Low-altitude security event automatic classification and response method based on cloud computing platform

By constructing an automatic classification and response method for low-altitude safety events based on a graph convolutional neural network and a spatiotemporal attention mechanism on a cloud computing platform, the problem of insufficient multi-source data modeling in existing technologies is solved, and high-precision, fast-response low-altitude safety event processing is achieved, improving the system's adaptability and automation level.

CN120877560AInactive Publication Date: 2025-10-31ZHANPENG JIAYE INFORMATION IND CO LTD
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202511070596.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-31
Publication Date
2025-10-31
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing low-altitude safety monitoring methods lack a unified modeling mechanism for multi-source heterogeneous data, making it impossible to efficiently construct airspace behavior maps and dynamic evolution paths. Event classification algorithms fail to fully incorporate graph neural networks and spatiotemporal attention mechanisms, resulting in the inability to accurately identify complex behavior patterns. The event level assessment process cannot be modeled in conjunction with event type identification. Policy functions lack vector-level response mapping capabilities, and the response command distribution chain and cloud computing platform are not deeply integrated, leading to low identification accuracy, slow response speed, and low automation.

Method used

A heterogeneous spatial behavior graph is constructed using graph convolutional neural networks and spatiotemporal attention mechanisms. A dual-branch intelligent classification network is used for event type identification and level evaluation. A multi-objective policy function is designed to generate response instructions. The instructions are then rapidly distributed and incrementally trained and optimized through a cloud computing platform to achieve end-to-end automatic event classification and response.

Benefits of technology

It significantly improves the accuracy and response speed of low-altitude safety incident identification, achieves high-precision discrimination of complex flight behavior, shortens the response chain, reduces human intervention, forms a closed-loop mechanism of identification-response-optimization, and enhances the system's adaptive capability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120877560A_ABST
    Figure CN120877560A_ABST
Patent Text Reader

Abstract

The invention discloses a low-altitude security event automatic classification and response method based on a cloud computing platform, and the method comprises the following steps: S1, collecting monitoring data in a low-altitude region, and carrying out the preprocessing of the monitoring data; s2, constructing a heterogeneous airspace behavior map, and extracting an evolution path and a behavior sequence pattern between nodes; s3, inputting the feature embedding matrix into a double-branch classification network, and respectively executing event type identification and event level evaluation; s4, constructing an event index, and designing a multi-objective strategy function; s5, the response instruction set is issued to the multi-source processing terminal through the cloud computing platform, and sealing control, interference, tracking or alarm operation is executed; and S6, performing incremental training on the graph convolutional neural network and the double-branch classification network, continuously updating network parameters and optimizing a multi-objective strategy function. According to the invention, high-precision identification and rapid intelligent response of low-altitude security events are realized, and the automation level and emergency disposal efficiency of airspace management and control are remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of low-altitude safety technology, and in particular to a method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform. Background Technology

[0002] In recent years, with the opening of urban airspace resources and the rapid development of the low-altitude aviation industry, a large number of unmanned aerial vehicles (UAVs) (such as multi-rotor UAVs, fixed-wing UAVs, and vertical takeoff and landing aircraft) have been widely used in various industries, including express delivery and logistics, urban inspection, agricultural remote sensing, and environmental monitoring. However, low-altitude safety incidents have also become increasingly frequent, including illegal flights, incursions into no-fly zones, prolonged loitering, communication link interruptions, and unauthorized mission execution. These abnormal behaviors not only threaten the lives and property of urban residents but may also disrupt the normal operation of civil aviation routes, sensitive areas, and core facilities. Therefore, how to achieve refined and intelligent real-time control of low-altitude airspace has become a significant technical challenge in the construction of smart air traffic control and smart cities.

[0003] Existing low-altitude safety monitoring methods primarily rely on fixed radar systems, video surveillance equipment, or distributed communication link monitoring technology. While these technologies can detect and identify some aircraft to a certain extent, they generally suffer from insufficient accuracy, poor data fusion capabilities, and slow response times. In terms of event classification, traditional methods often rely on rule-based matching, template library comparison, or human experience to determine whether aircraft behavior is abnormal. These methods heavily depend on prior knowledge, lack the ability to model complex behavioral evolution paths, and struggle to identify new and mutated abnormal flight patterns. Regarding response mechanisms, current systems mostly separate identification and response, resulting in long information flow chains, disconnected response command generation, and a lack of context awareness in policy execution. This leads to delayed responses in practical applications and a high risk of safety incident propagation.

[0004] Furthermore, traditional low-altitude safety management systems are generally centralized, relying on the computing power of local terminals or edge devices for data processing. This makes them ill-equipped to handle complex safety incidents involving large areas, cross regions, and multiple targets. Regarding multi-source data fusion, there is a lack of a unified modeling mechanism for multimodal information such as radar data, optical images, and communication signals. This makes it impossible to extract effective behavioral features and decision-making basis from heterogeneous data, affecting classification accuracy. In terms of algorithms, most existing models can only process static inputs, with limited ability to model the temporal dynamic evolution of aircraft behavior and a lack of in-depth exploration of sequence evolution patterns and cross-node behavioral dependencies. Especially in scenarios facing a surge in the number of aircraft, rapid changes in event severity, and unclear target states, traditional methods are increasingly inadequate, easily leading to misjudgments or omissions.

[0005] Current research is beginning to incorporate artificial intelligence and machine learning methods to assist in identification and response. Examples include using convolutional neural networks to analyze image information, employing long short-term memory networks to capture behavioral sequence features, and using Bayesian methods for risk probability assessment. However, these methods generally remain at the level of single-modal processing and single-task modeling, lacking a unified structure to support the entire end-to-end event handling process. Furthermore, event identification results are difficult to directly map to specific response strategies, often requiring manual intervention or the involvement of scheduling and control systems to generate handling plans, thus reducing automation and response efficiency.

