Low-altitude cooperative sensing data processing method and system based on space calculation

Abstract

The invention provides a low-altitude cooperative sensing data processing method and system based on space calculation. The method comprises the following steps: receiving original sensing point cloud data transmitted by a plurality of low-altitude aircrafts; performing space-time registration and semantic enhancement processing on the original sensing point cloud data to generate a space semantic feature set containing spatial domain object attributes and environment association features; inputting the spatial semantic feature set into a pre-constructed airspace digital twin model, and generating airspace state implicit representation through neural radiation field modeling; calling a multi-subject collaborative reasoning model to perform conflict risk prediction on the airspace state implicit representation, and outputting a conflict risk prediction map including the relative position relationship between the aircrafts; and generating a distributed cooperative obstacle avoidance strategy based on the conflict risk prediction map and a preset obstacle avoidance strategy library, encoding the distributed cooperative obstacle avoidance strategy into a control instruction sequence, and issuing the control instruction sequence to the corresponding low-altitude aircraft. According to the method, the reliability of low-altitude cooperative sensing data processing and the adaptive capacity to a complex airspace environment can be improved.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a low-altitude collaborative sensing data processing method and system based on spatial computing. Background Technology

[0002] With the expansion of low-altitude aircraft application scenarios, low-altitude collaborative sensing data processing technology has become a core technology supporting multi-aircraft collaborative operations. It primarily achieves airspace state modeling, conflict risk prediction, and obstacle avoidance strategy generation through the fusion processing of multi-source sensing data, ensuring the safe operation of aircraft swarms in complex airspace environments. In existing technologies, low-altitude collaborative sensing data processing typically employs traditional point cloud registration methods to perform coordinate transformation and time alignment on the raw data collected by multiple aircraft. It then constructs an airspace environment model through explicit 3D modeling, uses a collision detection algorithm based on distance thresholds to identify potential conflicts between aircraft, and generates control commands based on fixed heading adjustments or altitude avoidance. However, multi-source data processing lacks deep integration of semantic attributes and environmental features of airspace objects, resulting in fragmented data and insufficient semantic expression capabilities; explicit 3D modeling relies on dense sampling of discrete point clouds or grids, making it difficult to continuously represent the motion state of dynamic obstacles in the airspace and subtle changes in environmental disturbances; conflict risk prediction is prone to missing potential collaborative conflicts; obstacle avoidance strategy generation is based on a fixed rule base and does not combine real-time airspace status for dynamic adjustment, making it difficult to adapt to the performance differences of multiple aircraft and the needs of complex missions, leading to reduced strategy adaptability and execution reliability. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and system for low-altitude collaborative sensing data processing based on spatial computing.

[0004] According to one aspect of the present invention, a low-altitude collaborative sensing data processing method based on spatial computing is provided. The method includes: receiving raw sensing point cloud data transmitted by multiple low-altitude aircraft; performing spatiotemporal registration and semantic enhancement processing on the raw sensing point cloud data to generate a spatial semantic feature set containing airspace object attributes and environmental association features; inputting the spatial semantic feature set into a pre-constructed airspace digital twin model and generating an implicit representation of airspace state through neural radiation field modeling; calling a multi-agent collaborative reasoning model to perform conflict risk prediction on the implicit representation of airspace state and outputting a conflict risk prediction map containing the relative positional relationships between aircraft; generating a distributed collaborative obstacle avoidance strategy based on the conflict risk prediction map and a preset obstacle avoidance strategy library, and encoding the distributed collaborative obstacle avoidance strategy into a control command sequence and issuing it to the corresponding low-altitude aircraft.

[0005] According to another aspect of the present invention, a computer system is provided, comprising: a processor; and a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method as described above.

[0006] The low-altitude collaborative sensing data processing method based on spatial computing provided by this invention receives raw sensing point cloud data transmitted from multiple low-altitude aircraft and performs spatiotemporal registration and semantic enhancement processing to generate a spatial semantic feature set containing airspace object attributes and environmental correlation features. This method can transform multi-source heterogeneous raw sensing data into structured features with a unified spatiotemporal reference and semantic correlation, effectively solving the data fragmentation problem caused by differences in acquisition perspective and coordinate system among multi-aircraft data, and improving the consistency and semantic expression capability of data fusion. The spatial semantic feature set is input into a pre-constructed airspace digital twin model, and implicit representations of airspace state are generated through neural radiation field modeling. This method can break through the dependence of traditional explicit 3D modeling on discrete point clouds or meshes, and continuously and generalize the physical field attributes and semantic correlation information in the 3D space of the airspace in the form of implicit function parameterization, thus more accurately capturing dynamic obstacles in the airspace. Subtle changes in motion state and environmental disturbances provide a high-fidelity airspace state description for subsequent conflict risk prediction. A multi-agent collaborative reasoning model is invoked to predict conflict risks based on the implicit representation of the airspace state, outputting a conflict risk prediction map containing the relative positional relationships between aircraft. This multi-agent collaborative mechanism can construct a spatial relationship network between aircraft and obstacles, effectively capturing potential conflict risks during dynamic interactions between multiple aircraft and improving the comprehensiveness and real-time performance of conflict risk prediction. Based on the conflict risk prediction map and a pre-set obstacle avoidance strategy library, a distributed collaborative obstacle avoidance strategy is generated and encoded into a sequence of control commands, which is then sent to the corresponding low-altitude aircraft. This allows for the generation of obstacle avoidance strategies that consider the performance constraints and mission requirements of each aircraft in multi-aircraft collaborative scenarios, ensuring the adaptability and executability of the strategy in a multi-agent dynamic environment. Overall, this improves the reliability of low-altitude collaborative perception data processing and its adaptability to complex airspace environments. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the application scenario provided by the present invention; Figure 2 This is a flowchart illustrating a low-altitude collaborative sensing data processing method based on spatial computing provided by the present invention. Figure 3 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0008] To facilitate a clearer understanding of this invention, we will first introduce the application scenarios in which this invention is implemented, such as... Figure 1As shown, this application scenario includes a computer system 10 and a low-altitude aircraft cluster. The low-altitude aircraft cluster can include one or more low-altitude aircraft; the number of low-altitude aircraft is not limited here. Figure 1 As shown, the low-altitude aircraft cluster may specifically include low-altitude aircraft 1, low-altitude aircraft 2, ..., low-altitude aircraft n; it can be understood that low-altitude aircraft 1, low-altitude aircraft 2, low-altitude aircraft 3, ..., low-altitude aircraft n can all be network connected to computer system 10 so that each low-altitude aircraft can interact with computer system 10 through network connection.

[0009] Understandably, computer system 10 can refer to a device that executes the method of the present invention, such as a server. A server can be a single physical server, a server cluster or distributed system consisting of at least two physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. Low-altitude aircraft can refer to unmanned aerial vehicles (UAVs).

[0010] Further, please see Figure 2 This is a flowchart illustrating a low-altitude collaborative sensing data processing method based on spatial computing, provided in an embodiment of the present invention. Figure 2 As shown, this method can be derived from... Figure 1 The method is executed by a computer system in the space computing environment. The low-altitude collaborative sensing data processing method based on spatial computing may include the following steps: Step S100: Receive raw sensing point cloud data transmitted by multiple low-altitude aircraft.

[0011] Raw point cloud data is acquired by low-altitude aircraft using onboard sensors such as lidar and millimeter-wave radar. Lidar determines the distance to a target object by emitting a laser beam and measuring the time difference of the reflected light, thereby obtaining the three-dimensional coordinate information of a large number of discrete points on the target object's surface. These points, when collected together, constitute point cloud data. Millimeter-wave radar uses electromagnetic waves in the millimeter-wave frequency band to detect targets, and can similarly obtain information such as the target's position and velocity, converting it into point cloud data.

[0012] In practical applications, such as low-altitude logistics delivery areas in cities, multiple drones fly independently and perform tasks. The lidar on each drone continuously scans the surrounding environment, collecting point cloud data of surrounding buildings, other aircraft, trees, and other targets in real time. This raw perception point cloud data contains rich information about the airspace environment, but due to differences in the position, attitude, and scanning time of different aircraft, this data is scattered in time and space and lacks a unified standard.

[0013] Step S200: Perform spatiotemporal registration and semantic enhancement processing on the original sensing point cloud data to generate a spatial semantic feature set containing spatial object attributes and environmental association features.

[0014] Spatiotemporal registration and semantic enhancement are processes for in-depth mining and integration of raw perceptual point cloud data. The purpose of spatiotemporal registration is to address the inconsistencies in time and space between point cloud data collected by different aircraft. Due to differences in flight trajectories and acquisition times, the collected point cloud data may exhibit spatial positional deviations and temporal sequence differences. Spatiotemporal registration unifies these data to a common temporal and spatial reference frame, making the data comparable and consistent.

[0015] Semantic enhancement imbues point cloud data with semantic information, that is, it identifies different spatial objects in the point cloud data and adds corresponding semantic labels to them. For example, static buildings and trees are labeled as static features, while other flying aircraft are labeled as dynamic obstacles. This makes the data more readable and understandable, facilitating subsequent analysis and decision-making.

[0016] The spatial semantic feature set is a feature set generated after spatiotemporal registration and semantic enhancement processing. It integrates the attribute information of spatial objects (such as location, speed, size, shape, etc.) and environmental related features (such as terrain slope, electromagnetic interference intensity, meteorological disturbance, etc.).

[0017] As one implementation method, step S200 can be implemented as the following steps S210~S260: Step S210: Obtain the reference coordinate system parameters and time synchronization protocol data of each low-altitude aircraft, and perform coordinate transformation processing on the original sensing point cloud data based on the reference coordinate system parameters to obtain standardized point cloud data under a unified coordinate system.

[0018] The reference coordinate system parameters are crucial information defining the coordinate system of the airspace in which the aircraft resides. These include the origin offset, the axis rotation matrix, and the scaling factor. The origin offset describes the positional deviation of the global coordinate system origin relative to the origin of the aircraft's local coordinate system, reflecting the initial positional differences of the aircraft in space. The axis rotation matrix represents the rotation relationship between the coordinate axes of the global coordinate system and the coordinate axes of the local coordinate system. Since the aircraft may have different attitudes during flight, the orientation of its local coordinate system's coordinate axes may change. The axis rotation matrix can rotate the local coordinate system's coordinate axes to be consistent with the global coordinate system. The scaling factor is used to adjust the scale differences between the local and global coordinate systems, ensuring that data collected by different aircraft are consistent in scale.

[0019] Time synchronization protocol data is used to ensure the temporal consistency of point cloud data collected by different aircraft. In actual flight, the clocks of different aircraft may have some errors, resulting in time deviations in the collected data. The time synchronization protocol allows for the calibration of the clocks of each aircraft, ensuring that the collected data has a unified temporal reference.

[0020] As one implementation method, step S210 can be implemented as the following steps S211~S216: Step S211: Analyze the coordinate origin offset, coordinate axis rotation matrix and scale factor in the reference coordinate system parameters, and construct the coordinate transformation matrix.

[0021] Parsing the reference coordinate system parameters involves extracting the specific values ​​of the origin offset, axis rotation matrix, and scale factor from the received parameter information. This typically requires decoding and analyzing the parameter information to obtain accurate parameter values. For example, if the reference coordinate system parameters are transmitted in binary encoding, they need to be decoded according to predefined encoding rules to obtain the specific values ​​of the origin offset, axis rotation matrix elements, and scale factor.

[0022] Constructing a coordinate transformation matrix involves combining the analytically obtained origin offset, axis rotation matrices, and scaling factors into a single matrix according to specific mathematical rules. In 3D space, a 4×4 homogeneous coordinate transformation matrix is ​​typically used. Specifically, the origin offset affects the translation component of the matrix, the axis rotation matrix affects the rotation component, and the scaling factor affects the scaling component. By combining these components, a coordinate transformation matrix is ​​obtained that can transform point cloud data from a local coordinate system to a global coordinate system.

[0023] Step S212: Perform matrix multiplication on each 3D coordinate point in the original perceptual point cloud data and the coordinate transformation matrix to obtain the transformed intermediate coordinate point.

[0024] After obtaining the coordinate transformation matrix, each 3D coordinate point in the original perceptual point cloud data needs to be converted into homogeneous coordinates, that is, the 3D coordinates (x, y, z) are expanded into 4D homogeneous coordinates (x, y, z, 1). Then, the expanded homogeneous coordinates are multiplied by the coordinate transformation matrix. Through matrix multiplication, the 3D coordinate points in the local coordinate system can be transformed into the intermediate coordinate system, resulting in the transformed intermediate coordinate points.

[0025] Step S213: Extract outliers from intermediate coordinate points, identify outlier coordinate points that deviate from the overall distribution trend through neighborhood point distribution characteristics analysis, and remove outlier coordinate points.

[0026] After obtaining the intermediate coordinate points after transformation, some outliers may exist due to sensor errors, data transmission interference, and other reasons. These outliers are isolated coordinate points that deviate from the overall point cloud data distribution trend. If they are not processed, they will adversely affect the subsequent data analysis and processing results.

