Fixed-wing unmanned aerial vehicle flight path anomaly detection method and device, and medium

Abstract

The invention relates to a fixed-wing unmanned aerial vehicle track anomaly detection method and device and a medium, and the method comprises the steps: building a radar detection noise model based on a radar equation and a signal-to-noise ratio formula; on the basis of the motion equation, simulating the motion process in normal and abnormal states to generate a real track, and adding noise to the real track by using a radar detection noise model to obtain noise-containing simulation track data; constructing a dynamic graph based on the simulated track data; inputting the dynamic graph into a space-time diagram neural network model; the time-space diagram neural network model comprises a diagram attention network layer, a long-short term memory network layer and a linear layer which are connected in sequence; and performing feature extraction and classification on the dynamic graph through a time-space diagram neural network model, and outputting a track anomaly detection result. Therefore, precise detection of tiny track abnormalities of the fixed-wing unmanned aerial vehicle, especially gliding faults and the like, is realized.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to a method, device, and medium for detecting abnormal flight paths of fixed-wing UAVs. Background Technology

[0002] Fixed-wing unmanned aerial vehicles (UAVs) are widely used in surveying and exploration, power line inspection, emergency rescue, and logistics transportation due to their advantages such as long endurance, high payload efficiency, and good flight stability. With the rapid development of the low-altitude economy and unmanned systems, UAV operational safety has become a crucial aspect of low-altitude public safety management. If a UAV experiences an anomaly or malfunction during flight, it may not only lead to mission interruption and equipment damage but also cause a crash, posing a serious threat to ground personnel, buildings, and transportation facilities, especially in sensitive areas such as urban low-altitude zones and airport airspace.

[0003] Unlike manned aircraft, drones typically operate in low-altitude environments, making them susceptible to external interference such as complex terrain and sudden weather changes. Their autonomous control systems, power systems, and sensors are also vulnerable to electromagnetic interference and equipment aging, potentially leading to power failures, navigation drift, and control interruptions. Because drones lack real-time pilot intervention, initial signs of anomalies are often subtle, and they may even remain in a "passive gliding" state due to inertia, exhibiting flight characteristics that appear normal but are actually abnormal. This makes early identification and warning difficult using traditional manual observation or simple data analysis.

[0004] Currently, anomaly detection in UAV flight paths mainly relies on equipment such as radar and ADS-B to acquire data on position, speed, and heading, and employs traditional methods based on threshold judgment, statistical analysis, or rule-based reasoning for anomaly identification. However, these methods have significant shortcomings: Firstly, UAV flight paths exhibit spatiotemporal coupling characteristics of "short-period fluctuations and long-period trends." Traditional methods often process time or spatial dimension information separately, making it difficult to effectively capture spatiotemporal correlations, resulting in insufficient ability to identify subtle anomalies such as "gliding failures." Secondly, existing methods are mostly designed based on the flight path characteristics of manned aircraft, failing to fully consider the characteristics of UAVs, such as small size, low speed, and high maneuverability, leading to poor model adaptability and high false alarm rates.

[0005] Furthermore, existing technologies for radar detection noise modeling often employ Gaussian noise with a fixed standard deviation, failing to consider the actual impact of range attenuation and velocity-related noise, leading to discrepancies between simulation data and the real detection environment. In terms of graph structure construction, graph connections are typically established solely based on temporal adjacency, lacking physical rule guidance and key feature focus, thus limiting the model's ability to distinguish abnormal patterns. Regarding neural network structure design, most methods do not fully utilize the kinematic and aerodynamic characteristics of track data, and the loss function often only focuses on classification accuracy, lacking sufficient sensitivity to key fault features such as velocity anomalies and acceleration abrupt changes.

[0006] Therefore, existing technologies are insufficient to meet the requirements of accuracy, real-time performance, and high reliability for anomaly detection of UAV flight paths in low-altitude safety management. There is an urgent need for an anomaly detection method that can deeply integrate spatiotemporal information, embed prior physical knowledge, and adapt to the flight characteristics of UAVs, so as to achieve early detection and warning of various anomalies, especially the highly concealed "gliding faults". Summary of the Invention

[0007] This invention provides a method, device, and medium for detecting anomalies in the flight paths of fixed-wing UAVs. Its purpose is to solve the technical problems of existing detection methods, which suffer from weak ability to identify subtle anomalies such as "gliding failures," high false alarm rates, and insufficient real-time performance due to insufficient extraction of spatiotemporal features, lack of physical constraints, and inaccurate noise modeling.

