Multi-source perception load and edge storage coupling low-altitude situation identification method and system

Abstract

The application belongs to the technical field of low-altitude airspace monitoring and situation judgment, and proposes a low-altitude situation judgment method and system coupled with multi-source perception load and edge evidence storage. The multi-source load collects monitoring data and performs time alignment and coordinate unification, completes target detection, feature extraction and cross-sensor correlation, and constructs a space-time situation map; inputting the space-time situation map into a trained space-time graph Transformer model outputs a situation label and generates a situation map snapshot; after determining the certainty of the snapshot, KZG polynomial commitment is executed to obtain commitment value and proof, combined with hardware trust root trusted timestamp and signature of at least two loads and aggregated, the snapshot and commitment are written into edge evidence storage and uploaded to the center platform, realizing real-time judgment on the edge side and verifiable, tamper-proof and traceable evidence.

Description

Technical Field

[0001] This invention relates to the field of low-altitude airspace monitoring, and in particular to a method and system for low-altitude situational awareness that couples multi-source sensing payloads with edge evidence. Background Technology

[0002] In recent years, with the large-scale application of low-altitude aircraft such as drones and electric vertical takeoff and landing (eVTOL) aircraft in scenarios such as logistics delivery, emergency rescue, inspection, and surveying, the demand for refined governance of low-altitude airspace has grown rapidly. To achieve continuous monitoring and situational assessment of low-altitude targets, existing technologies typically employ multi-source sensing methods for joint perception, such as ground radar, photoelectric sensors, infrared sensors, radio frequency detection equipment, acoustic arrays, and broadcast automatic dependent surveillance information. These methods combine time synchronization and coordinate transformation to detect targets, extract features, and correlate across sensors. Fusion analysis is then performed at the central location, outputting target trajectories, risk warnings, or situational classification results. Simultaneously, situational understanding methods for complex airspaces are also developing. Some solutions introduce deep learning models for multi-target identification, trajectory prediction, and anomaly detection, and are gradually incorporating edge computing to reduce backhaul pressure and improve processing efficiency.

[0003] However, existing low-altitude situational awareness and discrimination schemes still have the following shortcomings:

[0004] First, most systems still use a centralized architecture for sensing and data acquisition, data transmission, central-side fusion and discrimination, and output results. The processing links are long and the central-side computing and transmission overhead is large, resulting in insufficient real-time situation judgment and alarm response, which makes it difficult to meet the requirements of low-altitude supervision for rapid response.

[0005] Second, the lack of reliable execution and evidence storage mechanisms in the collection, transmission and processing of multi-source monitoring data makes it easy for data to be tampered with, sources to be unverifiable, and time consistency and integrity to be difficult to prove, which makes the credibility of situation conclusions insufficient and difficult to support regulatory enforcement and accountability.

[0006] Third, when introducing distributed evidence storage or log retention, some solutions have high verification and storage overhead under the condition of limited resources on the edge side, and lack verifiable commitments and multi-source joint signature mechanisms for structured data such as situation maps, making it difficult to balance evidence collection capabilities and system efficiency.

[0007] Therefore, there is a need for a low-altitude situational awareness method and system that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0008] One objective of this invention is to propose a low-altitude situation assessment method and system that couples multi-source sensing payloads with edge evidence storage. Addressing the problems of existing technologies that generally rely on sensing data being transmitted back to a central location for fusion and assessment, resulting in long links, high transmission and central computational overhead, and insufficient real-time situation assessment response, as well as the lack of reliable execution and evidence storage mechanisms for monitoring data during acquisition, transmission, and processing, making it susceptible to tampering, difficult to verify source and time, and difficult to prove integrity, this invention proposes a method that performs time alignment and coordinate unification, target detection and feature extraction, and cross-sensor target association on multi-source monitoring data at the edge computing node to construct a spatiotemporal situation map. This method utilizes a trained time-space data processing module... The empty-map Transformer model outputs situational awareness results and generates a situational awareness snapshot containing a set of nodes, a set of edges, a time range identifier, and a discrimination label. The situational awareness snapshot is deterministically serialized and KZG polynomial commitment is executed to generate a commitment value and a commitment proof. Combined with the trusted timestamp provided by the hardware root of trust and the signatures of at least two sensing payloads, and aggregated signature, the snapshot, its commitment, and signature are appended to the edge storage and uploaded to the central platform. This technical solution achieves rapid edge-side discrimination of low-altitude situational awareness and ensures that the source of monitoring data and discrimination conclusions is verifiable, the content is tamper-proof, and traceable.

[0009] On the one hand, this invention provides a low-altitude situational awareness method coupled with multi-source sensing payloads and edge evidence storage, including:

[0010] S1. Monitoring low-altitude targets using multi-source sensing payloads generates multi-source monitoring data, performs time alignment and coordinate unification, and generates multi-source aligned data. S2. Target detection and feature extraction are performed on the multi-source aligned data, generating a multi-source feature sequence sorted by time. Cross-sensor target association is performed to determine the correspondence between target identifiers and observations, generating an associated feature sequence. S3. A spatiotemporal situation map is constructed based on the associated feature sequence, generating spatiotemporal situation map data. S4. The spatiotemporal situation map data is input into a trained spatiotemporal map Transformer model, which outputs low-altitude situation discrimination results. A situation map snapshot is generated based on the spatiotemporal situation map data and the low-altitude situation discrimination results, including a set of nodes and edges. S5. Deterministically serialize the situation map snapshot to obtain the snapshot data string, execute the KZG multinomial commitment algorithm to generate the commitment value and commitment proof; S6. Generate or load the signing private key from the hardware trust root and provide a trusted time source to generate a timestamp, generate the data to be signed based on the commitment value, commitment proof, timestamp and edge computing node identifier, generate the payload signature using the signing private keys corresponding to at least two sensing payloads, execute the aggregate signature algorithm to generate the aggregate signature, write the situation map snapshot, commitment value, commitment proof, timestamp, edge computing node identifier and aggregate signature into the edge evidence storage to form an evidence record, and upload it to the central platform.

[0011] Optionally, S1 includes:

[0012] The multi-source sensing payload acquires at least two monitoring data streams targeting the same low-altitude target, and writes time information and sensor identifiers into each monitoring data stream to form the multi-source monitoring data.

