Unmanned aerial vehicle threat intention prediction method and device and storage medium

Abstract

The invention relates to an unmanned aerial vehicle threat intention prediction method and device and a storage medium, and is applied to the technical field of anti-unmanned aerial vehicles. The method specifically comprises the steps that through multi-modal deep feature fusion and deep adversarial learning, the system can extract deeper and more abstract unmanned aerial vehicle behavior mode features from multi-source heterogeneous data, so that accurate prediction of the real threat intention of the unmanned aerial vehicle is achieved, the limitation that a traditional method is only based on surface feature judgment is overcome, and the accuracy of the unmanned aerial vehicle threatening intention prediction is improved. The transformation of threat assessment from post-event analysis to pre-event prediction is realized; in the deep adversarial learning framework, various complex unmanned aerial vehicle threat behavior modes including disguise, interference and novel attack strategies can be simulated, and meanwhile, the recognition capability is continuously improved in adversarial training, so that the system can still keep high accuracy and low false alarm rate when facing unknown or variable threats, and the safety of the system is improved. And the robustness and adaptability of the system are obviously enhanced.

Description

Technical Field

[0001] This invention relates to the field of counter-drone technology, specifically to a method, apparatus, and storage medium for predicting drone threat intent. Background Technology

[0002] With the rapid development and popularization of drone technology, while drones have shown great potential in the civilian field, they have also brought increasingly severe security challenges. Illegal intrusion, reconnaissance, carrying dangerous goods, and even launching attacks are frequent incidents, making the need for counter-drone technology particularly urgent.

[0003] Existing counter-drone systems typically include drone detection, identification, and tracking. Traditional systems mainly rely on radar, electro-optical, acoustic sensors, and radio spectrum monitoring to detect and identify drone targets. Although these technologies can capture the physical and communication characteristics of drones to some extent, they still have significant shortcomings in threat assessment and intent recognition, mainly in the following aspects:

[0004] The one-sidedness and lag of threat assessment: Existing counter-drone systems' threat assessment of drones is mostly limited to the analysis of their surface characteristics, including size, speed, heading, and type of communication signals. They usually cannot delve into the true "intentions" behind the drone's behavior. The system can identify that a drone is approaching a protected area, but it cannot determine whether it is unintentionally entering, conducting reconnaissance, attempting to interfere, or preparing to carry out a physical attack. This lack of understanding of "intentions" causes the threat assessment to remain at the level of "knowing what but not why," lacking the ability to predict the future behavior of drones.

[0005] Lack of robustness to complex threats: Existing threat assessment models are often based on preset rules or shallow machine learning models, which are difficult to effectively deal with complex strategies such as drone camouflage, jamming, multi-target coordination or swarm attacks. These threat patterns usually exhibit non-linear and dynamically changing behavioral characteristics, and traditional methods are unable to extract deep and abstract intent clues from them, resulting in high false alarm or false negative rates. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a method, device and storage medium for predicting the threat intent of drones, so as to solve the defects of existing anti-drone technology in threat assessment, such as one-sidedness, lag and lack of robustness, and difficulty in accurately discerning the true intent of drones.

[0007] According to a first aspect of the present invention, a method for predicting the threat intent of a drone is provided, the method comprising:

[0008] Acquire raw monitoring data from various drone monitoring sensors and preprocess the acquired raw monitoring data;

[0009] The preprocessed raw monitoring data are weighted and fused to obtain a multimodal fusion feature vector;

[0010] The multimodal fusion feature vector is input into a deep adversarial generative network framework. The deep adversarial generative network framework simulates the behavior sequence data of drones under different threat scenarios and compares it with the input multimodal fusion feature vector, thereby classifying the threat intent and regressing the threat level of the input multimodal fusion feature vector.

[0011] Preferably, it further includes:

[0012] Acquire various raw monitoring data, corresponding multimodal fusion feature vectors, and corresponding threat intent prediction results and threat levels;

[0013] The collected data undergoes quality checks, deduplication, and cleaning. For threat events that are not clearly labeled or are ambiguous in actual threat events, semi-supervised learning or active learning mechanisms are used to filter out the most valuable data and manually label it.