[0027] Neighborhood point distribution characteristic analysis is a method for identifying outlier coordinate points. It determines whether a point is an outlier based on the spatial relationships and distribution characteristics between each point and its neighbors in point cloud data. Specifically, a neighborhood range can be defined, and for each intermediate coordinate point, the number and distribution of points within its neighborhood are statistically analyzed. If the number of points in a point's neighborhood is significantly less than that of other points, or if its distance to its neighbors is significantly greater than the distances between other points and their neighbors, then that point is likely an outlier. For example, a density-based spatial clustering algorithm (DBSCAN) can be used for neighborhood point distribution characteristic analysis.

[0028] Step S214: Perform spatial grid division on the intermediate coordinate points after removing outlier coordinate points, store the coordinate points in layers based on the hierarchical index structure, and generate gridded coordinate data with hierarchical index.

[0029] Spatial meshing involves dividing the 3D space containing intermediate coordinates (after removing outliers) into several equal-sized grid cells. This is done to effectively organize and manage large amounts of point cloud data, improving data query and processing efficiency. By dividing the space into grid cells, point cloud data can be distributed across different grids, ensuring that the point cloud data within each grid has a certain spatial correlation.

[0030] Hierarchical indexing for storing coordinate points involves assigning each coordinate point to a corresponding grid cell and storing it according to its hierarchy. For example, in a 3D space, the entire space is first divided into a large cubic grid, serving as the top-level grid cell. This cubic grid is then further divided into smaller cubic grids, forming the second layer of grid cells. This process continues until the desired hierarchy is reached. For each coordinate point, its 3D coordinates determine its corresponding grid cell, which is then stored. Simultaneously, an index is created for each grid cell, recording its position within the hierarchical index structure and the coordinate information it contains. After generating gridded coordinate data with a hierarchical index, point cloud data for specific regions can be quickly located and accessed using the hierarchical index structure.

[0031] Step S215: Calculate the theoretical deviation between the gridded coordinate data and the unified coordinate system, perform trend analysis on the deviation value through the deviation fitting algorithm, and generate the deviation compensation coefficient.

[0032] After obtaining gridded coordinate data with hierarchical indexes, there may be some deviation between the gridded coordinate data and the unified coordinate system due to potential errors during coordinate transformation. Calculating the theoretical deviation between the gridded coordinate data and the unified coordinate system is to quantify the magnitude of this deviation.

[0033] By selecting reference points with known accurate locations, the difference between their positions in gridded coordinate data and their theoretical positions in a unified coordinate system can be calculated to obtain the deviation value. Deviation fitting algorithms are used to analyze the trend of the calculated deviation value. By fitting the trend of the deviation value changing with spatial position, a suitable mathematical model can be found to describe this deviation. Feasible deviation fitting algorithms include polynomial fitting and spline fitting. Taking polynomial fitting as an example, a polynomial relationship can be assumed between the deviation value and the spatial position. The coefficients of the polynomial can be solved using methods such as the least squares method to obtain the deviation fitting curve.

[0034] Step S216: The gridded coordinate data is fused with the deviation compensation coefficient to obtain standardized point cloud data in a unified coordinate system. The coordinate accuracy of the standardized point cloud data matches the accuracy level of the positioning system of the low-altitude aircraft.

[0035] After obtaining the deviation compensation coefficient, it is fused with the gridded coordinate data to correct the deviation in the gridded coordinate data. Specifically, for each coordinate point in the gridded coordinate data, the corresponding deviation compensation coefficient is obtained based on its spatial location, and then the deviation compensation coefficient is applied to the coordinate value of that coordinate point.

[0036] In this way, deviations in the gridded coordinate data can be corrected to obtain standardized point cloud data in a unified coordinate system. The coordinate accuracy of the standardized point cloud data matches the accuracy level of the positioning system of low-altitude aircraft, which means that the accuracy of the processed point cloud data can reflect the actual accuracy level of the low-altitude aircraft positioning system.

[0037] Step S220: Extract timestamp information from standardized point cloud data, perform timestamp alignment processing on point cloud data collected by different aircraft according to time synchronization protocol data, and generate a time-series point cloud sequence with temporal correlation.

[0038] The timestamp information in standardized point cloud data records the specific time each point cloud data point was acquired. In actual flight, the times at which different aircraft acquire point cloud data may differ, and even after calibration using time synchronization protocols, some time deviation may still exist. Therefore, it is necessary to perform timestamp alignment processing on point cloud data acquired by different aircraft to ensure data consistency in time.

[0039] Based on the time synchronization protocol data, the time deviation between different aircraft can be determined, and the timestamp information can be adjusted. For example, if the clock of one aircraft is ahead of the clocks of other aircraft by a certain amount of time, the timestamp of the point cloud data collected by that aircraft can be subtracted from the time deviation to align it with the timestamps of other aircraft.

[0040] Generating a temporally correlated point cloud sequence involves arranging the timestamped point cloud data in chronological order. This creates a continuous, temporally ordered sequence of point cloud data, enabling the analysis of dynamic changes in the spatial environment.

[0041] Step S230: Perform dynamic obstacle recognition processing on the temporal point cloud sequence, extract the point cloud subset of the moving object through neighborhood point density analysis, and calculate the centroid change trajectory and velocity vector parameters of the point cloud subset.

[0042] Dynamic obstacle recognition processing identifies moving objects, such as other flying aircraft and moving vehicles, from a time-series point cloud. The presence of these dynamic obstacles can threaten the flight safety of low-altitude aircraft, making accurate identification crucial. Neighborhood point density analysis is a method for extracting subsets of the point cloud containing moving objects. It determines whether a point belongs to a moving object based on the density changes between each point and its neighbors. In a time-series point cloud, the point cloud data of dynamic obstacles changes over time, manifesting as changes in neighborhood point density.

[0043] Neighborhood point density analysis can divide point cloud data in a time-series point cloud sequence into different subsets, including the subset containing moving objects—the target to be extracted. Calculating the centroid change trajectory and velocity vector parameters of the point cloud subsets is to further describe the motion state of the moving objects. The centroid is the geometric center of the point cloud subset. By calculating the centroid coordinates of the point cloud subset within each time window and arranging these coordinates in chronological order, the centroid change trajectory can be obtained. The velocity vector parameters represent the velocity and direction of the moving object. These parameters are obtained by calculating the difference between the centroid coordinates in adjacent time windows and dividing by the time interval.

[0044] As one implementation method, step S230 can be implemented as the following steps S231~S236: Step S231: Obtain a preset three-dimensional spatial neighborhood search range, perform neighborhood point search on each point cloud unit in the time-series point cloud sequence, and count the number of point clouds in the neighborhood as density feature value.

[0045] The preset 3D spatial neighborhood search range is a predefined spatial region used to determine the neighborhood range of each point cloud unit. In 3D space, the neighborhood search range can be represented by a cube, sphere, or other shaped region. Performing a neighborhood search for each point cloud unit in a temporal point cloud sequence means finding other point cloud units within the preset neighborhood search range. Spatial index structures (such as KD-trees, octrees, etc.) can be used to accelerate the neighborhood search process. During the neighborhood search, the coordinates of the point cloud unit can be used to quickly locate its region, and then neighborhood points can be found within that region and its adjacent regions.

[0046] The number of points in the neighborhood is used as a density feature value to quantify the neighborhood point density of each point cloud unit. A higher number of points indicates a higher neighborhood point density for that unit; conversely, a lower number of points indicates a lower neighborhood point density. By calculating the density feature value of each point cloud unit, the density distribution of the entire temporal point cloud sequence can be obtained.

[0047] Step S232: Perform differential operation on the point cloud density feature values ​​of continuous timestamps to generate a density change rate sequence. When the density change rate sequence meets the preset dynamic judgment conditions, mark the corresponding region as a potential dynamic region.

[0048] The purpose of performing a difference operation on the point cloud density feature values ​​of consecutive timestamps is to calculate the rate of change of point cloud density between adjacent timestamps. The difference operation can be implemented by calculating the difference between the density feature values ​​of adjacent timestamps.

[0049] By performing a difference operation on each point cloud cell in the entire temporal point cloud sequence, a density change rate sequence can be generated. This sequence reflects the change of point cloud density over time. The preset dynamic judgment condition is a pre-defined rule used to determine whether a region is a potential dynamic region. In practical applications, by analyzing the density change rate sequence and the preset dynamic judgment condition, regions in the temporal point cloud sequence that may contain dynamic obstacles can be quickly identified.

[0050] Step S233: Extract the point cloud set of potential dynamic regions, and perform clustering and segmentation of the point cloud set using a region clustering algorithm to generate multiple independent point cloud clusters, each point cloud cluster corresponding to a potential moving object.

[0051] After identifying potential dynamic regions, the point cloud set of these regions needs to be extracted and further processed. Region clustering algorithms are used to divide the point cloud set of potential dynamic regions into multiple independent point cloud clusters, each cluster representing a potential moving object. Examples of region clustering algorithms include K-means clustering and DBSCAN.

[0052] Region clustering algorithms can segment a point cloud set of potential dynamic regions into multiple independent point cloud clusters. Each point cloud cluster corresponds to a potential motion. For example, in the low-altitude area of ​​a city, the marked potential dynamic region may contain multiple moving aircraft. Region clustering algorithms can segment the point cloud data of these aircraft into different point cloud clusters, with each cluster corresponding to one aircraft. This allows for more accurate identification and tracking of the motion state of each dynamic obstacle.

[0053] Step S234: Calculate the three-dimensional centroid coordinates of each point cloud cluster under continuous timestamps, arrange the three-dimensional centroid coordinates in time order, and generate the centroid change trajectory.

[0054] After obtaining multiple independent point cloud clusters, the three-dimensional centroid coordinates of each point cloud cluster are calculated at consecutive timestamps. The centroid is the geometric center of the point cloud cluster, which can be obtained by calculating the average coordinates of all points in the cluster.

[0055] By arranging the 3D centroid coordinates of each point cloud cluster at consecutive timestamps in chronological order, a centroid change trajectory can be generated. The centroid change trajectory reflects the motion path of each potential moving object in the spatial domain.

[0056] Step S235: Based on the time interval of the centroid change trajectory, calculate the centroid coordinate difference between adjacent timestamps, and use the ratio of the difference to the time interval as the velocity vector parameter. The velocity vector parameter contains direction information and magnitude information.

[0057] After obtaining the trajectory of the center of mass change, the velocity vector parameters can be calculated based on the time interval of the trajectory. The time interval of the trajectory refers to the time difference between adjacent timestamps. The difference in the center of mass coordinates between adjacent timestamps is calculated, and the ratio of this difference to the time interval is used as the velocity vector parameter.

[0058] Step S236: Perform time-series smoothing on the velocity vector parameters to eliminate the influence of instantaneous fluctuations on velocity calculation and generate a smoothed velocity vector parameter sequence.

[0059] After the velocity vector parameters are calculated, they may fluctuate instantaneously due to sensor errors, data noise, and other reasons. These instantaneous fluctuations can affect the accurate judgment of the velocity of a moving object, so it is necessary to perform time-series smoothing on the velocity vector parameters.

[0060] Time-series smoothing can be achieved using various methods, such as moving average and Kalman filtering. Through time-series smoothing, instantaneous fluctuations in velocity vector parameters can be eliminated, resulting in a more stable and accurate velocity vector parameter sequence.

[0061] Step S240: Combining the centroid change trajectory and velocity vector parameters, add semantic labels to the static ground feature point cloud and dynamic obstacle point cloud in the time-series point cloud sequence to generate an initial semantic point cloud containing semantic classification information.

[0062] After obtaining the centroid change trajectory and velocity vector parameters, we can determine whether the points in the point cloud data belong to static features or dynamic obstacles. The centroid change trajectory of static features is usually fixed, and the velocity vector parameter is close to zero; while the centroid change trajectory of dynamic obstacles changes significantly over time, and the velocity vector parameter is not zero.

[0063] By combining the centroid change trajectory and velocity vector parameters, semantic labels are added to the point cloud data in the time-series point cloud sequence. Point clouds with a fixed centroid change trajectory and velocity vector parameters close to zero are labeled as static ground feature point clouds; point clouds with a clear centroid change trajectory and non-zero velocity vector parameters are labeled as dynamic obstacle point clouds.

[0064] Generating an initial semantic point cloud containing semantic classification information involves integrating point cloud data with added semantic labels. The initial semantic point cloud not only contains the three-dimensional coordinate information of the point cloud data, but also semantic classification information, i.e., whether each point belongs to a static feature or a dynamic obstacle.

[0065] Step S250: Perform environmental feature association processing on the initial semantic point cloud, extract terrain slope features, electromagnetic interference intensity features and meteorological disturbance features within the neighborhood of the point cloud, and generate an environmental association feature vector.

[0066] Environmental feature association processing involves associating the initial semantic point cloud with features of the surrounding environment to obtain more comprehensive airspace environmental information. Topographic slope features reflect the topographic undulations of the area where the point cloud is located, electromagnetic interference intensity features indicate the degree of electromagnetic interference in the area, and meteorological disturbance features reflect the impact of meteorological conditions (such as wind speed, wind direction, and air pressure) on the airspace environment.

