Unmanned aerial vehicle abnormal flight data security detection method for cloud server cluster

Abstract

The invention relates to the technical field of information, and discloses an unmanned aerial vehicle abnormal flight data safety detection method for a cloud server cluster, and the method comprises the following steps: S1, multi-modal data collection: collecting unmanned aerial vehicle data, and generating a high-dimensional feature matrix; s2, multi-modal interaction diagram construction: constructing an interaction diagram of the unmanned aerial vehicle cluster based on the high-dimensional feature matrix; s3, dimension reduction of high-dimensional data: mapping the high-dimensional feature matrix to a low-dimensional embedding space; s4, local anomaly detection: calculating a preliminary anomaly score of the unmanned aerial vehicle; s5, global anomaly propagation: generating a global anomaly detection result; and S6, cloud-side cooperative detection and response: issuing an abnormal response instruction to the unmanned aerial vehicle. A multi-unmanned aerial vehicle cluster cooperative detection mechanism is constructed through cooperation of a cloud server cluster and edge nodes in combination with dynamic interaction of unmanned aerial vehicles. Flight characteristics are captured by using a multi-modal interaction diagram, and high-precision global anomaly detection of the unmanned aerial vehicle cluster is realized through anomaly propagation and optimization.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and specifically to a method for detecting the security of abnormal flight data of unmanned aerial vehicles (UAVs) for a cloud server cluster. Background Art

[0002] In recent years, with the rapid development of UAV technology, UAV clusters have been widely used in fields such as logistics transportation, environmental monitoring, and disaster relief. UAV clusters have the advantages of high efficiency and flexibility when performing tasks. However, due to their large quantity, complex flight environment, and diverse tasks, higher requirements are put forward for the detection of their behavioral safety. Especially with the support of a cloud server cluster, by real-time monitoring of UAV flight data and quickly processing abnormal behaviors, the security and stability of UAV clusters can be significantly improved.

[0003] Currently, the research on multi-UAV collaborative detection technology is relatively scarce, and most existing methods are still limited to the behavioral analysis of a single UAV. These methods usually process the flight data of a single UAV through rule detection or machine learning models, and it is difficult to capture collaborative abnormal behaviors in UAV clusters. In addition, the data analysis methods based on a single device have poor adaptability to changes in the flight environment and the expansion of the cluster scale. Especially in the cloud server cluster environment, there is still a lack of a mature technical solution for effectively integrating the massive flight data of multiple UAVs for global analysis and anomaly detection. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for detecting the security of abnormal flight data of UAVs for a cloud server cluster, which solves the problems of the lack of a multi-UAV collaborative detection mechanism in the prior art and the difficulty in efficiently globally detecting anomalies in multi-UAV data in the cloud server cluster environment.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for detecting the security of abnormal flight data of UAVs for a cloud server cluster, including the following steps: S1. Multi-modal data acquisition: Collect the flight trajectory data, sensor data, and visual data of UAVs and fuse them to generate a high-dimensional feature matrix; S2. Construction of a multi-modal interaction graph: Based on the high-dimensional feature matrix, construct an interaction graph of the UAV cluster. The interaction graph includes nodes and edges. The nodes represent UAVs, and the edges represent the feature similarity between UAVs; S3. High-dimensional data dimensionality reduction: Based on the interaction graph, use the regularized variational dimensionality reduction method to map the high-dimensional feature matrix to a low-dimensional embedding space; S4. Local anomaly detection: Calculate the preliminary anomaly score of the UAV through the reconstruction error of the low-dimensional data; S5. Global anomaly propagation: Optimize the anomaly score based on the interaction graph of the UAV cluster to generate the global anomaly detection result; S6. Cloud-edge collaborative detection and response: The edge nodes complete local detection, the cloud server completes global optimization, and issues anomaly response instructions to the UAVs.

[0006] Preferably, the step S2 includes the following contents: Generate feature vectors using the multi-modal data of the UAVs; Determine the association strength between UAVs by calculating the similarity of feature vectors between UAVs, and use it as the edge weight of the interaction graph; The edge weight of the interaction graph is used to describe the feature association relationship between UAVs.

[0007] Preferably, in the step S3, the high-dimensional multi-modal data of the UAVs is mapped to a low-dimensional space by a regularization dimensionality reduction method, and the dimensionality reduction method has the following characteristics: Retain the local features of the multi-modal data while enhancing the global consistency of the data in the low-dimensional space; Reduce the redundancy of the UAV multi-modal data and ensure the separability of anomaly patterns in the low-dimensional space.

[0008] Preferably, during the high-dimensional data dimensionality reduction process, construct the association relationship between UAVs to ensure that the low-dimensional embedded data can reflect the multi-UAV interaction characteristics, and the association relationship is calculated based on the spatial similarity and communication frequency between UAVs.

[0009] Preferably, the step S4 includes the following contents: Reconstruct the multi-modal data of the UAVs using the reduced low-dimensional feature space; Calculate the difference between the original data and the reconstructed data. The larger the difference value, the more the flight behavior of the UAV deviates from the normal mode; Use this difference value as the preliminary anomaly score of the UAV.

[0010] Preferably, the step S5 includes the following contents: Based on the interaction graph of the UAV cluster, propagate the local anomaly score in the graph structure; During the propagation process, enhance the global consistency of the anomaly score and optimize the score result in combination with the topological structure of the graph; Through global propagation, generate the final anomaly score of each UAV. The higher the score value, the greater the possibility of abnormal behavior.

[0011] Preferably, add a sparsity optimization constraint to the global anomaly propagation to ensure that in a cluster environment, only a few UAVs are marked as abnormal, rather than determining the behaviors of all UAVs as abnormal, and enhance the accuracy and credibility of the detection.

[0012] Preferably, step S6 includes the following: The drone side collects multi-modal flight data in real time and uploads it to the edge node; The edge node completes local anomaly detection, including high-dimensional data dimensionality reduction and local anomaly scoring; The cloud server receives the data from all edge nodes, integrates the local anomaly scores, and completes the optimization of the global anomaly scores; The cloud server sends a response command to the drone according to the final anomaly detection result.

[0013] Preferably, the response command includes one or more of the following: Adjust the flight path to avoid the abnormal area; Send a return command to the abnormal drone; Remotely disconnect the communication link of the abnormal drone; Issue a no-fly zone warning and adjust the no-fly zone range.

[0014] Preferably, the method further includes a dynamic update mechanism, including the following: Based on the real-time collected flight data, the cloud server continuously optimizes the high-dimensional data dimensionality reduction model and the anomaly detection model; the optimized model parameters are sent to the drone side through the edge node to achieve the dynamic adaptation and continuous learning ability of the whole system.