[0014] The deep adversarial generative network framework is incrementally retrained or transferred to new data after annotation.

[0015] Preferably,

[0016] The raw monitoring data from the various UAV monitoring sensors include:

[0017] Data on the distance, azimuth, pitch angle, radial velocity, and radar cross section of the UAV detected by radar sensors;

[0018] Data on UAV shape recognition, payload recognition, flight attitude and behavior observation monitored by photoelectric sensors;

[0019] Acoustic signals from a drone monitored by an acoustic sensor array;

[0020] Radio signals acquired by radio spectrum monitoring equipment from drone remote control, image transmission, and data link.

[0021] Preferably,

[0022] The preprocessing of the acquired raw monitoring data includes:

[0023] The radar data monitored by the radar sensor is filtered, target detected, parameter extracted, and trajectory generated.

[0024] The photoelectric data monitored by the photoelectric sensor is subjected to denoising, image enhancement, target segmentation and feature extraction operations.

[0025] The acoustic signals monitored by the acoustic sensor array are used to locate the sound source by beamforming or generalized cross-correlation-phase transform algorithm, and the unique frequency of the UAV propeller is identified.

[0026] Broadband scanning, narrowband analysis, and signal demodulation are performed on the radio signals monitored by the radio spectrum monitoring equipment, and protocol analysis and fingerprint identification are also performed to obtain communication protocols, frequency hopping modes, and data traffic information.

[0027] Timestamp alignment, noise filtering, data normalization, and format unification are performed on various raw monitoring data.

[0028] By associating and fusing data from different sensors on the same observed target, the temporal and logical consistency of multi-dimensional data on the same UAV target can be ensured.

[0029] Preferably,

[0030] The weighted fusion of the preprocessed raw monitoring data to obtain a multimodal fusion feature vector includes:

[0031] Multiple preprocessed raw monitoring data are input into the Transformer encoder. The multi-head attention mechanism in the Transformer encoder dynamically learns and assigns importance weights of different modal features in the UAV state description, thereby generating the multimodal fusion feature vector based on the importance weights of different modal features.

[0032] Preferably,

[0033] The deep adversarial generative network framework includes: a generator and a discriminator;

[0034] The generator adopts a Transformer decoder-based structure. The generator generates simulated behavior sequence data through random noise vectors and preset threat context codes. The preset threat context codes represent one-hot encodings or dense embedding vectors of different threat intentions.

[0035] The discriminator employs a bidirectional long short-term memory network. It receives the multimodal fusion feature vector and the simulated behavior sequence data, distinguishes between real data and simulated data, and performs threat intent classification and threat level regression on the real data.

[0036] Preferably,

[0037] The generator and discriminator perform adversarial learning through iterative training. The discriminator's loss function LD includes a binary cross-entropy loss for authenticity judgment and losses for intent classification and threat level regression. The generator's loss function LG aims to maximize the probability that the discriminator misclassifies its generated data as real data.

[0038] According to a second aspect of the present invention, a drone threat intent prediction device is provided, the device comprising:

[0039] Data acquisition module: used to acquire raw monitoring data from various UAV monitoring sensors and preprocess the acquired raw monitoring data;

[0040] Feature fusion module: used to perform weighted fusion of various preprocessed raw monitoring data to obtain a multimodal fusion feature vector;

[0041] Threat prediction module: used to input the multimodal fusion feature vector into the deep adversarial generative network framework. The deep adversarial generative network framework simulates the behavior sequence data of UAVs under different threat scenarios and compares it with the input multimodal fusion feature vector, thereby classifying the threat intent and regressing the threat level of the input multimodal fusion feature vector.

[0042] According to a third aspect of the present invention, a storage medium is provided, the storage medium storing a computer program, which, when executed by a host controller, implements the steps of the above-described method.