[0067] Extracting terrain slope features, electromagnetic interference intensity features, and meteorological disturbance features within the neighborhood of a point cloud can be achieved using various sensors and data sources. For terrain slope features, lidar point cloud data can be used for terrain reconstruction to calculate the terrain slope within the point cloud neighborhood. For example, the terrain slope value can be obtained by calculating the elevation difference and horizontal distance between adjacent points in the point cloud data. For electromagnetic interference intensity features, electromagnetic sensors can be used to measure the electromagnetic signal strength within the point cloud neighborhood.

[0068] Generating an environmental feature vector involves integrating extracted terrain slope features, electromagnetic interference intensity features, and meteorological disturbance features. These features can be arranged in a specific order to form a vector.

[0069] Step S260: Integrate semantic classification information with environmental association feature vectors to construct a spatial semantic feature set that includes descriptions of spatial object attributes and multi-dimensional environmental influencing factors.

[0070] After obtaining the initial semantic point cloud containing semantic classification information and the environmental association feature vector, they are fused to construct a more comprehensive spatial semantic feature set.

[0071] Semantic classification information includes whether each point in the point cloud data belongs to a static feature or a dynamic obstacle, as well as the attributes of these objects (such as location, velocity, size, etc.). The environmental association feature vector includes multi-dimensional environmental influencing factors such as terrain slope features, electromagnetic interference intensity features, and meteorological disturbance features within the neighborhood of the point cloud.

[0072] The fusion of semantic classification information and environmentally related feature vectors can be achieved by concatenating these two types of feature vectors. Constructing a spatial semantic feature set that includes descriptions of airspace object attributes and multi-dimensional environmental influencing factors involves integrating the fused feature vectors from all point cloud data. This spatial semantic feature set can more comprehensively and accurately describe the environmental state of low-altitude airspace, providing a rich and accurate data foundation for subsequent construction of airspace digital twin models and conflict risk prediction.

[0073] Step S300: Input the spatial semantic feature set into the pre-constructed spatial digital twin model, and generate an implicit representation of the spatial state through neural radiation field modeling.

[0074] A spatial digital twin model is a model obtained by learning and modeling various data of the spatial environment, simulating the state and behavior of the real spatial environment. Pre-built spatial digital twin models are typically trained on a large amount of data and have the ability to accurately model and predict the spatial environment. Neural radiation field modeling is a method that uses neural networks to represent radiation fields in three-dimensional space. Radiation fields describe the propagation and interaction of light in space, including information such as light density and color. Through neural radiation field modeling, the set of spatial semantic features can be mapped to an implicit function, which can represent various attributes and characteristics of the spatial state.

[0075] A spatial semantic feature set is input into a pre-constructed airspace digital twin model. The model processes and analyzes the input data, generating an implicit representation of the airspace state through neural radiation field modeling. This implicit representation of the airspace state is a parameterized representation of the physical field attributes and semantic relationships in the three-dimensional airspace. It does not directly represent specific objects and features in the airspace, but rather describes the overall characteristics and changing patterns of the airspace state through implicit functions. This implicit representation of the airspace state can be used to simulate the airspace state of the region, predict aircraft flight trajectories and conflict risks, and provide decision support for the flight management of low-altitude aircraft.

[0076] As one implementation method, step S300 can be specifically implemented as the following steps S310~S360: Step S310: Analyze the network structure parameters of the spatial domain digital twin model, determine the voxel partitioning method and activation function type of the neural radiation field, perform three-dimensional mesh partitioning processing on the spatial semantic feature set based on the voxel partitioning method, and generate voxelized feature mesh.

[0077] The network structure parameters of a spatial domain digital twin model include the number of layers, the number of neurons in each layer, and the connection methods. Voxel partitioning methods include uniform and non-uniform partitioning. Uniform partitioning divides the three-dimensional space into voxel units of equal size, while non-uniform partitioning divides the space into voxel units of different sizes based on spatial characteristics and data distribution. Activation functions include, for example, the ReLU function and the Sigmoid function.

[0078] The three-dimensional meshing process for spatial semantic feature sets based on voxel partitioning involves dividing the spatial semantic feature set into voxel units. Each voxel unit contains the spatial semantic feature information within that region. Generating the voxelized feature mesh involves organizing and storing the partitioned voxel units to form a three-dimensional mesh structure.

[0079] Step S320: Extract the spatial object attribute information from the voxelized feature mesh, and map the spatial object attribute information to the initial values ​​of the density field parameters and color field parameters of the neural radiation field.

[0080] The voxelized feature mesh contains attribute information of spatial objects, such as position, velocity, size, and shape. Extracting this spatial object attribute information is for application in neural radiation field modeling. The density field parameter of the neural radiation field represents the density distribution of light rays in space, and the color field parameter represents the color distribution of light rays in space. Mapping the spatial object attribute information to the initial values ​​of the density and color field parameters of the neural radiation field determines the initial state of the density and color fields based on the spatial object attribute information.

[0081] In practice, mapping functions can be established to map the attribute information of airspace objects to the initial values ​​of density field parameters and color field parameters. For example, the density field parameter value can be calculated based on the speed and size of the aircraft; and the color field parameter value can be determined based on the color of the aircraft.

[0082] Step S330: Call the position encoding module of the spatial domain digital twin model to perform spatial feature expansion processing on the three-dimensional coordinates of the voxelized feature mesh, and generate an encoded coordinate vector with full-size spatial feature expression capability.

[0083] The position encoding module is used to perform spatial feature expansion processing on the 3D coordinates of the voxelized feature mesh. In neural radiation field modeling, directly using 3D coordinates may not be sufficient to express the spatial characteristics and variation patterns, therefore, it is necessary to expand the 3D coordinates.

[0084] Spatial feature extension processing can be achieved by encoding 3D coordinates using methods such as trigonometric functions and polynomials. The position encoding module generates an encoded coordinate vector with full-scale spatial feature representation capabilities by performing a series of transformations and combinations on the 3D coordinates. This encoded coordinate vector contains more spatial feature information and can better represent spatial position and feature changes. The dimension of the generated encoded coordinate vector matches the dimension of the input layer of the neural radiation field, ensuring that the encoded coordinate vector can be correctly input into the multilayer perceptron network of the neural radiation field for processing.

[0085] As one implementation method, step S330 can be implemented as the following steps S331~S336: Step S331: parse the coding dimension parameters and frequency band distribution range of the position coding module, and generate a combination of periodic functions containing different frequency band components based on the frequency band distribution range.

[0086] The coding dimension parameter of the location coding module determines the dimension of the coded coordinate vector, and the frequency band distribution range specifies the frequency range of the periodic function. Analyzing these parameters is to understand the specific settings and requirements of the location coding module.

[0087] Generating a combination of periodic functions containing different frequency band components based on the frequency band distribution range involves using periodic functions such as trigonometric functions to encode three-dimensional coordinates. For example, sine and cosine functions can be used to generate a combination of periodic functions.

[0088] Step S332: Combine each coordinate axis component of the three-dimensional coordinates of the voxelized feature mesh with a periodic function to generate the frequency band feature vector corresponding to each coordinate axis component.

[0089] Each coordinate component of the three-dimensional coordinates (x, y, z) of the voxelized feature mesh is combined with a periodic function to perform the operation. This involves substituting x, y, and z into the combination of periodic functions to obtain the frequency band feature vector corresponding to each coordinate component.

[0090] Step S333: Perform channel concatenation processing on the frequency band feature vectors, stacking the frequency band feature vectors corresponding to the three coordinate axis components according to the channel dimension to generate the initial coding vector.

[0091] Channel concatenation stacks the frequency band feature vectors corresponding to the three coordinate axis components along the channel dimension. The channel dimension refers to the dimension along which the vectors are concatenated, linking the three frequency band feature vectors into a longer vector. Through channel concatenation, the frequency band feature vectors of each coordinate axis component can be integrated into an initial encoded vector. This initial encoded vector contains all the feature information of the three-dimensional coordinates in different frequency bands.

[0092] Step S334: Extract the continuous spatial features and abrupt spatial features from the initial encoding vector, calculate the influence weights of the continuous spatial features and abrupt spatial features using the feature contribution analysis algorithm, and generate feature weight coefficients.

[0093] The initial encoding vector contains continuous spatial features and abrupt spatial features. Continuous spatial features represent features that change slowly in space, such as the gentle undulations of terrain; abrupt spatial features represent features that change abruptly in space, such as the edges of buildings.

[0094] Extracting continuous and abrupt spatial features from the initial coding vector can be achieved through signal processing and feature extraction methods. For example, a low-pass filter can be used to extract continuous spatial features, and a high-pass filter can be used to extract abrupt spatial features.

[0095] Feature contribution analysis algorithms are used to calculate the influence weights of features in continuous and abrupt spatial representations. By analyzing the degree to which these features contribute to the spatial feature representation, the importance of each feature is determined. Examples of feature contribution analysis algorithms include Principal Component Analysis (PCA) and Correlation Analysis. Generating feature weight coefficients involves integrating the calculated influence weights into a coefficient vector. This coefficient vector is used in subsequent weighting processes to highlight important features and suppress less important ones.

[0096] Step S335: Multiply the initial encoding vector element-wise with the feature weight coefficients to generate a weighted encoding vector.

[0097] Element-wise multiplication of the initial encoding vector with the feature weight coefficients involves multiplying each element of the initial encoding vector by the corresponding element of the feature weight coefficient. This element-wise multiplication weights the initial encoding vector, highlighting important spatial features and suppressing less important ones.

[0098] Step S336: Perform feature scale unification processing on the weighted encoding vector, constrain the vector feature scale within a preset range, and obtain an encoding coordinate vector with full-size spatial feature expression capability. The dimension of the encoding coordinate vector matches the input layer dimension of the neural radiation field.

[0099] Feature scale unification is performed to ensure that the feature scales in the weighted encoding vector are within a reasonable range. Since different spatial features may have different scales, this can affect the training performance of the neural network. Therefore, feature scale unification of the weighted encoding vector is necessary. Normalization methods, such as min-max normalization and z-score normalization, can be used to constrain the vector feature scales within a preset range.

[0100] By unifying the feature scale, an encoded coordinate vector with full-size spatial feature representation capability is obtained. The dimension of this encoded coordinate vector matches the input layer dimension of the neural radiation field, ensuring that the encoded coordinate vector can be correctly input into the multilayer perceptron network of the neural radiation field for processing.

[0101] Step S340: Input the encoded coordinate vector and the initial values ​​of the density field parameters and color field parameters into the multilayer perceptron network of the neural radiation field to perform radiation field parameter optimization training and generate an implicit function that characterizes the spatial location radiation characteristics.

[0102] A multilayer perceptron (MLP) network for neural radiation fields is a neural network composed of multiple neuron layers. It processes and learns the input encoded coordinate vector, density field parameters, and initial values ​​of the color field parameters. The encoded coordinate vector and the initial values ​​of the density and color field parameters are input into the MLP network. The network performs forward propagation calculations based on the input data and outputs predicted density and color field parameters. Radiation field parameter optimization training involves comparing the predicted density and color field parameters with the actual data (such as measured light density and color) and adjusting the weight parameters of the MLP network to make the predictions as close to the actual data as possible. Optimization training algorithms include stochastic gradient descent (SGD) and the Adam optimization algorithm.

[0103] The implicit function representing the spatial location radiation characteristics is obtained after the radiation field parameters have been optimized and trained. This implicit function can represent the spatial location radiation characteristics. Based on the input encoded coordinate vector, it can output the corresponding density field and color field parameters, thereby describing the propagation and interaction of light in the spatial domain.

[0104] Step S350: Perform ray projection sampling processing on the spatial domain based on implicit functions, calculate the radiation field rendering results under different observation perspectives, and adjust the implicit function parameters through error feedback with real perceived data.

[0105] Ray projection sampling processing is based on implicit functions to sample the spatial domain to obtain radiation field information from different observation perspectives. It simulates the process of light starting from a virtual camera, propagating in the spatial domain, and interacting with objects.

[0106] Based on implicit functions, spatial ray projection sampling is performed. First, the observation viewpoint parameters of the virtual camera are determined, including the camera intrinsic matrix, extrinsic matrix, and field of view. The camera intrinsic matrix describes the camera's internal parameters, such as focal length and principal point position; the extrinsic matrix describes the camera's position and attitude in the spatial domain; and the field of view defines the spatial range that the camera can observe. An initial ray projection direction vector is generated based on the observation viewpoint parameters. Starting from each pixel on the virtual camera's imaging plane, spatial sampling is performed along this initial ray projection direction vector. During sampling, the 3D coordinates of the sampling points are input into an implicit function, which outputs the density and color values ​​for each sampling point. The density and color values ​​are modulated by incorporating environmental interference intensity from local spatial features to account for the influence of environmental factors on light propagation.

[0107] Calculating the radiation field rendering results from different observation perspectives involves integrating the density and color values ​​of the sampled points to obtain the rendered color value for each pixel. By arranging the rendered color values ​​of all pixels, a radiation field rendering image from the observation perspective is generated.