[0008] To achieve the above objectives, the first aspect of the present invention provides a method for detecting anomalies in the flight path of a fixed-wing unmanned aerial vehicle, comprising the following steps: Based on the radar equation and signal-to-noise ratio formula, a radar detection noise model related to the target distance and speed of the UAV is established; Based on the motion equations of a fixed-wing UAV, its motion process under normal and abnormal conditions is simulated to generate a real track. Then, noise is added to the real track using the radar detection noise model to obtain noisy simulated track data. Based on the noisy simulated flight path data, a dynamic graph is constructed; wherein, the node features of the dynamic graph include position, velocity, acceleration and aerodynamic features, and the edge connections of the dynamic graph include temporal adjacency edges, physical constraint edges and dynamic attention edges; The dynamic graph is input into a spatiotemporal graph neural network model, which extracts and classifies features from the dynamic graph and outputs anomaly detection results. The spatiotemporal graph neural network model includes a graph attention network layer, a long short-term memory network layer, and a linear layer connected in sequence.

[0009] Furthermore, based on radar equations and signal-to-noise ratio formulas, methods for establishing radar detection noise models related to UAV target distance and velocity include: Calculate the distance between the UAV target and the radar based on the radar equations; Calculate the standard deviation of distance-related noise associated with the distance, wherein the standard deviation of distance-related noise is proportional to the square of the distance; Calculate the speed of the drone target; Calculate the speed-related noise standard deviation associated with the speed, wherein the speed-related noise standard deviation is proportional to the speed; The total noise standard deviation is obtained by summing the distance-related noise standard deviation and the velocity-related noise standard deviation by weighted square, thereby establishing the radar detection noise model.

[0010] Furthermore, the method for adding noise to the real trajectory using the radar detection noise model to obtain noisy simulated trajectory data includes: The actual flight track is obtained, which contains multiple trajectory points arranged in chronological order, and each trajectory point contains a timestamp and three-dimensional spatial coordinates; For each current trajectory point in the actual flight path, perform the following steps: Calculate the distance between the UAV and the radar corresponding to the current trajectory point; Calculate the speed of the drone corresponding to the current trajectory point; Based on the radar detection noise model, the total noise standard deviation is calculated based on the distance and the speed. Three-dimensional Gaussian noise is generated based on the total noise standard deviation; The three-dimensional Gaussian noise is added to the three-dimensional spatial coordinates of the current trajectory point to obtain a trajectory point containing noise; All noisy trajectory points are combined in chronological order to obtain the noisy simulated flight track data.

[0011] Furthermore, constructing the dynamic graph includes the node feature calculation step: For each track point in the noisy simulated track data, calculate its position, velocity, and acceleration; Based on the calculated velocity and acceleration, calculate the curvature, heading angle, and jerk. Obtain the altitude information of the waypoints; Based on the aerodynamic parameters and motion state of the fixed-wing UAV, the aerodynamic characteristics are calculated, including angle of attack, airspeed, lift-to-drag ratio and energy altitude. The calculated position, velocity, acceleration, curvature, heading angle, jerk, altitude, angle of attack, airspeed, lift-to-drag ratio, and energy altitude are combined to form the nodal feature vector of the waypoint.

[0012] Furthermore, the steps for constructing edge connections in a dynamic graph include: Construct temporal adjacency edges: connect adjacent nodes of each track point in the time series; Constructing physical constraint edges: For each track point, predict its expected position at the next moment based on Newtonian kinematics. If the deviation between the actual position of its subsequent track points and the expected position exceeds a preset threshold, then establish a connection between the current node and the node corresponding to the actual position. Constructing dynamic attention edges: Input all nodes of the dynamic graph into the graph attention network, calculate the attention coefficient between any two nodes, and filter and establish connections based on the attention coefficient.

