Unmanned aerial vehicle threat assessment method and device, computer equipment and storage medium

Abstract

The invention relates to the technical field of computers, and discloses an unmanned aerial vehicle threat assessment method and device, computer equipment and a storage medium, and the method comprises the steps: detecting a target unmanned aerial vehicle with an intrusion risk in a preset region; threat elements of the target unmanned aerial vehicle are obtained, spatial-temporal features of the threat elements are extracted through a time sequence mixer, and the threat elements at least comprise dynamic threat elements and static threat elements; and mapping the spatial-temporal characteristics to a threat prototype space, and determining the threat level of the target unmanned aerial vehicle according to preset prototype classification in the threat prototype space. According to the method, the problems of low accuracy and insufficient real-time performance of unmanned aerial vehicle threat assessment in a few-sample and dynamic complex scene in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically to methods, apparatus, computer equipment, and storage media for assessing drone threats. Background Technology

[0002] The rapid development of drone technology has led to its widespread application in aerial photography, logistics, agriculture, and other fields, but it has also brought new challenges to low-altitude security management and the protection of critical infrastructure. In important locations such as nuclear power plants, airports, and military bases, unauthorized drone intrusions can pose serious security threats. Counter-drone systems typically include detection, identification, threat assessment, and interception, with threat assessment being the core component, directly determining the rational allocation of defense resources and the effectiveness of response strategies. Currently, drone threat assessment mainly relies on expert experience rules, uncertainty reasoning, data-driven models, or multi-criteria optimization methods. These methods are effective to some extent under conditions of sufficient samples and stable scenarios, but they still have significant limitations.

[0003] Existing threat assessment methods generally face the following problems: First, assessment methods rely on a large number of historical samples for training, while in reality, drone attack events are scarce and sudden, leading to a significant decrease in model accuracy under conditions of few samples; Second, traditional methods have high computational complexity and poor real-time performance in dynamic and complex environments, making it difficult to meet the needs of immediate response to rapidly invading drones; Third, assessment models fail to effectively integrate the dynamic threat elements of drones (such as motion state and radar characteristics) with static threat elements (such as aircraft type and attack capabilities), resulting in insufficient assessment accuracy under varied attack modes. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method, apparatus, computer equipment, and storage medium for drone threat assessment, in order to solve the problems of low accuracy and insufficient real-time performance of existing technologies in drone threat assessment under conditions of few samples and dynamic, complex scenarios.

[0005] In a first aspect, embodiments of the present invention provide a method for assessing the threat of unmanned aerial vehicles (UAVs), the method comprising: Detect target drones that pose an intrusion risk within a pre-defined area; The threat elements of the target drone are acquired, and the spatiotemporal features of the threat elements are extracted using a time series mixer. The threat elements include at least dynamic threat elements and static threat elements. The spatiotemporal features are mapped to a threat prototype space, and the threat level of the target UAV is determined based on the preset prototype classification within the threat prototype space.

[0006] Furthermore, acquiring the threat elements of the target drone includes: Acquire monitoring data of the target drone within the preset area; The situational threat elements and radar threat elements of the target UAV are extracted from the monitoring data, and the situational threat elements and radar threat elements are used as the dynamic threat elements. The physical attribute elements of the target UAV are identified from the monitoring data, and these physical attribute elements are used as the static threat elements.

[0007] Furthermore, the situational threat elements include at least one of speed, heading angle, climb angle, altitude, radial distance, horizontal distance, azimuth angle, and elevation angle; the radar threat elements include at least one of pulse width, operating frequency, pulse repetition frequency, and radar cross-section area; and the physical attribute elements include at least one of aircraft type, attack capability, and maneuverability.

[0008] Furthermore, the extraction of the spatiotemporal features of the threat elements using a time-series mixer includes: The threat elements are mapped to a preset dimensional space through the channel mixing module in the time series mixer to obtain spatial mixing features; The time-series mixer uses a time-mixing module to establish the time dependency relationship between the dynamic features of UAVs in different time periods based on the spatial mixing characteristics. The spatial mixing features and the temporal dependencies are used as the spatiotemporal features.

[0009] Furthermore, establishing the temporal dependency relationship between the dynamic features of the UAV in different time periods based on the spatial mixing features includes: The query matrix, key matrix, and value matrix corresponding to the spatial hybrid features are calculated using the pre-trained weight matrix; Based on the query matrix and the key matrix, calculate the attention weights between features at different time steps; The value matrix is ​​weighted and fused based on the attention weights to obtain the fused temporal features; The fused temporal features are spliced ​​together to obtain the temporal dependencies between the dynamic features of UAVs in different time periods.

[0010] Furthermore, the step of mapping the spatiotemporal features to a threat prototype space and determining the threat level of the target UAV based on a preset prototype classification within the threat prototype space includes: Obtain the preset prototype classification within the threat prototype space; Query samples are constructed using the spatiotemporal features, and the similarity between the query samples and each of the prototype classifications is calculated. The category prototype with the lowest similarity is taken as the target category prototype to which the query sample belongs; Based on the mapping relationship between the prototype classification and the preset threat level, the target threat level corresponding to the target category prototype is determined as the threat level of the target UAV.

[0011] Furthermore, before mapping the spatiotemporal features to the threat prototype space, the method further includes: Acquire drone samples corresponding to multiple preset threat levels; Extract the spatiotemporal features of each UAV sample to obtain multiple sample feature vectors; For each preset threat level, the mean of the sample feature vectors belonging to the preset threat level is calculated, and the mean is used as the category prototype corresponding to the preset threat level; The threat prototype space is constructed using the category prototypes and the preset threat levels.