[0108] Adjusting the implicit function parameters through error feedback from real-world perceived data involves comparing the calculated radiation field rendering result with real-world perceived data (such as an actual photographed image) and calculating the error between the two. Based on the error feedback, an optimization algorithm is used to adjust the parameters of the implicit function, making the radiation field rendering result closer to the real-world perceived data.

[0109] As one implementation method, step S350 can be implemented as the following steps S351~S358: Step S351: Obtain the preset observation view parameters of the virtual camera, including the camera intrinsic parameter matrix, extrinsic parameter matrix and field of view range, and generate an initial ray projection direction vector based on the observation view parameters.

[0110] The preset observation angle parameters of the virtual camera are pre-defined parameters used to define the observation range and attitude of the virtual camera. The camera intrinsic parameter matrix describes the camera's internal optical characteristics, such as focal length and principal point position, which determines the camera's imaging method and image distortion. The camera extrinsic parameter matrix describes the camera's position and attitude in the spatial domain, including translation and rotation information. The field of view specifies the range of spatial angles that the camera can observe, determining the size of the camera's field of view.

[0111] Generating the initial ray projection direction vector based on the observation viewpoint parameters involves determining the ray projection direction originating from each pixel on the virtual camera's imaging plane, based on the camera's position, attitude, and field of view. The initial ray projection direction vector can be calculated using the camera's imaging model and geometric relationships.

[0112] Step S352: Call the local feature query interface of the implicit representation of the airspace state to extract the local airspace features on the initial ray projection path. The local airspace features include the spatial distribution information of dynamic obstacle density and environmental interference intensity.

[0113] The local feature query interface for implicit airspace state representation is a predefined interface for querying local airspace features. By calling this interface, local airspace features along the initial ray projection path can be obtained. Dynamic obstacle density reflects the number and distribution of dynamic obstacles (such as other aircraft) in a specific area, while environmental interference intensity reflects the impact of electromagnetic interference, meteorological disturbances, and other factors on light propagation and the airspace environment within that area.

[0114] Extracting local spatial features along the initial ray projection path can be achieved based on the previously generated implicit spatial state representation. This implicit representation already contains various spatial attributes and features. Through the local feature query interface, the spatial distribution information of dynamic obstacle density and environmental disturbance intensity along the initial ray projection path can be extracted based on its three-dimensional coordinates.

[0115] Step S353: Dynamically adjust the light projection direction vector according to the local features of the spatial domain. By analyzing the degree of feature difference between adjacent point cloud units, when the degree of feature difference meets the preset judgment conditions, mark the corresponding area as the sampling area to be enhanced, increase the light projection density in the sampling area to be enhanced, and decrease the light projection density in the static uniform area to generate an optimized set of light projection direction vectors.

[0116] Dynamically adjusting the ray projection direction vector based on local spatial features is to more accurately capture changes in the spatial environment. The degree of feature difference between adjacent point cloud units can be measured by comparing their local features such as dynamic obstacle density and environmental disturbance intensity. The preset judgment condition is a pre-set threshold used to determine whether the degree of feature difference reaches a level requiring special processing.

[0117] When the degree of feature difference meets the preset judgment conditions, it indicates that there may be significant feature changes in the region, such as the boundary of dynamic obstacles or abrupt changes in environmental interference. This region is then marked as a region to be enhanced for sampling. Increasing the ray projection density in the region to be enhanced allows for more detailed acquisition of the region's radiation field information, improving the accuracy of the rendering results. For static, uniform regions, due to their small feature changes, the ray projection density can be reduced, minimizing unnecessary computation and improving processing efficiency.

[0118] Generating an optimized set of ray projection direction vectors involves integrating the adjusted ray projection direction vectors.

[0119] Step S354: Starting from each pixel on the imaging plane of the virtual camera, perform spatial sampling along the optimized set of light projection direction vectors to generate a sampling point sequence. The distribution density of the sampling point sequence is positively correlated with the degree of dynamic change of local spatial features.

[0120] Starting from each pixel on the virtual camera's imaging plane, spatial sampling is performed along the optimized set of ray projection direction vectors. This involves sampling in the spatial domain according to the adjusted ray projection direction. The sampling point sequence is a series of three-dimensional coordinate points generated during the sampling process. These points reflect the propagation path of light in the spatial domain and their interaction positions with objects.

[0121] The distribution density of the sampling point sequence is positively correlated with the degree of dynamic change of local spatial features, meaning that the distribution density of sampling points will be higher in areas with greater dynamic changes, and lower in areas with less dynamic changes. This is because more sampling points are needed in areas with greater dynamic changes to accurately capture the changes and feature information of the radiation field.

[0122] Step S355: Input the three-dimensional coordinates of the sampling point sequence into the implicit function, output the density value and color value of each sampling point, and modulate the density value and color value by combining the environmental interference intensity in the local spatial features.

[0123] The three-dimensional coordinates of the sampling point sequence are input into an implicit function, which uses a previously generated implicit function characterizing the spatial location radiation properties to calculate the density and color values ​​of each sampling point. Based on the input three-dimensional coordinates, the implicit function outputs the corresponding density and color field parameters, i.e., the density and color values ​​of the sampling point.

[0124] Modulating density and color values ​​by incorporating environmental interference intensity within local spatial features takes into account the impact of environmental interference on light propagation and object color. Environmental interference intensity can cause phenomena such as light scattering and absorption, thus affecting the density and color values ​​of sampling points. For example, in a spatial region with strong electromagnetic interference, the high intensity of interference can hinder light propagation, potentially reducing the density value and altering the color value of the sampling point. A pre-established modulation model can be used to adjust density and color values ​​based on the intensity of environmental interference. For instance, a linear modulation model can be used to multiply the density and color values ​​by a coefficient related to the intensity of environmental interference to achieve modulation. This more accurately reflects the impact of the spatial environment on light propagation and object color, improving the realism of the radiation field rendering.

[0125] Step S356: Calculate the opacity parameter of the sampling point based on the modulated density value. The opacity parameter increases with the increase of environmental interference intensity.

[0126] The opacity parameter represents the ability of a sampling point to block light and is closely related to the density value of that sampling point. The opacity parameter of a sampling point can be calculated based on the modulated density value using a predefined mapping function. Generally, the higher the density value, the higher the opacity parameter of the sampling point.

[0127] The intensity of environmental interference affects the propagation and scattering of light. As the intensity of interference increases, light is more easily scattered and absorbed, and the sampling point's ability to block light also increases. Therefore, the opacity parameter increases with the intensity of environmental interference. For example, in a foggy airspace, the environmental interference intensity is high, and light will be heavily scattered and absorbed by the fog during propagation, resulting in a corresponding increase in the opacity parameter of the sampling point. The opacity parameter calculated based on the density value can be adjusted by establishing a correction coefficient related to the intensity of environmental interference. For example, multiplying the opacity parameter by a correction coefficient that increases with the intensity of environmental interference achieves the effect of the opacity parameter increasing with the intensity of environmental interference.

[0128] Step S357: Perform temporal correlation integration on the opacity parameter and the modulated color value. The weight of the integration operation increases as the distance between the sampling point and the dynamic obstacle decreases, generating the dynamic rendering color value of the pixel.

[0129] Temporal correlation integration calculates the integration of the opacity parameter and the modulated color value over time to account for the cumulative effect of light propagation. The weight of the integration operation increases as the distance between the sampling point and the dynamic obstacle decreases. This is because the interaction between light and the obstacle is stronger in areas closer to the obstacle, resulting in a greater impact on the color value of the pixel.

[0130] For example, in an airspace where aircraft are flying, sampling points closer to the aircraft contribute more to the pixel's color value. A distance-based weighting function can be used to determine the weights for the integration operation. For each sampling point, a weight value is calculated based on its distance from dynamic obstacles; the smaller the distance, the larger the weight value. Then, the opacity parameter and the modulated color value are multiplied by the corresponding weight value, and the integration operation is performed.

[0131] Step S358: Arrange the dynamic rendering color values ​​of all pixels according to the position of the imaging plane to generate a radiation field rendering image from the observation perspective. The radiation field rendering image contains the dynamic visual feature distribution of the spatial domain and the visual representation of the influence of environmental interference.

[0132] Arranging the dynamic rendering color values ​​of all pixels according to their positions on the imaging plane involves organizing and arranging the dynamic rendering color values ​​of each pixel according to its position on the virtual camera's imaging plane, forming a two-dimensional image matrix. This image matrix is ​​the radiation field rendering image from the observation viewpoint.

[0133] The radiation field rendering image includes a dynamic visual distribution of spatial features and a visual representation of the impact of environmental interference. The dynamic visual feature distribution reflects the movement and distribution of dynamic obstacles (such as aircraft) in the airspace, represented by different colors and brightness. The visual representation of the impact of environmental interference reflects the influence of environmental factors such as electromagnetic interference and meteorological disturbances on light propagation and the airspace environment. For example, in areas with strong electromagnetic interference, the image may appear blurry or distorted in color.

[0134] Step S360: When the error between the radiation field rendering result and the real perceived data meets the preset convergence condition, save the current implicit function parameters as the implicit representation of the spatial state. The implicit representation of the spatial state represents the physical field properties and semantic association information in the three-dimensional space of the spatial domain through implicit function parameterization.

[0135] The error between the radiation field rendering result and the real-world perceived data is calculated by comparing the rendered radiation field image with the actual acquired spatial images (such as images taken by a camera). For example, image error metrics such as mean squared error (MSE) and structural similarity index (SSIM) are used to calculate the error. A preset convergence condition is a pre-defined error threshold. When the calculated error is less than this threshold, it indicates that the radiation field rendering result is very close to the real-world perceived data and has reached acceptable accuracy.

[0136] Saving the current implicit function parameters as an implicit representation of the spatial state involves storing these implicit function parameters at this time. These parameters can accurately describe the physical field properties (such as light density, color distribution, etc.) and semantic association information (such as the classification of dynamic obstacles and static features) in the three-dimensional spatial space.

[0137] Step S400: Call the multi-agent collaborative reasoning model to predict the conflict risk of the implicit representation of the airspace state and output a conflict risk prediction map containing the relative positional relationships between aircraft.

[0138] Multi-agent collaborative reasoning models are used to handle the interaction and reasoning between multiple agents (such as multiple low-altitude aircraft). They can comprehensively consider the state information and airspace environment information of multiple aircraft to predict the risk of conflict between aircraft.

[0139] Using a multi-agent collaborative reasoning model to predict conflict risk based on the implicit representation of airspace state involves taking the implicit representation of airspace state as input. The model analyzes factors such as the trajectories, speeds, and positions of multiple aircraft based on the input information to predict the likelihood of conflict between them. Conflicts between aircraft may include collisions, flight path intersections, and other similar situations.

[0140] The output includes a conflict risk prediction map showing the relative positions of the aircraft, presenting the conflict risk prediction results in a visual way. The conflict risk prediction map can use different colors, markers, etc., to represent different levels of conflict risk, while also displaying the relative positions of the aircraft.

[0141] As one implementation method, step S400 can be implemented as the following steps S410~S460: Step S410: Parse the input interface protocol of the multi-agent collaborative reasoning model, and convert the implicit representation of the spatial state into a model-compatible feature tensor format. The feature tensor format includes the implicit function parameters of the three-dimensional spatial grid and the motion state vector of the dynamic obstacle.

[0142] The input interface protocol for multi-agent collaborative reasoning models is a predefined protocol used to standardize the format and requirements of the model's input data. Parsing this protocol is to understand the model's specific requirements for input data, including data type, dimensions, and format.

[0143] Converting the implicit representation of the spatial state into a model-compatible feature tensor format requires data conversion based on the parsed input interface protocol. The feature tensor format is a multi-dimensional data structure that can effectively organize and represent various information about the spatial state. The implicit function parameters of the three-dimensional spatial mesh describe the physical field properties and structural information of the spatial domain, while the motion state vectors of dynamic obstacles contain motion information such as the position, velocity, and acceleration of the dynamic obstacles (e.g., aircraft).

[0144] For example, implicit function parameters in the implicit representation of spatial state can be arranged and organized according to certain rules to form a multidimensional tensor. For the motion state vectors of dynamic obstacles, the motion information of each obstacle (such as position coordinates, velocity components, etc.) can be combined into a vector, and the motion state vectors of multiple obstacles can be arranged in a certain order to form a matrix or tensor.

[0145] Step S420: Extract the current flight path planning data of each low-altitude aircraft, calculate the predicted track point sequence within a preset time period based on the flight path planning data, and convert the predicted track point sequence into a three-dimensional coordinate path in the implicit representation of airspace status.

[0146] The current flight path planning data for each low-altitude aircraft consists of its pre-defined flight route and plan, including information such as origin, destination, waypoints, and flight speed. This data can be extracted through communication with the aircraft's flight control system.

[0147] Calculating a predicted sequence of waypoints for a predetermined time period based on flight path planning data involves predicting the aircraft's flight trajectory over a future period based on its current position, speed, and flight path planning. This can be done using kinematic models; for example, assuming the aircraft is moving at a constant velocity or with uniform acceleration, the position at different time points can be calculated based on its current speed and acceleration.