[0015] The present invention provides a method for secure detection of abnormal flight data of drones for a cloud server cluster. It has the following beneficial effects: 1. Through the collaborative work of the cloud server cluster and the edge node, combined with the dynamic interaction relationship between drones, the present invention constructs a collaborative detection mechanism for multi-drone clusters, captures flight characteristics among drones using a multi-modal interaction graph, and based on an efficient anomaly propagation and optimization method, achieves high precision and high robustness in global anomaly detection, especially having significant advantages in the identification of abnormal collaborative behaviors of drone clusters.

[0016] 2. Through multi-modal data collection and fusion, the present invention constructs a high-dimensional feature matrix and uses a regularization dimensionality reduction method to map the data to a low-dimensional embedding space. While reducing data redundancy and complexity, it retains the local and global features in flight characteristics, making abnormal patterns clearer in the low-dimensional space and providing an efficient and accurate basis for subsequent anomaly detection.

[0017] 3. Through the design of the cloud-edge collaborative architecture, the present invention distributes the execution of complex global anomaly detection tasks, uses edge nodes to complete local processing, and the cloud server to complete global optimization, and generates corresponding response instructions (such as path adjustment, flight prohibition, etc.). This architecture effectively reduces the latency of data transmission, realizes rapid response and precise intervention for abnormal behaviors, and greatly improves the safety and operation efficiency of the UAV cluster. Brief Description of the Drawings

[0018] Figure 1 is the flowchart of the method of the present invention; Detailed Embodiments

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the specification of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] To better understand the present invention, the above content will be described in detail below in conjunction with specific embodiments. Please refer to the attached Figure 1 The embodiment of the present invention provides a method for detecting the security of abnormal flight data of UAVs for a cloud server cluster, which is characterized by including the following steps: S1. Multi-modal data collection: Collect and fuse the flight trajectory data, sensor data, and visual data of the UAV to generate a high-dimensional feature matrix; In this embodiment, the multi-modal data collection method aims to realize real-time monitoring and fusion processing of multi-dimensional information during the flight of the UAV, and form an input data basis suitable for high-dimensional feature analysis. Specifically, the trajectory data, sensor data, and visual data generated during the flight of the UAV are comprehensively collected from multiple dimensions such as spatial position, motion state, and environmental perception. Subsequently, data of different modalities are constructed into a high-dimensional feature matrix through specific fusion methods, laying a data foundation for subsequent dimensionality reduction processing and anomaly detection.

[0021] It should be noted that the multi-modal data collection process covers a variety of sensor devices and data processing technologies, including but not limited to GPS modules, IMU sensors, cameras, and preprocessing methods such as data denoising and format conversion. In some embodiments of the present invention, by constructing a specific multi-modal feature space, the relevance and applicability of data in dimensionality reduction and anomaly detection can be effectively improved.

[0022] Among them: 1. Trajectory data collection: In the implementation of the present invention, the trajectory data is obtained through the global positioning system (GPS) module carried by the drone, recording the spatial position information of the drone during flight to form a three-dimensional coordinate point sequence: X GPS ={(x t ,y t ,z t )|t=1,2,…,T} Among them, X GPS Represents the GPS coordinate sequence matrix; (x t ,y t ,z t ) represent the coordinates of the drone in directions x, y, and z at time t, respectively; T is the total number of sampling moments; To improve the continuity and reliability of position information, in the case where the GPS signal is weak or blocked, an inertial navigation system (INS) is introduced in this embodiment to generate a supplementary trajectory by calculating the flight path.

[0023] In some embodiments, the flight speed and acceleration are also collected and the calculation formula is: Among them, v t represents the speed of the UAV at time t; and are the velocity components in the x, y, and z directions respectively; a t represents the acceleration of the drone at time t; Indicates the rate of change of velocity over time (acceleration) By collecting trajectory data, the flight path, speed and motion status of the drone can be accurately reflected, providing precise location information input for multimodal data.

[0024] 2. Sensor data collection Sensor data collection mainly relies on the UAV's inertial measurement unit (IMU), and the collected features include angular velocity, linear acceleration, pitch angle, yaw angle, etc. The generated sensor feature matrix is ​​expressed as: X IMU ={ω x ,ω y ,ω z ,α x ,α y ,α z} Among them, ω x ,ω y ,ω z Respectively represent the angular velocity around the x, y, and z axes; α x ,α y ,α zrespectively represent the linear accelerations along the x, y, and z axes; X IMU represents the UAV's time-series sensor data matrix.

[0025] To reduce noise interference, in this embodiment, the sliding window filtering method is used to preprocess the IMU data, and the filtered sensor data can better reflect the true motion state of the UAV.

[0026] In some embodiments, a barometer and a magnetometer are also introduced: The barometer is used to estimate the flight altitude of the UAV; The magnetometer is used to correct the heading angle of the UAV.

[0027] Therefore, the sensor data can accurately reflect the attitude changes and motion state of the UAV, providing dynamic motion characteristics input for the anomaly detection model.

[0028] 3. Visual data acquisition The camera mounted on the UAV collects environmental visual data during flight, mainly used to perceive surrounding obstacles and environmental changes. The collected image data sequence is expressed as: I t ={I 1 , I 2 , …, I T} where I t represents the image sequence; I T represents a single-frame image collected at time t.

[0029] The image features are extracted by a pre-trained convolutional neural network (CNN), and the generated visual features are expressed as: X visual = CNN(I t ) where X visual represents the visual feature matrix; CNN(.) represents the feature extraction process based on the convolutional neural network.

[0030] In one implementation, the visual data also includes edge detection features and object detection results for dynamic obstacle recognition.

[0031] Therefore, the visual data can supplement the information of the UAV's surrounding environment perception, especially of great significance for flight safety, providing environmental interaction characteristics input for anomaly detection.

[0032] 4. Multimodal feature fusion Fuse the above three types of data to form a high-dimensional feature matrix, expressed as: X = [X GPS , X IMU , X visual ​ wherein, it represents the fused multi-modal high-dimensional feature matrix; To eliminate the differences in the eigenvalue ranges between modalities, the present invention performs a normalization process on the feature matrix X, and the normalization formula is: wherein, X ′ represents the normalized feature matrix; μ is the feature mean; σ is the feature standard deviation.

[0033] In a possible implementation manner, a time series alignment process is also performed on the feature matrix X to ensure the consistency of the modal data at the same time step.