[0012] Secondly, embodiments of the present invention provide a drone threat assessment device, the device comprising: The detection module is used to detect target drones that pose an intrusion risk within a preset area; The acquisition module is used to acquire the threat elements of the target UAV and extract the spatiotemporal features of the threat elements through a time series mixer, wherein the threat elements include at least dynamic threat elements and static threat elements; The determination module is used to map the spatiotemporal features to a threat prototype space and determine the threat level of the target UAV based on the preset prototype classification within the threat prototype space.

[0013] Thirdly, embodiments of the present invention provide a computer device, including: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect or any corresponding embodiment thereof.

[0014] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions that cause a computer to perform the method described in the first aspect or any of its corresponding embodiments.

[0015] The method provided in this application has the following beneficial effects: The method provided in this application achieves accurate threat assessment of intruding drones under conditions of scarce samples and complex environments by sequentially executing target drone detection, spatiotemporal feature extraction, and prototype mapping classification. By detecting target drones with intrusion risk within a preset area, it can proactively lock onto the objects to be assessed from a complex airspace background, ensuring the targeting of the assessment process and avoiding wasted resources on misjudging non-threatening targets. By acquiring threat elements containing both dynamic and static dimensions and extracting their spatiotemporal features using a time-series mixer, it can simultaneously characterize the instantaneous state and inherent attributes of drones. Furthermore, the time-series mixer analyzes the complex patterns of multi-dimensional features in temporal evolution and spatial correlation, thereby forming a more comprehensive and in-depth representation of drone behavior. By mapping high-dimensional spatiotemporal features to a threat prototype space and determining the threat level based on the distance to a preset prototype, it leverages the advantages of prototype networks in few-sample learning. Efficient classification can be achieved with only a small number of prototypes for each category, overcoming the bottleneck of scarce historical attack data and improving the model's classification accuracy and real-time decision-making capabilities under limited data conditions. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the drone threat assessment method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of threat assessment elements according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatiotemporal feature extraction architecture according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the classification mechanism of the threat prototype space according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating another drone threat assessment method according to an embodiment of the present invention; Figure 6 This is a schematic diagram of a drone threat assessment architecture according to an embodiment of the present invention; Figure 7 This is a structural block diagram of a drone threat assessment device according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] According to embodiments of the present invention, a method, apparatus, computer device, and storage medium for assessing unmanned aerial vehicle (UAV) threats are provided. It should be noted that the steps shown in the flowcharts in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0020] This embodiment provides a method for assessing drone threats. Figure 1 This is a flowchart of a drone threat assessment method according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Detect target drones with intrusion risk within a preset area.

[0021] In this embodiment, drone detection equipment (such as radar and photoelectric sensors) deployed in critical infrastructure areas such as nuclear power plants, communication base stations, airports, or military bases continuously monitors pre-defined geographical areas or airspaces requiring key protection (i.e., preset areas). When a flight target matching the characteristics of a drone appears in the monitoring data, its flight behavior, trajectory characteristics, and preset risk rules (such as unauthorized intrusion, abnormal approach, evasive maneuvers, etc.) are analyzed and judged in real time. Target drones with intrusion risk are selected from all monitored targets, thus identifying the specific objects to be dealt with in subsequent assessment processes. Through dynamic monitoring and real-time assessment of intruding drones, reliable threat level information can be provided to the defense command center, thereby enabling reasonable resource allocation and defense decisions.

[0022] Step S102: Obtain the threat elements of the target UAV and extract the spatiotemporal characteristics of the threat elements through a time series mixer. The threat elements include at least dynamic threat elements and static threat elements.

[0023] In this embodiment of the application, acquiring the threat elements of the target drone includes: Step A1: Obtain monitoring data of the target drone within the preset area.

[0024] Specifically, after identifying a target drone posing an intrusion risk, multiple detection sensors deployed in a pre-defined area (such as radar, radio spectrum detection equipment, and photoelectric tracking systems) are continuously invoked to collect and aggregate the raw observation data stream of the specific target drone at fixed time intervals (e.g., multiple times per second), forming its monitoring data. This monitoring data consists of unprocessed, low-level signals and readings directly output by the sensors. Its content comprehensively covers multi-dimensional raw information such as the target drone's real-time spatial location, motion parameters, radar scattering characteristics, communication signal characteristics, and external morphology, forming the core data foundation for subsequent threat element extraction and quantitative analysis. Continuous, multi-dimensional, and synchronous data collection from identified threat targets ensures the real-time nature and completeness of the input information to the threat assessment model.

[0025] Step A2: Extract the situational threat elements and radar threat elements of the target UAV from the monitoring data, and treat the situational threat elements and radar threat elements as dynamic threat elements.

[0026] Specifically, such as Figure 2 As shown, the threat elements in each time segment consist of static and dynamic threat elements. Dynamic threat elements can be further divided into situational threat elements and radar threat elements, and their evaluation indicators include speed, heading angle, climb angle, altitude, radial range, horizontal range, azimuth angle, elevation angle, pulse width, operating frequency, repetition frequency, and stealth capability. Dynamic threat elements mainly reflect the UAV's motion characteristics, onboard radar status, and deployment in the actual scenario, providing important information for real-time prediction of its attack path and evasion capabilities.

[0027] The relationship between dynamic threat elements and threat levels is analyzed in the table below:

[0028] Step A3: Identify the physical attribute elements of the target UAV from the monitoring data and treat the physical attribute elements as static threat elements.