[0013] Based on the time information, the multi-source monitoring data is mapped to a unified time axis to complete time alignment;

[0014] The multi-source monitoring data is converted to a unified coordinate system based on pre-calibrated coordinate transformation parameters to achieve coordinate unification.

[0015] Calculate data quality indicators for time-aligned and coordinate-uniformed data, and remove or mark low-quality data based on the data quality indicators;

[0016] Noise suppression processing is performed on the data after it has been removed or marked to generate the multi-source aligned data.

[0017] Optionally, S2 includes:

[0018] Target detection is performed on the multi-source aligned data according to the sensor identifier to obtain target detection results, the target detection results including at least the target location and the target detection confidence level;

[0019] Based on the target detection results, feature vectors representing each target are extracted and associated with the corresponding time information to generate a multi-source feature sequence sorted by time.

[0020] Based on the feature similarity between feature vectors corresponding to different sensors in the multi-source feature sequence and the motion consistency of the target position changing over time, the association score is calculated, and the cross-sensor observation correspondence is determined according to the association score. The same target identifier is assigned to the observation correspondence that meets the preset association conditions, and an association feature sequence containing the target identifier is generated.

[0021] Optionally, S3 includes:

[0022] Using the associated feature sequence as input, the feature vectors corresponding to the same low-altitude target at different times are aggregated according to the target identifier to generate a node for representing the low-altitude target, and the feature vector, the position parameters corresponding to the feature vector, and the target detection confidence are written into the node attributes of the node;

[0023] Based on the position parameters corresponding to different target identifiers in the associated feature sequence, calculate the spatial distance and relative motion parameters between target entities, and when the preset relationship conditions are met, establish an edge between the corresponding nodes to represent the relationship between target entities and write the edge attribute.

[0024] Based on the pre-configured airspace elements and the location parameters, the spatial relationship between the low-altitude target and the airspace elements is determined, and an edge is established between the corresponding node and the airspace element to represent the relationship between the target entity and the airspace element.

[0025] Write time indices determined by the time information to the nodes and edges to characterize time evolution and generate spatiotemporal situation map data containing node sets, edge sets, node attributes, edge attributes and time indices.

[0026] Optionally, S4 includes:

[0027] Extract the node set and edge set corresponding to the preset time window from the spatiotemporal situation map data, and determine the start time and end time of the preset time window according to the time index to generate a time range identifier, thereby obtaining the window spatiotemporal situation map data;

[0028] The spatiotemporal situation map data of the window is input into the trained spatiotemporal map Transformer model to output the discrimination score corresponding to each situation category, and the discrimination label and its confidence level of the low-altitude situation discrimination result are determined based on the discrimination score.

[0029] Using the spatiotemporal situation map data of the window as the graph data content and the discrimination label as the discrimination label field, a situation map snapshot is generated, wherein the situation map snapshot includes at least the node set, the edge set, the time range identifier and the discrimination label.

[0030] Furthermore, the trained spatiotemporal graph Transformer model is obtained through the following training method:

[0031] Acquire historical multi-source monitoring data and the situation labels corresponding to the historical multi-source monitoring data;

[0032] Time alignment and coordinate unification are performed on the historical multi-source monitoring data, and noise suppression is performed on the aligned data to generate multi-source aligned data for training.

[0033] Target detection and feature extraction are performed on the multi-source alignment data for training to generate a multi-source feature sequence for training. Cross-sensor target association is performed based on the multi-source feature sequence for training to generate an association feature sequence for training.

[0034] Based on the training-related feature sequences, a training spatiotemporal situation map is constructed and training spatiotemporal situation map data is generated;

[0035] Using the spatiotemporal situation map data for training as the model input, and the situation labels as the supervision signal, a loss function is constructed and the model parameters are iteratively updated to obtain a trained spatiotemporal map Transformer model.

[0036] Optionally, S5 includes:

[0037] The situation map snapshot is subjected to deterministic serialization processing according to a preset field order to convert the node set, the edge set, the time range identifier, and the discrimination label into a snapshot data string;

[0038] Based on pre-configured KZG polynomial commitment public parameters, the snapshot data string is mapped to a polynomial coefficient sequence to obtain a snapshot polynomial;

[0039] Perform the KZG polynomial commitment algorithm on the snapshot polynomial to generate the commitment value;

[0040] The commitment proof is generated based on the snapshot polynomial and the preset proof points.

[0041] Furthermore, the preset proof points are a set of proof points generated by the central platform based on random challenges. After receiving the random challenge, the edge computing node generates multiple commitment proofs corresponding to the set of proof points for the same situational map snapshot and uploads them to the central platform to enable verifiable spot checks on the snapshot data string.

[0042] Optionally, S6 includes:

[0043] The hardware root of trust generates or loads a signed private key corresponding to at least two sensing payloads, and generates a timestamp based on a trusted time source provided by the hardware root of trust.

[0044] Perform a hash calculation on the situational map snapshot to obtain the snapshot hash value;

[0045] Data to be signed is generated based on the commitment value, the commitment proof, the timestamp, the edge computing node identifier, and the snapshot hash value;

[0046] Generate payload signatures for the data to be signed using the signature private keys corresponding to the at least two sense payloads respectively;

[0047] An aggregate signature is generated by performing an aggregate signature algorithm on the payload signature;

[0048] The situational map snapshot, the commitment value, the commitment proof, the timestamp, the edge computing node identifier, the snapshot hash value, and the aggregate signature are appended to the edge evidence storage to form an evidence record;

[0049] A hash calculation is performed on the evidence storage record to obtain an evidence storage digest, and the evidence storage digest is uploaded to the central platform.

[0050] On the other hand, the present invention also provides a low-altitude situational awareness system coupled with multi-source sensing payloads and edge evidence storage, comprising:

[0051] A multi-source sensing payload includes at least two sensing payloads that acquire monitoring data for the same low-altitude target.

[0052] An edge computing node, communicatively connected to the multi-source sensing payload, comprises:

[0053] The alignment preprocessing module is used to receive the multi-source monitoring data generated by the multi-source sensing payload, and to perform time alignment and coordinate unification on the multi-source monitoring data to generate multi-source aligned data.

[0054] The detection and feature generation module is used to perform target detection and feature extraction on the multi-source aligned data and generate a multi-source feature sequence sorted by time.

[0055] The cross-sensor association module is used to perform cross-sensor target association based on the multi-source feature sequence to determine the correspondence between target identifiers and observations, and to generate an association feature sequence.