[0013] Furthermore, in the step of constructing dynamic attention edges, the construction of the dynamic attention edges includes: Calculate the attention coefficients between all node pairs using a graph attention network; Based on the calculated attention coefficient, select and retain the node pairs whose attention weight ranks in the top 30%; An edge connection is established between the preserved node pairs as the dynamic attention edge.

[0014] Furthermore, the method of inputting the dynamic graph into a spatiotemporal graph neural network model, extracting features and classifying the dynamic graph through the spatiotemporal graph neural network model, and outputting track anomaly detection results includes: The dynamic graph is input into the first graph attention network layer of the spatiotemporal graph neural network model to extract spatial features from the nodes and edges of the dynamic graph. The extracted spatial features are input into the second graph attention network layer of the spatiotemporal graph neural network model for deep extraction and fusion of spatial features. The features output from the second graph attention network layer are input into the bidirectional long short-term memory network layer of the spatiotemporal graph neural network model to capture the evolution pattern of the trajectory data in the time series. The temporal features output by the bidirectional long short-term memory network layer are sequentially input into the first linear layer and the second linear layer of the spatiotemporal graph neural network model for physical constraint verification and classification. Based on the output of the second linear layer, track anomaly detection results are generated and output.

[0015] Furthermore, the method also includes training the spatiotemporal graph neural network model, wherein the training of the spatiotemporal graph neural network model employs a composite loss function, which consists of the following parts: The first part is the cross-entropy loss between the predicted category and the true label; The second part is the penalty for the difference between the predicted waypoint speed and the actual speed; The third part is the penalty for sudden changes in acceleration at the waypoint.

[0016] To achieve the above objectives, a second aspect of the present invention provides an electronic device including a memory and a processor, the memory being used to store a program that supports the processor in executing the fixed-wing unmanned aerial vehicle trajectory anomaly detection method, and the processor being configured to execute the program stored in the memory.

[0017] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the fixed-wing unmanned aerial vehicle (UAV) trajectory anomaly detection method.

[0018] The beneficial effects of this invention are: Compared with existing technologies, the present invention provides a method, device, and medium for detecting anomalies in the flight paths of fixed-wing UAVs. By constructing a Gaussian radar noise model with a distance-velocity dual-factor correlation, it significantly improves the realism and fidelity of simulated flight path data. Furthermore, in the dynamic graph construction stage, it integrates a three-layer connection strategy of temporal adjacency edges, physical constraint edges, and dynamic attention edges, and designs node representations that include position, velocity, acceleration, and aerodynamic features. This not only strengthens the modeling of the temporal continuity of flight path data, but also achieves efficient differentiation between "physically feasible normal fluctuations" and "abnormal deviations caused by faults" by introducing Newtonian kinematics priors and adaptive attention mechanisms. In particular, it enhances the detection of subtle and hidden anomaly patterns such as "gliding faults." Finally, a spatiotemporal graph neural network (STGNN) model was designed, which uses a two-layer graph attention network (GAT) for deep extraction of spatial features, a bidirectional long short-term memory network (LSTM) to capture temporal evolution patterns, and a linear layer with physical constraint verification for classification. This model is further optimized by combining a composite loss function that integrates classification loss, velocity constraint loss, and acceleration mutation penalty term. This enables the model to maintain high classification accuracy while being more sensitive to key fault features such as abnormal velocity fluctuations and acceleration mutations. As a result, the model achieves more accurate, real-time, and lower false alarm detection and early warning of fixed-wing UAV trajectory anomalies, especially "passive gliding" anomalies that are difficult to identify by traditional methods. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0020] Figure 1 This is a flowchart of a method for detecting abnormal flight paths of a fixed-wing unmanned aerial vehicle (UAV) according to an embodiment of the present invention.

[0021] Figure 2This is a spatiotemporal graph neural network (STGNN) model architecture diagram disclosed in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] According to embodiments of the present invention, it should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the following manufacturing method, in some cases the steps shown or described may be performed in a different order than that shown here.