[0029] Specifically, static threat elements include at least the type of flying object, attack capabilities, and maneuverability, primarily used to describe the physical attributes of drones. By accurately extracting these elements and combining them with actual drone intrusion scenarios, a comprehensive assessment of the threat level of drones can be achieved, thus providing theoretical support for optimizing defense strategies. In practical applications of static data, discrete encoding processing can be performed for different threat levels.

[0030] The relationship between static threat elements and threat levels is analyzed in the table below:

[0031] By acquiring monitoring data, a real-time and original source of observational data is provided for the entire assessment process, ensuring the objectivity and completeness of the assessment basis. By extracting situational threat elements and radar threat elements as dynamic threat elements, the movement status, trajectory intentions and radar signal characteristics of UAVs in the airspace can be captured in real time, providing a direct basis for quantifying their instantaneous attack behavior and evasion capabilities. By identifying physical attribute elements as static threat elements, the inherent attack potential and physical limitations of UAVs can be effectively assessed, providing a stable reference for determining their basic threat level.

[0032] In this application embodiment, the situational threat elements include at least one of speed, heading angle, climb angle, altitude, radial distance, horizontal distance, azimuth angle, and elevation angle; the radar threat elements include at least one of pulse width, operating frequency, pulse repetition frequency, and radar cross-sectional area; and the physical attribute elements include at least one of aircraft type, attack capability, and maneuverability.

[0033] Specifically, situational threat elements are real-time status parameters extracted from the continuous movement trajectory and spatial relationship of the target UAV. Among them, speed refers to the displacement change per unit time; the faster the speed, the shorter the reaction time and the higher the threat. Heading angle refers to the angle between the UAV's flight direction and the direction pointing towards the defense point; flying towards the defense point poses a higher threat. Climb angle refers to the angle between the flight trajectory and the horizontal plane; a larger angle may indicate that it is rapidly approaching the target or performing evasive maneuvers. Altitude refers to the vertical distance of the UAV relative to the ground; low-altitude flight provides strong concealment and poses a higher threat, but too low an altitude limits maneuverability and may reduce the threat. Radial distance and horizontal distance refer to the straight-line distance between the UAV and the defense point and their projection distance on the horizontal plane, respectively; the closer they are, the higher the threat. Azimuth angle refers to the angle of the UAV on the horizontal plane relative to the sensor's reference direction; if it is in the sensor's blind spot or electronic interference area, the threat increases. Pitch angle refers to the angle between the sensor's line of sight and the horizontal plane, used to determine the target's altitude; a larger pitch angle usually indicates a higher target, and the threat may change accordingly.

[0034] Radar threat factors are characteristic parameters extracted from radar echo signals: pulse width refers to the duration of the radar pulse; a wider pulse width may indicate a longer detection time and a higher threat. Operating frequency refers to the center frequency of the radar signal; different frequency bands affect the detection and anti-jamming capabilities of UAVs, thus relating to the threat level. Pulse repetition frequency refers to the number of pulses emitted per unit time; a higher frequency is more suitable for tracking fast-moving targets and usually poses a higher threat. Radar cross-sectional area is a physical quantity that characterizes the target's ability to scatter radar waves; a smaller area indicates better stealth performance of the UAV and a higher threat.

[0035] Physical attributes are static parameters used to describe the inherent characteristics of UAVs: Object type refers to whether the target is a military-grade, commercial-grade UAV, or a non-UAV object, identified by signal characteristics or shape; different types correspond to different basic threat levels. Attack capability refers to the potential destructive power assessed based on its payload (such as cameras, projectiles, and electromagnetic interference equipment), categorized as strong, medium, or weak. Maneuverability refers to the flexibility of movement that a UAV can achieve based on its aerodynamic design and propulsion system, such as high-speed maneuvering, hovering, and sharp turns, categorized as high, medium, or low. These elements together constitute a multi-layered threat assessment system, from dynamic behavior and signal characteristics to static attributes.

[0036] Situational threat factors (such as speed and heading angle) transform the spatial motion state of the UAV into a series of calculable geometric and kinematic parameters, enabling precise quantification of the contribution of its flight intentions (such as approach and climb) to the threat level; radar threat factors (such as pulse width and radar cross-section area) transform the electromagnetic scattering and signal characteristics of the UAV into measurable physical parameters, enabling explicit modeling of the correlation between its stealth capability, detection strength, and threat level; physical attribute factors (such as flight object type and attack capability) transform the platform attributes of the UAV into discrete or hierarchical labels, enabling standardized processing of the basic threat differences between different models and payloads.

[0037] In this embodiment of the application, the spatiotemporal features of threat elements are extracted using a time series mixer, including: Step B1: Using the channel mixing module in the time series mixer, threat elements are mapped to a preset dimensional space to obtain spatial mixing features.