[0006] In summary, existing technologies still face numerous bottlenecks and shortcomings in the automatic identification and response to low-altitude safety incidents. These mainly manifest in the following ways: First, the lack of a unified modeling mechanism for multi-source heterogeneous data hinders the efficient construction of airspace behavior maps and dynamic evolution paths. Second, the failure of event classification algorithms to fully incorporate graph neural networks and spatiotemporal attention mechanisms results in the inaccurate identification of complex behavioral patterns. Third, the inability to coordinate event level assessment with event type identification limits the accuracy of multi-level policy responses. Fourth, the lack of vector-level response mapping capabilities in policy functions prevents precise matching and policy generation to specific event contexts. Fifth, the failure to deeply integrate response command distribution links and feedback mechanisms with cloud computing platforms restricts the realization of response loops and continuous model optimization.

[0007] Therefore, how to provide an automatic classification and response method for low-altitude safety incidents based on a cloud computing platform is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose an automatic classification and response method for low-altitude safety events based on a cloud computing platform. This invention fully integrates graph convolutional neural networks, spatiotemporal attention mechanisms, and dual-branch intelligent classification networks. It describes in detail how to extract features, identify events, and assess levels from low-altitude monitoring data by constructing heterogeneous airspace behavior maps, and generates a dynamic response instruction set by combining context-aware multi-objective policy functions. It has the advantages of high recognition accuracy, fast response speed, and strong adaptability.

[0009] The automatic classification and response method for low-altitude safety incidents based on a cloud computing platform according to embodiments of the present invention includes the following steps:

[0010] S1. Collect monitoring data in the low-altitude region and perform preprocessing;

[0011] S2. Input the preprocessed monitoring data into the graph convolutional neural network to construct a heterogeneous spatial behavior map, and combine the spatiotemporal attention mechanism to extract the evolution path and behavior sequence pattern between nodes to generate a feature embedding matrix.

[0012] S3. Input the feature embedding matrix into the dual-branch classification network, perform event type recognition and event level evaluation respectively, and output event type label and level score vector;

[0013] S4. Based on the event type labels and rating vectors, construct an event index and design a multi-objective strategy function based on vector matching and contextual constraints to generate a set of response instructions;

[0014] S5. Send the set of response instructions to the multi-source processing terminal through the cloud computing platform to perform blocking, interference, tracking or alarm operations, and collect terminal feedback information and regional status change data;

[0015] S6. Based on terminal feedback information and regional state change data, incremental training is performed on the graph convolutional neural network and the dual-branch classification network to continuously update network parameters and optimize the multi-objective policy function.

[0016] Optionally, the monitoring data includes radar data, optical image data, and communication link data.

[0017] Optionally, the preprocessing includes noise filtering, missing value imputation, outlier removal, timestamp alignment, spatial projection mapping, and data standardization.

[0018] Optionally, the dual-branch classification network includes an event type recognition branch and an event level evaluation branch. The event type recognition branch consists of residual convolutional units and a fully connected classifier, used to extract temporal behavioral features from the feature embedding matrix and output event category labels. The event level evaluation branch adopts a multi-scale attention fusion mechanism to jointly model local feature weights and global context information, and outputs an event level score vector through a multilayer perceptron. The two branches share the underlying encoder and are trained end-to-end through a joint loss function, which includes cross-entropy loss and mean squared error loss.

[0019] Optionally, S2 specifically includes:

[0020] S21. Based on the preprocessed monitoring data, construct a node set V = {v1, v2, ..., v...} n} and the edge set E={e ij |i,j=1,2,…,n}, where each node v i Indicates at time t i The target state unit is mapped from radar data, optical image data, and communication link data, with side e ij Indicates v i With v j Behavioral dependencies in time series;

[0021] S22. Based on the node set and edge set, construct a heterogeneous spatial behavior graph G = (V, E, A, W), where A is the adjacency matrix, defined as A(i, j) = 1 representing node v. i With v j There are information connections, A(i,j)=0 indicates no connection, W is the edge weight matrix, and the element W(i,j) represents the connection weight calculated by the joint metric function of radar echo intensity and communication interference;

[0022] S23. Input the heterogeneous spatial behavior graph G into the graph convolutional neural network and perform multiple rounds of graph convolution propagation. Calculate the node feature matrix H of the (l+1)th layer after each round of propagation using the following formula. (l+1) :

[0023]

[0024] Among them, H (l) Let D represent the feature matrix of the nodes in the l-th layer, and let D be the degree matrix of the adjacency matrix A. Let σ represent the trainable weight matrix of layer l, σ be the non-linear activation function, α be the global fusion coefficient, P be the number of spatiotemporal attention paths, and β be the training weight matrix of layer l. p T represents the spatiotemporal attention weight of the p-th path. p Let W be the transition matrix of the p-th path, representing the normalized relative time intervals between nodes within the path. p This is the path transformation matrix;

[0025] S24, Regarding the node characteristic matrix H (l) Perform behavioral sequence modeling and path extraction, and construct a feature embedding matrix:

[0026]

[0027] γ ij =exp(-κ·d ij )·(1-η·ξ ij );

[0028] Where Z represents the feature embedding matrix, φ represents the embedding compression function used to extract stable representations, and n represents the number of nodes. Represents node v i v j eigenvectors, δ ij Indicates v i v j The time interval difference between them, ψ is the joint function used to fuse node features and time interval, γ ij For node v i v j The mutual influence coefficient is calculated based on spatial projection and communication loss. exp(·) represents the natural exponential function with base e, and dij ξ represents spatial distance. ij This represents the normalized value of the communication interference intensity, where κ and η are regular scaling factors.