[0024] like Figure 1 As shown, the present invention provides a method for detecting anomalies in the flight path of a fixed-wing unmanned aerial vehicle (UAV), comprising the following steps: Step S100: Based on the radar equation and signal-to-noise ratio formula, establish a radar detection noise model related to the target distance and speed of the UAV; In step S100, firstly, based on the basic physical principles of radar detection, the signal power received by the radar is... Modeled as a function related to parameters such as target range and radar cross-section (RCS), its radar equation is:

[0025] in, For transmission power, and These are the transmit and receive antenna gains, respectively. For radar wavelength, For the target radar cross-section, The distance between the target and the radar. This refers to the system loss factor (including atmospheric attenuation, hardware loss, etc.).

[0026] In practical detection, the key indicator for measuring radar performance is the signal-to-noise ratio (SNR), and its formula is:

[0027] in, For noise power, Boltzmann's constant, The system noise temperature, For receiver bandwidth, This represents the noise figure.

[0028] The position offset (noise) detected by radar is mainly determined by measurement error, which is closely related to the signal-to-noise ratio (SNR). Radar measurement error is generally inversely proportional to the SNR; the higher the SNR, the smaller the measurement error. The specific relationship can be expressed as:

[0029] According to the radar formula, the relationship between measurement error and distance is as follows:

[0030] Position offset can be seen as a direct manifestation of measurement error; therefore, position offset is proportional to the square of the distance. Based on this physical relationship, this step specifically models the detection noise as a function of distance. and speed Gaussian noise in two-factor dynamic correlation: Standard deviation of distance from relevant noise The design is proportional to the square of the distance, that is:

[0031] in, This is a constant characterizing the sensitivity of a radar system, reflecting the accumulation of errors caused by signal attenuation with distance; Standard deviation of velocity-related noise Designed to match the target speed Proportional, that is:

[0032] in, The constant is dependent on the radar Doppler resolution and is designed to simulate the Doppler effect and scintillation noise caused by high-speed targets.

[0033] Finally, by assigning weights to the distance and velocity noise components... and The total noise standard deviation of the injected track was obtained by using a weighted sum of squares. :

[0034] The noise model is expressed as follows: This modeling process enables the simulated noisy flight path data to more realistically reflect the characteristics of "greater noise in long-range, high-speed target detection" in actual radar detection environments.

[0035] Step S200: Based on the motion equations of the fixed-wing UAV, simulate its motion process under normal and abnormal conditions, generate a real track, and add noise to the real track using the radar detection noise model to obtain noisy simulated track data. The equations of motion for fixed-wing unmanned aerial vehicles (UAVs) are based on Newtonian mechanics and flight aerodynamics, and are typically divided into a six-degree-of-freedom (6-DOF) model, including three translational degrees of freedom (along the x, y, and z axes) and three rotational degrees of freedom (roll, pitch, and yaw). The complete six-DOF equations of motion include translational and rotational dynamics, with the translational dynamics equations being:

[0036] The equation of rotational dynamics is:

[0037] in, For the quality of drones, For velocity vectors, The inertia matrix, It is the angular velocity vector. For aerodynamic vectors, The gravity vector For thrust vector, It is the aerodynamic torque vector. The thrust torque vector; the aerodynamic force vector can be further expressed as:

[0038] Aerodynamic torque can be further expressed as:

[0039] in, For resistance ( ), It is a lateral force. For lift ( ), For rolling torque, For pitching moment, For yaw moment, air density, For wing area, The drag coefficient, This is the lift coefficient.

[0040] Based on this model, different flight states are simulated by adjusting the force and torque balance equations: during normal level flight, the lift of the UAV... With gravity Balance, thrust With resistance Equilibrium, the force equilibrium equation is:

[0041] The equation of motion is:

[0042] in, It is the flight path angle, and also the pitch angle; it is 0 during level flight.

[0043] When climbing or tilting down, the drone's flight path angle Lift With thrust When the components change, the force balance equation becomes:

[0044] The equations of motion remain unchanged.

[0045] When turning, the drone generates lateral force through roll, achieving yaw motion. The force balance equation is:

[0046] The equation of motion is:

[0047] in, For roll angle, The turning radius, This is the yaw angle.