[0034] Therefore, through multi-modal data acquisition and fusion, the present invention can generate a structured high-dimensional feature matrix X, covering the flight path, motion state, and environmental perception information of the unmanned aerial vehicle. This matrix significantly improves the data correlation and robustness in subsequent dimensionality reduction processing, laying a foundation for high-precision anomaly detection. The formula and technical processing are fully disclosed and have clear industrial realizability.

[0035] S2. Construction of multi-modal interaction graph: Based on the high-dimensional feature matrix, construct an interaction graph of the unmanned aerial vehicle cluster. The interaction graph includes nodes and edges. The nodes represent unmanned aerial vehicles, and the edges represent the feature similarity between unmanned aerial vehicles; In the present invention, in order to more comprehensively reflect the dynamic association relationship among multiple unmanned aerial vehicles in the unmanned aerial vehicle cluster, a multi-modal interaction graph of the unmanned aerial vehicle cluster will be constructed based on the multi-modal feature matrix. The interaction graph models the unmanned aerial vehicle and its feature association in the form of nodes and edges, where the nodes represent the unmanned aerial vehicle itself, and the edges represent the feature similarity between unmanned aerial vehicles. By constructing this interaction graph, the global topological structure and local feature connections of the unmanned aerial vehicle cluster can be intuitively expressed, providing a structured basis for subsequent data dimensionality reduction and anomaly detection.

[0036] It should be noted that the construction of the interaction graph not only depends on the individual features of the unmanned aerial vehicle, but also comprehensively considers the distance, motion state, and communication characteristics between unmanned aerial vehicles. The edge weights in the interaction graph reflect the association strength between unmanned aerial vehicles, and through these edge weights, the possible cooperative behaviors and abnormal distribution characteristics of multiple unmanned aerial vehicles during the flight mission can be captured.

[0037] S3. High-dimensional data dimensionality reduction: Based on the interaction graph, use the regularized variational dimensionality reduction method to map the high-dimensional feature matrix to a low-dimensional embedding space; In the present invention, high-dimensional data reduction aims to solve the problems of data redundancy and noise in the multi-modal feature matrix of unmanned aerial vehicles (UAVs), while preserving the core features of the multi-modal data and improving the efficiency and accuracy of subsequent anomaly detection. The present invention maps the original high-dimensional feature matrix to a low-dimensional embedding space through a regularization variational data reduction method based on an interaction graph, so that the data reduction result can not only reflect the independent characteristics of each UAV, but also retain the interaction associations between multiple UAVs in the cluster.

[0038] It should be noted that the regularization method is adopted in the data reduction process, which not only considers the local characteristics of the multi-modal data of UAVs (such as the change trend of flight trajectories and the correlation between sensor values), but also makes full use of the global graph information of the interaction between multiple nodes in the UAV cluster. This method is particularly suitable for high-dimensional complex data scenarios and can significantly improve the applicability and robustness of the data reduction result.

[0039] S4. Local anomaly detection: Calculate the preliminary anomaly score of the UAV through the reconstruction error of the low-dimensional data; The local anomaly detection method of the present invention aims to calculate the preliminary anomaly score of each UAV by analyzing the reconstruction error of the low-dimensional embedded data of the UAV. Specifically, after the high-dimensional multi-modal data is reduced, the obtained low-dimensional embedded features retain the key information of the multi-modal features, while significantly reducing redundancy and noise interference. Using these low-dimensional features, efficient data reconstruction can be performed, and whether the flight behavior of the UAV deviates from the normal mode can be judged by the difference between the original multi-modal data and the reconstructed data.

[0040] It should be noted that the result of local anomaly detection serves as the initial input for subsequent global anomaly propagation, providing a distributed and localized basis for anomaly detection. In some embodiments of the present invention, the stability and accuracy of the anomaly score can be further improved by normalizing the reconstruction error of the low-dimensional data and performing clustering analysis.

[0041] S5. Global anomaly propagation: Optimize the anomaly score based on the interaction graph of the UAV cluster to generate the global anomaly detection result; The global anomaly propagation of the present invention is an important part of the present invention, which is used to optimize the local anomaly score based on the interaction graph of the UAV cluster and generate the global anomaly detection result. In this step, by combining the interaction relationship of the multi-modal features in the UAV cluster, the constructed interaction graph is used to propagate and optimize the local anomaly scores of each UAV, further improving the global consistency and robustness of the anomaly detection result. Through global propagation, collaborative anomaly behaviors that may exist between UAV clusters, such as multi-UAV linkage anomalies and cross-regional anomaly diffusion patterns, can be effectively captured.

[0042] It should be noted that the global anomaly propagation combines the spatial relationships among drones, communication interaction characteristics, and local detection results, and gradually generates the global anomaly scores of each drone through iterative optimization on the graph structure. This score not only reflects the individual abnormal behaviors but also reveals the overall abnormal patterns of the drone swarm.

[0043] S6. Cloud-edge collaborative detection and response: The edge nodes complete local detection, the cloud server completes global optimization, and issues anomaly response instructions to the drones.

[0044] The cloud-edge collaborative detection and response steps of the present invention aim to make full use of the collaborative capabilities of the cloud server and edge nodes to achieve efficient detection and real-time response to the abnormal behaviors of drones. Specifically, by completing local anomaly detection at the edge nodes, uploading the preliminary detection results to the cloud server, where global optimization and decision-making are carried out by the cloud, and then timely issuing the anomaly response instructions to the corresponding drones. This method can significantly improve the real-time performance and accuracy of anomaly detection, while reducing the latency and computational burden brought by data transmission.

[0045] It should be noted that the cloud-edge collaborative detection and response not only include the real-time detection of abnormal behaviors but also cover the formulation and execution of dynamic decision-making and response strategies. The response measures combine the operating status and task requirements of the drones, and can quickly intervene and control the abnormal drones on the premise of ensuring the overall safety of the swarm.

[0046] In the present invention, step S2 includes the following contents: Generate feature vectors using the multi-modal data of the drones; By calculating the similarity of feature vectors among drones, determine the association strength between drones, and use it as the edge weight of the interaction graph; The edge weight of the interaction graph is used to describe the feature association relationship between drones.

[0047] In this embodiment, the construction of the interaction graph is divided into three steps: feature extraction, association strength calculation, and edge weight assignment. In the feature extraction stage, high-dimensional feature vectors are extracted by combining the trajectory data, sensor data, and visual data of the drones. In the association strength calculation stage, the association strength between drones is calculated based on the feature similarity of multi-modal data. In the edge weight assignment stage, the association strength is mapped to the edge weight of the interaction graph to describe the feature association relationship between drones.