[0029] Optionally, during the graph convolution propagation process, the heterogeneous spatial behavior graph G constructs a dual-weighted fusion graph convolutional unit, embeds an edge weight constraint matrix and a path evolution matrix to control the feature propagation direction, superimposes a step path attention gating structure to adjust the information transmission frequency between nodes, and maintains the spatial topological consistency of the graph through a structure preservation factor.

[0030] Optionally, S3 specifically includes:

[0031] S31. Input the feature embedding matrix Z into the shared bottom encoder of the dual-branch classification network, perform residual convolution operation and temporal compression processing to obtain the intermediate feature matrix U:

[0032]

[0033] Where ρ is the combination of linear transformation and normalization function, ω l The fusion weight for the l-th residual channel is... It is a two-channel convolutional unit, θ l Here are the parameters for the convolution kernel of the l-th channel, * indicates the convolution operation, and L represents the total number of residual stacking layers;

[0034] S32. Input the intermediate feature matrix U into the event type recognition branch and the event level evaluation branch respectively, and extract the context matrix ξ = [ξ1, ξ2, ..., ξ]. K ], where each ξ k Let K represent the context vector derived from the intermediate feature matrix U, and K represent the dimension of the context vector, satisfying k∈{1,2,...,K}. The event type identification branch outputs the event type label y through a fully connected classifier, and the event level evaluation branch uses a multi-scale attention mechanism to generate a level rating vector s. The level rating vector is calculated as follows:

[0035]

[0036] Where Ω is the mapping function, M is the scale number, and λ mk U represents the attention weight of the k-th channel at the m-th scale. mk χ represents the feature representation of channel k at the m-th scale. k μ represents the context encoding vector. m Let represent the scale gating factor, π be the cross-scale feature aggregation function, ν be the global attention adjustment coefficient, and ∈ be the context compression factor. U represents the weighting factor for the j-th feature dimension, m represents the number of feature dimensions, and U represents the weighting factor for the j-th feature dimension.j Let represent the feature representation of the j-th feature dimension, and tanh(·) represent the hyperbolic tangent function;

[0037] S33. Based on the event type label y and the rating vector s, the dual-branch classification network is trained end-to-end using a joint loss function, namely cross-entropy loss. With mean square error loss Weighted sum:

[0038]

[0039] in, This represents the joint loss function, where τ1 and τ2 are the loss weighting coefficients, C is the total number of event types, and δ... i Let be the i-th indicator variable of the true label, and let log2(·) denote the natural exponential function with base 2. i Let be the predicted probability for the i-th label, D be the number of dimensions in the rating vector, and s be the predicted probability. d For the d-th dimension score, The corresponding target score.

[0040] Optionally, S4 specifically includes:

[0041] S41. Based on the event type label y and the rating vector s, jointly map the type label and the rating value to construct the event index matrix q. The vector dimension of the event index matrix is ​​C+D, where C is the total number of event types and D is the number of dimensions of the rating vector.

[0042] S42. Construct a candidate vector set for strategies Θ = {θ1, θ2, ..., θ} T}, where T is the number of policy templates, and each candidate policy vector θ t It is a real-valued vector of length C+D+K, where K represents the dimension of the context vector;

[0043] S43. Concatenate the event index matrix q and the context matrix ξ directly to form the target event matrix ζ = [q‖ξ], and calculate the multi-target matching score ρ between each candidate policy vector and the target event matrix. t This generates a set of response instructions. The multi-target matching score is calculated as follows:

[0044]

[0045] Where, γ t Let ζ represent the t-th response instruction vector that matches the event. i Let θ represent the target event matrix of the i-th vector dimension. t,i Let λ represent the t-th candidate policy vector of the i-th vector dimension. iLet ψ represent the policy priority coefficient of the i-th vector dimension, ψ be the joint function used to non-linearly combine type, level, and context, and σ be the policy priority coefficient of the i-th vector dimension. i Let φ represent the context modulation coefficient of the i-th vector dimension, φ represent the embedding compression function, and ξ represent the context modulation coefficient of the i-th vector dimension. j χ represents the j-th context vector derived from the intermediate feature matrix U. j Let be the aggregate weight of the j-th context vector, and ∈ be a stabilizing term to prevent the denominator from being zero. This is the activation threshold for the policy.

[0046] Optionally, the event index matrix q is constructed based on the event type label y and the rating vector s. By introducing a label vector encoder and a rating weight mapping function, the outputs corresponding to y and s are concatenated, and a unified vector representation Q is generated through a nonlinear projection network. This Q is then combined with the context matrix ξ to construct the event index matrix. in The outer product operation is used to express the bidirectional interactive relationship between the event and the context matrix, enhancing the distinctiveness and context sensitivity of the event index.

[0047] Optionally, S5 specifically includes:

[0048] S51. Perform structured encoding on the response instruction set Γ to construct an instruction transmission vector stream. Each vector corresponds to an event response unit, which includes a policy index, target area encoding, operation weight vector, and time scheduling identifier field.

[0049] S52. The instruction transmission vector stream is pushed in parallel through the distributed cloud scheduling platform. A dynamic instruction distribution strategy based on task priority and geographical coverage is adopted to push each response vector to the processing terminal. The terminal type is automatically matched by the static mapping table of the central node according to the event category label.

[0050] S53. After receiving the instruction, each terminal performs blockade, interference transmission, target tracking or audible and visual alarm operations based on the local status parsing operation field. During the execution, it continuously collects environmental feedback signals and regional status change data, and transmits the disposal process data back to the cloud platform.

[0051] S54. After receiving feedback information and regional status change data uploaded by all terminals, the cloud platform performs timestamp alignment and low-altitude area archiving, constructs a feedback record stream, and writes the feedback record stream into the task feedback database for incremental training and response strategy evaluation of graph convolutional neural networks and dual-branch classification networks.