[0048] During a stall, the drone's lift is insufficient to counteract gravity, causing a rapid descent. At this point, the drone's equations of motion need to account for the imbalance between gravity and aerodynamic forces. The force equations are:

[0049] The equations of motion still satisfy the constraints for level flight:

[0050] During stall, lift coefficient A sharp drop leads to lift. Significantly reduced. The flight path angle of the drone. A negative value indicates a decrease.

[0051] The equations of motion for fixed-wing UAVs vary depending on the flight state (level flight, climb, turn, stall), but they are all based on Newtonian mechanics and aerodynamics. In simulations, these equations can be used to model the dynamic behavior of the UAV, and combined with radar detection noise, a more realistic trajectory simulation can be achieved.

[0052] For example, during the simulation, initial parameters of the drone (such as mass) can be set. Initial airspeed Initial position (etc.) and time step The velocity, position, and attitude angle (roll angle) are updated iteratively through numerical integration. Pitch angle Yaw angle This generates a noise-free, realistic track sequence. Subsequently, the radar detection noise model established in step S100 is invoked, and for each trajectory point in the real flight path, the distance between it and the radar is calculated based on its real-time distance. and instantaneous speed Calculate the standard deviation of total noise And generate three-dimensional Gaussian noise. The data is superimposed onto the coordinates, and the final output includes simulated trajectory data containing timestamps and noisy spatial coordinates.

[0053] Step S300: Based on the noisy simulated flight path data, construct a dynamic graph; wherein, the node features of the dynamic graph include position, velocity, acceleration and aerodynamic features, and the edge connections of the dynamic graph include temporal adjacency edges, physical constraint edges and dynamic attention edges; Dynamic graph data construction mainly includes node feature design and edge connection design, node features The system incorporates 17 dimensions of features, including the drone's trackpoint position, velocity, acceleration, and aerodynamic characteristics, and is designed as follows:

[0054] Among them, the first 9 dimensions are basic kinematic features (three-dimensional position). (Unit: meters), three-dimensional velocity (Unit: meters per second) and three-dimensional acceleration (Unit: m / s²) Subsequent features include: curvature at time step t (Reflects the curvature of the trajectory, calculated from velocity and acceleration), heading angle (radians), jerk (The derivative of acceleration, unit: m / s³), height (meters), and the key aerodynamic characteristic—angle of attack. (radians), airspeed m / s), lift-to-drag ratio (Dimensionless) and energy level (Meters, reflecting the drone's remaining gliding capability).

[0055] These aerodynamic characteristics are based on the aerodynamic parameters of fixed-wing UAVs (such as lift coefficient). drag coefficient Wing area The aerodynamic state and gliding efficiency are calculated based on the current motion state (such as airspeed and attitude) and other parameters, aiming to comprehensively characterize the aerodynamic state and gliding efficiency.

[0056] Next, a three-layer edge connection construction strategy is executed to establish associations between nodes: 1) Temporal adjacency edges: connecting each node with a fixed weight (e.g., weight = 1). Its direct successor node in the time series This ensures that the model can capture the fundamental temporal continuity; 2) Physically Constrained Edges: Introducing Newtonian kinematics as prior knowledge for nodes... Based on its current velocity and acceleration, predict its expected position at the next moment. If in subsequent actual data, the actual location of the corresponding time node Deviation from predicted value Exceeding the preset threshold Then at node and An edge is established between them, and its weight can be related to the magnitude of the deviation, thereby marking track segments that clearly violate physical laws as potential anomaly regions; 3) Dynamic Attention Edges: Input all node features into a Graph Attention Network (GAT), and compute the dynamic attention edge between any two nodes. and Attention coefficient between :

[0057] in, It is a normalized exponential function; It is a non-linear activation function; This represents a learnable weight matrix; and Each node and nodes The original node feature vector; This indicates a vector concatenation operation.

[0058] Then, based on the calculated attention coefficient, only the edges with the top 30% weight ranking are connected, so that the model can adaptively focus on key node pairs that are strongly correlated or unusually related in time and space (such as non-local dependencies between sharp turns, stall start points and subsequent trajectory points).

[0059] Through this step, the original temporal track data is transformed into a dynamic graph structure rich in multi-level semantics (temporal, physical, and attention).