[0048] It should be noted that the core of the construction of the interaction graph lies in accurately depicting the similarity among multi-modal data while retaining the spatial, communication, and feature associations between drones. This provides a solid data foundation for subsequent high-dimensional data dimensionality reduction and anomaly detection.

[0049] 1. Generating feature vectors from drone multi-modal data In the present invention, each unmanned aerial vehicle extracts multimodal features from trajectory, sensor, and visual data to form a unified feature vector, where: X i =[[X GPS ,[[X IMU ,[[X visual In the formula, X i represents the high-dimensional feature vector of the i-th unmanned aerial vehicle; X GPS is the trajectory data feature; X IMU is the sensor data feature; X visual is the visual data feature; 2. Calculate the similarity of feature vectors between unmanned aerial vehicles Based on the multimodal feature vectors between unmanned aerial vehicles, calculate the feature similarity between any two unmanned aerial vehicles i and j. The specific calculation formula is: where Sim(X i ,[[X j ) is the similarity between feature vectors; X i is the feature vector of unmanned aerial vehicle i; X j is the feature vector of unmanned aerial vehicle j; ∥X i -[[X j ∥ represents the Euclidean distance between two feature vectors; is the scale parameter of the Gaussian kernel, which is used to adjust the sensitivity of the feature vector distance.

[0050] As an option, to improve the robustness of the feature similarity calculation, the present invention also considers a modal weighting mechanism, that is, weights are assigned to features of different modalities. The weighted form of the feature vector is: where w k is the weight of the k-th modality; X i,k is the feature of the k-th modality. By calculating the feature vector similarity, the correlation between unmanned aerial vehicles in the multimodal feature space can be accurately captured.

[0051] 3. Construct an interaction graph and assign edge weights Construct an interaction graph based on the feature vector similarity between unmanned aerial vehicles where: v represents the node set of unmanned aerial vehicles, and each node corresponds to an unmanned aerial vehicle; ε represents the edge set between unmanned aerial vehicles, and the weight of each edge describes the feature association strength between unmanned aerial vehicles; The calculation of the edge weight is based on the feature vector similarity and is expressed as: W ij =Sim(X i ,[[X j ) ​Among them, W ij is the edge weight between node i and node j.

[0052] In a possible implementation, the present invention also adjusts the edge weight by combining the spatial adjacency and communication frequency of the drones. The adjusted edge weight is expressed as: W ij = α·Sim(X i , X j ) + β·Comm ij Among them, α and β are weight adjustment coefficients; Comm ij represents the communication frequency between drones i and j. Therefore, the construction of the interaction graph reflects the multi-modal association strength between drones through the edge weight, providing a graph structure basis for subsequent data dimensionality reduction and anomaly detection; In some embodiments, to reduce the complexity of the interaction graph, the edge weight can be thresholded by W ij clipped, and only the edges greater than the set threshold are retained to form a sparse interaction graph.

[0053] As an option, the interaction graph can also adopt a dynamic update mechanism, that is, recalculate the edge weight according to the real-time collected drone data to adapt to the changes in the relationship between drones.

[0054] In another possible implementation, the node weight of the interaction graph can be dynamically adjusted by combining the individual characteristics of the drones (such as anomaly scores) to better reflect the global behavior characteristics.

[0055] Through the above steps, an interaction graph reflecting the multi-modal association relationship between drones is successfully constructed. This interaction graph not only captures the similarity of multi-modal feature vectors, but also enhances the correlation of collaborative behaviors between drones through edge weights. The interaction graph provides an efficient graph data structure for subsequent dimensionality reduction and anomaly detection, and has good adaptability and scalability. The formula part is disclosed in detail and is applicable to actual application scenarios.

[0056] In the present invention, in step S3, the high-dimensional multi-modal data of the drones is mapped to a low-dimensional space through a regularization dimensionality reduction method. The dimensionality reduction method has the following characteristics: Retain the local features of the multi-modal data while enhancing the global consistency of the data in the low-dimensional space; Reduce the redundancy of the multi-modal data of the drones and ensure the separability of abnormal patterns in the low-dimensional space.

[0057] During the high-dimensional data dimensionality reduction process, the association relationship between drones is constructed to ensure that the low-dimensional embedded data can reflect the multi-drone interaction characteristics. The association relationship is calculated based on the spatial similarity and communication frequency between drones.

[0058] In step S3 of the present invention, through a regularization dimensionality reduction method, the high-dimensional multi-modal data of the unmanned aerial vehicle (UAV) is mapped into a low-dimensional space. This dimensionality reduction method has the following characteristics: Local feature preservation: It can effectively preserve the local geometric characteristics of the multi-modal data, reflecting the differences in the individual characteristics of UAVs; Global consistency enhancement: By constructing the association relationship between UAVs, it ensures that the global characteristics of the UAV cluster in the low-dimensional embedding space are preserved; Redundancy reduction and anomaly separation: By eliminating the redundant information in the high-dimensional data, it improves the separability of the data, making it easier to identify abnormal patterns in the low-dimensional space.

[0059] Specifically, the present invention constructs the association relationship between UAVs, utilizes their spatial similarity and communication frequency characteristics, defines a regularization objective function, and realizes the embedding mapping from high-dimensional data to low-dimensional space through an optimization method.

[0060] It should be noted that the dimensionality reduction process of the present invention not only provides low-dimensional feature data for subsequent anomaly detection, but also strengthens the collaborative detection ability between UAVs by constructing an association relationship, improving the overall accuracy of anomaly detection.

[0061] In this embodiment 1. Input of high-dimensional features of multi-modal data The multi-modal data of the UAV includes flight trajectory data, sensor data, and visual data. These data are collected in real time through specific acquisition devices, and multi-dimensional features are extracted respectively. For each UAV, its multi-modal data can be represented as a feature vector X i The overall feature matrix is composed of the data of all UAVs: where X is the multi-modal high-dimensional feature matrix; represents the high-dimensional feature vector of the i-th UAV; N is the number of UAVs; D is the total dimension of the multi-modal data, including the sum of the dimensions of trajectory data, sensor data, and visual data; 2. Composition of multi-modal data The multi-modal feature vector X of each UAV i consists of three parts: X i =[X i,GPS , X i,IMU , X i,visual where X i,GPS is the flight trajectory data, including spatial position, speed, and acceleration; X i,IMU is the sensor data, including angular velocity, linear acceleration, pitch angle, yaw angle, etc.; X i,visual is the visual data, including environmental image features and depth information.