[0038] It should be noted that spatiotemporal feature extraction refers to extracting the temporal and spatial information features from real-time UAV threat element data, using a time-series mixer architecture, such as... Figure 3As shown, the system comprises a channel blending module and a temporal blending module. The channel blending module is responsible for feature interaction and enhancement of the input multi-dimensional threat elements. Its processing flow begins with a reshape operation on the input tensor to accommodate the computation of subsequent linear transformation layers. Then, an initial projection is performed through a linear transformation layer, feeding the features into a two-layer multilayer perceptron (MLP). The core of this MLP is a sequence structure containing linear transformation layers, ReLU activation functions, and linear transformation layers. Its function is to map the original threat elements to a higher-dimensional latent space through nonlinear transformations, achieving information fusion and enhancement between different feature channels. The processed features are finally restored to a temporal tensor format through a reshape operation, and the output is a spatial blending feature. The temporal blending module is responsible for modeling the long-range dependencies of the sequence data in the temporal dimension. Its input is the spatial blending feature output from the channel blending module. First, the tensor dimension is adjusted through a transpose operation to accommodate the computational requirements of the attention mechanism. Subsequently, the features are fed in parallel into three independent linear transformation layers to generate query (Q), key (K), and value (V) matrices, respectively. These three matrices are fed into the Scaled Dot-Product Attention mechanism. Attention weights are obtained by calculating the dot product of Q and K, scaling, and softmax normalization. These weights are then summed with respect to V to capture the correlation between features at different time steps. Multiple such attention heads are computed in parallel, and their results are aggregated through a concatenation operation. A final linear transformation layer is then applied for integration and dimensionality reduction, resulting in the output of Multi-Head Attention. This output is finally transposed to restore its dimensionality, generating temporal features representing complex temporal dependencies. The channel mixing and temporal mixing modules are sequentially connected to form a time-series mixer, ultimately outputting spatiotemporal features that simultaneously contain semantic information and long-range temporal patterns.

[0039] Specifically, the extracted structured threat elements (including dynamic and static threat elements, with dimension D) are used as input and fed into the channel mixing module of the time series mixer for processing. The core of this module is a two-layer multilayer perceptron (MLP), whose function is to perform nonlinear transformation and interactive integration on each channel (i.e., each threat element dimension) of the input features. This MLP first performs a linear transformation on the first layer (with parameters...). and The input threat elements are mapped from the original dimension D to a higher-dimensional intermediate latent space, then nonlinearity is introduced through the ReLU activation function; finally, a second linear transformation (with parameters...) is applied. and The latent space representation is further mapped to a specified preset dimension space (dimension H). This process can be represented as:

[0040]

[0041] in, For threat elements organized by batch (B) and time step (T), B is the batch size, T is the time step size set to 10 steps, and D is the feature dimension size for each time step, which is 15. This is the output of the multilayer perceptron; For activation functions; This is the weight matrix learned during the first-level linear transformation, used to map the input features from dimension D to the intermediate latent space; This is the bias vector learned during the first layer of linear transformation; Spatial blending features obtained after channel blending; The weight matrix learned during the second-level linear transformation is used to further map the latent space representation to a preset dimension H; This is the bias vector learned during the training in the second-level linear transformation.

[0042] By increasing the feature dimensions and fusing multi-source threat information, the model's ability to represent complex threat patterns is enhanced, providing a richer feature foundation for subsequent time modeling.

[0043] Step B2: Using the time mixing module in the time series mixer, establish the time dependency relationship between the dynamic features of the UAV in different time periods based on the spatial mixing characteristics.

[0044] In this embodiment of the application, the temporal dependency relationship between the dynamic characteristics of UAVs in different time periods is established based on spatial mixing characteristics, including: Step B201: Calculate the query matrix, key matrix, and value matrix corresponding to the spatial hybrid features using the pre-trained weight matrix.

[0045] Specifically, the obtained spatial mixing features (denoted as...) As input, the parameters are linearly transformed with the pre-trained weight matrix (i.e., the parameter matrix learned during model training). The weight matrices are: the weight matrix used to generate the query matrix (Q). The weight matrix used to generate the key matrix (K) and the weight matrix used to generate the value matrix (V). The calculation process is as follows:

[0046] in, , and It is the weight matrix obtained through training, representing the linear transformation of query, key, and value; For query matrix; The key matrix; It is a value matrix.

[0047] The query matrix (Q) carries the information to be determined at the current time step, the key matrix (K) stores the identification information of the features at each time step, and the value matrix (V) contains the actual feature content at each time step. Linear projection transforms the original spatial mixture features into a structured representation suitable for subsequent attention calculations, thus laying the foundation for modeling the relationships between features at different time steps.

[0048] Step B202: Calculate the attention weights between features at different time steps based on the query matrix and the key matrix.

[0049] Specifically, after obtaining the query matrix (Q) and key matrix (K), matrix operations are used to quantify the correlation strength between features at different time steps: first, the transpose of the query matrix and the key matrix is ​​calculated (Q). The dot product between the query vector and all key vectors is used to obtain the original matching score matrix for each query vector and all key vectors; then, this score matrix is ​​scaled, i.e., divided by . (in (Using the feature dimension as a reference), a scaled score matrix is ​​obtained to prevent excessively large inner product values ​​from causing gradient instability. Finally, the scaled score matrix is ​​normalized using a softmax function, transforming all scores corresponding to each query into a probability distribution. This probability distribution represents the attention weights, with each element explicitly characterizing the degree to which the current time step should focus on features from any other time step in the history when fusing information, thereby dynamically establishing dependencies across time steps.

[0050] Step B203: The value matrix is ​​weighted and fused based on attention weights to obtain the fused temporal features.

[0051] Specifically, after calculating the attention weight matrix, it is multiplied with the value matrix (V) to achieve selective integration of information along the time dimension. The attention weight matrix, normalized by Softmax (whose elements represent the degree of attention that should be given at each time step), is multiplied by the value matrix (containing the feature content of each time step), i.e.:

[0052] in, It is a scaled dot product attention, and the softmax operation is used to compute the weights; This is the attention weight matrix; The final output of the self-attention mechanism is the fused temporal feature.