[0056] The spatiotemporal situation map construction module is used to construct a spatiotemporal situation map based on the associated feature sequence and generate spatiotemporal situation map data.

[0057] The situation discrimination module includes a model storage unit that stores a trained spatiotemporal graph Transformer model. It is used to input the spatiotemporal situation map data into the trained spatiotemporal graph Transformer model, output a low-altitude situation discrimination result, and generate a situation map snapshot based on the spatiotemporal situation map data and the low-altitude situation discrimination result. The situation map snapshot includes a node set, an edge set, a time range identifier, and a discrimination label.

[0058] The snapshot serialization and commitment module is used to deterministically serialize the situation map snapshot to obtain a snapshot data string, and execute the KZG polynomial commitment algorithm on the snapshot data string to generate a commitment value and a commitment proof;

[0059] The hardware trust root module is used to generate or load signed private keys and provide a trusted time source to generate timestamps;

[0060] The signing and aggregation module is used to generate data to be signed based on the commitment value, the commitment proof, the timestamp, and the edge computing node identifier; generate payload signatures for the data to be signed using the signing private keys corresponding to at least two sensing payloads; and execute an aggregation signature algorithm on the payload signatures to generate an aggregation signature.

[0061] The edge evidence storage module is used to write the situation map snapshot, the commitment value, the commitment proof, the timestamp, the edge computing node identifier, and the aggregate signature into the edge evidence storage memory to form an evidence storage record;

[0062] The central platform interface module is used to upload the evidence storage records to the central platform.

[0063] The beneficial effects of this invention are:

[0064] 1. Improve the real-time performance of low-altitude situation assessment: By completing multi-source data alignment preprocessing, target detection and cross-sensor correlation, spatiotemporal situation map construction, and situation assessment based on spatiotemporal map Transformer at the edge computing node, the burden of monitoring data backhaul and central-side fusion calculation is reduced, the processing link is shortened, thereby improving the response speed of situation assessment and alarm output.

[0065] 2. Enhance the credibility and tamper resistance of monitoring data and discrimination results: Generate a situation map snapshot from spatiotemporal situation map data and discrimination labels and perform deterministic serialization. Use KZG polynomial commitment to generate commitment values ​​and commitment proofs, so that the integrity of the snapshot content can be verified and tampering can be detected, thereby improving the credibility of situation conclusions.

[0066] 3. Enhance the ability to obtain evidence and trace responsibility: Combine the trusted timestamp provided by the hardware root of trust with the digital signatures of at least two sensing payloads and perform aggregated signatures. The snapshot, commitment and signature are appended to form an edge evidence record and uploaded to the central platform, making the data source, generation time and processing node verifiable, which facilitates regulatory enforcement, evidence fixation and responsibility tracing. Attached Figure Description

[0067] 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:

[0068] Figure 1 This is a flowchart of the low-altitude situational awareness method that couples multi-source sensing payloads with edge evidence, as proposed in this invention. Detailed Implementation

[0069] 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.

[0070] refer to Figure 1 A low-altitude situational awareness method coupled with multi-source sensing payloads and edge evidence storage includes:

[0071] S1. Monitoring low-altitude targets using multi-source sensing payloads generates multi-source monitoring data, performs time alignment and coordinate unification, and generates multi-source aligned data. S2. Target detection and feature extraction are performed on the multi-source aligned data, generating a multi-source feature sequence sorted by time. Cross-sensor target association is performed to determine the correspondence between target identifiers and observations, generating an associated feature sequence. S3. A spatiotemporal situation map is constructed based on the associated feature sequence, generating spatiotemporal situation map data. S4. The spatiotemporal situation map data is input into a trained spatiotemporal map Transformer model, which outputs low-altitude situation discrimination results. A situation map snapshot is generated based on the spatiotemporal situation map data and the low-altitude situation discrimination results, including a set of nodes and edges. S5. Deterministically serialize the situation map snapshot to obtain the snapshot data string, execute the KZG multinomial commitment algorithm to generate the commitment value and commitment proof; S6. Generate or load the signing private key from the hardware trust root and provide a trusted time source to generate a timestamp, generate the data to be signed based on the commitment value, commitment proof, timestamp and edge computing node identifier, generate the payload signature using the signing private keys corresponding to at least two sensing payloads, execute the aggregate signature algorithm to generate the aggregate signature, write the situation map snapshot, commitment value, commitment proof, timestamp, edge computing node identifier and aggregate signature into the edge evidence storage to form an evidence record, and upload it to the central platform.

[0072] In this specific embodiment, S1 includes:

[0073] The multi-source sensing payload consists of a millimeter-wave radar sensing payload and a three-dimensional lidar sensing payload deployed at the same site. The edge computing node receives frame-level multi-source monitoring data reported by the two payloads via Ethernet. Each frame of multi-source monitoring data includes a frame header and a frame body, wherein the frame header contains time information. With sensor identification The Nanosecond counting, calculated from Coordinated Universal Time (UTC), is used, and the edge computing node serves as the IEEE 1588PTP master clock to synchronize the two payload clocks, ensuring that the maximum deviation between the local time of any payload and the reference time of the edge computing node does not exceed 1ms. Each pre-set load is uniquely numbered and bound to its installation orientation calibration parameters;

[0074] The edge computing nodes establish a unified timeline and set a starting point for the timeline. With fixed time step , will receive any frame pass and Integer quantization is mapped to time index and with To aggregate radar and laser frames in a circular buffer to form a multi-source aligned record under the same time index, a missing payload is marked for missing frames. And retain empty frame bodies to ensure alignment during subsequent processing;

[0075] In coordinate unification processing, the edge computing nodes are pre-stored to a unified coordinate system. The two load extrinsic parameters are represented, the For any load identifier, a northeast-central coordinate system is established with the installation reference point as the origin. The corresponding frame point set, placing each point in the load coordinate system The position vector below The result is obtained by transforming the rigid body into a unified coordinate system. The rigid body transformation satisfies ,in For load from arrive The rotation matrix, For load exist The translation vector in the text, the and During the installation phase, at least 6 known... The coordinates of the reflection target point are collected into a point cloud and the external parameters are obtained by the least squares solution process. The external parameters are then stored in a fixed manner with a version number and CRC check code to avoid parameter drift during runtime.