[0062] Example: Assume that the trajectory data of each drone has 6 dimensions (position x, y, z, speed, direction, flight time), the sensor data has 9 dimensions (including angular velocity and linear acceleration), and the visual data extracts 128 - dimensional features. Then the dimension D of the multi - modal feature vector X i is: D = 6 (trajectory data)+9 (sensor data)+128 (visual data)=143 3. Normalization processing of multi - modal features To ensure the consistency of the numerical ranges between different modal features, the present invention performs normalization processing on each feature component. The normalized feature vector is represented as: where, X ′ i is the normalized feature vector; μ is the feature mean, and σ is the feature standard deviation. Through the normalization processing, the influence of the numerical range differences between different modal features on subsequent dimensionality reduction and anomaly detection can be reduced, ensuring the balance of feature input.

[0063] Based on the high - dimensional feature input of multi - modal data of the drone flight trajectory, sensor, and visual data, a unified feature matrix is constructed. After normalization processing, this matrix has higher numerical consistency and structural characteristics, providing an efficient input basis for dimensionality reduction processing and anomaly detection. The formula details the meaning of each parameter and is suitable for practical engineering applications.

[0064] In the present invention, step S4 includes the following content: Reconstruct the multi - modal data of the drone using the reduced - dimensional low - dimensional feature space; Calculate the difference between the original data and the reconstructed data. The larger the difference value, the more the flight behavior of the drone deviates from the normal mode; Take this difference value as the preliminary anomaly score of the drone.

[0065] In step S4 of the present invention, the multi - modal data of the drone is reconstructed through the reduced - dimensional low - dimensional feature space, and by comparing the difference between the original data and the reconstructed data, it is determined whether the flight behavior of the drone is abnormal. Specifically, the embedded features in the low - dimensional space represent the main information of the multi - modal data in the high - dimensional space. The low - dimensional features are restored to the high - dimensional space using the reconstruction function, and then compared with the original data.

[0066] It should be noted that the greater the reconstruction error, the higher the likelihood that the flight behavior of the drone deviates from the normal mode. The present invention uses this difference value as the preliminary anomaly score of the drone, providing input for subsequent global anomaly propagation. In some embodiments, by designing the reconstruction function and optimizing the error calculation method, the sensitivity and robustness of local anomaly detection are further improved.

[0067] 1. Reconstruction of Low-Dimensional Embedding Features The multi-modal data of the drone is reconstructed using the low-dimensional feature matrix Y after dimensionality reduction. Specifically, each row of the low-dimensional feature matrix Y i corresponds to the low-dimensional features of the i-th drone, and its formula for reconstructing to the high-dimensional feature space is: where, represents the high-dimensional feature vector reconstructed from the low-dimensional embedding features; P(.) represents the reconstruction function, which is used to restore the high-dimensional features from the low-dimensional features Y i In a possible implementation, the reconstruction function can also adopt a non-linear mapping model, such as a neural network model based on a multi-layer perceptron (MLP); 2. Calculate the Difference between the Original Data and the Reconstructed Data The original high-dimensional data X i is compared with the reconstructed data to calculate the difference value (reconstruction error) between the two. The formula for the reconstruction error is: where, S i is the reconstruction error of the i-th drone; ∥.∥ represents the Euclidean norm, which is used to measure the distance between the original data and the reconstructed data. It should be noted that the larger the reconstruction error S i is, the more difficult it is for the multi-modal data of the i-th drone to be explained by the low-dimensional features, indicating possible abnormal behavior.

[0068] In some embodiments, in order to enhance the sensitivity to abnormal patterns, the reconstruction error can be weighted: where, w k is the weight of different modal features; X i,k and respectively represent the original data and the reconstructed data of the k-th mode of the i-th drone.

[0069] 3. Generate a Preliminary Anomaly Score The calculated reconstruction error S i is used as the preliminary anomaly score of the i-th drone to determine whether the flight behavior of the drone deviates from the normal mode.

[0070] In a possible implementation, the preliminary anomaly score can be binarized according to a preset threshold: where A i is the preliminary anomaly label of the i-th drone; τ is the threshold of the reconstruction error, which needs to be set according to historical data and experimental results.

[0071] In a possible implementation, other distance metric methods, such as cosine distance or Mahalanobis distance, can be used to calculate the reconstruction error to improve the adaptability to different feature distributions. As an option, a regularization term can be added when calculating the reconstruction error to suppress the influence of noise in high-dimensional features.

[0072] Through the above steps, the present invention realizes local anomaly detection based on low-dimensional embedded features. The calculation of the reconstruction error provides an effective preliminary anomaly score, which provides a quantitative index for whether the flight behavior of the drone deviates from the normal mode. By using low-dimensional space reconstruction and reconstruction error analysis, the redundancy of high-dimensional data can be significantly reduced, while the sensitivity and recognition ability to abnormal behaviors are improved. The formula is disclosed in detail, with integrity and feasibility, and is applicable to the anomaly detection scenario of drone swarms In the present invention, step S5 includes the following contents: Based on the interaction graph of the drone swarm, the local anomaly score is propagated in the graph structure; During the propagation process, the global consistency of the anomaly score is enhanced, and the scoring result is optimized in combination with the topological structure of the graph; through global propagation, the final anomaly score of each drone is generated, and the higher the score value, the greater the possibility of abnormal behavior.

[0073] Among them, sparsity optimization constraints are added in the global anomaly propagation to ensure that in a cluster environment, only a few drones are marked as abnormal, rather than judging the behaviors of all drones as abnormal, enhancing the accuracy and credibility of the detection.

[0074] In step S5 of the present invention, through global anomaly propagation based on the interaction graph of the drone swarm, the global consistency of the local anomaly score is optimized and the anomaly detection result is further optimized in combination with the topological structure of the graph. Specifically, the propagation method is based on the graph structure, and the local anomaly scores of the drones are propagated among nodes to capture the correlation and cooperative behavior patterns among the drones. At the same time, by introducing sparsity optimization constraints, it is ensured that the final anomaly score is sparse, and only a few drones are marked as abnormal, avoiding the generalization problem of anomaly detection.

[0075] It should be noted that the global anomaly propagation not only utilizes the spatial structure and communication characteristics of the UAV swarm, but also makes global consistency adjustments to the local anomaly scores, enhancing the robustness and credibility of the detection.