[0053] Attention weights are used to sum the feature vectors of all time steps in the value matrix, generating a new feature representation for each time step. This new feature representation is the fused temporal feature, which is no longer isolated information from a single time step, but integrates the features of all relevant time steps in the historical time period and dynamically weights them according to their importance, thereby more effectively capturing and strengthening key patterns and long-range dependencies in the time series.

[0054] Step B204 involves multi-head stitching of the fused temporal features to obtain the temporal dependencies between the dynamic features of the UAV in different time periods.

[0055] Specifically, in the multi-head attention mechanism, the fused temporal features corresponding to N different attention heads are independently calculated and obtained. First, through a concatenation operation, the output features of these N heads are connected along the feature dimension (i.e., the H dimension) to form a composite feature tensor that aggregates information from multiple subspaces. Subsequently, this composite feature tensor is passed through a linear transformation layer (the final linear transformation matrix). The projection and dimensionality reduction are performed, and the final output is the integrated feature representation, which is the time dependency:

[0056] in, The output features of the multi-head attention mechanism; Concat represents concatenating N independent attention heads along the feature dimension. The output vectors of the ) are concatenated to integrate attention information from different subspaces; It is a learnable weight matrix used to perform linear transformation and integration on the concatenated features. Its function is to map the high-dimensional concatenated features back to the target output dimension, forming a unified time-dependent representation rich in multi-view information.

[0057] This temporal dependency integrates the complex dependency patterns among the dynamic features of drones across multiple time periods captured by different attention heads from diverse subspaces and interactive perspectives, thus providing sufficient temporal contextual information for subsequent threat level determination.

[0058] Simply put, the temporal fusion module is composed of multiple self-attention layers stacked together, with the output of each layer serving as the input to the next. Each layer further fuses the relationships between time steps to capture more complex temporal dependencies.

[0059]

[0060] in, This represents the output feature at time step t in the i-th layer attention mechanism; Indicates that it comes from the previous level i 1. In the previous moment t 1. Input features.

[0061] The layer-by-layer propagation process of the multi-head attention mechanism in the temporal fusion module: Each layer takes the output of the previous layer as input, and further integrates and refines the dependencies between time steps through the multi-head self-attention mechanism, thereby gradually enhancing the model's ability to model long-range nonlinear patterns in time series.

[0062] By utilizing pre-trained weight matrices to compute query, key, and value matrices, spatial hybrid features are transformed into standardized representations suitable for attention computation, laying the foundation for feature interaction. By computing attention weights, the correlation strength between features at any two time steps can be dynamically and quantitatively evaluated, enabling soft selection and focusing on long-range dependencies. Through weighted fusion of value matrices based on attention weights, key information from historical time steps is selectively integrated, generating context-rich temporal features. Finally, by multi-head concatenation of the fused features, diverse dependency patterns captured in multiple subspaces (attention heads) are integrated, forming comprehensive temporal dependencies. This series of steps collectively achieves accurate modeling of nonlinear, long-range temporal patterns.

[0063] Step B3: Spatial mixing features and temporal dependencies are used as spatiotemporal features.

[0064] Specifically, after generating spatial hybrid features (representing the fusion and enhancement results of multi-dimensional threat elements in a high-dimensional feature space) via the channel hybrid module and temporal dependencies (representing long-term nonlinear correlation patterns between UAV dynamic features at different time periods) via the temporal hybrid module, these two outputs are combined to form the final feature representation used for subsequent threat level determination, namely, spatiotemporal features. Spatial hybrid features provide semantically rich static and dynamic attribute information that has undergone deep nonlinear transformation at each time step; temporal dependencies provide context-aware dynamic evolution information across consecutive time steps. The combination of the two allows the spatiotemporal features to simultaneously contain the richness of threat elements in space (feature dimension) and the continuity in time (sequence evolution), thus providing a comprehensive and structured high-level input for subsequent prototype spatial mapping.

[0065] By mapping threat elements to a preset dimensional space through the channel fusion module, nonlinear transformation and deep fusion of multi-source heterogeneous features are achieved, enhancing the discriminative power and representational richness of features in high-dimensional space. By establishing temporal dependencies between features in different time periods through the temporal fusion module, the long-range correlation and evolutionary patterns of UAV behavior in the temporal dimension (such as continuous maneuvering and cooperative intent) can be actively captured, overcoming the dependence of traditional methods on short-term local features. By using spatial fusion features and temporal dependencies together as spatiotemporal features, the output features are ensured to contain both enhanced semantic information and a temporal context that describes dynamic evolution, providing a comprehensive and structured high-level input for subsequent prototype classification.

[0066] Step S103: Map the spatiotemporal features to the threat prototype space, and determine the threat level of the target UAV based on the preset prototype classification within the threat prototype space.

[0067] In this embodiment of the application, spatiotemporal features are mapped to a threat prototype space, and the threat level of the target UAV is determined based on a preset prototype classification within the threat prototype space, including: Step C1: Obtain the preset prototype classifications within the threat prototype space.

[0068] Specifically, before determining the real-time threat level, it is necessary to retrieve predefined prototype classifications from the constructed and stored threat prototype space. This threat prototype space is pre-constructed during the model training phase and contains prototype classifications (i.e., category prototype vectors) that strictly correspond to different threat levels (e.g., high, medium, and low threats) and serve as typical representatives of each category. Each prototype classification (e.g., The similarity vector is obtained by calculating the mean of the spatiotemporal feature vectors of all training samples under that threat level. Essentially, it represents the center or anchor point of that threat pattern in the feature space. Therefore, the specific implementation method is as follows: during the initialization of the evaluation process, the prototype vector set bound to a fixed threat level is directly loaded to provide a benchmark for subsequent real-time calculation of the similarity between query samples and each category.