[0043] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0044] This application utilizes multimodal deep feature fusion and deep adversarial learning to extract deeper and more abstract drone behavior patterns from multi-source heterogeneous data. This enables accurate prediction of the true threat intent of drones, overcoming the limitations of traditional methods that rely solely on surface feature judgments. It transforms threat assessment from post-event analysis to pre-event prediction. The deep adversarial learning framework can simulate various complex drone threat behavior patterns, including camouflage, jamming, and novel attack strategies. Furthermore, its recognition capabilities are continuously improved during adversarial training, ensuring the system maintains high accuracy even when facing unknown or mutated threats. The low false alarm rate significantly enhances the system's robustness and adaptability. The fused multimodal features and predicted threat intent / level provide operators with comprehensive and intuitive UAV situational awareness information. The intent prediction results provide important intelligent support for subsequent countermeasure decisions, enabling countermeasures to be deployed more accurately, efficiently, and rationally, avoiding overreaction or underresponse. Accurate threat intent prediction can effectively distinguish between different types of UAV behavior, such as accidental intrusion and malicious intrusion, avoiding unnecessary countermeasures against harmless targets, thereby significantly reducing the false alarm rate, saving countermeasure resources, and improving the system's economy and efficiency.

[0045] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0046] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0047] Figure 1 This is a flowchart illustrating a method for predicting the threat intent of a drone according to an exemplary embodiment;

[0048] Figure 2 This is a system schematic diagram of a drone threat intent prediction device according to another exemplary embodiment;

[0049] In the attached diagram: 1-Data acquisition module, 2-Feature fusion module, 3-Threat prediction module. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0051] Example 1

[0052] Figure 1 This is a flowchart illustrating a method for predicting the threat intent of a drone according to an exemplary embodiment, such as... Figure 1 As shown, the method includes:

[0053] S1: Acquire raw monitoring data from various UAV monitoring sensors and preprocess the acquired raw monitoring data.

[0054] S2, weighted fusion of various preprocessed raw monitoring data to obtain a multimodal fusion feature vector;

[0055] S3, the multimodal fusion feature vector is input into the deep adversarial generative network framework. The deep adversarial generative network framework simulates the behavior sequence data of the UAV under different threat scenarios and compares it with the input multimodal fusion feature vector, thereby classifying the threat intent and regressing the threat level of the input multimodal fusion feature vector.

[0056] It is understood that this embodiment collects raw data from various heterogeneous sensors (such as radar, photoelectric, acoustic, and radio spectrum monitoring equipment) in real time. After preprocessing, temporal alignment, and multimodal deep feature fusion, these raw data form a unified high-dimensional feature vector, which is then input into a deep adversarial generative network framework. The deep adversarial generative network framework performs in-depth analysis on the fused features and predicts the potential threat intent and threat level of the UAV in real time and dynamically.

[0057] The raw data from various heterogeneous sensors includes:

[0058] Radar sensor: used to detect the distance, azimuth, pitch angle, radial velocity and radar cross section of the UAV. X-band or Ku-band phased array radar is preferred to achieve simultaneous tracking of multiple targets and high-precision positioning. The radar data is filtered, target detected, parameter extracted and trajectory generated by the signal processing unit 111.

[0059] Photoelectric sensors: including visible light cameras and infrared thermal imagers, are used for UAV shape recognition, payload recognition, flight attitude and behavior observation, and perform noise reduction, image enhancement, target segmentation and feature extraction operations on the acquired photoelectric data;

[0060] Acoustic sensor array: Composed of multiple high-sensitivity microphones, it uses beamforming or generalized cross-correlation-phase transform algorithms to locate the sound source of the acquired acoustic signals, identify the unique frequency of the drone propeller, and help determine the drone model and approximate altitude.

[0061] Radio spectrum monitoring equipment: used to detect and identify radio signals from UAV remote control, image transmission, data link, etc., and to perform broadband scanning, narrowband analysis, signal demodulation, protocol analysis and fingerprint identification on the acquired radio signals to obtain information such as communication protocols, frequency hopping modes, and data traffic;

[0062] The acquired heterogeneous raw data are time-stamped with high precision, and noise is filtered out, data is normalized and format is unified. In addition, a multi-target tracking algorithm (multi-hypothesis tracking based on Kalman filtering or particle filtering) is used to associate and fuse observations of the same target from different sensors to ensure the temporal and logical consistency of multi-dimensional data of the same UAV target.