[0076] In the data quality processing stage, the edge computing node calculates the time deviation index for each aligned record. With points indicators ,in From the payload frame within the record and The absolute difference is obtained and the threshold is set to The number of valid points retained after coordinate unification, with a threshold set to [value]. ,when or The payload frame is then marked as low quality. The load point set is removed from the frame body of the alignment record, but the frame header is retained for traceable alignment.

[0077] In noise suppression processing, targeting The payload frame volume point set is first downsampled based on a voxel grid to the voxel side length. Centroid replacement is performed on points within the same voxel to suppress dense noise, followed by statistical outlier removal based on the number of neighboring points. With standard deviation threshold Calculate the distance from each point to its The average distance to nearest neighbors is calculated and then removed if the average distance exceeds the global mean. Points with a global standard deviation of 10 times are used to suppress scattered echoes, while a range threshold is applied to the millimeter-wave radar point set. With radial velocity threshold To remove obvious out-of-bounds points, the final output of the aligned record set after time alignment, coordinate unification, low-quality processing and noise suppression is the multi-source aligned data.

[0078] In this specific embodiment, S2 includes:

[0079] Edge computing nodes align multi-source data by sensor identifier Target detection and feature extraction are performed separately, and cross-sensor target association results are generated. Indicates the millimeter-wave radar sensing payload and Representing the 3D lidar sensing payload, edge computing nodes are indexed by time. Process the records aligned at the same time and obtain a set of target detection results for each payload. ;

[0080] against The 3D lidar sensing payload, the edge computing node first performs height threshold filtering on the input point cloud to remove ground points and then uses voxel size ( The voxelization process is performed, with a maximum of 32 points per voxel and a maximum of 20,000 voxels per frame. The voxelization results are then input into the 3D object detection model PointPillars to output 3D candidate boxes and their class confidence scores. PointPillars consists of a cylinder feature encoder, a 2D convolutional backbone network, and a detection head. The cylinder feature encoder maps the intra-voxel features to 64 channels via a fully connected layer and then performs ReLU activation and batch normalization to obtain the cylinder feature map. The 2D convolutional backbone network contains three sets of residual blocks with 64 output channels each. Multi-scale feature maps are obtained through top-down feature fusion. The detection head employs an anchor-frame-based classification and regression branch, with the anchor frame size set to [value missing]. and Two discrete angles are set on the heading angle. The edge computing node performs confidence threshold filtering on the output of the detection head and sets the target detection confidence threshold to 0.5. Then, three-dimensional non-maximum suppression is performed and the cross-union threshold is set to 0.2 to obtain the final target detection result.

[0081] against The millimeter-wave radar sensing payload, with edge computing nodes performing two-dimensional constant false alarm rate (CFAR) detection on the input range-Doppler angle cubic data and setting the CFAR window size to... And the false alarm rate is set to For echo points detected by constant false alarm rate, DBSCAN clustering is used in a unified coordinate system with a neighborhood radius of 1.5m and a minimum number of points of 5. The centroid of each cluster is used as the target location and the normalized value of the echo signal-to-noise ratio within the cluster is used as the target detection confidence.

[0082] Edge computing nodes will connect any sensor In time index The next Each detection target is represented as and for Write to target location Confidence of target detection ,in To be a three-dimensional position vector in a unified coordinate system, with its three components corresponding to the position components in the northeast-sky direction, the... This is used to detect the confidence scalar output of the model or obtained by normalizing the radar signal-to-noise ratio;

[0083] In feature extraction, edge computing nodes are for each Constructing feature input vectors And fix its field to the target location. Target size Target speed Target detection confidence With sensor identification ,in The dimensions of the 3D candidate bounding box are calculated from the bounding box of the clustered point cloud in the radar detection results. The results were obtained by inter-frame pairing of nearest neighbor targets with adjacent time indices from the same sensor, and... Remove bits for time step calculation Obtain the three-dimensional velocity vector;

[0084] Edge computing nodes will Input feature embedding model to generate feature vectors representing the target The feature embedding model is a three-layer multilayer perceptron with layer widths of [missing information]. Each hidden layer uses ReLU activation and layer normalization is applied after each hidden layer. The output layer is then subjected to... implement Normalization to make and will Time information corresponding to the detection target After association, press Write the multi-source feature sequences in ascending order;

[0085] In cross-sensor target association, edge computing nodes index simultaneously. Internal and For each target to be detected, candidate pairs are constructed and a correlation score is calculated. ,in For radar target detection index and For laser-detected target indexing, the association score is determined by feature similarity and motion consistency and satisfies the following:

[0086] ;

[0087] in The weighting coefficient has a value of The feature vector of the radar target detection and Let be the feature vector of the laser-detected target, with symbol . Indicates transpose and Represents the L2 norm, The three-dimensional velocity vector of the target detected by the radar and This represents the three-dimensional velocity vector of the target detected by the laser. The velocity scale parameter has a value of Used to map the speed difference as Consistency score of the interval;

[0088] Edge computing nodes simultaneously apply spatial thresholds to ensure that candidate pairings satisfy... And only candidate pairings that pass the threshold are retained. Subsequently constructed with The score matrix of the elements is used, and the Hungarian algorithm is employed to solve the one-to-one maximum total score matching to determine the observation correspondence. Observation correspondences with a matching score not less than 0.65 are assigned the same target identifier ID, and the ID is maintained within the edge computing nodes up to the latest target location. ,speed eigenvectors Information with last update time The state table is used to assign a new ID to unmatched detection targets and write them as new targets into the state table, ultimately resulting in an associated feature sequence containing the target identifier ID and the correspondence between cross-sensor observations.