[0148] Converting the predicted waypoint sequence into a 3D coordinate path in the implicit representation of airspace state involves associating the 3D coordinate information of the predicted waypoint sequence with the implicit representation of airspace state. This allows the 3D coordinates of the waypoint sequence to be mapped onto a 3D airspace grid, forming a continuous 3D coordinate path.

[0149] Step S430: Input the implicit representation of the three-dimensional coordinate path and the airspace state into the relationship modeling module of the multi-agent collaborative reasoning model to construct a spatial relationship graph network of aircraft-aircraft and aircraft-obstacles. The nodes of the spatial relationship graph network represent airspace objects, and the edges represent the relative positional relationship and distance parameters between objects.

[0150] The relationship modeling module of the multi-agent collaborative reasoning model is used to construct spatial relationships between airspace objects. By inputting 3D coordinate paths and implicit representations of airspace states into this module, it constructs a spatial relationship graph network of aircraft-aircraft and aircraft-obstacles based on this information.

[0151] A spatial relationship graph network is a graph structure in which nodes represent spatial objects, such as aircraft and obstacles (buildings, other aircraft, etc.), and edges represent the relative positional relationships and distance parameters between objects. Relative positional relationships can be represented by information such as azimuth and altitude difference, while distance parameters represent the spatial distance between two objects.

[0152] As one implementation method, step S430 can be implemented as the following steps S431~S436: Step S431: Take each low-altitude aircraft and dynamic obstacle as an airspace object node, and extract the attribute features of each node, including the current position coordinates, motion direction vector and size parameters.

[0153] Treating low-altitude aircraft and dynamic obstacles as airspace object nodes involves abstracting these entities into nodes in a graph network. Each node represents an airspace object and has its own attributes and characteristics.

[0154] Extract the attribute features of each node, including its current position coordinates, motion direction vector, and size parameters. The current position coordinates describe the node's specific location in the airspace and can be represented using three-dimensional coordinates. The motion direction vector represents the node's direction and velocity, containing both directional and size information. The size parameters reflect the size and shape of the object the node represents, such as the length, width, and height of an aircraft.

[0155] Step S432: Calculate the spatial distance parameter between any two nodes, determine the connection relationship between nodes based on the spatial distance parameter, and establish a connection edge between the corresponding nodes when the spatial distance parameter meets the preset connection conditions.

[0156] The spatial distance between any two nodes can be calculated using the Euclidean distance formula. The connection relationship between nodes is determined based on this spatial distance parameter, with a pre-defined distance threshold as the default connection condition. When the spatial distance between two nodes is less than this threshold, it indicates that the two nodes are relatively close and may have a mutual influence relationship; therefore, a connection edge is established between the corresponding nodes.

[0157] Step S433: Extract the relative positional relationship between the nodes at both ends of the connecting edge, including azimuth information and height difference information, and encode the azimuth information and height difference information into edge attribute parameters.

[0158] The relative positional relationship between the nodes at both ends of the connecting edge is extracted. Azimuth information describes the direction of one node relative to another, while height difference information represents the vertical height difference between the two nodes. Azimuth information can be calculated using trigonometric functions and vector operations. For example, for the position coordinates of two nodes (x1, y1, z1) and (x2, y2, z2), their projection vectors on the horizontal plane can be calculated, and then the azimuth angle can be calculated using the formula for the angle between vectors. Height difference information can be directly obtained from the difference in the z-coordinates of the two nodes.

[0159] Encoding azimuth and elevation difference information as edge attribute parameters involves converting this information into a data format suitable for graph network processing. Azimuth and elevation difference information can be combined into a vector or matrix to serve as edge attribute parameters. For example, azimuth and elevation difference information can be arranged in a specific order to form a two-dimensional vector, which can then be used as the attribute parameters for connecting edges.

[0160] Step S434: Combine distance parameters and edge attribute parameters to construct a spatial relationship graph network containing node attribute matrices and edge attribute matrices. The row dimension of the node attribute matrix corresponds to the node number, and the column dimension corresponds to the attribute feature dimension.

[0161] Combining distance parameters and edge attribute parameters involves integrating the previously calculated spatial distance parameters between nodes and edge attribute parameters (azimuth information and height difference information).

[0162] A spatial relational graph network is constructed, comprising node attribute matrices and edge attribute matrices. The node attribute matrix stores the attribute features of each node, with its rows corresponding to node numbers and its columns corresponding to attribute feature dimensions. For example, each row of the node attribute matrix represents a node, and each column represents an attribute feature (such as current position coordinates, motion direction vector, size parameters, etc.). The edge attribute matrix stores the attribute parameters of connecting edges, with its rows and columns corresponding to the start and end node numbers of the connecting edges, respectively. By constructing such a matrix structure, the attribute information of nodes and edges in the spatial relational graph network can be effectively organized and represented.

[0163] Step S435: Optimize the spatial relationship graph network by removing redundant connection edges that contribute less than a preset threshold to the prediction of conflict risk, retaining key connection relationships, and generating a simplified spatial relationship graph network.

[0164] Structural optimization of spatial relationship graph networks aims to reduce network complexity and improve the efficiency of conflict risk prediction. Redundant edges refer to those edges that contribute little to conflict risk prediction, possibly due to unreasonable distance threshold settings or other unnecessary connections.

[0165] The preset threshold is a pre-defined standard used to determine whether the contribution of a connecting edge is sufficiently high. Its contribution can be assessed by calculating the degree of influence of the connecting edge on the conflict risk prediction model. For example, feature importance analysis can be used to analyze the influence weight of each connecting edge's attribute parameters on the conflict risk prediction results. If the influence weight of a connecting edge is lower than the preset threshold, the connecting edge is considered redundant and is removed.

[0166] Preserving key connections means retaining those edges that significantly impact conflict risk prediction. These key connections reflect important relationships between objects in the airspace, such as connections between aircraft that may clash. By removing redundant connections, a simplified spatial relationship graph network is generated.

[0167] Step S436: Associate and bind the simplified spatial relationship graph network with the implicit representation of the spatial domain state, so that the graph network nodes can call the implicit representation parameters of the corresponding spatial domain location in real time.

[0168] Associating a simplified spatial relationship graph network with an implicit representation of spatial state involves linking nodes in the graph network to corresponding positions in the implicit representation of spatial state. Through this association, graph network nodes can obtain implicit representation parameters of their corresponding spatial positions in real time, such as physical field properties and semantic association information.

[0169] Association and binding can be achieved by establishing a mapping relationship. For example, each node in a graph network can be assigned a unique identifier, and this identifier can be mapped to the corresponding three-dimensional coordinate position in the implicit representation of the spatial state. When a graph network node needs to obtain the implicit representation parameters of the corresponding spatial position, it can query through this mapping relationship.

[0170] Step S440: Call the graph attention mechanism layer of the relational graph network to calculate the influence weights of different airspace object nodes on the current aircraft conflict risk and generate an attention weight matrix.

[0171] The graph attention mechanism layer of a relational graph network is a mechanism used to learn the importance of nodes in a graph network. It can calculate the influence weight of different airspace object nodes on the current aircraft conflict risk based on the connection relationship and attribute information between nodes.

[0172] The graph attention mechanism layer is invoked by first inputting the simplified spatial relationship graph network and node attribute information into this layer. The graph attention mechanism layer determines the importance of different nodes to the current aircraft by calculating the attention coefficients between nodes. The attention coefficients can be calculated using the attribute features of the nodes and the attribute parameters of the connecting edges. For example, an attention function can be used, taking the attribute features of the nodes (such as position, velocity, size, etc.) and the attribute parameters of the connecting edges (such as azimuth, altitude difference, distance, etc.) as input to calculate the attention coefficients between the nodes.

[0173] Generating the attention weight matrix involves normalizing the calculated attention coefficients to form a matrix. The rows and columns of the attention weight matrix correspond to the node numbers in the graph network, and each element in the matrix represents the influence weight of one node on another.

[0174] Step S450: Update the spatial relationship graph network by message passing based on the attention weight matrix, aggregate the state features of neighboring nodes, and generate node feature vectors containing global spatial relationship information.

[0175] Message passing updates in spatial graph networks based on attention weight matrices utilize these matrices to guide information transmission and updates between nodes. Message passing refers to a node transmitting its state characteristics to its neighboring nodes and receiving information from them.

[0176] Aggregating the state features of neighboring nodes involves weighted summation of the state features of the current node's neighboring nodes according to the attention weight matrix. The elements in the attention weight matrix represent the influence weights of neighboring nodes on the current node; weighted summation highlights the feature information of important neighboring nodes. For example, for a node in a graph network, its neighboring nodes include other nodes connected to it by edges. Based on the attention weight matrix, the state features of neighboring nodes (such as position, velocity, size, etc.) are multiplied by their corresponding attention weights, and then summed to obtain the aggregated neighborhood features.

[0177] Generating node feature vectors that contain global spatial domain relationship information involves fusing the original state features of the current node with the aggregated neighborhood features. Through multiple iterative message passing updates, the node's feature vectors can gradually incorporate more global spatial domain relationship information.

[0178] As one implementation method, step S450 can be implemented as the following steps S451~S456: Step S451: Initialize the state feature vector of each node, and use the attribute features of the node as the initial state features.

[0179] Initializing the state feature vector of each node involves assigning an initial feature vector to each node in the graph network. The node's attribute features are used as initial state features because these attributes (such as position, velocity, and size) reflect the node's basic state and characteristics.

[0180] For example, a node representing an aircraft might have attributes including its current position coordinates (x, y, z), motion direction vector (vx, vy, vz), and size parameters (length, width, height). These attributes are arranged in a specific order to form a vector, which serves as the node's initial state feature vector. Initializing each node's state feature vector provides the foundation for subsequent message passing and feature updates.

[0181] Step S452: Determine the set of neighboring nodes and the corresponding weight coefficients for each node based on the attention weight matrix. The set of neighboring nodes includes related nodes that have a significant impact on the current node.

[0182] The attention weight matrix determines the set of neighboring nodes and their corresponding weight coefficients for each node. Elements in the attention weight matrix represent the influence weight of one node on another. For each node, nodes with larger influence weights can be selected as its neighboring nodes based on the attention weight matrix.

[0183] The neighborhood node set contains associated nodes that significantly influence the current node. A weight threshold can be set, and nodes corresponding to elements in the attention weight matrix that exceed this threshold can be considered neighborhood nodes. Simultaneously, the attention weights corresponding to these neighborhood nodes are used as their respective weight coefficients. For example, in an airspace scenario containing multiple aircraft and obstacles, for a node representing an aircraft, its neighborhood node set is determined using the attention weight matrix, including other aircraft that may conflict with it and nearby obstacle nodes. Furthermore, the weight coefficients corresponding to these neighborhood nodes are obtained; these weight coefficients reflect the importance of the neighborhood nodes to the current node.

[0184] Step S453: Perform weighted aggregation processing on the state feature vectors of the neighborhood node set. Multiply the state feature vectors of each neighborhood node by their corresponding weight coefficients and sum them to generate neighborhood aggregation features.

[0185] Weighted aggregation of the state feature vectors of the neighborhood node set involves multiplying the state feature vector of each neighborhood node by its corresponding weight coefficient, and then summing these weighted vectors. This weighted aggregation process highlights the feature information of important neighborhood nodes while suppressing interference from unimportant ones.

[0186] Step S454: Perform feature fusion processing on the initial state features and the neighborhood aggregation features, and generate updated node state features through feature combination algorithm.

[0187] Feature fusion, which combines initial state features with neighborhood aggregation features, integrates these two features to obtain more comprehensive and accurate node state characteristics. Feature combination algorithms can employ various methods, such as concatenation and weighted summation.

[0188] Feature fusion processing can enable a node's state features to include its own basic attributes and important information from neighboring nodes, thereby improving the expressive power of the node's state features.

[0189] Step S455: Repeat the neighborhood aggregation and feature fusion process until the node state features no longer change significantly, and a stable node state feature vector is obtained.

[0190] Repeatedly executing the neighborhood aggregation and feature fusion process ensures that the state features of a node can fully absorb information from its neighboring nodes, continuously updating and improving them. In each iteration, the set of neighboring nodes and their corresponding weight coefficients are first determined based on the new attention weight matrix. Then, a weighted aggregation of the state feature vectors of the neighboring nodes is performed. Finally, the initial state features and the aggregated neighborhood features are fused together.

[0191] The process continues until the node state characteristics no longer change significantly, determined by setting a threshold value. After each iteration, the degree of change in the node state characteristic vector (e.g., the Euclidean distance of the vector) is calculated. If the degree of change is less than the set threshold, the node state characteristics are considered stable.

[0192] Step S456: Perform global information integration processing on the stable node state feature vector, incorporate the overall airspace environment features, and generate a node feature vector containing global airspace relationship information. The dimension of the node feature vector matches the input dimension of the risk prediction layer of the multi-agent collaborative reasoning model.