[0060] Step S400: Input the dynamic graph into the spatiotemporal graph neural network model, extract features and classify the dynamic graph through the spatiotemporal graph neural network model, and output the trajectory anomaly detection result; wherein, the spatiotemporal graph neural network model includes a graph attention network layer, a long short-term memory network layer and a linear layer connected in sequence.

[0061] The architecture of the Spatiotemporal Graph Neural Network (STGNN) model is as follows: Figure 2 As shown, it sequentially includes a graph attention network layer, a long short-term memory network layer, and a linear layer. Dynamic graph data (containing 17-dimensional node features and edge features) is input into the first graph attention network layer (…). ), which is defined as:

[0062] in, This is an identifier for the convolutional layer of a graph attention network; Represents the number of attention heads; The dimension representing the edge features; This layer aggregates spatial features of nodes and their adjacent edges (including temporal, physical constraints, and dynamic attention edges) through a multi-head attention mechanism, and outputs features after ReLU activation.

[0063] Subsequently, the features are fed into the second graph attention network layer ( ), defined as:

[0064] Further, deep extraction and fusion of spatial features are performed. Next, the node feature sequence at each time step is input into a bidirectional long short-term memory (Bi-LSTM) network layer, defined as follows:

[0065] in, It is the standard identifier for Long Short-Term Memory (LSTM) network modules; It is a Boolean parameter that controls whether the LSTM is bidirectional; For parameters The specific value to be set. This means that a bidirectional LSTM (Bi-LSTM) will be enabled to capture the long-term dependencies and evolution patterns of track data over time.

[0066] Then, the temporal features output by the LSTM are sequentially passed through the first linear layer ( ) and ReLU activation function, and the second linear layer ( The process is performed on the output, which outputs a 6-dimensional vector corresponding to the predicted score of the preset anomaly category (such as normal, power failure, navigation drift, etc.). Finally, by performing Softmax normalization or taking argmax operation on the output, the anomaly detection result of each waypoint or flight segment is obtained.

[0067] To train the model, a defined composite loss function is used for optimization, which consists of three parts: The first part is the cross-entropy loss between the predicted category and the true label. .

[0068] The second part is the L2 norm penalty term between the predicted value and the actual value. .

[0069] The third part is about the sudden acceleration. Penalty items .

[0070] The total loss is:

[0071] in, This represents the total loss value that needs to be optimized during model training; This represents the probability distribution of the predicted categories in the model output. Indicates the true category label of the sample; This represents the weighting hyperparameters used to balance the contribution of this loss term; This represents the speed value predicted by the model based on the flight path data; This represents the actual velocity value corresponding to the predicted point; This represents the weight hyperparameter used to balance the contribution of this penalty term; This represents the change in acceleration between adjacent time points; This is a linear rectified function used to ensure that a penalty is applied only when the change exceeds a threshold of 0.5; the weight is set according to actual conditions, and can be set empirically to... .

[0072] This loss function design enables the model to optimize classification accuracy while sensitively capturing deviations between predicted and actual trajectory values, as well as abrupt changes in acceleration. This effectively enhances the ability to identify anomalous patterns such as "gliding failures" accompanied by subtle kinematic changes. Through the synergistic training of the aforementioned model structure and the customized loss function, this method ultimately achieves accurate and reliable detection of trajectory anomalies for fixed-wing UAVs.

[0073] The method for detecting anomalies in the flight paths of fixed-wing UAVs provided by this invention is specifically reflected in the following three aspects: Firstly, in radar noise modeling, this invention breaks through the traditional simplified model using Gaussian noise with a fixed standard deviation, and constructs a Gaussian noise model dynamically correlated with both target range and velocity. Traditional methods fail to reflect the actual physical law that "the farther the distance and the faster the speed, the greater the noise." This invention, on the one hand, establishes a relationship between measurement error and the square of the distance based on radar equations (…). This simulates the error accumulation caused by long-distance signal attenuation; on the other hand, for the Doppler effect of high-speed targets, a noise component proportional to the velocity is constructed ( Finally, the total noise is obtained through weighted fusion, which makes the generated simulated trajectory data highly consistent with the real radar detection environment, providing a high-fidelity data foundation for model training.