[0089] In this specific embodiment, S3 includes:

[0090] Edge computing nodes construct spatiotemporal situation maps and generate spatiotemporal situation map data using associated feature sequences as input, wherein the associated feature sequences are indexed by time. Grouped, and each observation record must contain at least the target identifier ID and the sensor identifier. Target location Target speed Target detection confidence With feature vectors ,in For the same sensor In time index The detection target number below and and All are based on the unified coordinate system of step S1 Express;

[0091] Edge computing nodes indexed at each time Internally, observation records from different sensors are aggregated according to their IDs to generate target nodes. For any Its node attributes are defined to include fused feature vectors. , fusion position Fusion speed Fusion confidence Time Index and sensor-assisted masking ,in and Indicates time index The target identifier ID exists from the sensor. The observation records are collected and aggregated, and the fused feature vector is determined by a confidence-weighted average and satisfies the following:

[0092] ;

[0093] in For time index The observation set corresponding to the target identifier ID and the set elements Indicates source from sensor The Each detection target is associated with an ID. The target detection confidence level for this observation and The fusion location is the feature vector of the observation. With fusion speed Adopted and The same Weighted rules in Internally respectively to and The fusion confidence level is calculated by weighting the average. Pick Inside The maximum value is used to reflect the strongest observational support under that time index;

[0094] Edge computing nodes are indexed at the same time Any two target nodes and Constructing the relationship edge between targets ,in Edge computing nodes calculate the fusion position of two nodes. and The Euclidean distance is used, and undirected edges are established when the distance is no greater than 300m to represent spatial proximity. At the same time, the fusion speed of the two nodes is calculated. and The velocity difference modulus is not greater than the modulus. At that time, the edge attribute field rel_type is set to "formation / cooperation", and the distance and velocity difference modulus are written to the edge attribute fields dist and dvel respectively, and the time index is set. Write the edge attribute field 'time' to represent the evolution of the relationship over time;

[0095] Edge computing nodes are pre-configured with a set of spatial features. Each spatial element is treated as a static node. Add to the node set, where Identify airspace elements and The node attributes include a feature type field (type) and a geometry field (geom), where the type value is one of "no-fly zone", "airway corridor", or "take-off and landing point", and the geoom value is in a unified coordinate system. The data is stored as a 3D polygon patch or a 3D polyline plus radius corridor tube with a height range. For 3D determination, edge computing nodes are used for each target node. Based on its fusion location The `geom` field of the spatial feature performs point determination within the polygon and calculates the shortest distance from the point to the geometric boundary, and the target height component falls within the polygon. Furthermore, when the point is within the no-fly zone polygon, establish a relationship edge from the target to the no-fly zone and set the edge attribute field rel_type to "enter no-fly zone". The shortest distance from the target point to the no-fly zone boundary must not exceed 50m, and the altitude component must fall within the no-fly zone boundary. Establish a relation edge and set rel_type to "approaching no-fly zone". The shortest distance from the target point to the centerline of the flight corridor is not greater than the corridor radius and the altitude component falls within the corridor. Establish a relation edge and set rel_type to "in the airway corridor", write the shortest distance into the edge attribute field dist, and add a time index. Write the time attribute field to the edge;

[0096] Edge computing nodes will target all nodes With all airspace element nodes This is aggregated into a node set, and all inter-target relationship edges and target-to-spatial feature relationship edges are aggregated into an edge set. Simultaneously, the time index of each node and each edge is recorded. The data is written as a time index field, thereby generating a spatiotemporal situation map data containing a set of nodes, a set of edges, node attributes, edge attributes, and a time index.

[0097] In this specific embodiment, S4 includes:

[0098] After receiving spatiotemporal situation map data, the edge computing node performs situation assessment and generates a situation map snapshot within a fixed time window. The length of the fixed time window is set to... And the time step follows the same as step S1. Edge computing nodes are indexed by the end time corresponding to the current output time. Determine the window start time index And extract time index fields from the spatiotemporal situation map data. The target node set and the spatial feature node set are merged into a window node set. Simultaneously extract fields that satisfy the time index. The edges representing relationships between targets and the edges representing relationships between targets and spatial elements are merged into a window edge set. And calculate the window time range identifier based on the unified time axis in step S1. ,in and To unify the starting point of the timeline;

[0099] Edge computing nodes map each node in the window node set to a model input token and construct its initial feature vector for the target node. fused feature vectors in node attributes , fusion position Fusion speed Fusion confidence Sensor-based masking and normalized time characteristics The dimension is obtained by concatenating the fields in a fixed order and then linearly mapping it. The node embedding vector, for spatial feature nodes Encode the feature type field (type) into a discrete type index and query the type embedding vector with a dimension of 64. Simultaneously, calculate the geometric centroid position from the geometry field (geom). With geometric scale and After concatenation, linear mapping yields the same dimension. _ node embedding vector, where The lengths of the geometric bounding box along the three axes of northeast, zenith, and celestial to ensure that the input dimensions are fixed;

[0100] Edge computing nodes will set up window edges The edge attribute field rel_type is encoded as a discrete relation index and the relation embedding vector with a dimension of 32 is queried. The dist and dvel are normalized according to the dimensions and then concatenated with the relation embedding vector to form edge features. These features are used as attention bias inputs in the model so that attention introduces relation constraints between node pairs with edge connections.

[0101] The trained spatiotemporal graph Transformer model is obtained through offline training on the central platform and then distributed to edge computing nodes for permanent storage. The training data consists of historical multi-source monitoring data and corresponding situation labels, with the number of situation categories set to [number to be specified]. The five situation category labels are defined as "normal flight", "entering the no-fly zone", "approaching the no-fly zone", "formation / coordination", and "abnormal loitering". The central platform sequentially performs the same time alignment, coordinate unification, noise suppression, target detection, feature extraction, cross-sensor target association, and spatiotemporal situation map construction as steps S1 to S3 on historical multi-source monitoring data to obtain training spatiotemporal situation map data, and follows the same steps as in this step. Training samples are obtained by slicing according to window rules;

[0102] The spatiotemporal graph Transformer model adopts an encoder-only structure and includes... There are Transformer coding layers, and each layer has 1 multi-head self-attention head. Each feedforward network has a hidden layer width of 1024 and uses ReLU activation. Inter-layer residual connections and layer normalization are enabled, and the dropout rate is fixed at 0.1. The model output uses a graph-level readout method, averaging the embeddings of all nodes within the window to obtain a graph embedding vector, which is then output through a fully connected classification layer. Dimensional category logit;

[0103] The AdamW optimizer was used during training, and the learning rate was set to [value missing]. The weight decay coefficient was set to 0.01, the batch size was set to 16, the number of training rounds was set to 50, the loss function was multi-class cross-entropy, and manually labeled situational tags were used as supervision signals to iteratively update the model parameters until the training was completed and a well-trained spatiotemporal graph Transformer model was obtained.

[0104] During online inference, edge computing nodes embed window nodes and edge features into the trained spatiotemporal graph Transformer model to obtain the logit for each category and convert it into a discriminant score. Let the first... The class logit is and , order the Class discrimination score but:

[0105] ;

[0106] in It is an exponential function and For category summation index, edge computing nodes are... Determine the discrimination label and use it As the confidence level of this discrimination label;

[0107] Edge computing nodes use windowed spatiotemporal situational map data as graph data content and... As a time range identifier field, As a discriminant label field, As a confidence field, a situation map snapshot is generated, and the situation map snapshot contains at least a set of window nodes. Window edge set Time range identifier and discrimination labels .