[0193] Global information integration of stable node state feature vectors involves incorporating the overall environmental characteristics of the airspace into these vectors. These overall airspace environmental characteristics include the global distribution of dynamic obstacle density, the overall intensity of environmental disturbances, and topographical information. This integration can be achieved by concatenating or weighting the stable node state feature vectors with the overall airspace environmental characteristics. For example, representing the overall airspace environmental characteristics as a vector and then weighting and summing it with the stable node state feature vectors according to a certain ratio yields a new vector containing global information.

[0194] A node feature vector containing global spatial domain relationship information is generated. This vector reflects the node's position in the entire spatial domain, its relationship with other nodes, and its impact on the overall spatial environment. The dimension of the node feature vector matches the input dimension of the risk prediction layer of the multi-agent collaborative reasoning model to ensure that the node feature vector is correctly input into the risk prediction layer for processing.

[0195] Step S460: Call the risk prediction layer of the multi-agent collaborative reasoning model to perform conflict risk assessment on the node feature vectors and generate a conflict risk prediction map containing risk distribution characteristics.

[0196] The risk prediction layer of the multi-agent collaborative reasoning model is invoked by inputting node feature vectors containing global airspace relationship information. The risk prediction layer is the part of the multi-agent collaborative reasoning model used to assess conflict risk, and it can evaluate the conflict risk of each node (i.e., each aircraft) based on the information in the node feature vectors.

[0197] Conflict risk assessment can be implemented based on machine learning models or statistical methods. For example, a classifier model can be used, taking node feature vectors as input and outputting the conflict risk level (e.g., low risk, medium risk, high risk) for each node. Alternatively, a regression model can be used to predict the probability of a node engaging in conflict.

[0198] Generating a conflict risk prediction map that includes risk distribution characteristics is a way to visually present the results of conflict risk assessment. The conflict risk prediction map can use different colors, markers, etc., to represent different risk levels or probabilities.

[0199] Step S500: Generate a distributed cooperative obstacle avoidance strategy based on the conflict risk prediction map and the preset obstacle avoidance strategy library, and encode the distributed cooperative obstacle avoidance strategy into a control command sequence and send it to the corresponding low-altitude aircraft.

[0200] A distributed cooperative obstacle avoidance strategy is generated based on a conflict risk prediction map and a pre-set obstacle avoidance strategy library. The conflict risk prediction map provides conflict risk information for different locations and aircraft in the airspace, while the pre-set obstacle avoidance strategy library contains various obstacle avoidance strategies for different conflict scenarios. Based on the risk information in the conflict risk prediction map, appropriate obstacle avoidance strategies can be selected from the pre-set obstacle avoidance strategy library, combined, and optimized to generate a distributed cooperative obstacle avoidance strategy.

[0201] Encoding the distributed cooperative obstacle avoidance strategy into a sequence of control commands involves converting the obstacle avoidance strategy into control commands that the aircraft can recognize and execute. These control commands include attitude adjustment parameters, power system parameters, and navigation mode switching commands. These commands are then arranged in chronological order of execution to form a control command sequence.

[0202] Control commands are transmitted to the corresponding low-altitude aircraft via wireless communication. Upon receiving the control command sequence, the aircraft performs flight operations according to the instructions, achieving obstacle avoidance and cooperative flight.

[0203] As one implementation method, step S500 can be implemented as the following steps S510~S560: Step S510: Analyze the risk area coordinates and risk level parameters in the conflict risk prediction map, and match the corresponding basic obstacle avoidance strategy type from the preset obstacle avoidance strategy library based on the risk level parameters. The basic obstacle avoidance strategy types include altitude adjustment, heading change and speed control.

[0204] This paper analyzes the risk area coordinates and risk level parameters in a conflict risk prediction map. Different colors and markers are used to represent different risk areas and risk levels. Through image recognition and data analysis techniques, the three-dimensional coordinate information of the risk areas and their corresponding risk level parameters can be extracted. Risk level parameters can be represented by numerical values ​​or classification labels, such as low risk, medium risk, and high risk.

[0205] Based on the risk level parameter, the system matches the corresponding basic obstacle avoidance strategy type from a pre-built obstacle avoidance strategy library. This library is a pre-established database containing various obstacle avoidance strategies. Basic obstacle avoidance strategy types include altitude adjustment, heading change, and speed control. A matching rule can be formulated based on the risk level parameter to select a suitable basic obstacle avoidance strategy type from the pre-built library.

[0206] As one implementation method, in step S510, matching the corresponding basic obstacle avoidance strategy type from the preset obstacle avoidance strategy library based on the risk level parameter can be implemented as the following steps S511~S517: Step S511: Parse the strategy metadata of the preset obstacle avoidance strategy library. The strategy metadata includes the strategy applicable scenario characteristics, historical execution effect indicators and the correlation weight matrix with risk characteristics.

[0207] The metadata of the preset obstacle avoidance strategy library is analyzed. Each strategy in the library has corresponding metadata. The applicable scenario characteristics of the strategy describe the airspace scenarios in which the strategy is applicable, such as near airports, urban high-rise areas, open areas, etc. Historical performance indicators record the effectiveness of the strategy in past use, such as obstacle avoidance success rate, execution time, energy consumption, etc. The correlation weight matrix with risk characteristics represents the degree of correlation between the strategy and different risk characteristics (such as the scope of risk impact, risk evolution trend, risk urgency, etc.).

[0208] Parsing this metadata can be achieved by reading data files or database records from the policy repository. For example, policy metadata can be stored in a structured database, and relevant information can be extracted using SQL queries.

[0209] Step S512: Extract the risk impact range, risk evolution trend and risk urgency characteristics from the risk level parameters, construct a risk feature vector, and adjust the risk feature vector in real time as the implicit representation of the airspace state is updated.

[0210] Extract the risk impact range, risk evolution trend, and risk urgency characteristics from the risk level parameters. The risk impact range describes the size and spatial distribution of the risk area, the risk evolution trend indicates how the risk changes over time (e.g., whether the risk increases or decreases), and the risk urgency reflects the degree of urgency required to take immediate obstacle avoidance measures.

[0211] Constructing a risk feature vector involves combining these features into a single vector. For example, feature values ​​such as the area of ​​risk impact, the rate of change of risk evolution trends, and the level of risk urgency can be arranged in a specific order to form a vector. The risk feature vector is adjusted in real time as the implicit representation of the airspace state is updated, because the airspace state is dynamic, and the risk features will change accordingly.

[0212] Step S513: Calculate the matching score between the risk feature vector and the applicable scenario features of each strategy based on the correlation weight matrix. The matching score is obtained by combining the inner product operation of the feature vector with the feedback correction of historical execution effect indicators.

[0213] The matching score between the risk feature vector and the applicable scenario features of each strategy is calculated based on the association weight matrix. The association weight matrix represents the degree of association between each strategy and different risk features. An inner product operation can be performed on the risk feature vector and the applicable scenario feature vectors of each strategy to obtain a preliminary matching score. The inner product operation of feature vectors can measure the similarity between two vectors.

[0214] Incorporating feedback from historical performance metrics for adjustment takes into account the actual effectiveness of the strategy in past use. An adjustment coefficient can be assigned to each strategy based on historical performance metrics (such as obstacle avoidance success rate, execution time, energy consumption, etc.). The initial matching score is multiplied by the adjustment coefficient to obtain the final matching score.

[0215] Step S514: Sort the matching scores in descending order and select the type corresponding to the strategy with the highest score as the candidate basic obstacle avoidance strategy type.

[0216] Sort the matching scores in descending order, which sorts the calculated matching scores of each strategy from largest to smallest. By sorting, the degree of matching between each strategy and the current risk characteristics can be clearly seen.

[0217] The strategy with the highest score is selected as the candidate basic obstacle avoidance strategy type. The strategy with the highest score represents the strategy that best matches the current risk characteristics. For example, in the low-altitude area of ​​a city, after calculation and sorting, a certain altitude adjustment strategy is found to have the highest matching score, so altitude adjustment is selected as the candidate basic obstacle avoidance strategy type. By selecting candidate basic obstacle avoidance strategy types, the direction of obstacle avoidance strategy suitable for the current risk situation can be preliminarily determined.

[0218] Step S515: Call the airspace digital twin model to quickly simulate the candidate basic obstacle avoidance strategy types, input the implicit representation of the current airspace state and the aircraft performance constraint parameters, and predict the trend of conflict risk changes after the strategy is executed.

[0219] A digital twin model of the airspace is invoked to quickly simulate candidate basic obstacle avoidance strategy types. This model is capable of simulating airspace states and aircraft behavior. The implicit representation of the current airspace state and aircraft performance constraint parameters are input into the model to simulate the execution process of candidate basic obstacle avoidance strategy types.

[0220] Aircraft performance constraints include maneuverability parameters (such as maximum turning angle and maximum acceleration), speed regulation range, and load limitations. These parameters restrict the obstacle avoidance strategies that the aircraft can execute. By simulating the changing trends of conflict risk after the strategy is executed, the effectiveness of the strategy can be evaluated.

[0221] Step S516: When the trend of conflict risk changes shows that the risk reduction meets the preset improvement conditions, the candidate basic obstacle avoidance strategy type is determined as the final matching basic obstacle avoidance strategy type; otherwise, the strategy type with the second highest score is selected for simulation verification again until the basic obstacle avoidance strategy type that meets the improvement conditions is determined.

[0222] When the trend of conflict risk changes shows that the risk reduction meets the preset improvement conditions, which are pre-defined risk reduction thresholds, the candidate basic obstacle avoidance strategy type is determined as the final matching basic obstacle avoidance strategy type if the simulation results show that the conflict risk reduction exceeds this threshold. Otherwise, the strategy type with the second highest score is selected for re-simulation verification. If the simulation results of the candidate basic obstacle avoidance strategy type do not meet the preset improvement conditions, it indicates that the strategy may not be optimal. In this case, the strategy type with the second highest matching score is selected, and the simulation process in step S515 is repeated until a basic obstacle avoidance strategy type that meets the improvement conditions is found. In this way, it can be ensured that the selected basic obstacle avoidance strategy type can effectively reduce conflict risk.

[0223] Step S517: Based on the determined basic obstacle avoidance strategy type, extract the corresponding strategy parameter template from the preset obstacle avoidance strategy library. The strategy parameter template contains the trajectory adjustment parameter items to be filled and the parameter constraint range.

[0224] Based on the determined basic obstacle avoidance strategy type, the corresponding strategy parameter template is extracted from the preset obstacle avoidance strategy library. The strategy parameter template is a predefined template used to guide the implementation of the obstacle avoidance strategy, containing the track adjustment parameter items to be filled and the parameter constraint range.

[0225] The trajectory adjustment parameters vary depending on the basic obstacle avoidance strategy type. For example, if the basic obstacle avoidance strategy is altitude adjustment, the trajectory adjustment parameters might include the adjusted altitude value and the speed of altitude adjustment; if it's a heading change, the trajectory adjustment parameters might include the changed heading angle and the angular velocity of the turn. Parameter constraints define the range of values ​​for these parameters to ensure that the aircraft does not exceed its performance limits when executing obstacle avoidance strategies.

[0226] Step S520: Extract the performance constraint parameters of each low-altitude aircraft, including maneuverability parameters, speed adjustment range and load limit conditions. Based on the performance constraint parameters, perform feasibility screening of basic obstacle avoidance strategy types and generate a set of feasible obstacle avoidance strategies.

[0227] The performance constraint parameters for each low-altitude aircraft are extracted. Maneuverability parameters describe the aircraft's turning, acceleration, and deceleration performance, such as maximum turning angle, maximum acceleration, and maximum deceleration. The speed adjustment range specifies the minimum and maximum speeds the aircraft can achieve. Load limitations indicate the performance constraints of the aircraft when carrying different loads.

[0228] These performance constraint parameters can be extracted by communicating with the aircraft's flight control system. Feasibility screening of basic obstacle avoidance strategy types based on these performance constraint parameters involves checking whether each basic obstacle avoidance strategy type is within the aircraft's performance range. For example, if a basic obstacle avoidance strategy type requires the aircraft to perform a turn exceeding its maximum turning angle, then that strategy type is not feasible.

[0229] Generating a set of feasible obstacle avoidance strategies involves combining basic obstacle avoidance strategy types that meet the aircraft's performance constraints after feasibility screening. For example, in an airspace scenario containing multiple drones of different types, for each drone, the basic obstacle avoidance strategy types are screened based on its performance constraints to generate a set of feasible obstacle avoidance strategies for each drone.

[0230] Step S530: Perform conflict resolution simulation for each strategy in the set of feasible obstacle avoidance strategies, input the trajectory adjustment parameters corresponding to the strategy into the airspace digital twin model, and predict the probability of conflict between the adjusted aircraft trajectory and other airspace objects.

[0231] Conflict resolution simulation is performed for each strategy in the set of feasible obstacle avoidance strategies to evaluate whether each strategy can effectively avoid conflicts during actual execution. The trajectory adjustment parameters corresponding to the strategy are input into the airspace digital twin model, which can then simulate the aircraft's trajectory adjustment process based on these parameters.