[0074] Secondly, in terms of dynamic graph construction, a three-layer edge connection strategy integrating temporal, physical constraints, and attention is proposed, overcoming the limitations of traditional methods that rely solely on temporal adjacency edges. This invention not only ensures temporal continuity through temporal adjacency edges with a weight of 1, but also introduces two types of functional edges: first, physical constraint edges, which are automatically established based on Newtonian kinematics priors when the deviation between the actual trajectory position and the dynamically predicted position exceeds a threshold, thus directly transforming physically infeasible motion into anomaly signals in the network; second, dynamic attention edges, which adaptively calculate the association between nodes through a graph attention network (GAT) and retain only the top 30% of edges with the highest attention weights, enabling the model to focus on key nodes (such as sharp turns and stall points) and capture non-local spatiotemporal dependencies. Combined with 17-dimensional node features covering position, velocity, acceleration, angle of attack, lift-to-drag ratio, etc., this dynamic graph structure provides rich spatiotemporal semantic information for accurately distinguishing between "normal fluctuations" and "abnormal deviations."

[0075] Finally, in terms of model and loss function design, dual optimization of spatiotemporal feature extraction and physical constraints was achieved. The model adopts an architecture of "two-layer GAT + bidirectional LSTM + physically constrained linear layer": a two-layer graph attention network (... and The system deeply extracts spatial features; a bidirectional LSTM captures long- and short-term temporal evolution patterns; subsequent linear layers are dedicated to physical constraint verification. Simultaneously, a fusion of classification loss and prediction bias penalty is designed. ) and acceleration mutation penalty ( The composite loss function (based on empirical assumptions) This design forces the model to impose physical constraints on the continuity of its predicted velocities and the smoothness of its accelerations while pursuing classification accuracy, thereby significantly enhancing its sensitivity to subtle kinematic anomalies such as "failure gliding".

[0076] In summary, this invention achieves accurate, real-time, and reliable detection of anomalies in the flight paths of fixed-wing UAVs, especially the highly concealed "gliding faults," through the synergy of noise modeling, graph structure construction, and learning model.

[0077] According to another aspect of the embodiments of this application, an electronic device is also provided, including a processor and a memory, wherein the processor is configured to implement the steps of the method when executing a computer program stored in the memory.

[0078] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0079] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0080] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0081] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0082] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in the flight path of a fixed-wing unmanned aerial vehicle (UAV), characterized in that, Includes the following steps: Based on the radar equation and signal-to-noise ratio formula, a radar detection noise model related to the target distance and velocity of the UAV is established; Based on the motion equations of a fixed-wing UAV, its motion process under normal and abnormal conditions is simulated to generate a real track. Then, noise is added to the real track using the radar detection noise model to obtain noisy simulated track data. Based on the noisy simulated flight path data, a dynamic graph is constructed; wherein, the node features of the dynamic graph include position, velocity, acceleration and aerodynamic features, and the edge connections of the dynamic graph include temporal adjacency edges, physical constraint edges and dynamic attention edges; The dynamic graph is input into a spatiotemporal graph neural network model, which extracts and classifies features from the dynamic graph and outputs anomaly detection results. The spatiotemporal graph neural network model includes a graph attention network layer, a long short-term memory network layer, and a linear layer connected in sequence.

2. The method for detecting anomalies in the flight path of a fixed-wing UAV as described in claim 1, characterized in that, Methods for establishing radar detection noise models related to UAV target distance and velocity based on radar equations and signal-to-noise ratio formulas include: Calculate the distance between the UAV target and the radar based on the radar equations; Calculate the standard deviation of distance-related noise associated with the distance, wherein the standard deviation of distance-related noise is proportional to the square of the distance; Calculate the speed of the drone target; Calculate the speed-related noise standard deviation associated with the speed, wherein the speed-related noise standard deviation is proportional to the speed; The total noise standard deviation is obtained by summing the distance-related noise standard deviation and the velocity-related noise standard deviation by weighted square, thereby establishing the radar detection noise model.