[0069] Step C2: Construct query samples using spatiotemporal features and calculate the similarity between the query samples and each prototype classification.

[0070] Specifically, the spatiotemporal features (i.e., the final feature vector of the target UAV to be evaluated) are directly used as query samples (denoted as...). Then, in the feature space, a distance metric is used to quantitatively evaluate the relationship between the query sample and each acquired prototype classification (i.e., the prototype vector corresponding to each threat level). The similarity is calculated as the degree of closeness between the query sample and each prototype vector. In practice, similarity is quantified by calculating the Euclidean distance between the query sample and each prototype vector; a smaller distance value indicates higher similarity. The distance calculation formula is:

[0071] in, For query samples; Classify the c-th prototype; For query samples With the c-th prototype classification The Euclidean distance between them is used to quantify the similarity between the two.

[0072] By systematically calculating distances, a quantitative similarity score is generated for each candidate threat level, thus providing a direct basis for determining the final threat level in the next step.

[0073] Step C3: The prototype of the category with the lowest similarity is taken as the target category prototype to which the query sample belongs.

[0074] Specifically, after calculating the Euclidean distance (i.e., the similarity metric) between the query sample and all prototype categories, the query sample is compared with each prototype category. Distance value The process involves identifying the category prototype corresponding to the smallest distance (i.e., the highest similarity). This prototype is then determined as the most likely category to which the current query sample belongs, and is called the target category prototype. The formula for determining the target category prototype is as follows:

[0075] in, For query samples With the c-th prototype classification The Euclidean distance between them is used to quantify the similarity between the two. For query samples Predicted threat categories; Indicates selecting the distance The category index c that yields the minimum value is used. This decision-making process is completed by a function, whose output is the prototype vector of the preset threat level that is closest to the features of the query sample.

[0076] It should be noted that, Figure 4 This is a schematic diagram illustrating the core classification mechanism of the threat prototype space (i.e., prototype network) in few-sample threat assessment, as shown below. Figure 4As shown, the threat prototype space is divided into three regions, corresponding to three preset threat levels: low, medium, and high. Each level is represented by the position of its category prototype (c1, c2, c3) in the feature space. Query Sample (i.e., the spatiotemporal feature vector of the current UAV) is placed in this space, and its distance to each prototype is measured by Euclidean distance (usually represented by lines or contour lines in the diagram). Based on the nearest neighbor principle, the query sample is classified into the threat level corresponding to the prototype with the smallest distance (e.g., ...). (Classified as threat in c2). By performing distance comparisons in a space composed of a small number of typical samples (prototypes), this method achieves efficient and interpretable threat level determination, thus highlighting its effectiveness and practicality under conditions of scarce data.

[0077] Step C4: Based on the mapping relationship between prototype classification and preset threat level, determine the target threat level corresponding to the target category prototype as the threat level of the target UAV.

[0078] Specifically, after determining the target category prototype, a mapping lookup operation is performed based on the fixed binding relationship between each prototype category (i.e., the category prototype vector) and the preset threat level (e.g., high, medium, low threat), which has been pre-established and stored during the model building phase. This mapping relationship is determined when the threat prototype space is constructed (e.g., prototype...). Corresponding to "high threat", the prototype Corresponding to the "medium threat", the prototype (Corresponding to "low threat"). By identifying the specific preset threat level associated with the target category prototype in the mapping relationship and directly outputting it as the target threat level, the final qualitative judgment of the threat level of the target UAV is completed. The similarity comparison results in the feature space are transformed into intuitive and actionable threat level information for subsequent defense decisions. Based on the nearest neighbor principle, distance metrics in the feature space are used to achieve efficient and intuitive classification and adjudication under conditions of few samples.

[0079] By acquiring pre-defined prototype classifications, a pre-learned and representative category reference benchmark is provided for real-time classification, which is a prerequisite for few-shot learning. By constructing query samples and calculating their similarity to each prototype, the classification problem is transformed into a distance metric problem in the feature space, achieving complete quantification and automation of classification decisions. By determining the prototype with the lowest similarity as the target category prototype, an intuitive and fast classification decision is achieved based on the nearest neighbor principle, with a simple and efficient algorithm. By determining the target threat level based on mapping relationships, the internal feature space distance results are mapped to the final understandable threat level output, completing the closed loop from data to decision. These steps together ensure the feasibility of the few-shot evaluation process and the credibility of the results.

[0080] In this embodiment of the application, before mapping the spatiotemporal features to the threat prototype space, such as Figure 5 As shown, the method also includes: Step S201: Obtain drone samples corresponding to multiple preset threat levels.

[0081] In this embodiment, during the offline training or initialization phase of constructing the threat prototype space, a batch of drone instance data with known threat levels is collected and organized from historical monitoring records, simulation data, or real-world scenario data annotated by experts. This data must cover all preset threat levels (e.g., high, medium, and low threats), and each level must contain a certain number of drone samples. Each sample consists of the drone's original time-series data (including dynamic and static features) within a complete monitoring period and its annotated real threat level label. This provides a sufficient and balanced foundation of representative category data for subsequent prototype calculations, ensuring that each threat level can be effectively represented by its typical samples, thereby supporting the prototype network to achieve reliable classification under conditions of few samples.

[0082] Step S202: Extract the spatiotemporal features of each UAV sample to obtain multiple sample feature vectors.