[0052] The beneficial effects of this invention are:

[0053] First, this invention, by introducing graph convolutional neural networks and a spatiotemporal attention mechanism, enables joint modeling of multi-source heterogeneous low-altitude monitoring data (including radar data, optical image data, and communication link data) to construct a dynamically evolving airspace behavior map. This effectively extracts the temporal dependencies and spatial structural correlations between aircraft behaviors, significantly improving the ability to identify abnormal flight patterns. Compared to existing methods that rely on rule templates or static feature matching, this invention achieves higher discrimination accuracy in complex flight behaviors and multi-target scenarios.

[0054] Secondly, this invention employs a dual-branch classification network structure to collaboratively model event type identification and event level assessment. This achieves fine-grained discrimination of low-altitude safety event categories and introduces a multi-scale attention mechanism to model event severity, making the output results more hierarchical and targeted. Simultaneously, a multi-objective policy matching function based on event labels, level scores, and contextual information is constructed, enabling a direct mapping from event identification to response policy generation, significantly shortening the response chain and reducing manual intervention.

[0055] Finally, leveraging the distributed processing capabilities of a cloud computing platform, this invention can rapidly distribute response command sets to multi-source processing terminals. By collecting terminal feedback data and monitoring regional state changes, it incrementally trains the network, enabling dynamic updates of network parameters and continuous optimization of policy functions, forming a closed-loop mechanism of identification-response-optimization. Overall, this invention effectively improves the perception accuracy, response efficiency, and system adaptability of low-altitude airspace safety incidents, overcoming the limitations of existing technologies in terms of real-time performance, accuracy, and system coordination. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of the automatic classification and response method for low-altitude safety incidents based on a cloud computing platform proposed in this invention;

[0058] Figure 2 This is a schematic diagram illustrating the behavior graph construction and feature extraction of the automatic classification and response method for low-altitude safety incidents based on a cloud computing platform proposed in this invention.

[0059] Figure 3 This is a flowchart of the dual-branch classification network structure of the automatic classification and response method for low-altitude safety events based on a cloud computing platform proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figure 1-3 An automatic classification and response method for low-altitude safety incidents based on a cloud computing platform includes the following steps:

[0062] S1. Collect monitoring data in the low-altitude region and perform preprocessing;

[0063] S2. Input the preprocessed monitoring data into the graph convolutional neural network to construct a heterogeneous spatial behavior map, and combine the spatiotemporal attention mechanism to extract the evolution path and behavior sequence pattern between nodes to generate a feature embedding matrix.

[0064] S3. Input the feature embedding matrix into the dual-branch classification network, perform event type recognition and event level evaluation respectively, and output event type label and level score vector;

[0065] S4. Based on the event type labels and rating vectors, construct an event index and design a multi-objective strategy function based on vector matching and contextual constraints to generate a set of response instructions;

[0066] S5. Send the set of response instructions to the multi-source processing terminal through the cloud computing platform to perform blocking, interference, tracking or alarm operations, and collect terminal feedback information and regional status change data;

[0067] S6. Based on terminal feedback information and regional state change data, incremental training is performed on the graph convolutional neural network and the dual-branch classification network to continuously update network parameters and optimize the multi-objective policy function.

[0068] This invention proposes an end-to-end event classification and response method for low-altitude monitoring scenarios, which integrates a closed-loop mechanism of data collection, identification, evaluation, response and self-learning to achieve fully automated processing of low-altitude safety events, significantly improving the automation level and handling efficiency of low-altitude airspace event perception.

[0069] In this embodiment, the monitoring data includes radar data, optical image data, and communication link data.

[0070] This invention, by jointly inputting radar data, optical image data, and communication link data, fully integrates multimodal information, effectively enhancing the accuracy of monitoring complex aircraft behavior and the ability to express features, and improving the robustness and generalization of event recognition.

[0071] In this embodiment, the preprocessing includes noise filtering, missing value imputation, outlier removal, timestamp alignment, spatial projection mapping, and data standardization.

[0072] This invention eliminates data noise and structural defects by standardizing and preprocessing the raw monitoring data, ensuring the stability and effectiveness of subsequent mapping, modeling, graph neural network propagation, and classification learning processes, and providing a high-quality data foundation for accurate modeling.

[0073] In this embodiment, the dual-branch classification network includes an event type recognition branch and an event level evaluation branch. The event type recognition branch consists of residual convolutional units and a fully connected classifier, used to extract temporal behavioral features from the feature embedding matrix and output event category labels. The event level evaluation branch adopts a multi-scale attention fusion mechanism to jointly model local feature weights and global context information, and outputs an event level score vector through a multilayer perceptron. The two branches share the underlying encoder and are trained end-to-end through a joint loss function, which includes cross-entropy loss and mean squared error loss.

[0074] This invention constructs a dual-branch classification network architecture, which models event type and event level separately. By sharing the underlying encoder and designing a joint loss function, it improves the consistency and accuracy of classification, and realizes the multi-dimensional expression and efficient differentiation of complex events.