3. The method for detecting anomalies in the flight path of a fixed-wing UAV as described in claim 1, characterized in that, The method for adding noise to the real flight track using the radar detection noise model to obtain noisy simulated flight track data includes: The actual flight track is obtained, which contains multiple trajectory points arranged in chronological order, and each trajectory point contains a timestamp and three-dimensional spatial coordinates; For each current trajectory point in the actual flight path, perform the following steps: Calculate the distance between the UAV and the radar corresponding to the current trajectory point; Calculate the speed of the drone corresponding to the current trajectory point; Based on the radar detection noise model, the total noise standard deviation is calculated based on the distance and the speed. Three-dimensional Gaussian noise is generated based on the total noise standard deviation; The three-dimensional Gaussian noise is added to the three-dimensional spatial coordinates of the current trajectory point to obtain a trajectory point containing noise; All noisy trajectory points are combined in chronological order to obtain the noisy simulated flight track data.

4. The method for detecting anomalies in the flight path of a fixed-wing UAV as described in claim 1, characterized in that, Constructing the dynamic graph includes the following node feature calculation steps: For each track point in the noisy simulated track data, calculate its position, velocity, and acceleration; Based on the calculated velocity and acceleration, calculate the curvature, heading angle, and jerk. Obtain the altitude information of the waypoints; Based on the aerodynamic parameters and motion state of the fixed-wing UAV, the aerodynamic characteristics are calculated, including angle of attack, airspeed, lift-to-drag ratio and energy altitude. The calculated position, velocity, acceleration, curvature, heading angle, jerk, altitude, angle of attack, airspeed, lift-to-drag ratio, and energy altitude are combined to form the nodal feature vector of the waypoint.

5. The method for detecting anomalies in the flight path of a fixed-wing UAV as described in claim 4, characterized in that, The steps for constructing edge connections in a dynamic graph include: Construct temporal adjacency edges: connect adjacent nodes of each track point in the time series; Constructing physical constraint edges: For each track point, predict its expected position at the next moment based on Newtonian kinematics. If the deviation between the actual position of its subsequent track points and the expected position exceeds a preset threshold, then establish a connection between the current node and the node corresponding to the actual position. Constructing dynamic attention edges: Input all nodes of the dynamic graph into the graph attention network, calculate the attention coefficient between any two nodes, and filter and establish connections based on the attention coefficient.

6. The method for detecting anomalies in the flight path of a fixed-wing UAV as described in claim 5, characterized in that, In the step of constructing dynamic attention edges, the construction of the dynamic attention edges includes: Calculate the attention coefficients between all node pairs using a graph attention network; Based on the calculated attention coefficient, select and retain the node pairs whose attention weight ranks in the top 30%; An edge connection is established between the retained node pairs as the dynamic attention edge.

7. The method for detecting anomalies in the flight path of a fixed-wing UAV as described in claim 1, characterized in that, The method of inputting the dynamic graph into a spatiotemporal graph neural network model, extracting features and classifying the dynamic graph through the spatiotemporal graph neural network model, and outputting track anomaly detection results includes: The dynamic graph is input into the first graph attention network layer of the spatiotemporal graph neural network model to extract spatial features from the nodes and edges of the dynamic graph. The extracted spatial features are input into the second graph attention network layer of the spatiotemporal graph neural network model for deep extraction and fusion of spatial features. The features output from the second graph attention network layer are input into the bidirectional long short-term memory network layer of the spatiotemporal graph neural network model to capture the evolution pattern of the trajectory data in the time series. The temporal features output by the bidirectional long short-term memory network layer are sequentially input into the first linear layer and the second linear layer of the spatiotemporal graph neural network model for physical constraint verification and classification. Based on the output of the second linear layer, track anomaly detection results are generated and output.

8. The method for detecting anomalies in the flight path of a fixed-wing UAV as described in claim 1, characterized in that, The method further includes training the spatiotemporal graph neural network model, wherein the training of the spatiotemporal graph neural network model employs a composite loss function, which consists of the following parts: The first part is the cross-entropy loss between the predicted category and the true label; The second part is the penalty for the difference between the predicted waypoint speed and the actual speed; The third part is the penalty for sudden changes in acceleration at the waypoint.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the fixed-wing UAV trajectory anomaly detection method according to any one of claims 1-8, and the processor is configured to execute the programs stored in the memory.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the fixed-wing UAV trajectory anomaly detection method according to any one of claims 1-8.

Images