[0232] The airspace digital twin model predicts the likelihood of conflicts between the adjusted aircraft's trajectory and other airspace objects. By simulating the trajectory adjustment process, it calculates the distances and relative positions of the adjusted aircraft's trajectory with other aircraft, obstacles, and other airspace objects. Based on preset conflict determination criteria (e.g., a distance less than a certain threshold is considered a conflict), the model predicts the probability of conflict. By calculating the distances of the trajectory to other aircraft and buildings, it predicts whether a conflict will occur. Through conflict resolution simulation, strategies that effectively avoid conflicts can be further selected, improving the safety of obstacle avoidance strategies.

[0233] Step S540: Calculate the execution cost of the obstacle avoidance strategy when the conflict probability meets the preset safety conditions. The execution cost includes the change in range, the time adjustment, and the change in energy consumption, and generate a strategy cost evaluation matrix.

[0234] The execution cost of an obstacle avoidance strategy that meets a preset safety condition is calculated. The preset safety condition is a pre-defined threshold for the likelihood of conflict; only obstacle avoidance strategies with a conflict likelihood below this threshold are considered. The execution cost includes changes in range, time adjustment, and energy consumption.

[0235] The change in range refers to the difference between the aircraft's planned flight range after implementing the obstacle avoidance strategy and the original planned flight range. The time adjustment represents the difference between the time it takes for the aircraft to reach its destination after implementing the obstacle avoidance strategy and the original planned time. The change in energy consumption is the difference between the aircraft's planned energy consumption during the implementation of the obstacle avoidance strategy and the original planned energy consumption.

[0236] The process of an aircraft executing obstacle avoidance strategies can be simulated using an airspace digital twin model, and the costs of these actions can be calculated. For example, during the simulation, the aircraft's flight path length, flight time, and energy consumption are recorded and compared with the original plan to obtain changes in range, time adjustment, and energy consumption.

[0237] The strategy cost evaluation matrix is ​​generated by organizing and arranging the execution cost of each obstacle avoidance strategy that meets the preset safety conditions into a matrix. The rows of the matrix correspond to different obstacle avoidance strategies, and the columns correspond to different execution cost indicators (range change, time adjustment, energy consumption change).

[0238] Step S550: Based on the strategy cost evaluation matrix, a multi-objective optimization algorithm is used to select the best obstacle avoidance strategy, generating a distributed cooperative obstacle avoidance strategy that takes into account both safety and mission execution efficiency. The distributed cooperative obstacle avoidance strategy includes the trajectory adjustment sequence of each aircraft and the execution timestamp.

[0239] Obstacle avoidance strategies are selected using multi-objective optimization algorithms based on the strategy cost evaluation matrix. Multi-objective optimization algorithms are used to weigh and optimize multiple objectives (such as safety, range, time, energy consumption, etc.). Common multi-objective optimization algorithms include genetic algorithms and particle swarm optimization algorithms.

[0240] Taking genetic algorithms as an example, the algorithm iteratively optimizes strategy combinations by simulating the biological evolution process. First, an initial set of strategy combinations is randomly generated as a population. Then, the fitness value of each strategy combination is calculated based on the strategy cost evaluation matrix. The fitness value comprehensively considers objectives such as safety, range variation, time adjustment, and energy consumption variation. Next, new strategy combinations are generated through operations such as selection, crossover, and mutation. This process is repeated until the optimal strategy combination is found.

[0241] A distributed cooperative obstacle avoidance strategy is generated that balances safety and mission efficiency. This strategy considers the coordinated operations of multiple aircraft to ensure that no new conflicts are triggered during obstacle avoidance, while minimizing range changes, time adjustments, and energy consumption. The distributed cooperative obstacle avoidance strategy includes the trajectory adjustment sequence of each aircraft and execution timestamps. The trajectory adjustment sequence describes the position and attitude adjustments of the aircraft at different points in time, and the execution timestamps specify the execution time of each adjustment operation.

[0242] Step S560: Convert the trajectory adjustment sequence into the control command format of the low-altitude aircraft, including attitude adjustment parameters, power system parameters and navigation mode switching commands, and arrange them in the order of execution timestamps to generate a control command sequence.

[0243] The trajectory adjustment sequence is converted into a control command format for low-altitude aircraft. The trajectory adjustment sequence describes the aircraft's position and attitude adjustments at different points in time, and this adjustment information needs to be converted into control commands that the aircraft can recognize and execute. Attitude adjustment parameters include the aircraft's pitch angle, roll angle, yaw angle, etc., power system parameters involve engine thrust, speed, etc., and navigation mode switching commands are used to switch the aircraft's navigation mode (such as autopilot, manual control, etc.).

[0244] Based on the interface protocol of the aircraft's flight control system, information from the trajectory adjustment sequence can be mapped to corresponding control command parameters. For example, position and attitude information in the trajectory adjustment sequence can be converted into specific values ​​for attitude adjustment parameters and power system parameters.

[0245] The control command sequence is generated by arranging the converted control commands according to their execution timestamps. This ensures that the aircraft executes the control commands in the correct time sequence. Finally, the control command sequence is transmitted to the aircraft wirelessly, and the aircraft performs flight operations according to the command sequence to achieve obstacle avoidance and cooperative flight.

[0246] As one implementation method, step S560 can be implemented as the following steps S561~S568: Step S561: Obtain the real-time status data of each low-altitude aircraft. The real-time status data includes the remaining power percentage, equipment health indicators and communication link quality parameters, and is updated in real time through the aircraft status monitoring interface.

[0247] The system acquires real-time status data for each low-altitude aircraft. The remaining battery percentage indicates the remaining battery power of the aircraft. The equipment health index describes the health status of various components of the aircraft (such as engines and sensors). The communication link quality parameters reflect the reliability of communication between the aircraft and the ground control station or other aircraft.

[0248] These data can be updated in real time via the aircraft condition monitoring interface. The aircraft condition monitoring interface is a predefined interface used to acquire real-time status information of the aircraft. For example, sensors and communication modules can be used to collect this data and transmit it to the ground control station via wireless communication.

[0249] Step S562: Parse the control command protocol specification of the low-altitude aircraft to determine the structure definition, field data types and transmission priority rules of the command frame. The structure definition includes the command header, parameter segment and check segment.

[0250] This section analyzes the control command protocol specification for low-altitude aircraft. The control command protocol specification is a set of rules defined by aircraft manufacturers or flight control system developers to standardize the format and transmission of control commands. The structure definition of a command frame describes the components of a control command, typically including a command header, a parameter section, and a checksum section. The command header identifies the type and version of the command, the parameter section contains specific control parameters (such as attitude adjustment parameters, propulsion system parameters, etc.), and the checksum section verifies the completeness and correctness of the command.

[0251] Field data types define the data format of each field, such as integers, floating-point numbers, and Boolean values. Transmission priority rules determine the transmission priority of different types of commands; for example, emergency obstacle avoidance commands have a higher transmission priority than ordinary status reporting commands. By parsing the control command protocol specification, the format and requirements of control commands can be accurately understood, providing a foundation for subsequent command generation and transmission. For example, detailed information about the control command protocol specification can be obtained and parsed by reading protocol documents or communicating with the aircraft manufacturer. The length, data type, and meaning of each part of the command frame, as well as the transmission priority order of different commands, must be clearly defined.

[0252] Step S563: Adjust the attitude adjustment parameters in the trajectory adjustment sequence based on the equipment health index in the real-time status data. When the equipment health index is lower than the preset threshold, reduce the change amplitude of the attitude adjustment parameters to avoid exceeding the equipment's tolerance range.

[0253] Equipment health indicators reflect the working status and performance of various components of an aircraft. Attitude adjustment parameters include the magnitude and rate of adjustment of the aircraft's pitch, roll, and yaw angles. When equipment health indicators fall below preset thresholds, it means that the equipment may have a certain degree of malfunction or performance degradation. If a large-scale attitude adjustment is made at this time, it may cause further damage to the equipment or even lead to a flight accident.

[0254] Based on real-time acquired equipment health indicators, attitude adjustment parameters in the flight path adjustment sequence are dynamically adjusted. For example, if the equipment health indicator shows a minor malfunction in a certain servo of the aircraft, with a preset threshold of 80%, and the current equipment health indicator is 70%, which is below the threshold, then in the flight path adjustment sequence, for attitude adjustment parameters involving the control of that servo, such as the pitch angle adjustment range, which was originally planned to be adjusted by 10°, it is now reduced to 5° to reduce the workload of the equipment and avoid exceeding the equipment's tolerance range.

[0255] Step S564: Dynamically adjust the power system parameters based on the remaining battery percentage, optimize the thrust adjustment parameters while ensuring obstacle avoidance performance, and make the energy consumption change match the remaining battery percentage.

[0256] The remaining battery percentage directly affects an aircraft's endurance and the duration of its mission. Power system parameters mainly include adjustment parameters such as engine thrust magnitude and thrust direction. When performing obstacle avoidance maneuvers, a certain amount of energy is required to change the aircraft's flight state, but the remaining battery level must also be considered to avoid excessive power consumption that could prevent the aircraft from completing its mission or returning safely.

[0257] The power system parameters are dynamically adjusted based on the real-time remaining battery percentage. For example, when the remaining battery percentage is low, such as below 30%, the engine thrust is appropriately reduced while ensuring effective obstacle avoidance, thus minimizing the fluctuation range of the thrust adjustment parameters. Where thrust would normally be increased to 80% for rapid obstacle avoidance, it can be adjusted to 60% to reduce energy consumption. Conversely, when the remaining battery is sufficient, such as above 70%, the thrust adjustment parameters can be appropriately increased to achieve faster and more flexible obstacle avoidance. In this way, the change in energy consumption is matched to the remaining battery percentage, improving the aircraft's energy efficiency and ensuring that the aircraft can safely and efficiently complete obstacle avoidance tasks.

[0258] Step S565: Fill the adjusted attitude adjustment parameters, power system parameters and navigation mode switching commands into the corresponding fields according to the command frame structure definition to generate the initial command frame set.

[0259] After adjusting the attitude adjustment parameters and power system parameters, and determining the navigation mode switching command, this information needs to be filled in according to the previously parsed command frame structure definition. The command frame structure definition clearly defines the specific location and format requirements of the command header, parameter segment, and check segment.

[0260] The adjusted attitude adjustment parameters and power system parameters are converted into their corresponding data types and arranged according to the parameter segment format requirements of the command frame. For example, if the parameter segment requires the attitude adjustment parameter values ​​to be stored in a specific byte order, the adjusted parameters need to be converted to the appropriate byte order. Navigation mode switching commands are also filled into the command frame according to the prescribed format and position. Simultaneously, the command type, version, and other information are filled into the command header. Finally, the value of the check segment is calculated according to the command frame's verification rules and filled into the corresponding position. In this way, all control information is accurately filled into the various fields of the command frame, generating the initial command frame set.

[0261] Step S566: Perform command conflict detection on the initial command frame set, analyze the parameter correlation between different command frames, and identify the core control commands and auxiliary status reporting commands involving aircraft attitude adjustment and power system adjustment; when parameter contradictions exist, resolve the conflict based on the command function correlation rules, retain the parameter values ​​of the core control commands that have a direct impact on the execution of real-time obstacle avoidance actions, and temporarily suspend the execution of auxiliary status reporting commands.

[0262] The initial set of command frames may contain multiple command frames of different types, and there may be parameter conflicts between these command frames. For example, one command frame may require the aircraft to make a large attitude adjustment, while another command frame may simultaneously require it to maintain the current attitude, which results in a parameter conflict.

[0263] When performing command conflict detection, the parameter correlation between different command frames is analyzed first. Core control commands involving aircraft attitude adjustment and power system regulation directly affect the aircraft's flight status and obstacle avoidance performance, and therefore require special attention. Auxiliary status reporting commands are mainly used to transmit aircraft status information, such as battery level and equipment temperature, to ground control stations or other equipment, and have no direct impact on the execution of real-time obstacle avoidance actions.

[0264] When parameter conflicts are detected, conflict resolution is performed based on command function correlation rules. Priority is given to preserving parameter values ​​for core control commands that directly impact real-time obstacle avoidance actions. For example, if there is a parameter conflict between an attitude adjustment command and a status reporting command, to ensure the aircraft can avoid obstacles promptly and effectively, the parameter value of the attitude adjustment command is preserved, and the execution of the status reporting command is temporarily suspended. The suspended command frames can be temporarily stored and processed after the obstacle avoidance task is completed.

[0265] Step S567: Select the corresponding instruction encoding method according to the communication link quality parameters. When the link bit error rate exceeds the preset bit error threshold or the signal strength attenuation value is lower than the preset strength threshold, use an anti-interference encoding method with parity bit redundancy to enhance the stability of instruction transmission; otherwise, use a simplified encoding method that removes unnecessary parity fields to reduce the amount of transmitted data.

[0266] Communication link quality parameters directly affect the reliability and efficiency of command transmission. The link bit error rate reflects the probability of data errors occurring during communication, while the signal strength attenuation value indicates the degree to which the signal strength weakens during transmission.