[0075] In this embodiment, S2 specifically includes:

[0076] S21. Based on the preprocessed monitoring data, construct a node set V = {v1, v2, ..., v...} n} and the edge set E={e ij |i,j=1,2,…,n}, where each node v i Indicates at time t i The target state unit is mapped from radar data, optical image data, and communication link data, with side e ij Indicates v i With v j Behavioral dependencies in time series;

[0077] S22. Based on the node set and edge set, construct a heterogeneous spatial behavior graph G = (V, E, A, W), where A is the adjacency matrix, defined as A(i, j) = 1 representing node v. i With v j There are information connections, A(i,j)=0 indicates no connection, W is the edge weight matrix, and the element W(i,j) represents the connection weight calculated by the joint metric function of radar echo intensity and communication interference;

[0078] S23. Input the heterogeneous spatial behavior graph G into the graph convolutional neural network and perform multiple rounds of graph convolution propagation. Calculate the node feature matrix H of the (l+1)th layer after each round of propagation using the following formula.(l+1) :

[0079]

[0080] Among them, H (l) Let D represent the feature matrix of the nodes in the l-th layer, and let D be the degree matrix of the adjacency matrix A. Let σ represent the trainable weight matrix of layer l, σ be the non-linear activation function, α be the global fusion coefficient, P be the number of spatiotemporal attention paths, and β be the training weight matrix of layer l. p T represents the spatiotemporal attention weight of the p-th path. p Let W be the transition matrix of the p-th path, representing the normalized relative time intervals between nodes within the path. p This is the path transformation matrix;

[0081] S24, Regarding the node characteristic matrix H (l) Perform behavioral sequence modeling and path extraction, and construct a feature embedding matrix:

[0082]

[0083] γ ij =exp(-κ·d ij )·(1-η·ξ ij );

[0084] Where Z represents the feature embedding matrix, φ represents the embedding compression function used to extract stable representations, and n represents the number of nodes. Represents node v i v j eigenvectors, δ ij Indicates v i v j The time interval difference between them, ψ is the joint function used to fuse node features and time interval, γ ij For node v i v j The mutual influence coefficient is calculated based on spatial projection and communication loss. exp(·) represents the natural exponential function with base e, and d ij ξ represents spatial distance. ij This represents the normalized value of the communication interference intensity, where κ and η are regular scaling factors.

[0085] This invention constructs a heterogeneous spatial behavior graph and embeds a spatiotemporal attention mechanism, introduces a structured graph convolutional propagation method, and extracts behavioral paths and node evolution features, significantly enhancing the graph convolutional neural network's ability to express and understand temporal dependencies and spatial structure changes.

[0086] In this embodiment, during the graph convolution propagation process, the heterogeneous spatial behavior graph G constructs a dual-weighted fusion graph convolution unit, embeds an edge weight constraint matrix and a path evolution matrix to control the feature propagation direction, superimposes a step path attention gating structure to adjust the information transmission frequency between nodes, and maintains the spatial topological consistency of the graph through a structure preservation factor.

[0087] This invention embeds edge weight constraints and path evolution mechanisms on the basis of graph convolution, and introduces gating and structure preservation factors to control the propagation frequency and consistency of graph convolution, thereby enhancing the structural adaptability and representation ability of graph neural networks in complex spatial behavior graphs.

[0088] In this embodiment, S3 specifically includes:

[0089] S31. Input the feature embedding matrix Z into the shared bottom encoder of the dual-branch classification network, perform residual convolution operation and temporal compression processing to obtain the intermediate feature matrix U:

[0090]

[0091] Where ρ is the combination of linear transformation and normalization function, ω l The fusion weight for the l-th residual channel is... It is a two-channel convolutional unit, θ l Here are the parameters for the convolution kernel of the l-th channel, * indicates the convolution operation, and L represents the total number of residual stacking layers;

[0092] S32. Input the intermediate feature matrix U into the event type recognition branch and the event level evaluation branch respectively, and extract the context matrix ξ = [ξ1, ξ2, ..., ξ]. K ], where each ξ k Let K represent the context vector derived from the intermediate feature matrix U, and K represent the dimension of the context vector, satisfying k∈{1,2,...,K}. The event type identification branch outputs the event type label y through a fully connected classifier, and the event level evaluation branch uses a multi-scale attention mechanism to generate a level rating vector s. The level rating vector is calculated as follows:

[0093]

[0094] Where Ω is the mapping function, M is the scale number, and λ mk U represents the attention weight of the k-th channel at the m-th scale. mk χ represents the feature representation of channel k at the m-th scale. k μ represents the context encoding vector. m Let represent the scale gating factor, π be the cross-scale feature aggregation function, ν be the global attention adjustment coefficient, and ∈ be the context compression factor. U represents the weighting factor for the j-th feature dimension, m represents the number of feature dimensions, and U represents the weighting factor for the j-th feature dimension. j Let represent the feature representation of the j-th feature dimension, and tanh(·) represent the hyperbolic tangent function;

[0095] S33. Based on the event type label y and the rating vector s, the dual-branch classification network is trained end-to-end using a joint loss function, namely cross-entropy loss. With mean square error loss Weighted sum:

[0096]

[0097] in, This represents the joint loss function, where τ1 and τ2 are the loss weighting coefficients, C is the total number of event types, and δ... i Let be the i-th indicator variable of the true label, and let log2(·) denote the natural exponential function with base 2. i Let be the predicted probability for the i-th label, D be the number of dimensions in the rating vector, and s be the predicted probability. d For the d-th dimension score, The corresponding target score.

[0098] This invention enhances the ability to model and abstract behavior-embedded features by constructing residual convolution and multi-scale attention structures, achieving more context-aware and behavior-accurate classification and discrimination, and optimizes overall performance by using end-to-end training.

[0099] In this embodiment, S4 specifically includes:

[0100] S41. Based on the event type label y and the rating vector s, jointly map the type label and the rating value to construct the event index matrix q. The vector dimension of the event index matrix is ​​C+D, where C is the total number of event types and D is the number of dimensions of the rating vector.

