate features according to a unified format.

[0055] It should be noted that temporal correlation structure refers to a network structure used to establish correlations between state features across consecutive time scales, and is used to simultaneously capture the dependencies of cross-modal UAV state features in both short-term changes and long-term evolution.

[0056] S4.3: Input the temporal intermediate features and the cross-modal UAV state feature sequence into the attention weighting structure, assign time weights and modal weights, and generate attention weighted state feature vectors; Furthermore, the intermediate temporal features and the cross-modal UAV state feature sequence are jointly input into the attention weighting structure. The temporal dimension attention generation structure in the attention weighting structure is used to obtain the correlation degree of each time scale in the UAV motion trend, and time weights are assigned according to the temporal correlation threshold. At the same time, the modal attention generation structure in the attention weighting structure is used to assign modal weights to the contribution of image features, phased array radar motion features and acoustic features in the current state determination. By applying the time weights and modal weights to the corresponding feature records, the weighted fusion of cross-time and cross-modal features is achieved, resulting in an attention-weighted state feature vector formed under the combined effect of continuous change mode and modal correlation.

[0057] It should be noted that the time correlation threshold (example range: 0.3 to 0.7) is set with the lower limit based on the weakest but still identifiable time correlation in the short-term maneuver of the UAV, and the upper limit based on the maximum permissible correlation to avoid over-concentration on a single time scale and loss of information on continuous changes across time.

[0058] S4.4: Input the attention-weighted state feature vector into the output mapping structure, output the UAV state prediction result through multi-level feature recombination, and decompose and combine the components to output the countermeasure action parameters; Furthermore, the attention-weighted state feature vector is input to the output mapping structure according to a fixed input format. The multi-level feature recombination structure in the output mapping structure is used to recombine the image features, phased array radar motion features, and acoustic features within the attention-weighted state feature vector layer by layer, so that the cross-modal features form a comprehensive expression that can be used for UAV state inference in the spatial and temporal dimensions. After the multi-level feature recombination is completed, the state determination structure extracts the position change trend, velocity change trend, and attitude change trend of the UAV based on the feature distribution and organizes them into UAV state prediction results. After the UAV state prediction results are generated, the component decomposition structure decomposes the spatiotemporal changes in the results and separates the parameter set used for countermeasure execution. Then, the action parameter combination structure recombines the parameter set to include the action control values ​​corresponding to the frequency dimension, time dimension, and spatial dimension, generating countermeasure action parameters.

[0059] It should be noted that the cross-modal countermeasure prediction deep learning network is used to perform temporal construction, attention weighting, and action parameter inference on the state feature sequence of cross-modal UAVs. The temporal construction structure extracts continuous patterns of position changes, velocity changes, attitude changes, and link changes. The attention weighting structure obtains temporal weights and modal weights, so that the contribution of image features, phased array radar motion features, and acoustic features in different scenarios can be dynamically adjusted. The weighted cross-modal features are recombined through the output mapping structure to generate UAV state prediction results and countermeasure action parameters.

[0060] The training of the cross-modal countermeasure prediction deep learning network is achieved by constructing a cross-modal training dataset covering different UAV types, environmental backgrounds, flight actions, and interference conditions. Image features, phased array radar motion features, and acoustic features are organized into a cross-modal UAV state feature sequence according to the time scale, and the actual state changes and countermeasure action parameters are labeled. The temporal construction structure, attention weighting structure, and output mapping structure are jointly updated by the supervision signal, and noise samples and weak signal samples are added to enhance robustness. This enables the trained network to generate stable state prediction results and countermeasure action parameters under complex conditions.

[0061] Output mapping structure refers to the feature mapping and recombination structure used to convert attention-weighted cross-modal state features into UAV state prediction results and further decompose and combine them to generate countermeasure action parameters.

[0062] Multi-level feature reorganization refers to the phased rearrangement and fusion of attention-weighted state feature vectors according to a hierarchical relationship from coarse to fine and from global to local. It is usually set to at least three levels of reorganization: The first level of reorganization is based on modality, and internal feature aggregation is performed on image features, phased array radar motion features and acoustic features to form a stable representation of each modality; The second level of reorganization is based on time sequence, and the features of each modality at the same time scale are aligned and combined to highlight the consistency of position changes, velocity changes and attitude changes in the time dimension; The third level of reorganization is based on task, and the cross-modal time sequence features obtained in the first two levels are further rearranged into a set of functional features for state prediction and countermeasure execution. For example, features related to spatial changes are aggregated into state prediction inputs, and features related to frequency, time and space control are aggregated into countermeasure action parameter inputs, thereby realizing the step-by-step mapping and reorganization from cross-modal state expression to executable countermeasure parameters.