[0063] Multiple heterogeneous raw data after preprocessing are input into a Transformer encoder for feature fusion. The Transformer encoder architecture includes multiple self-attention layers and feedforward network layers, specifically:

[0064] Generating time-series trajectory feature vectors from radar data Where T is the time step, and d r For radar feature dimensions;

[0065] Visual feature vectors V for drone appearance and payload identification are extracted from photoelectric images using a convolutional neural network (CNN). optical ∈R d ;

[0066] Extracting communication fingerprint feature vectors from spectrum data using a one-dimensional convolutional neural network (1D-CNN).

[0067] These heterogeneous feature vectors are transformed into a unified embedding representation and then input into the Transformer encoder. The multi-head attention mechanism in the encoder can dynamically learn and assign importance weights to different modal features in the UAV state description, generating a unified, high-dimensional, semantically rich fusion feature vector F. target ∈R D This fused feature vector comprehensively and accurately describes the drone's current state, behavioral patterns, and potential intent cues.

[0068] The fusion process can be represented as:

[0069] F target =

[0070] TransformerEncoder([Embed(V radar Embed(V) optical Embed(V) spectrum)]) where [·;·;·] represents feature concatenation, and Embed(·) represents mapping the original features to the embedding space that the Transformer can process;

[0071] The deep adversarial generative network framework dynamically predicts the threat intent and threat level of drones through continuous learning and adversarial training, specifically including:

[0072] Generator: Designed to learn and generate simulated data that highly resembles the threat behavior of real drones;

[0073] Architecture: The generator adopts a Transformer-based decoder structure;

[0074] Input: a random noise vector z ~ N(0,1) and a predefined threat context code C threat The contextual encoding is a one-hot encoding or dense embedding vector representing different threat intentions;

[0075] Output: Generate a series of simulated drone behavior sequence data G(z,C) threat These simulation data include simulated flight trajectories, speed changes, attitude adjustments, virtual payload delivery events, and simulated changes in communication signal characteristics, aiming to simulate the behavior patterns of drones as realistically as possible under various threat scenarios.

[0076] Discriminator: Responsible for distinguishing between real and simulated data, and accurately identifying threat intent and regression threat level from real data;

[0077] Architecture: The discriminator employs a bidirectional long short-term memory (Bi-LSTM) network to effectively capture the contextual information and temporal dependencies of the input sequence. The network ends with a fully connected layer and outputs two key results:

[0078] Authenticity judgment: A scalar probability value D realfake ∈[0,1] (via the Sigmoid activation function), indicating whether the input data comes from a real drone (approaching 1) or from a generator simulation (approaching 0);

[0079] Intent Classification and Threat Level Regression: For data judged as real, output a multi-class probability distribution P. intent ∈[0,1] K (Using the Softmax activation function) indicates the drone's threat intent, where K is the number of intent categories, and outputs a continuous threat level value R. threat ∈[0,10] (through a linear activation function), the higher the value, the greater the threat level;

[0080] Discriminator output function: D(X) = (D realfake (X),Pintent (x),R threat (X) where X represents the input data, which can be the real fused feature vector sequence Ftarget or the simulated behavior sequence data G(z,C) threat );

[0081] Adversarial training process: The generator and discriminator undergo adversarial learning through iterative training, specifically including:

[0082] Discriminator Loss Function (LD): The training objective of the discriminator is to maximize its ability to correctly distinguish between real and simulated data, and to maximize its ability to correctly predict intent and regress threat levels on real data. Where Lintent is the intent classification loss (cross-entropy loss), and Lthreat is the threat level regression loss (mean squared error loss);

[0083] Generator loss function L G The goal of the generator is to generate data that "fools" the discriminator as much as possible, that is, to maximize the probability that the discriminator misclassifies its generated data as real threat data. Furthermore, the generator's training can incorporate an intent-based auxiliary loss to ensure that the generated simulated data truly conforms to the pre-defined threat context encoding C. threat ;

[0084] Through this dynamic adversarial process, the deep adversarial generative network framework can continuously optimize its intent prediction model, especially when facing complex, highly disguised or new drone threats, and can improve its identification ability and robustness.