[0101] S42. Construct a candidate vector set for strategies Θ = {θ1, θ2, ..., θ} T}, where T is the number of policy templates, and each candidate policy vector θ t It is a real-valued vector of length C+D+K, where K represents the dimension of the context vector;

[0102] S43. Concatenate the event index matrix q and the context matrix ξ directly to form the target event matrix ζ = [q‖ξ], and calculate the multi-target matching score ρ between each candidate policy vector and the target event matrix. t This generates a set of response instructions. The multi-target matching score is calculated as follows:

[0103]

[0104] Where, γ t Let ζ represent the t-th response instruction vector that matches the event. i Let θ represent the target event matrix of the i-th vector dimension. t,i Let λ represent the t-th candidate policy vector of the i-th vector dimension. i Let ψ represent the policy priority coefficient of the i-th vector dimension, ψ be the joint function used to non-linearly combine type, level, and context, and σ be the policy priority coefficient of the i-th vector dimension. i Let φ represent the context modulation coefficient of the i-th vector dimension, φ represent the embedding compression function, and ξ represent the context modulation coefficient of the i-th vector dimension. j χ represents the j-th context vector derived from the intermediate feature matrix U. j Let be the aggregate weight of the j-th context vector, and ∈ be a stabilizing term to prevent the denominator from being zero. This is the activation threshold for the policy.

[0105] This invention designs a multi-objective strategy function based on vector matching and context constraints, realizing direct mapping from event recognition results to response strategies, improving the targeting and real-time performance of strategy generation, and effectively opening up the closed-loop process of "recognition-response".

[0106] In this embodiment, the event index matrix q is constructed based on the event type label y and the rating vector s. By introducing a label vector encoder and a rating weight mapping function, the outputs corresponding to y and s are concatenated, and a unified vector representation Q is generated through a nonlinear projection network. The event index matrix is ​​then constructed by combining this Q with the context matrix ξ. in The outer product operation is used to express the bidirectional interactive relationship between the event and the context matrix, enhancing the distinctiveness and context sensitivity of the event index.

[0107] This invention introduces a nonlinear projection network through label vector encoding and scoring weight mapping mechanism, and combines it with a context matrix to achieve highly expressive event indexing, thereby enhancing the ability of event vectors to distinguish semantic differences and behavioral details.

[0108] In this embodiment, S5 specifically includes:

[0109] S51. Perform structured encoding on the response instruction set Γ to construct an instruction transmission vector stream. Each vector corresponds to an event response unit, which includes a policy index, target area encoding, operation weight vector, and time scheduling identifier field.

[0110] S52. The instruction transmission vector stream is pushed in parallel through the distributed cloud scheduling platform. A dynamic instruction distribution strategy based on task priority and geographical coverage is adopted to push each response vector to the processing terminal. The terminal type is automatically matched by the static mapping table of the central node according to the event category label.

[0111] S53. After receiving the instruction, each terminal performs blockade, interference transmission, target tracking or audible and visual alarm operations based on the local status parsing operation field. During the execution, it continuously collects environmental feedback signals and regional status change data, and transmits the disposal process data back to the cloud platform.

[0112] S54. After receiving feedback information and regional status change data uploaded by all terminals, the cloud platform performs timestamp alignment and low-altitude area archiving, constructs a feedback record stream, and writes the feedback record stream into the task feedback database for incremental training and response strategy evaluation of graph convolutional neural networks and dual-branch classification networks.

[0113] This invention constructs a structured encoding, distributed push, and multi-terminal execution feedback mechanism for response instructions, enabling rapid distribution, accurate execution, and dynamic acquisition of process data for response instructions, thus providing support for continuous system learning and adaptive optimization.

[0114] Example 1:

[0115] To verify the feasibility of this invention in practice, it was applied to a typical low-altitude security command center dispatch platform. This platform covers a wide area and monitors complex targets, with over 350 targets flying online simultaneously under normal conditions. It experiences frequent communication signal interruptions, aircraft overstaying, abnormal trajectory deviations, and incursions into no-fly zones—typical low-altitude security incidents. Traditional dispatch methods rely on fixed rules and manual intervention, resulting in delayed responses and frequent missed detections. There is an urgent need for an adaptive event handling mechanism with multimodal recognition capabilities, intelligent classification, and rapid response. The automatic classification and response method for low-altitude security incidents based on a cloud computing platform proposed in this invention is integrated into the core process of this dispatch platform, replacing the original rule-engine-based classification model and manual response strategy matching mechanism.

[0116] In practical applications, the platform first acquires monitoring data from radar arrays, ground optical monitoring, and wireless communication acquisition nodes. After noise filtering, temporal alignment, and spatial mapping by the preprocessing module of this invention, a heterogeneous airspace behavior map is constructed, forming a behavior network structure that can be used for graph convolutional propagation. The system then extracts the path evolution features of low-altitude flying targets through a graph convolutional neural network combined with a spatiotemporal attention mechanism and generates a feature embedding matrix. In the map structure, the connection weights between nodes integrate radar echo intensity and communication interference indicators, which can accurately reflect the mutual interference relationship between aircraft and the group behavior pattern. During deployment, the system automatically processes and identifies typical event types from multiple regions, such as "illegal crossing of no-fly zones," "stationary hovering exceeding the threshold," "periodic track deviation," and "recurrence after signal interruption." A dual-branch classification network is used to complete category labeling and level scoring, and response strategies are generated in real time.

[0117] During a verification process, the system identified 42 medium-to-high-level security incidents, 26 of which were multi-factor overlapping incidents (such as flight anomalies + communication anomalies). Using the multi-objective policy function of this invention, the system quickly matches and generates a set of response instructions from the policy template library, and then sends these instructions to the corresponding interference terminals and audible / visual alarm modules via the cloud platform. Platform records show that the average response time was reduced from 28.4 seconds to 6.9 seconds, the success rate of response instruction issuance increased from 82.7% to 97.4%, and the data collection completeness rate from the linked terminals to the cloud database increased from 61.3% to 92.8%. In the dynamic learning phase, the system automatically performs incremental optimization of the graph convolutional network and classification network based on the feedback stream uploaded by the terminals. Under repeated scenario testing the following day, the network recognition accuracy further improved by 3.2%, reaching 92.5%.