[0063] S5: Select countermeasures based on the UAV status prediction results and the current availability of the equipment, convert the countermeasure parameters, and generate a UAV detection and countermeasure instruction set; S5.1: Read the current available status of the drone's equipment, bind the drone status prediction results with the countermeasure action parameters, and generate a countermeasure action filtering input set; Furthermore, the frequency capability range, spatial pointing capability range, power output capability range, and response delay range in the available status record of the current UAV equipment are read. The position change trend, speed change trend, and attitude change trend in the UAV status prediction result are bound to the action control set in the countermeasure action parameters according to a fixed format. This makes the available status record of the equipment, the UAV status prediction result, and the countermeasure action parameters form a corresponding combination structure at the same time scale. The countermeasure action filtering input set is generated by organizing the combination structure according to the field order.

[0064] S5.2: Compare the countermeasure action parameters in the countermeasure action screening input set with the current availability status of the equipment item by item, eliminate countermeasure actions that do not meet the equipment capability conditions, and generate a target countermeasure action record set; Furthermore, the action control set and equipment availability status record in the countermeasure action filtering input set are read in a fixed field order. The frequency control field, spatial pointing control field, time control field, and power control field in the action control set are compared item by item with the frequency capability range, spatial pointing capability range, time response range, and power output capability range in the equipment availability status record. The action control field is determined to fall within the corresponding capability range based on the action feasibility threshold. If any control field does not meet the capability range condition, the corresponding countermeasure action is marked as an unexecutable action and removed during the filtering process. After comparing all action control sets, the action control sets that meet all capability conditions are organized into a target countermeasure action record set according to the time scale.

[0065] It should be noted that the action feasibility threshold (example range: 80% to 90%) is set with the lower limit based on the minimum matching degree that ensures the countermeasure action has basic executability within the device's capabilities, and the upper limit based on the highest reasonable matching degree that avoids selecting overly stringent options that would result in an insufficient number of executable actions.

[0066] S5.3: The frequency band range, time window, spatial pointing and power parameters in the target countermeasure action record set are disassembled and recombined into executable control fields to generate a UAV detection and countermeasure instruction set; Furthermore, the action control set in the target countermeasure action record set is read according to a fixed field order. The frequency range, time window, spatial pointing, and power parameters in the action control set are decomposed item by item according to the field recombination rules. The frequency range is decomposed into the corresponding start and end frequency points, the time window is decomposed into the action start time and end time, the spatial pointing is decomposed into the horizontal angle control field and the pitch angle control field, and the power parameter is decomposed into the power amplitude control field and the power adjustment step size control field. After the decomposition is completed, all control fields are recombined into an executable control field set according to the execution order. The control field set is then organized into a UAV detection and countermeasure command set by combining the time scale of the target countermeasure action record set.

[0067] This embodiment also provides a computer device applicable to the case of a deep learning-based drone detection and countermeasure method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning-based drone detection and countermeasure method proposed in the above embodiment.

[0068] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0069] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the deep learning-based UAV detection and countermeasure method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0070] In summary, this invention achieves high-dimensional structured extraction of weak UAV signals based on deep learning by combining electromagnetic situation reconstruction and channel separation enhancement, accurately screening candidate targets and improving the signal-to-noise ratio of electromagnetic domain detection; and achieves synchronous generation of state prediction and countermeasure parameters based on deep learning through cross-modal temporal modeling and attention weighting, supporting intelligent decision-making and achieving adaptive and timely countermeasure actions.

[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting and countering unmanned aerial vehicles (UAVs) based on deep learning, characterized in that: include, Collect UAV detection data, preprocess it, and align it to the same time axis and spatial reference frame to generate a multi-source sensing time-series dataset; Joint electromagnetic situation reconstruction is performed on multi-source sensing time-series datasets to obtain electromagnetic situation feature tensors. Channel separation and key component enhancement processing are then performed to generate candidate electromagnetic feature packages for UAVs. Extract phased array radar motion features, image features, and acoustic features from candidate electromagnetic feature packages of UAVs to generate a cross-modal UAV state feature sequence; The state feature sequence of the cross-modal UAV is input into the cross-modal countermeasure prediction deep learning network, and time-series modeling and attention weighting are performed to output the UAV state prediction result and countermeasure action parameters. Based on the UAV status prediction results and the current availability of the equipment, countermeasures are selected, and the countermeasure parameters are converted to generate a UAV detection and countermeasure command set.

2. The deep learning-based drone detection and countermeasure method as described in claim 1, characterized in that: The UAV detection data includes phased array radar echo data, radio spectrum sampling data, optical or infrared image frame data, and acoustic waveform data. The preprocessing includes noise suppression, outlier removal, missing fragment imputation, and time stamp normalization.