[0076] 1. Global Propagation Based on Interaction Graph In the present invention, the local anomaly score S i is propagated among nodes based on the interaction graph of the UAV swarm The propagation formula is: where represents the anomaly score after propagation; are the local anomaly scores of nodes i and j respectively; η is the propagation step; N(i) represents the set of adjacent nodes of node i. Through the propagation operation, the anomaly scores are adjusted among neighboring nodes, reflecting the collaborative behavior characteristics among UAVs and enhancing the global consistency of the anomaly scores.

[0077] 2. Optimizing the Score by Combining with the Topological Structure of the Graph During the anomaly propagation process, the score results are globally optimized by combining with the topological structure of the graph. Specifically, the optimization objective function is defined as: where is the global consistency loss; S i , S j are the anomaly scores of nodes i and j; W ij is the edge weight. The physical meaning of the optimization objective is to make the anomaly scores of adjacent nodes as close as possible to reflect the global characteristics of the graph structure.

[0078] 3. Introducing Sparsity Optimization Constraints To avoid all UAVs being marked as anomalies, the present invention adds sparsity optimization constraints during the global anomaly propagation process. The objective function of the sparsity optimization is defined as: where is the sparsity loss; S i is the anomaly score of node i. The physical meaning of the sparsity loss is to encourage the anomaly score distribution to be sparser, so that only a few nodes are marked as anomalies.

[0079] Combining the global consistency and sparsity optimizations, the final optimization objective is: where is the total objective function; μ is the sparsity weight, used to balance the influences of global consistency and sparsity, Represents the global consistency loss, which is used to optimize the global consistency of the UAV anomaly score on the interaction graph.

[0080] 4. Generate the final anomaly score Generate the final anomaly score for each UAV by optimizing the total objective function. The higher the score value, the greater the likelihood of abnormal behavior of the UAV.

[0081] In a possible implementation, the anomaly score is converted into an anomaly label A i , and the definition of the anomaly label A i is: where τ is the threshold for global anomaly detection.

[0082] Through the above steps, the present invention realizes global anomaly propagation and optimization based on the interaction graph. Introducing the sparsity optimization constraint ensures that the anomaly detection results are more accurate and reliable. The finally generated anomaly score not only enhances the global consistency but also effectively reduces the false alarm situation, providing reliable technical support for the anomaly detection of UAV swarms. The formulas and optimization methods are fully disclosed and suitable for actual engineering implementation.

[0083] In the present invention, step S6 includes the following contents: The UAV side collects multi-modal flight data in real time and uploads it to the edge node; The edge node completes local anomaly detection, including high-dimensional data reduction and local anomaly scoring; The cloud server receives the data of all edge nodes, integrates the local anomaly scores, and completes the optimization of the global anomaly score; The cloud server sends response instructions to the UAVs according to the final anomaly detection results.

[0084] In step S6 of the present invention, the anomaly detection and response of the UAV are realized through a cloud-edge collaborative architecture. Specifically, the UAV side is responsible for collecting multi-modal data in real time and uploading it to the edge node. The edge node performs local anomaly detection, generates the dimensionality-reduced feature data and preliminary anomaly scores, and then uploads them to the cloud server. The cloud server integrates the data of all edge nodes, generates the final detection results through global anomaly score optimization, and issues corresponding response instructions to the UAVs according to the results.

[0085] It should be noted that this cloud-edge collaborative architecture makes full use of the real-time processing ability of edge computing and the global optimization ability of cloud computing, and can significantly improve the efficiency and accuracy of anomaly detection. The design of the response instructions ensures the rapid processing of abnormal behaviors and effectively enhances the security of the UAV swarm.

[0086] In this embodiment 1. Real-time data collection on the drone side The drone is equipped with multi-modal data collection devices to obtain trajectory data, sensor data, and visual data during flight in real time.

[0087] The content of data collection includes: Position coordinates (such as GPS data); Motion state (such as IMU sensor data, including angular velocity and linear acceleration); Environmental perception information (such as images and visual features collected by the camera).

[0088] In a possible implementation, the collected data is uploaded to the edge node at fixed time intervals after preliminary denoising and formatting.

[0089] It should be noted that the drone side is only responsible for data collection and simple preprocessing, and complex computing tasks are completed by the edge node and the cloud server to reduce the burden on a single machine.

[0090] 2. The edge node completes local anomaly detection and dimensionality reduction processing: The edge node receives the high-dimensional multi-modal data uploaded by the drone and maps it to a low-dimensional feature space through a regularization dimensionality reduction method: The dimensionality reduction processing retains the main information of the multi-modal data and removes redundant features and noise; The dimensionality-reduced feature data Y has higher separability, providing an optimized input for subsequent anomaly detection.

[0091] Local anomaly scoring: In the low-dimensional feature space, calculate the local anomaly score S through the reconstruction error i : The local anomaly score reflects the degree to which the flight behavior of a single drone deviates from the normal mode.

[0092] As an option, the edge node can also perform binary processing (marking anomalies and normals) on the preliminary anomaly score and upload it to the cloud server together with the dimensionality-reduced data.

[0093] Exemplary implementation: Assume that there are 10 drones in the management area of a certain edge node. After dimensionality reduction processing, each drone generates a 5-dimensional feature vector and obtains the corresponding local anomaly score through the reconstruction error.

[0094] 3. The cloud server integrates the local anomaly scores The cloud server receives the dimensionality-reduced data and local anomaly scores from all edge nodes and globally integrates the multi-modal data of the drone cluster.

[0095] Global anomaly score optimization: The cloud server utilizes the interaction graph structure of the UAV cluster, and through anomaly score propagation and sparsity optimization, generates globally consistent anomaly scores. The optimized anomaly scores are more accurate and can reflect the collaborative behavior patterns among UAVs.

[0096] In a possible implementation, the cloud server also performs trend analysis on the anomaly score results in combination with historical data to predict potential anomaly risks.

[0097] It should be noted that the global optimization process of the cloud server makes full use of computing resources, can handle complex anomaly detection tasks, and generates more valuable detection results.

[0098] 4. The cloud server issues response instructions Based on the optimized global anomaly scores, the cloud server generates corresponding response instructions and sends them to the UAV side through the communication link. The response instructions include but are not limited to the following: Adjust the flight path: Re-plan the flight path of the anomalous UAV to avoid potential risks; No-fly instruction: Issue a no-fly instruction to the UAV with serious anomalous behavior to restrict its activity range; Return-to-base command: Instruct the UAV to return to a specified safe location; Communication interruption: Cut off the communication link of the anomalous UAV to prevent it from being maliciously controlled.