[0118] Meanwhile, a statistical comparison of low-altitude safety incidents processed by the platform over the past week revealed that: after the implementation of this invention, the overall event false negative rate decreased from 8.6% to 1.7%, and the false positive rate decreased from 5.4% to 2.2%; the system's automatic identification capability was significantly enhanced, and the platform's manual intervention rate decreased from 47.9% to 12.3%. Particularly in handling complex behavioral clustering events (such as illegal formation flying + signal-induced interference), the behavioral graph and strategy response mechanism constructed by this invention demonstrated stronger correlation identification and decision-making capabilities. Finally, through feedback closed-loop training, network iterative propagation achieved rapid convergence and formed a self-evolving mechanism adaptable to various scenarios. The following is a summary table comparing the core indicators during the actual operation of the system:

[0119] Table 1 Comparison of Low-Altitude Safety Incident Classification and Response System Performance

[0120]

[0121] In summary, the implementation results show that the present invention can achieve efficient abnormal behavior identification and multi-source strategy linkage response in complex low-altitude airspace. It not only significantly improves response speed and accuracy, but also greatly reduces the cost of human intervention, and has good engineering adaptability and deployment and promotion value.

[0122] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform, characterized in that, Includes the following steps: S1. Collect monitoring data in the low-altitude region and perform preprocessing; S2. Input the preprocessed monitoring data into the graph convolutional neural network to construct a heterogeneous spatial behavior map, and combine the spatiotemporal attention mechanism to extract the evolution path and behavior sequence pattern between nodes to generate a feature embedding matrix. S3. Input the feature embedding matrix into the dual-branch classification network, perform event type recognition and event level evaluation respectively, and output event type label and level score vector; S4. Based on the event type labels and rating vectors, construct an event index and design a multi-objective strategy function based on vector matching and contextual constraints to generate a set of response instructions; S5. Send the set of response instructions to the multi-source processing terminal through the cloud computing platform to perform blocking, interference, tracking or alarm operations, and collect terminal feedback information and regional status change data; S6. Based on terminal feedback information and regional state change data, incremental training is performed on the graph convolutional neural network and the dual-branch classification network to continuously update network parameters and optimize the multi-objective policy function.

2. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 1, characterized in that, The monitoring data includes radar data, optical image data, and communication link data.

3. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 1, characterized in that, The preprocessing includes noise filtering, missing value imputation, outlier removal, timestamp alignment, spatial projection mapping, and data standardization.

4. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 1, characterized in that, The dual-branch classification network includes an event type recognition branch and an event level evaluation branch. The event type recognition branch consists of residual convolutional units and a fully connected classifier, used to extract temporal behavioral features from the feature embedding matrix and output event category labels. The event level evaluation branch adopts a multi-scale attention fusion mechanism to jointly model local feature weights and global context information, and outputs an event level score vector through a multilayer perceptron. The two branches share the underlying encoder and are trained end-to-end through a joint loss function, which includes cross-entropy loss and mean squared error loss.

5. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 1, characterized in that, S2 specifically includes: S21. Based on the preprocessed monitoring data, construct a node set V = {v1, v2, ..., v...} n } and the edge set E={e ij |i,j=1,2,…,n}, where each node v i Indicates at time t i The target state unit is mapped from radar data, optical image data, and communication link data, with side e ij Indicates v i With v j Behavioral dependencies in time series; S22. Based on the node set and edge set, construct a heterogeneous spatial behavior graph G = (V, E, A, W), where A is the adjacency matrix, defined as A(i, j) = 1 representing node v. i With v j There are information connections, A(i,j)=0 indicates no connection, W is the edge weight matrix, and the element W(i,j) represents the connection weight calculated by the joint metric function of radar echo intensity and communication interference; S23. Input the heterogeneous spatial behavior graph G into the graph convolutional neural network and perform multiple rounds of graph convolution propagation. Calculate the node feature matrix H of the (l+1)th layer after each round of propagation using the following formula. (l+1) : Among them, H (l) Let D represent the feature matrix of the nodes in the l-th layer, and let D be the degree matrix of the adjacency matrix A. Let σ represent the trainable weight matrix of layer l, σ be the non-linear activation function, α be the global fusion coefficient, P be the number of spatiotemporal attention paths, and β be the training weight matrix of layer l. p T represents the spatiotemporal attention weight of the p-th path. p Let W be the transition matrix of the p-th path, representing the normalized relative time intervals between nodes within the path. p This is the path transformation matrix; S24, Regarding the node characteristic matrix H (l) Perform behavioral sequence modeling and path extraction, and construct a feature embedding matrix: c ij =exp(-κ·d ij )·(1-η·ξ ij ); Where Z represents the feature embedding matrix, φ represents the embedding compression function used to extract stable representations, and n represents the number of nodes. Represents node v i v j eigenvectors, δ ij Indicates v i v j The time interval difference between them, ψ is the joint function used to fuse node features and time interval, γ ij For node v i v j The mutual influence coefficient is calculated based on spatial projection and communication loss. exp(·) represents the natural exponential function with base e, and d ij ξ represents spatial distance. ij This represents the normalized value of the communication interference intensity, where κ and η are regular scaling factors.

6. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 5, characterized in that, During the graph convolution propagation process, the heterogeneous spatial behavior graph G constructs a dual-weighted fusion graph convolution unit, embeds an edge weight constraint matrix and a path evolution matrix to control the feature propagation direction, superimposes a step path attention gating structure to adjust the information transmission frequency between nodes, and maintains the spatial topological consistency of the graph through a structure preservation factor.

7. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 1, characterized in that, S3 specifically includes: S31. Input the feature embedding matrix Z into the shared bottom encoder of the dual-branch classification network, perform residual convolution operation and temporal compression processing to obtain the intermediate feature matrix U: Where ρ is the combination of linear transformation and normalization function, ω l The fusion weight for the l-th residual channel is... It is a two-channel convolutional unit, θ l Here are the parameters for the convolution kernel of the l-th channel, * indicates the convolution operation, and L represents the total number of residual stacking layers; S32. Input the intermediate feature matrix U into the event type recognition branch and the event level evaluation branch respectively, and extract the context matrix ξ = [ξ1, ξ2, ..., ξ]. K ], where each ξ k Let K represent the context vector derived from the intermediate feature matrix U, and K represent the dimension of the context vector, satisfying k∈{1,2,...,K}. The event type identification branch outputs the event type label y through a fully connected classifier, and the event level evaluation branch uses a multi-scale attention mechanism to generate a level rating vector s. The level rating vector is calculated as follows: Where Ω is the mapping function, M is the scale number, and λ mk U represents the attention weight of the k-th channel at the m-th scale. mk χ represents the feature representation of channel k at the m-th scale. k μ represents the context encoding vector. m Let represent the scale gating factor, π be the cross-scale feature aggregation function, ν be the global attention adjustment coefficient, and ∈ be the context compression factor. U represents the weighting factor for the j-th feature dimension, m represents the number of feature dimensions, and U represents the weighting factor for the j-th feature dimension. j Let represent the feature representation of the j-th feature dimension, and tanh(·) represent the hyperbolic tangent function; S33. Based on the event type label y and the rating vector s, the dual-branch classification network is trained end-to-end using a joint loss function, namely cross-entropy loss. With mean square error loss Weighted sum: in, This represents the joint loss function, where τ1 and τ2 are the loss weighting coefficients, C is the total number of event types, and δ... i Let be the i-th indicator variable of the true label, and let log2(·) denote the natural exponential function with base 2. i Let be the predicted probability for the i-th label, D be the number of dimensions in the rating vector, and s be the predicted probability. d For the d-th dimension score, The corresponding target score.

8. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 1, characterized in that, S4 specifically includes: S41. Based on the event type label y and the rating vector s, jointly map the type label and the rating value to construct the event index matrix q. The vector dimension of the event index matrix is ​​C+D, where C is the total number of event types and D is the number of dimensions of the rating vector. S42. Construct a candidate vector set for strategies Θ = {θ1, θ2, ..., θ} T }, where T is the number of policy templates, and each candidate policy vector θ t It is a real-valued vector of length C+D+K, where K represents the dimension of the context vector; S43. Concatenate the event index matrix q and the context matrix ξ directly to form the target event matrix ζ = [q‖ξ], and calculate the multi-target matching score ρ between each candidate policy vector and the target event matrix. t This generates a set of response instructions. The multi-target matching score is calculated as follows: Where, γ t Let ζ represent the t-th response instruction vector that matches the event. i Let θ represent the target event matrix of the i-th vector dimension. t,i Let λ represent the t-th candidate policy vector of the i-th vector dimension. i Let ψ represent the policy priority coefficient of the i-th vector dimension, ψ be the joint function used to non-linearly combine type, level, and context, and σ be the policy priority coefficient of the i-th vector dimension. i Let φ represent the context modulation coefficient of the i-th vector dimension, φ represent the embedding compression function, and ξ represent the context modulation coefficient of the i-th vector dimension. j χ represents the j-th context vector derived from the intermediate feature matrix U. j Let be the aggregate weight of the j-th context vector, and ∈ be a stabilizing term to prevent the denominator from being zero. This is the activation threshold for the policy.

9. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 8, characterized in that, The event index matrix q is constructed based on event type labels y and rating vectors s. By introducing a label vector encoder and a rating weight mapping function, the outputs corresponding to y and s are concatenated, and a unified vector representation Q is generated through a nonlinear projection network. This Q is then combined with the context matrix ξ to construct the event index matrix. in The outer product operation is used to express the bidirectional interactive relationship between the event and the context matrix, enhancing the distinctiveness and context sensitivity of the event index.

10. The method for automatic classification and response to low-altitude safety incidents based on a cloud computing platform according to claim 1, characterized in that, S5 specifically includes: S51. Perform structured encoding on the response instruction set Γ to construct an instruction transmission vector stream. Each vector corresponds to an event response unit, which includes a policy index, target area encoding, operation weight vector, and time scheduling identifier field. S52. The instruction transmission vector stream is pushed in parallel through the distributed cloud scheduling platform. A dynamic instruction distribution strategy based on task priority and geographical coverage is adopted to push each response vector to the processing terminal. The terminal type is automatically matched by the static mapping table of the central node according to the event category label. S53. After receiving the instruction, each terminal performs blockade, interference transmission, target tracking or audible and visual alarm operations based on the local status parsing operation field. During the execution, it continuously collects environmental feedback signals and regional status change data, and transmits the disposal process data back to the cloud platform. S54. After receiving feedback information and regional status change data uploaded by all terminals, the cloud platform performs timestamp alignment and low-altitude area archiving, constructs a feedback record stream, and writes the feedback record stream into the task feedback database for incremental training and response strategy evaluation of graph convolutional neural networks and dual-branch classification networks.

Citation Information

Cited By

  • Intelligent dynamic traffic management method and system for future urban air traffic

    CN121171068A

  • Target identification and damage evaluation method and system based on photoelectric pod multi-mode perception

    CN122049752A

  • Target recognition and damage assessment method and system based on multi-modal perception of photoelectric pod

    CN122049752B