[0108] In this specific embodiment, S5 includes:

[0109] After generating a situational map snapshot in step S4, the edge computing node performs deterministic serialization and KZG polynomial commitment to obtain the commitment value and commitment proof.

[0110] The deterministic serialization encodes the situation map snapshot in a preset field order and outputs a unique snapshot data string. The fields are ordered as follows: time range identifier, discrimination label, node set, and edge set, where the time range identifier is... Write and and All use nanosecond counts calculated from Coordinated Universal Time (UTC) and are encoded in unsigned 64-bit big-endian integer order. The discrimination labels are encoded in unsigned 16-bit big-endian integer order, and the confidence levels are encoded in IEEE 754 binary 32-bit floating-point big-endian order.

[0111] The node set is first partitioned by node type and sorted according to fixed rules before being encoded, with target nodes initially indexed by time. Sort by ascending order, then by target identifier ID in ascending order; spatial element nodes are sorted by spatial element identifier. Sort in ascending order. Each node is encoded in a TLV structure with a length prefix followed by its content. The node attribute fields are in a fixed order: node identifier, node type, time index, and merged position. Fusion speed Fusion confidence fusion feature vectors Sensor-based masking ,in and Each component is encoded in IEEE 754 binary 32-bit floating-point big-endian and the following is prohibited: To ensure serialization determinism, a value of infinity is used, and if such a node appears, it is marked as invalid and removed from the node set. The 128 components are written in IEEE 754 binary 32-bit floating-point big-endian order, and the order is the same as that output in step S2. The dimensional order is consistent. Encoded using a two-bit bitmap with a fixed bit order. ;

[0112] The edge set is sorted and encoded in ascending lexicographical order by the starting node identifier, ending node identifier, and time index. The edge attribute fields are in a fixed order: relation type rel_type, distance dist, velocity difference dvel, and time index. rel_type uses a pre-defined enumeration table to map to an unsigned 16-bit integer to ensure consistency across devices;

[0113] Edge computing nodes will capture snapshot data strings The data is divided into fixed 31-byte blocks, and each block is interpreted as a scalar field element in big-endian order and written into the coefficient sequence. The scalar field is the scalar field of the BLS12-381 curve. Furthermore, each block, if less than 31 bytes, is padded with zeros to reach 31 bytes to form a definitive final block, thus obtaining the snapshot polynomial. The coefficient sequence and make The maximum number of times does not exceed ,in The total number of blocks;

[0114] The KZG polynomial commitment public parameters are generated and distributed by the central platform using a structured reference string (SRS). The SRS contains... ,in and These are two cyclic groups on BLS12-381. It is a bilinear pairing mapping. For the group generator and For the group generator, It is a secret scalar that is used only when generating the SRS and is not leaked to edge computing nodes. Exponent elements for exponentiation and cover the highest degree of the snapshot polynomial to ensure the commitment is computable;

[0115] Edge computing nodes based on SRS snapshot polynomial Calculate the commitment value And for the pre-set proof points Generate Proof of Commitment ,in Depend on Modulo the hash value To ensure that the commitment value is bound to and reproducible with the snapshot, the commitment value and the commitment proof satisfy:

[0116] ;

[0117] in Indicates the secret scalar Substituting the polynomial The obtained scalar value, This indicates that the point to be proved Substituting the polynomial The obtained scalar value, Indicates in Multiplicative inverse on and fractional representation on scalar division on This indicates that the generator is represented by the scalar as an exponent. Perform scalar exponentiation;

[0118] When the central platform initiates a random challenge, it sends a random number seed to the edge computing nodes. Size of the set of proof points Edge computing nodes The input is derived via HKDF-SHA256. One proof point and the same commitment value for the same situation map snapshot. Generate multiple corresponding commitment proofs respectively Each of them All as described above The generation rules will Replace with Obtain and Edge computing nodes will capture snapshot data strings Corresponding commitment value , Promise of Pre-set Proof Points and the set of proof points corresponding to random challenges It is output together with multiple sets of commitment proofs for signing and uploading to the central platform for verifiable spot checks.

[0119] In this specific embodiment, S6 includes:

[0120] The hardware root of trust module of the edge computing node consists of a secure processor with secure boot and secure storage and an independent trusted clock. The secure processor has a built-in true random number generator and generates signature private keys for the two sensing payloads during the initial deployment. These keys can only be used within the secure processor. and ,in Corresponding sensor identifier and Corresponding sensor identifier The and The key is stored in the security processor's key storage area in a non-exportable manner and bound to the edge computing node identifier. With sensor identification To prevent cross-node replication;

[0121] Edge computing nodes obtain commitment values ​​from step S5. With Proof of Commitment And obtain the snapshot data string from step S4. Regarding the above Execute SHA-256 hash to obtain snapshot hash value ,in It is a 32-byte digest and is used to bind the signature object to a deterministically serialized snapshot of the situation map;

[0122] The hardware trust root module reads the current nanosecond time from an independent trusted clock and generates a timestamp. ,in It uses an unsigned 64-bit integer representation calculated from Coordinated Universal Time and is generated by a secure processor and... They are jointly encapsulated into a non-reversible, trusted time record to ensure consistency. The time interval increases monotonically.