[0085] This embodiment also discloses an adaptive update and iteration mechanism for the deep adversarial generative network framework, specifically including:

[0086] All data generated during operation is collected in real time and securely stored, including raw sensor data and fused feature vectors from various heterogeneous sensors, as well as intent prediction results and threat levels from deep adversarial generative network frameworks. All data will be structured and stored in a secure data lake and encrypted and version controlled.

[0087] The collected data undergoes quality checks, deduplication, and cleaning. For threat events that are not clearly labeled or are ambiguous in actual threat events, semi-supervised learning or active learning mechanisms are used to select the most valuable data and guide experts to perform manual labeling in order to form a high-quality supervised learning dataset.

[0088] Using newly labeled data, the generator and discriminator in the deep adversarial generative network framework are incrementally retrained or transferred to adapt to new drone behavior patterns and threat strategies, thereby improving the accuracy of intent prediction.

[0089] After rigorous offline testing and validation, the newly trained model is safely and seamlessly deployed to the online system through methods such as canary release or blue-green deployment to replace the old model.

[0090] This embodiment also discloses the construction of a highly realistic virtual UAV adversarial simulation platform. This platform can accurately simulate various UAV types, flight modes, attack strategies, payload behavior, and complex electromagnetic and meteorological conditions. The generator in the deep adversarial generative network framework can generate a large amount of synthetic threat data in this simulated environment for the pre-training and stress testing of the discriminator, greatly expanding the training dataset, especially when real threat data is scarce. Based on the new UAV features, communication protocols, attack methods, and successful threat prediction experience identified during operation, the UAV feature library and threat pattern knowledge base are automatically updated to form a continuously growing intelligent knowledge system.

[0091] Example 2:

[0092] Figure 2 This is a system schematic diagram of a drone threat intent prediction device according to another exemplary embodiment, the device comprising:

[0093] Data acquisition module 1: Used to acquire raw monitoring data from various UAV monitoring sensors and preprocess the acquired raw monitoring data;

[0094] Feature fusion module 2: used to perform weighted fusion of various preprocessed raw monitoring data to obtain a multimodal fusion feature vector;

[0095] Threat prediction module 3: used to input the multimodal fusion feature vector into the deep adversarial generative network framework. The deep adversarial generative network framework simulates the behavior sequence data of UAVs under different threat scenarios and compares it with the input multimodal fusion feature vector, thereby classifying the threat intent and regressing the threat level of the input multimodal fusion feature vector.

[0096] Example 3:

[0097] This embodiment provides a storage medium storing a computer program, which, when executed by a host controller, implements the various steps in the above method.

[0098] It is understood that the storage medium mentioned above can be a read-only memory, a hard disk, or an optical disk, etc.

[0099] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0100] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a small number of sparsely distributed" means at least two.

[0101] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or fewer sparsely distributed executable instructions for implementing a particular logical function or process, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0102] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, a small number of sparsely distributed steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0103] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0104] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0105] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0106] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or a small number of sparsely distributed embodiments or examples.

[0107] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting the threat intent of unmanned aerial vehicles (UAVs), characterized in that, The method includes: Acquire raw monitoring data from various drone monitoring sensors and preprocess the acquired raw monitoring data; The preprocessed raw monitoring data are weighted and fused to obtain a multimodal fusion feature vector; The multimodal fusion feature vector is input into a deep adversarial generative network framework. The deep adversarial generative network framework simulates the behavior sequence data of drones under different threat scenarios and compares it with the input multimodal fusion feature vector, thereby classifying the threat intent and regressing the threat level of the input multimodal fusion feature vector.