[0083] In this embodiment, for each acquired UAV sample (i.e., complete time-series data with known threat level labels), a feature extraction operation consistent with the real-time assessment process is performed. First, the raw monitoring data of the sample is preprocessed and threat element transformed to obtain its structured dynamic and static threat element sequences. Then, these time-series threat elements are input into a defined time-series mixer (containing a channel mixing module and a time mixing module). Channel mixing is performed sequentially to improve feature dimensionality and representational capability, and time mixing is performed to model the temporal dependencies within the sequence. Finally, the fused features output by the time-series mixer are the spatiotemporal features of the sample. After this processing, each sample generates a high-level feature representation with fixed dimensions that integrates its multi-dimensional attributes and temporal evolution patterns; this representation is the sample feature vector. By batch processing all samples, multiple sample feature vector sets corresponding to each preset threat level are obtained, providing input for subsequent calculation of prototype vectors for each category.

[0084] Step S203: For each preset threat level, calculate the mean of the sample feature vectors belonging to the preset threat level, and use the mean as the category prototype corresponding to the preset threat level.

[0085] In this embodiment, after obtaining the sample feature vectors corresponding to all threat levels, the samples are grouped according to threat level. For each preset threat level (e.g., high, medium, low threat), all feature vectors with tags belonging to that level are selected from multiple sample feature vectors to form a sample set for that category. Subsequently, the mean of all feature vectors in this sample set is calculated, that is, the element-wise arithmetic mean of these vectors is taken. The calculation formula is:

[0086] in, Represents the category prototype of category c; This represents the number of samples belonging to category c; This represents the set of all samples belonging to category c; To represent the sample set The feature vector of the i-th sample.

[0087] The calculated mean vector is the unique category prototype (i.e., prototype vector) corresponding to the preset threat level. This category prototype represents the central or typical feature pattern of the corresponding threat level in the feature space. By aggregating the common information of similar samples, it summarizes the essential attributes of this type of threat, providing a stable and representative reference benchmark for subsequent classification by distance comparison under conditions of few samples.

[0088] Step S204: Construct a threat prototype space using category prototypes and preset threat levels.

[0089] In this embodiment, after calculating the category prototypes corresponding to all preset threat levels, a stable mapping relationship is established between these category prototypes (i.e., the feature space center vector of each threat level) and their corresponding preset threat levels (such as high, medium, and low threats). This mapping relationship is organized and stored in the form of structured data (e.g., a lookup table or vector set containing prototype vectors and their level labels). The resulting complete data structure is the threat prototype space, which essentially encapsulates the pattern knowledge (represented by prototype vectors) and their semantic labels (threat levels) of different threat levels in a unified and efficiently accessible reference framework. Once this space is constructed, it can be directly invoked during the real-time evaluation phase, providing a benchmark reference system for similarity comparison and classification decisions for query samples, thereby supporting the entire system to achieve accurate threat level determination under conditions of few samples.

[0090] By acquiring drone samples corresponding to multiple preset threat levels, necessary and balanced labeled data was collected for prototype learning, ensuring the representativeness of the prototype. By extracting the spatiotemporal features of each sample, the raw data was transformed into a high-level feature representation suitable for metric learning, unifying the data format. The mean of the feature vectors of each threat level sample was calculated as a category prototype, and typical features for each threat level were constructed using the concept of class centers as the core of the prototype network. A structured knowledge base was formed by constructing a threat prototype space using category prototypes and threat levels. This pre-construction process is the fundamental guarantee for effective execution in subsequent real-time evaluation, improving the practicality and deployability of the entire system.

[0091] This embodiment provides a drone threat assessment architecture, such as Figure 6 As shown, the architecture includes: first, detecting and collecting data on intruding threat targets, and transforming their raw information into physical features (static threat elements), situational features, and radar features (both constituting dynamic threat elements); these multi-source heterogeneous threat elements are then input into a time-series mixer, which, through its channel mixing and time mixing modules, performs deep nonlinear transformations and cross-time-step dependency modeling on the input features, thereby extracting high-level spatiotemporal features that comprehensively characterize the target's behavioral patterns; these spatiotemporal features then enter a prototype network, which calculates the Euclidean distance between the target and pre-defined threat level categories (high, medium, and low threats) in the threat prototype space, and classifies the target to the threat level corresponding to the prototype with the smallest distance based on the nearest neighbor principle; finally, the architecture outputs a clear threat assessment result, i.e., low threat, medium threat, or high threat level, providing a direct basis for subsequent defense decisions. This architecture, through its end-to-end processing flow from raw data to threat level decision-making, highlights the effectiveness and integrity of combining a time-series mixer and a prototype network in low-sample, real-time scenarios.

[0092] This embodiment also provides a drone threat assessment device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0093] This embodiment provides a drone threat assessment device, such as... Figure 7 As shown, it includes: Detection module 71 is used to detect target drones with intrusion risk within a preset area; The acquisition module 72 is used to acquire the threat elements of the target UAV and extract the spatiotemporal characteristics of the threat elements through a time series mixer. The threat elements include at least dynamic threat elements and static threat elements. The determination module 73 is used to map spatiotemporal features to a threat prototype space and determine the threat level of the target UAV based on the preset prototype classification within the threat prototype space.

[0094] In this embodiment of the application, the acquisition module 72 is specifically used to acquire monitoring data of the target UAV within a preset area; extract the situational threat elements and radar threat elements of the target UAV from the monitoring data, and use the situational threat elements and radar threat elements as dynamic threat elements; identify the physical attribute elements of the target UAV from the monitoring data, and use the physical attribute elements as static threat elements.