[0099] In a possible implementation, the cloud server also sends the anomaly detection results and optimized model parameters to the edge nodes for updating the local detection process in the next round.

[0100] Example: Suppose a UAV is marked as high-risk in anomaly detection, and its anomaly score S i final exceeds the preset threshold. The cloud server sends it a return-to-base command and at the same time notifies other UAVs to adjust their paths to avoid the risk area.

[0101] Through the above steps, the present invention realizes UAV anomaly detection and response based on a cloud-edge collaborative architecture. The real-time data collection at the UAV side, the local anomaly detection at the edge nodes, and the global optimization at the cloud server are clearly divided, and the computing capabilities of each layer are fully utilized. The design of the response instructions ensures the rapid handling of anomalous behaviors and the safety of the UAV cluster operation. The formulas and technical details are fully disclosed, and it has good industrial realizability.

[0102] In the present invention, the response instruction includes one or more of the following: adjusting the flight path to avoid the abnormal area; sending a return command to the abnormal drone; remotely disconnecting the communication link of the abnormal drone; issuing a no-fly zone warning and adjusting the scope of the no-fly zone.

[0103] In the present invention, as part of the cloud-edge collaborative detection and response, the response instruction is a specific operation command generated according to the global abnormal score result, which is used to promptly handle the abnormal behavior of the drone and ensure the overall safety of the drone cluster. The response instruction has the characteristics of diversification and flexibility, and different operation methods can be selected according to the type and severity of the abnormal behavior, or a combination of multiple operation methods can be executed.

[0104] In this embodiment 1. Adjusting the flight path to avoid the abnormal area When the abnormal behavior of a certain drone is related to the environmental risk of a specific area, the cloud server will generate a response instruction to adjust the flight path.

[0105] Exemplary implementation method: If it is detected that the drone is approaching an abnormal area (such as an obstacle, a signal blind area or a meteorological risk area) during flight, the cloud server will re-plan its flight path to guide the drone to avoid the abnormal area.

[0106] The adjusted path information is sent to the drone side through the communication link, and the drone independently executes the change of the flight path.

[0107] It should be noted that the calculation of the path adjustment is based on the global abnormal detection result, and combines the current position and the target point of the drone to ensure the safety and optimality of the new path.

[0108] 2. Sending a return command to the abnormal drone For drones with a relatively high abnormal score and a relatively large behavior risk, the cloud server can directly issue a return instruction, requiring the drone to return to a specified safe location (such as the take-off point or the relay station).

[0109] Specifically: The cloud server calculates the return path according to the real-time position of the drone and sends the path information as part of the instruction to the drone; After receiving the return command, the drone pauses the current task and immediately executes the return operation.

[0110] Example: During the flight of a certain drone, communication anomalies occur and it deviates from the preset flight trajectory. The cloud server judges that its behavior has a high risk through the abnormal score, so it issues a return command to guide it to a safe area.

[0111] 3. Remotely disconnect the communication link of the abnormal drone If the cloud server detects potential malicious manipulation or abnormal behavior in a certain drone (such as deviating from the flight plan, unauthorized communication behavior, etc.), it can immediately issue an instruction to disconnect the communication link and cut off its communication with the outside world.

[0112] Specific implementation method: The cloud server sends a link disconnection instruction through the network node identifier of the drone; After receiving the instruction, the drone stops the operation of the communication module and enters the isolation mode.

[0113] It should be noted that the disconnection of the communication link is an important means to handle malicious behavior or major abnormalities, and is usually used in conjunction with other response instructions (such as return flight command or no-fly zone command).

[0114] 4. Issue a no-fly zone warning and adjust the scope of the no-fly zone When the abnormal behavior involves the safety risks of a specific flight area (such as a high-density drone aggregation area, above important facilities, etc.), the cloud server can issue a no-fly zone warning and dynamically adjust the scope of the no-fly zone.

[0115] Specifically: The cloud server delimits a new no-fly zone scope according to the abnormal detection result and broadcasts the updated no-fly zone information to all drones; After receiving the no-fly zone warning, the drone will automatically adjust the flight plan to avoid entering the no-fly zone.

[0116] Example: If there are high abnormal score behaviors of multiple drones in a certain area, the cloud server marks this area as a potential high-risk area, delimits the no-fly zone scope in real time, and notifies all drones in the cluster to update the flight path.

[0117] As a key part of the present invention, the response instruction effectively deals with the abnormal behavior of the drone through flexible and diverse operation methods. In a dynamic and safe operating environment, response measures such as adjusting the flight path, issuing a return flight command, disconnecting the communication link, and issuing a no-fly zone warning provide multi-level guarantees for the security of the drone cluster. The execution mechanism of the instruction is publicly sufficient, with practical operability and good industrial application prospects.

[0118] In the present invention, the method further includes a dynamic update mechanism, including the following: Based on the flight data collected in real time, the cloud server continuously optimizes the high-dimensional data dimensionality reduction model and the abnormal detection model; the optimized model parameters are sent to the drone side through the edge node to achieve the dynamic adaptation and continuous learning ability of the entire system.

[0119] The method of the present invention also includes a dynamic update mechanism, which aims to improve the system's adaptability to complex flight environments and new abnormal behaviors by continuously optimizing the high-dimensional data dimensionality reduction model and the anomaly detection model. Specifically, the dynamic update mechanism is based on the flight data collected in real time, utilizes the powerful computing power of the cloud server to optimize the dimensionality reduction model and the detection model, and distributes the optimized model parameters to the UAV side through the edge nodes. This mechanism effectively realizes the dynamic adaptation and continuous learning ability of the entire system.

[0120] It should be noted that the dynamic update mechanism is applicable to various changes in flight scenarios, including changes in environmental characteristics, the introduction of new risk patterns, and the dynamic expansion of the UAV cluster scale, and can ensure that the system always maintains efficient and reliable anomaly detection capabilities.

[0121] In this embodiment 1. Model optimization based on flight data collected in real time Real-time data collection: The multi-modal flight data collected by the UAV side, including trajectory data, sensor data, and visual data, is uploaded to the edge node and finally aggregated to the cloud server.

[0122] The real-time data is not only used for anomaly detection in the current round but also used to update the training data set of the model.

[0123] Model optimization objectives: Based on the newly collected data, the cloud server continuously optimizes the following models: High-dimensional data dimensionality reduction model: Optimize the dimensionality reduction mapping function so that the reduced feature space can better retain the local characteristics and global structure of multi-modal data; Anomaly detection model: Optimize the reconstruction function and the anomaly score calculation method to enhance the model's ability to identify new abnormal behaviors.