[0123] Edge computing node pairs and The data to be signed is deterministically concatenated with fixed field order and fixed byte order, and then formed inside the security processor. and call the security processor to use respectively and Generate payload signature and Subsequently and Performing BLS aggregate signing yields the aggregate signature. Its calculation process satisfies:

[0124] ;

[0125] in This represents the SHA-256 hash function and its output as... Input, symbols This indicates that byte strings are concatenated according to a preset field order, and each field uses big-endian order and a length prefix to eliminate ambiguity. A compressed byte string representing a commitment value and an element of the BLS12-381 curve group. This represents a compressed encoded byte string that represents a commitment proof and is an element of the BLS12-381 curve group. Indicates a trusted timestamp. This represents the edge computing node identifier and is encoded using an unsigned 32-bit integer. Represents snapshot data string SHA-256 digest, Indicates will Mapping to the BLS12-381 curve group The hash-to-curve function uses the BLS12-381 standard of the IETF hash-to-curve specification. Map and fix the DST to the string "LASS-SIGN-V1". Indicates the first The scalar of the signature private key corresponding to the road-aware payload and For the scalar field of BLS12-381, the symbol • denotes scalar multiplication of a scalar with respect to group elements. Indicates the first Load signature of road-sensing load, This represents the set of perceptual payloads participating in the signature, with a set size of 2, and the symbol... In the group The product operation is performed on the product and the result is an aggregate signature. ;

[0126] Edge computing nodes will capture situational snapshots and commit values. Proof of Commitment timestamp Edge computing node identifier Snapshot hash value With aggregate signature Evidence records are composed according to a fixed field order. And it is written to the edge evidence storage memory in an append-only manner, wherein the edge evidence storage memory is a sequential log area with append-only protection enabled, and each entry... Each record has an incrementing sequence number and a digest of the previous record to form a hash chain, thus enabling deletion and insertion actions to be detected;

[0127] Edge computing nodes for the stored evidence records Perform SHA-256 hash calculation to obtain the evidence digest. And Record number, timestamp and edge computing node identifiers Upload to the central platform via the central platform interface module to complete central platform record keeping and subsequent verification.

[0128] 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.

[0129] This invention employs a combined scheme of "spatiotemporal graph Transformer situational awareness assessment, graph snapshot commitment, and multi-payload aggregation signature," directly addressing two technical issues: real-time performance and reliable evidence collection in low-altitude situational awareness assessment. On one hand, multi-source monitoring data undergoes time alignment, coordinate unification, and cross-sensor correlation at the edge to construct a spatiotemporal situational graph. The spatiotemporal graph Transformer performs edge-side inference on the graph structure data, outputting discrimination labels. This transforms situational awareness assessment from post-processing at the center to on-site assessment at the edge, shortening the data backhaul and central fusion computation link, reducing transmission and central computing overhead, and improving response speed. On the other hand, the spatiotemporal situational graph data and assessment results are solidified into a situational graph snapshot and deterministically serialized. A KZG polynomial commitment is performed on the snapshot to obtain a commitment value and a commitment proof. This is then combined with a trusted timestamp provided by the hardware root of trust and signatures from at least two sensing payloads, and aggregated to form an evidence record. This ensures the integrity of the snapshot content is verifiable, tampering is detectable, and the source and generation time are verifiable, thereby improving the reliability and verifiability of the situational awareness conclusions.

[0130] Furthermore, this case addresses the aforementioned technical issues by improving and constraining the algorithm structure for engineering implementation: First, it generates "situation map snapshots" from the continuously evolving spatiotemporal situation map according to preset time windows, enabling model inference and evidence storage objects to form a stable windowed granularity. This is beneficial for edge-side streaming processing and low-latency output, and also facilitates verification of evidence within a specific time range by the central side. Second, it transforms graph structure data into unique data strings through a deterministic serialization mechanism, and uses KZG commitments to perform lightweight and verifiable binding of structured snapshots, avoiding the storage and verification burden caused by traditional full on-chain or full backhaul. Third, it achieves multi-source joint endorsement without significantly increasing link overhead through a multi-payload signature aggregation mechanism, and allows unified verification of commitment proofs and aggregated signatures by the central side, thereby more effectively balancing the technical effects of real-time discrimination and trusted evidence storage.

Claims

1. A low-altitude situational awareness method coupled with multi-source sensing payloads and edge evidence storage, characterized in that, include: S1. The low-altitude target is monitored by the multi-source sensing payload, multi-source monitoring data is generated, time alignment and coordinate unification are performed, and multi-source aligned data is generated. S2. Perform target detection and feature extraction on the multi-source aligned data, generate a multi-source feature sequence sorted by time, perform cross-sensor target association to determine the correspondence between target identifiers and observations, and generate an associated feature sequence; S3. Construct a spatiotemporal situation map based on the associated feature sequence and generate spatiotemporal situation map data; S4. Input the spatiotemporal situation map data into the trained spatiotemporal map Transformer model, output the low-altitude situation discrimination result, and generate a situation map snapshot based on the spatiotemporal situation map data and the low-altitude situation discrimination result, including the node set, edge set, time range identifier and discrimination label; S5. Perform deterministic serialization on the situation map snapshot to obtain the snapshot data string, execute the KZG polynomial commitment algorithm, and generate the commitment value and commitment proof; S6. Generate or load a signing private key from the hardware trust root and provide a trusted time source to generate a timestamp. Generate data to be signed based on the commitment value, commitment proof, timestamp, and edge computing node identifier. Generate payload signatures for the data to be signed using the signing private keys corresponding to at least two sensing payloads. Execute the aggregate signature algorithm to generate an aggregate signature. Write the situational map snapshot, commitment value, commitment proof, timestamp, edge computing node identifier, and aggregate signature into the edge evidence storage to form an evidence record, and upload it to the central platform.

2. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 1, characterized in that, S1 includes: The multi-source sensing payload acquires at least two monitoring data streams targeting the same low-altitude target, and writes time information and sensor identifiers into each monitoring data stream to form the multi-source monitoring data. Based on the time information, the multi-source monitoring data is mapped to a unified time axis to complete time alignment; The multi-source monitoring data is converted to a unified coordinate system based on pre-calibrated coordinate transformation parameters to achieve coordinate unification. For data that has been time-aligned and has unified coordinates, calculate data quality indicators and, based on these indicators, remove or mark low-quality data. Noise suppression processing is performed on the data after it has been removed or marked to generate the multi-source aligned data.

3. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 1, characterized in that, S2 include: Target detection is performed on the multi-source aligned data according to the sensor identifier to obtain target detection results, which include at least the target location and the target detection confidence level. Based on the target detection results, feature vectors representing each target are extracted and associated with the corresponding time information to generate a multi-source feature sequence sorted by time. Based on the feature similarity between feature vectors corresponding to different sensors in the multi-source feature sequence and the motion consistency of the target position changing over time, the association score is calculated, and the cross-sensor observation correspondence is determined according to the association score. The same target identifier is assigned to the observation correspondence that meets the preset association conditions, and an association feature sequence containing the target identifier is generated.

4. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 1, characterized in that, S3 includes: Using the associated feature sequence as input, the feature vectors corresponding to the same low-altitude target at different times are aggregated according to the target identifier to generate a node for representing the low-altitude target, and the feature vector, the position parameters corresponding to the feature vector, and the target detection confidence are written into the node attributes of the node; Based on the position parameters corresponding to different target identifiers in the associated feature sequence, calculate the spatial distance and relative motion parameters between target entities, and when the preset relationship conditions are met, establish an edge between the corresponding nodes to represent the relationship between target entities and write the edge attribute. Based on the pre-configured airspace elements and the location parameters, the spatial relationship between the low-altitude target and the airspace elements is determined, and an edge is established between the corresponding node and the airspace element to represent the relationship between the target entity and the airspace element. Write time indices determined by the time information to the nodes and edges to characterize time evolution, and generate spatiotemporal situation map data containing node sets, edge sets, node attributes, edge attributes and time indices.

5. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 1, characterized in that, S4 includes: Extract the node set and edge set corresponding to the preset time window from the spatiotemporal situation map data, and determine the start time and end time of the preset time window according to the time index to generate a time range identifier, thereby obtaining the window spatiotemporal situation map data; The spatiotemporal situation map data of the window is input into the trained spatiotemporal map Transformer model to output the discrimination score corresponding to each situation category, and the discrimination label and its confidence level of the low-altitude situation discrimination result are determined based on the discrimination score. Using the spatiotemporal situation map data of the window as the graph data content and the discrimination label as the discrimination label field, a situation map snapshot is generated, wherein the situation map snapshot includes at least the node set, the edge set, the time range identifier and the discrimination label.

6. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 1, characterized in that, S5 include: The situation map snapshot is subjected to deterministic serialization processing according to a preset field order to convert the node set, the edge set, the time range identifier, and the discrimination label into a snapshot data string; Based on pre-configured KZG polynomial commitment public parameters, the snapshot data string is mapped to a sequence of polynomial coefficients to obtain a snapshot polynomial; Perform the KZG polynomial commitment algorithm on the snapshot polynomial to generate the commitment value; The commitment proof is generated based on the snapshot polynomial and the preset proof points.

7. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 1, characterized in that, S6 include: The hardware root of trust generates or loads a signed private key corresponding to at least two sensing payloads, and generates a timestamp based on a trusted time source provided by the hardware root of trust. Perform a hash calculation on the situational map snapshot to obtain the snapshot hash value; Data to be signed is generated based on the commitment value, the commitment proof, the timestamp, the edge computing node identifier, and the snapshot hash value; Generate payload signatures for the data to be signed using the signature private keys corresponding to the at least two sense payloads respectively; An aggregate signature is generated by performing an aggregate signature algorithm on the payload signature; The situational map snapshot, the commitment value, the commitment proof, the timestamp, the edge computing node identifier, the snapshot hash value, and the aggregate signature are appended to the edge evidence storage to form an evidence record; A hash calculation is performed on the evidence storage record to obtain an evidence storage digest, and the evidence storage digest is uploaded to the central platform.

8. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 1, characterized in that, The trained spatiotemporal graph Transformer model was obtained through the following training method: Acquire historical multi-source monitoring data and the situation labels corresponding to the historical multi-source monitoring data; Time alignment and coordinate unification are performed on the historical multi-source monitoring data, and noise suppression is performed on the aligned data to generate multi-source aligned data for training. Target detection and feature extraction are performed on the multi-source alignment data used for training to generate a multi-source feature sequence used for training. Cross-sensor target association is performed based on the multi-source feature sequence used for training to generate an association feature sequence used for training. Based on the training-related feature sequences, a training spatiotemporal situation map is constructed and training spatiotemporal situation map data is generated; Using the spatiotemporal situation map data for training as the model input, and the situation labels as the supervision signal, a loss function is constructed and the model parameters are iteratively updated to obtain a trained spatiotemporal map Transformer model.

9. The low-altitude situational awareness method coupled with multi-source sensing payload and edge evidence storage according to claim 6, characterized in that, The preset proof points are a set of proof points generated by the central platform based on random challenges. After receiving the random challenge, the edge computing node generates multiple commitment proofs corresponding to the set of proof points for the same situational map snapshot and uploads them to the central platform to enable verifiable spot checks on the snapshot data string.

10. A low-altitude situational awareness system coupled with multi-source sensing payloads and edge evidence storage, used to execute the low-altitude situational awareness method coupled with multi-source sensing payloads and edge evidence storage as described in any one of claims 1 to 9, characterized in that, include: A multi-source sensing payload includes at least two sensing payloads that acquire monitoring data for the same low-altitude target. An edge computing node, communicatively connected to the multi-source sensing payload, comprises: The alignment preprocessing module is used to receive the multi-source monitoring data generated by the multi-source sensing payload, and to perform time alignment and coordinate unification on the multi-source monitoring data to generate multi-source aligned data. The detection and feature generation module is used to perform target detection and feature extraction on the multi-source aligned data and generate a multi-source feature sequence sorted by time. The cross-sensor association module is used to perform cross-sensor target association based on the multi-source feature sequence to determine the correspondence between target identifiers and observations, and to generate an association feature sequence. The spatiotemporal situation map construction module is used to construct a spatiotemporal situation map based on the associated feature sequence and generate spatiotemporal situation map data. The situation discrimination module includes a model storage unit that stores a trained spatiotemporal graph Transformer model. It is used to input the spatiotemporal situation map data into the trained spatiotemporal graph Transformer model, output a low-altitude situation discrimination result, and generate a situation map snapshot based on the spatiotemporal situation map data and the low-altitude situation discrimination result. The situation map snapshot includes a node set, an edge set, a time range identifier, and a discrimination label. The snapshot serialization and commitment module is used to deterministically serialize the situation map snapshot to obtain a snapshot data string, and execute the KZG polynomial commitment algorithm on the snapshot data string to generate a commitment value and a commitment proof; The hardware trust root module is used to generate or load signed private keys and provide a trusted time source to generate timestamps; The signing and aggregation module is used to generate data to be signed based on the commitment value, the commitment proof, the timestamp, and the edge computing node identifier; generate payload signatures for the data to be signed using at least two signature private keys corresponding to the sensing payloads; and execute an aggregation signature algorithm on the payload signatures to generate an aggregate signature. The edge evidence storage module is used to write the situation map snapshot, the commitment value, the commitment proof, the timestamp, the edge computing node identifier, and the aggregate signature into the edge evidence storage memory to form an evidence storage record; The central platform interface module is used to upload the evidence storage records to the central platform.

Images