3. The deep learning-based drone detection and countermeasure method as described in claim 2, characterized in that: The steps for generating the multi-source sensing time-series dataset are as follows: A standard time scale is constructed according to a unified timeline and aligned with the preprocessed UAV detection data to generate a time-aligned dataset. The location information and observation direction parameters attached to the time-aligned dataset are subjected to coordinate transformation and combined into a multi-source sensing time-series dataset within a unified spatial reference frame.

4. The deep learning-based drone detection and countermeasure method as described in claim 3, characterized in that: The steps for jointly reconstructing the electromagnetic situation from the multi-source sensing time-series dataset and obtaining the electromagnetic situation feature tensor are as follows: Frequency sub-interval power distribution and spatial node echo features are extracted from multi-source sensing time-series datasets according to a unified time scale and combined into joint electromagnetic situation reconstruction input nodes. The input nodes of the joint electromagnetic situation reconstruction are mapped to an electromagnetic situation grid in the time, frequency and spatial dimensions, and then normalized and scaled to generate a multidimensional electromagnetic situation feature tensor.

5. The deep learning-based drone detection and countermeasure method as described in claim 4, characterized in that: The process of performing channel separation and key component enhancement to generate candidate electromagnetic feature packets for the UAV is as follows: The multidimensional electromagnetic situation feature tensor is input into the electromagnetic situation channel separation deep learning network. Channel separation is performed through multi-layer convolutional structure and attention structure, and the electromagnetic channel separation feature set is output. Based on the electromagnetic channel separation feature set, the response amplitude in the multidimensional electromagnetic situation feature tensor is enhanced, and the feature set is combined according to the time scale and spatial reference frame to generate candidate electromagnetic feature packages for UAVs.

6. The deep learning-based drone detection and countermeasure method as described in claim 5, characterized in that: The steps for extracting phased array radar motion features, image features, and acoustic features from candidate electromagnetic feature packets of UAVs to generate a cross-modal UAV state feature sequence are as follows. Based on the candidate electromagnetic feature packages of UAVs, image regions are cropped from the multi-source sensing time-series dataset, phased array radar echo segments are extracted, and acoustic segments are truncated to generate a multimodal UAV candidate dataset. Image regions from the multimodal drone candidate dataset are input into an AI visual recognition deep learning network to extract drone contour shape, size, attitude, and image confidence features, generating image features. The phased array radar echo segments are used to extract the UAV's position coordinates, radial velocity, motion trend, and radar signal confidence to generate phased array radar motion characteristics. Rotor frequency, harmonic structure, and energy change characteristics are extracted from acoustic segments and combined with image features and phased array radar motion characteristics to generate a cross-mode UAV state feature sequence.

7. The deep learning-based drone detection and countermeasure method as described in claim 6, characterized in that: The steps for inputting the cross-modal UAV state feature sequence into the cross-modal countermeasure prediction deep learning network, performing temporal modeling and attention weighting, and outputting the UAV state prediction result and countermeasure action parameters are as follows. The state feature sequences of cross-modal UAVs are spliced ​​and normalized in chronological order, and then divided into cross-modal UAV state time segments according to time windows. Input the time segments of the cross-modal UAV state into the temporal construction structure of the cross-modal countermeasure prediction deep learning network, extract the continuous change pattern of the UAV state, and generate temporal intermediate features. The intermediate temporal features and the cross-modal UAV state feature sequence are input into the attention weighting structure, and time weights and modal weights are assigned to generate attention weighted state feature vectors. The attention-weighted state feature vector is input into the output mapping structure. Through multi-level feature recombination, the UAV state prediction result is output, and the components are decomposed and combined to output the countermeasure action parameters.

8. The deep learning-based drone detection and countermeasure method as described in claim 7, characterized in that: The steps for selecting countermeasures based on the UAV status prediction results and the current availability of the equipment, converting the countermeasure parameters, and generating a UAV detection and countermeasure command set are as follows: Read the current available status of the drone's equipment, bind the drone status prediction results with the countermeasure action parameters, and generate a countermeasure action filtering input set; The countermeasure action parameters in the input set are compared with the current availability status of the equipment, and countermeasure actions that do not meet the equipment capability conditions are eliminated to generate a target countermeasure action record set. The frequency band range, time window, spatial pointing and power parameters in the target countermeasure action record set are disassembled and recombined into executable control fields to generate a UAV detection and countermeasure instruction set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based drone detection and countermeasure method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based drone detection and countermeasure method according to any one of claims 1 to 8.

Images