2. The method according to claim 1, characterized in that, Also includes: Acquire various raw monitoring data, corresponding multimodal fusion feature vectors, and corresponding threat intent prediction results and threat levels; The collected data undergoes quality checks, deduplication, and cleaning. For threat events that are not clearly labeled or are ambiguous in actual threat events, semi-supervised learning or active learning mechanisms are used to filter out the most valuable data and manually label it. The deep adversarial generative network framework is incrementally retrained or transferred to new data after annotation.

3. The method according to claim 2, characterized in that, The raw monitoring data from the various UAV monitoring sensors include: Data on the distance, azimuth, pitch angle, radial velocity, and radar cross section of the UAV detected by radar sensors; Data on UAV shape recognition, payload recognition, flight attitude and behavior observation monitored by photoelectric sensors; Acoustic signals from a drone monitored by an acoustic sensor array; Radio signals acquired by radio spectrum monitoring equipment from drone remote control, image transmission, and data link.

4. The method according to claim 3, characterized in that, The preprocessing of the acquired raw monitoring data includes: The radar data monitored by the radar sensor is filtered, target detected, parameter extracted, and trajectory generated. The photoelectric data monitored by the photoelectric sensor is subjected to denoising, image enhancement, target segmentation and feature extraction operations. The acoustic signals monitored by the acoustic sensor array are used to locate the sound source by beamforming or generalized cross-correlation-phase transform algorithm, and the unique frequency of the UAV propeller is identified. Broadband scanning, narrowband analysis, and signal demodulation are performed on the radio signals monitored by the radio spectrum monitoring equipment, and protocol analysis and fingerprint identification are also performed to obtain communication protocols, frequency hopping modes, and data traffic information. Timestamp alignment, noise filtering, data normalization, and format unification are performed on various raw monitoring data. By associating and fusing data from different sensors on the same observed target, the temporal and logical consistency of multi-dimensional data on the same UAV target can be ensured.

5. The method according to claim 4, characterized in that, The weighted fusion of the preprocessed raw monitoring data to obtain a multimodal fusion feature vector includes: Multiple preprocessed raw monitoring data are input into the Transformer encoder. The multi-head attention mechanism in the Transformer encoder dynamically learns and assigns importance weights of different modal features in the UAV state description, thereby generating the multimodal fusion feature vector based on the importance weights of different modal features.

6. The method according to claim 5, characterized in that, The deep adversarial generative network framework includes: a generator and a discriminator; The generator adopts a Transformer decoder-based structure. The generator generates simulated behavior sequence data through random noise vectors and preset threat context codes. The preset threat context codes represent one-hot encodings or dense embedding vectors of different threat intentions. The discriminator employs a bidirectional long short-term memory network. It receives the multimodal fusion feature vector and the simulated behavior sequence data, distinguishes between real data and simulated data, and performs threat intent classification and threat level regression on the real data.

7. The method according to claim 6, characterized in that, The generator and discriminator perform adversarial learning through iterative training. The discriminator's loss function LD includes a binary cross-entropy loss for authenticity judgment and losses for intent classification and threat level regression. The generator's loss function LG aims to maximize the probability that the discriminator misclassifies its generated data as real data.

8. A device for predicting the threat intent of unmanned aerial vehicles (UAVs), characterized in that, The device includes: Data acquisition module: used to acquire raw monitoring data from various UAV monitoring sensors and preprocess the acquired raw monitoring data; Feature fusion module: used to perform weighted fusion of various preprocessed raw monitoring data to obtain a multimodal fusion feature vector; Threat prediction module: used to input the multimodal fusion feature vector into the deep adversarial generative network framework. The deep adversarial generative network framework simulates the behavior sequence data of UAVs under different threat scenarios and compares it with the input multimodal fusion feature vector, thereby classifying the threat intent and regressing the threat level of the input multimodal fusion feature vector.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by the main controller, implements the various steps of the drone threat intent prediction method as described in any one of claims 1-7.

Images