[0095] In this embodiment of the application, the acquisition module 72 is specifically used to map the threat elements to a preset dimension space through the channel mixing module in the time series mixer to obtain spatial mixing features; to establish the time dependency relationship between the dynamic features of UAVs in different time periods based on the spatial mixing features through the time mixing module in the time series mixer; and to use the spatial mixing features and the time dependency relationship as spatiotemporal features.

[0096] In this embodiment, the acquisition module 72 is specifically used to calculate the query matrix, key matrix, and value matrix corresponding to the spatial hybrid features using the pre-trained weight matrix; calculate the attention weights between features at different time steps based on the query matrix and key matrix; perform weighted fusion of the value matrix based on the attention weights to obtain the fused time features; and perform multi-head splicing on the fused time features to obtain the time dependency relationship between the UAV dynamic features at different time periods.

[0097] In this embodiment of the application, the determining module 73 is specifically used to obtain the preset prototype classification within the threat prototype space; construct a query sample using spatiotemporal features and calculate the similarity between the query sample and each prototype classification; take the prototype of the category with the lowest similarity as the target category prototype to which the query sample belongs; and determine the target threat level corresponding to the target category prototype as the threat level of the target UAV based on the mapping relationship between the prototype classification and the preset threat level.

[0098] In this embodiment of the application, the device further includes: a construction module, configured to acquire multiple drone samples corresponding to multiple preset threat levels; extract the spatiotemporal features of each drone sample to obtain multiple sample feature vectors; for each preset threat level, calculate the mean of the sample feature vectors belonging to the preset threat level, and use the mean as the category prototype corresponding to the preset threat level; and construct a threat prototype space using the category prototype and the preset threat level.

[0099] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 8 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system).

[0100] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0101] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0102] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device as shown by a landing page for an app. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0103] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0104] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0105] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0106] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for assessing the threat of unmanned aerial vehicles (UAVs), characterized in that, The method includes: Detect target drones that pose an intrusion risk within a pre-defined area; The threat elements of the target drone are acquired, and the spatiotemporal features of the threat elements are extracted using a time series mixer. The threat elements include at least dynamic threat elements and static threat elements. The spatiotemporal features are mapped to a threat prototype space, and the threat level of the target UAV is determined based on the preset prototype classification within the threat prototype space.

2. The method according to claim 1, characterized in that, The acquisition of threat elements of the target drone includes: Acquire monitoring data of the target drone within the preset area; The situational threat elements and radar threat elements of the target UAV are extracted from the monitoring data, and the situational threat elements and radar threat elements are used as the dynamic threat elements. The physical attribute elements of the target UAV are identified from the monitoring data, and these physical attribute elements are used as the static threat elements.

3. The method according to claim 2, characterized in that, The situational threat elements include at least one of speed, heading angle, climb angle, altitude, radial distance, horizontal distance, azimuth angle, and elevation angle; the radar threat elements include at least one of pulse width, operating frequency, pulse repetition frequency, and radar cross-sectional area; the physical attribute elements include at least one of aircraft type, attack capability, and maneuverability.

4. The method according to claim 1, characterized in that, The extraction of the spatiotemporal features of the threat elements using a time-series mixer includes: The threat elements are mapped to a preset dimensional space through the channel mixing module in the time series mixer to obtain spatial mixing features; The time-series mixer uses a time-mixing module to establish the time dependency relationship between the dynamic features of UAVs in different time periods based on the spatial mixing characteristics. The spatial mixing features and the temporal dependencies are used as the spatiotemporal features.

5. The method according to claim 4, characterized in that, The step of establishing the temporal dependency relationship between the dynamic features of UAVs in different time periods based on the spatial mixing features includes: The query matrix, key matrix, and value matrix corresponding to the spatial hybrid features are calculated using the pre-trained weight matrix; Based on the query matrix and the key matrix, calculate the attention weights between features at different time steps; The value matrix is ​​weighted and fused based on the attention weights to obtain the fused temporal features; The fused temporal features are spliced ​​together to obtain the temporal dependencies between the dynamic features of UAVs in different time periods.

6. The method according to claim 1, characterized in that, The step of mapping the spatiotemporal features to a threat prototype space and determining the threat level of the target UAV based on a preset prototype classification within the threat prototype space includes: Obtain the preset prototype classification within the threat prototype space; Query samples are constructed using the spatiotemporal features, and the similarity between the query samples and each of the prototype classifications is calculated. The category prototype with the lowest similarity is taken as the target category prototype to which the query sample belongs; Based on the mapping relationship between the prototype classification and the preset threat level, the target threat level corresponding to the target category prototype is determined as the threat level of the target UAV.

7. The method according to claim 1, characterized in that, Before mapping the spatiotemporal features to the threat prototype space, the method further includes: Acquire drone samples corresponding to multiple preset threat levels; Extract the spatiotemporal features of each UAV sample to obtain multiple sample feature vectors; For each preset threat level, the mean of the sample feature vectors belonging to the preset threat level is calculated, and the mean is used as the category prototype corresponding to the preset threat level; The threat prototype space is constructed using the category prototypes and the preset threat levels.

8. A drone threat assessment device, characterized in that, The device includes: The detection module is used to detect target drones that pose an intrusion risk within a preset area; The acquisition module is used to acquire the threat elements of the target UAV and extract the spatiotemporal features of the threat elements through a time series mixer, wherein the threat elements include at least dynamic threat elements and static threat elements; The determination module is used to map the spatiotemporal features to a threat prototype space and determine the threat level of the target UAV based on the preset prototype classification within the threat prototype space.

9. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

Images