[0124] Optimization method: The cloud server uses the batch incremental training method to optimize the model, specifically including: Online training of incremental data to ensure that the model can quickly adapt to the newly collected data; Regular retraining, combining historical data and incremental data to further improve the global performance of the model.

[0125] It should be noted that the model optimization process combines supervised learning and unsupervised learning, which can not only utilize existing labeled data (such as abnormal samples) but also extract potential patterns from new data.

[0126] 2. Generation of optimized model parameters Parameter Optimization of High-Dimensional Data Dimensionality Reduction Model: By optimizing the regularization objective function, a new dimensionality reduction mapping matrix W is generated to improve the separability and feature retention ability of low-dimensional embedded data.

[0127] Parameter Optimization of Anomaly Detection Model: The parameters of the reconstruction function (such as the weights of the neural network) and the scoring calculation parameters (such as the sparsity constraint weights) are retrained based on new data to generate updated versions.

[0128] 3. Distribution and Application of Model Parameters Distribution Mechanism: The cloud server packs the optimized model parameters and sends them to the drone side through the edge node; The edge node, as a relay device, ensures that the model parameters can be transmitted to all drones efficiently and with low latency.

[0129] Dynamic Loading of the Model: After receiving the new model parameters, the drone side loads the updated model in real time for the next round of anomaly detection; To avoid interruption of the detection task, the drone side adopts a dynamic switching mechanism, that is, it continues to run the old model when loading the new model and switches to the new model after the transition is completed. 4. Effects of the Dynamic Update Mechanism Enhanced Adaptive Ability: As the flight environment changes (such as weather, communication conditions, risk level of the flight area, etc.), the system can automatically adapt to new characteristics through model updates; When the scale of the drone cluster expands or shrinks, the updated model can quickly adapt to the new operating state.

[0130] Improved Continuous Learning Ability: The system gradually accumulates the ability to recognize new abnormal behaviors through iterative training of real-time data, improving the accuracy of anomaly detection.

[0131] Therefore, through the dynamic update mechanism, the present invention realizes the continuous optimization of the high-dimensional data dimensionality reduction model and the anomaly detection model, enhancing the dynamic adaptability and continuous learning ability of the system. Through the unified management of the cloud server and the efficient transmission of the edge node, the updated model can be quickly deployed to the drone side, providing higher accuracy and robustness for the anomaly detection of the drone cluster. The mechanism is fully disclosed and has good industrial applicability and technological innovation.

[0132] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting abnormal flight data of unmanned aerial vehicles in a cloud server cluster, characterized in that: The following steps are involved: S1. Multimodal data collection: collect the flight trajectory data, sensor data and visual data of the drone and fuse them to generate a high-dimensional feature matrix; S2. Construction of multimodal interaction graph: constructing an interaction graph of the drone cluster based on a high-dimensional feature matrix, wherein the interaction graph includes nodes and edges, where nodes represent drones and edges represent feature similarities between drones; S3, high-dimensional data dimensionality reduction: based on the interaction graph, a regularized variational dimensionality reduction method is used to map the high-dimensional feature matrix to a low-dimensional embedding space; S4, local anomaly detection: calculate the preliminary anomaly score of the drone through the reconstruction error of low-dimensional data; S5, global anomaly propagation: optimize anomaly scores based on the interaction graph of the drone cluster and generate global anomaly detection results; S6, Cloud-edge collaborative detection and response: The edge node completes local detection, the cloud server completes global optimization, and sends abnormal response instructions to the drone.

2. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 1 is characterized in that: The step S2 includes the following contents: Generate feature vectors using multimodal data from drones; By calculating the similarity of feature vectors between drones, the association strength between drones is determined and used as the edge weight of the interaction graph; The edge weights of the interaction graph are used to describe the feature association relationships between drones.

3. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 1 is characterized in that: In step S3, the high-dimensional multimodal data of the drone is mapped to a low-dimensional space by a regularized dimensionality reduction method, and the dimensionality reduction method has the following characteristics: Preserve the local characteristics of multimodal data while enhancing the global consistency of low-dimensional spatial data; Reduce the redundancy of UAV multimodal data and ensure the separability of abnormal patterns in low-dimensional space.

4. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 3 is characterized in that: In the process of dimensionality reduction of high-dimensional data, the association relationship between drones is constructed to ensure that the low-dimensional embedded data can reflect the interaction characteristics of multiple drones. The association relationship is calculated based on the spatial similarity and communication frequency between drones.

5. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 1 is characterized in that: The step S4 includes the following contents: The multimodal data of the UAV is reconstructed using the low-dimensional feature space after dimensionality reduction; Calculate the difference between the original data and the reconstructed data. The larger the difference, the more the flight behavior of the drone deviates from the normal mode. The difference value is used as the preliminary anomaly score of the drone.

6. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 1 is characterized in that: The step S5 includes the following contents: Based on the interaction graph of the drone cluster, the local anomaly score is propagated in the graph structure; During the propagation process, the global consistency of anomaly scores is enhanced, and the scoring results are optimized in combination with the topological structure of the graph; Through global propagation, the final anomaly score of each drone is generated, and the higher the score value, the greater the possibility of abnormal behavior.

7. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 6 is characterized in that: The sparsity optimization constraint is added to the global anomaly propagation to ensure that in a cluster environment, only a few drones are marked as abnormal, and the behavior of all drones is not judged as abnormal, thereby enhancing the accuracy and credibility of detection.

8. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 1 is characterized in that: The step S6 includes the following contents: The drone collects multimodal flight data in real time and uploads it to the edge node; Edge nodes complete local anomaly detection, including high-dimensional data dimensionality reduction and local anomaly scoring; The cloud server receives data from all edge nodes and integrates local anomaly scores to complete global anomaly score optimization; The cloud server sends response instructions to the drone based on the final anomaly detection results.

9. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 1, characterized in that: The response instruction includes one or more of the following contents: Adjust flight paths to avoid unusual areas; Issue a return order to the abnormal drone; Remotely disconnect the communication link of abnormal drones; Issue no-fly zone warnings and adjust no-fly zone scopes.

10. The method for detecting abnormal flight data of unmanned aerial vehicles for cloud server clusters according to claim 1, characterized in that: The method also includes a dynamic update mechanism, including the following: Based on the real-time collected flight data, the cloud server continuously optimizes the high-dimensional data dimensionality reduction model and anomaly detection model; The optimized model parameters are sent to the drone end through the edge node to achieve dynamic adaptation and continuous learning capabilities of the entire system.

Images