[0267] The preset bit error rate threshold and preset signal strength threshold are pre-defined judgment criteria. When the link bit error rate exceeds the preset bit error rate threshold or the signal strength attenuation value is lower than the preset strength threshold, it indicates that the communication link quality is poor and there is a high risk of interference and errors. In this case, an anti-interference coding method that includes parity bit redundancy, such as Cyclic Redundancy Check (CRC) coding, is adopted.

[0268] Conversely, when the link error rate is lower than the preset error threshold and the signal strength attenuation is higher than the preset strength threshold, it indicates that the communication link quality is good and the risk of interference and errors is low. In this case, a simplified encoding method that removes unnecessary check fields is adopted.

[0269] Step S568: Add verification information from the real-time status data to the encoded instruction frame, generate a control instruction sequence containing a status feedback field, and strictly align the timestamp of the control instruction sequence with the execution timestamp of the track adjustment sequence to ensure that the instructions are executed in sequence.

[0270] After the command encoding is completed, in order to further improve the reliability and accuracy of command transmission, it is necessary to add verification information from the real-time status data to the encoded command frame. The verification information in the real-time status data can be a verification and confirmation of the aircraft's current status, such as check codes for equipment health indicators, check values ​​for remaining battery percentage, etc.

[0271] These verification details are added to the encoded command frame to form a control command sequence that includes a status feedback field. The status feedback field allows the receiver to verify the completeness and accuracy of the command based on the verification information, and also to understand the current status of the aircraft.

[0272] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0273] Please see Figure 3 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Figure 3 As shown, the computer system 10 described above may include: a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the computer system 10 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The user interface 1003 may include a display screen and a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the processor 1001. Figure 3 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0274] exist Figure 3 In the computer system 10 shown, the network interface 1004 provides network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the methods provided in the above embodiments.

Claims

1. A method for low-altitude cooperative perception data processing based on spatial computation, characterized in that, The method comprises: receiving original perception point cloud data transmitted by a plurality of low-altitude aircrafts; performing spatiotemporal registration and semantic enhancement processing on the original perception point cloud data to generate a spatial semantic feature set containing spatial object attributes and environment-related features; inputting the spatial semantic feature set into a pre-constructed spatial digital twin model to generate spatial state implicit representation through neural radiance field modeling; calling a multi-agent collaborative reasoning model to perform conflict risk prediction on the spatial state implicit representation, and outputting a conflict risk prediction graph containing relative position relationships between aircrafts; generating a distributed collaborative obstacle avoidance strategy based on the conflict risk prediction graph and a preset obstacle avoidance strategy library, and encoding the distributed collaborative obstacle avoidance strategy into a control instruction sequence and issuing the control instruction sequence to corresponding low-altitude aircrafts.

2. The method of claim 1, wherein, The spatiotemporal registration and semantic enhancement processing on the original perception point cloud data to generate a spatial semantic feature set containing spatial object attributes and environment-related features comprises: obtaining reference coordinate system parameters and time synchronization protocol data of each low-altitude aircraft, performing coordinate conversion processing on the original perception point cloud data based on the reference coordinate system parameters to obtain standardized point cloud data in a unified coordinate system; extracting timestamp information in the standardized point cloud data, performing timestamp alignment processing on point cloud data collected by different aircrafts according to the time synchronization protocol data to generate time sequence point cloud sequences with time sequence correlation; performing dynamic obstacle identification processing on the time sequence point cloud sequences, extracting point cloud subsets of moving objects through neighborhood point density analysis, calculating centroid change trajectories and velocity vector parameters of the point cloud subsets; combining the centroid change trajectories and the velocity vector parameters, adding semantic labels to static terrain point clouds and dynamic obstacle point clouds in the time sequence point cloud sequences respectively to generate initial semantic point clouds containing semantic classification information; performing environment feature correlation processing on the initial semantic point clouds, extracting terrain slope features, electromagnetic interference intensity features and meteorological disturbance features within a point cloud neighborhood range to generate environment-related feature vectors; and fusing the semantic classification information and the environment-related feature vectors to construct a spatial semantic feature set containing spatial object attribute descriptions and multi-dimensional environment influence factors.

3. The method of claim 2, wherein, The coordinate conversion processing of the original perception point cloud data based on the reference coordinate system parameters to obtain standardized point cloud data in a unified coordinate system includes: analyzing the coordinate origin offset, coordinate axis rotation matrix and scale factor in the reference coordinate system parameters to construct a coordinate conversion matrix; performing matrix multiplication operation on each three-dimensional coordinate point in the original perception point cloud data and the coordinate conversion matrix to obtain converted intermediate coordinate points; extracting outliers in the intermediate coordinate points, identifying outlier coordinate points deviating from the overall distribution trend through neighborhood point distribution characteristic analysis, and removing the outlier coordinate points; performing spatial grid division on the intermediate coordinate points after removing the outlier coordinate points, storing the coordinate points based on a hierarchical index structure, and generating grid coordinate data with hierarchical index; calculating the theoretical deviation value of the grid coordinate data and the unified coordinate system, performing trend analysis on the deviation value through a deviation fitting algorithm, and generating a deviation compensation coefficient; fusing the grid coordinate data and the deviation compensation coefficient to obtain standardized point cloud data in a unified coordinate system, and the coordinate precision of the standardized point cloud data matches the positioning system precision level of the low-altitude flying vehicle.

4. The method of claim 3, wherein, The dynamic obstacle identification processing of the time sequence point cloud sequence includes extracting a point cloud subset of a moving object through neighborhood point density analysis, calculating the centroid change trajectory and velocity vector parameters of the point cloud subset, including: obtaining a preset three-dimensional space neighborhood search range, performing neighborhood point search on each point cloud unit in the time sequence point cloud sequence, and counting the number of neighborhood point clouds as a density feature value; performing difference operation on the point cloud density feature values of consecutive time stamps to generate a density change rate sequence, and when the density change rate sequence meets a preset dynamic judgment condition, marking the corresponding area as a potential dynamic area; extracting a point cloud set of the potential dynamic area, performing clustering segmentation on the point cloud set through a region clustering algorithm to generate multiple independent point cloud clusters, and each point cloud cluster corresponds to a potential moving object; calculating the three-dimensional centroid coordinates of each point cloud cluster at consecutive time stamps, arranging the three-dimensional centroid coordinates in chronological order to generate a centroid change trajectory; based on the time interval of the centroid change trajectory, calculating the centroid coordinate difference value of adjacent time stamps, and taking the ratio of the difference value and the time interval as a velocity vector parameter, which contains direction information and size information; performing time sequence smoothing processing on the velocity vector parameter to eliminate the influence of instantaneous fluctuations on velocity calculation, and generating a smoothed velocity vector parameter sequence.

5. The method of claim 1, wherein, The inputting the spatial semantic feature set into the pre-constructed space domain digital twin model generates a space domain state implicit representation through neural radiance field modeling, including: analyzing network structure parameters of the space domain digital twin model, determining voxel division mode and activation function type of the neural radiance field, performing three-dimensional grid division processing on the spatial semantic feature set based on the voxel division mode to generate a voxelized feature grid; extracting space domain object attribute information in the voxelized feature grid, and mapping the space domain object attribute information into initial values of density field parameters and color field parameters of the neural radiance field; calling a position encoding module of the space domain digital twin model to perform spatial feature expansion processing on three-dimensional coordinates of the voxelized feature grid to generate an encoded coordinate vector with full-size space feature expression capability; inputting the encoded coordinate vector, the initial values of the density field parameters and the color field parameters into a multi-layer perceptron network of the neural radiance field to perform radiance field parameter optimization training, and generating an implicit function representing space domain position radiance characteristics; performing light ray projection sampling processing on the space domain space based on the implicit function, calculating radiance field rendering results under different observation viewing angles, and adjusting implicit function parameters through error feedback with real perception data; when the error between the radiance field rendering results and the real perception data meets a preset convergence condition, saving the current implicit function parameters as the space domain state implicit representation, and the space domain state implicit representation parameterizes the physical field attributes and semantic association information in the space domain three-dimensional space through the implicit function.

6. The method of claim 5, wherein, The calling the position encoding module of the space domain digital twin model to perform spatial feature expansion processing on three-dimensional coordinates of the voxelized feature grid to generate an encoded coordinate vector with full-size space feature expression capability includes: analyzing encoding dimension parameters and frequency band distribution range of the position encoding module, generating a periodic function combination containing different frequency band components based on the frequency band distribution range; performing operation on each coordinate axis component in the three-dimensional coordinates of the voxelized feature grid and the periodic function combination respectively to generate a frequency band feature vector corresponding to each coordinate axis component; performing channel splicing processing on the frequency band feature vector, stacking the frequency band feature vectors corresponding to the three coordinate axis components according to channel dimensions to generate an initial encoding vector; extracting continuous space features and abrupt space features in the initial encoding vector, calculating influence weights of the continuous space features and the abrupt space features through a feature contribution analysis algorithm to generate feature weight coefficients; performing element-by-element multiplication on the initial encoding vector and the feature weight coefficients to generate a weighted encoding vector; performing feature scale unification processing on the weighted encoding vector to constrain vector feature scale within a preset range to obtain an encoded coordinate vector with full-size space feature expression capability, and the dimension of the encoded coordinate vector matches the input layer dimension of the neural radiance field.

7. The method of claim 6, wherein, The ray casting sampling process based on the implicit function is performed on the airspace to calculate the radiation field rendering result under different observation angles, including: obtaining the observation angle parameters of a preset virtual camera to generate an initial ray casting direction vector; calling a local feature query interface of the implicit representation of the airspace state to extract the airspace local features on the initial ray casting path, the airspace local features including the spatial distribution information of the dynamic obstacle density and the environmental interference intensity; dynamically adjusting the ray casting direction vector according to the airspace local features, analyzing the feature difference degree of adjacent point cloud units, and when the feature difference degree meets a preset determination condition, marking the corresponding region as a to-be-enhanced sampling region, increasing the ray casting density of the to-be-enhanced sampling region, reducing the ray casting density of a static uniform region, and generating an optimized ray casting direction vector set; starting from each pixel point of the virtual camera imaging plane, performing airspace sampling along the optimized ray casting direction vector set to generate a sampling point sequence, the distribution density of the sampling point sequence being positively correlated with the dynamic change degree of the airspace local features; inputting the three-dimensional coordinates of the sampling point sequence into the implicit function to output the density value and the color value of each sampling point, and modulating the density value and the color value in combination with the environmental interference intensity in the airspace local features; calculating the opacity parameter of the sampling point based on the modulated density value, the opacity parameter increasing with the increase of the environmental interference intensity; performing time sequence correlation integral operation on the opacity parameter and the modulated color value, the weight of the integral operation increasing with the decrease of the distance between the sampling point and the dynamic obstacle to generate the dynamic rendering color value of the pixel point; arranging the dynamic rendering color values of all pixel points according to the imaging plane positions to generate the radiation field rendering image under the observation angle, the radiation field rendering image including the dynamic visual feature distribution of the airspace and the visual representation of the environmental interference influence.

8. The method of claim 1, wherein, The calling multi-agent collaborative reasoning model performs conflict risk prediction on the airspace state implicit representation, and outputs a conflict risk prediction graph containing the relative position relationship between aircrafts, including: analyzing the input interface protocol of the multi-agent collaborative reasoning model, converting the airspace state implicit representation into a model-compatible feature tensor format, which contains the implicit function parameters of the airspace three-dimensional grid and the motion state vector of the dynamic obstacle; extracting the current route planning data of each low-altitude aircraft, calculating the predicted track point sequence in the future preset period based on the route planning data, and converting the predicted track point sequence into a three-dimensional coordinate path in the airspace state implicit representation; inputting the three-dimensional coordinate path and the airspace state implicit representation into the relationship modeling module of the multi-agent collaborative reasoning model, constructing an aircraft-aircraft and aircraft-obstacle spatial relationship graph network, the nodes of the spatial relationship graph network represent airspace objects, and the edges represent the relative position relationship and distance parameters between objects; calling the graph attention mechanism layer of the spatial relationship graph network to calculate the influence weight of different airspace object nodes on the conflict risk of the current aircraft, and generating an attention weight matrix; performing message passing update on the spatial relationship graph network based on the attention weight matrix, aggregating the state features of the neighbor nodes, and generating a node feature vector containing global airspace relationship information; calling the risk prediction layer of the multi-agent collaborative reasoning model to perform conflict risk assessment on the node feature vector, and generating a conflict risk prediction graph containing risk distribution features, which is used to represent the conflict possibility distribution of different positions in the airspace.

9. The method of claim 8, wherein, The construction of the aircraft-aircraft and aircraft-obstacle spatial relationship graph network includes: taking each low-altitude aircraft and dynamic obstacle as an airspace object node, extracting the attribute features of each node, including the current position coordinates, motion direction vector and size parameters; calculating the spatial distance parameters between any two nodes, determining the connection relationship between the nodes based on the spatial distance parameters, and establishing a connection edge between the corresponding nodes when the spatial distance parameters meet the preset connection conditions; extracting the relative position relationship between the nodes at both ends of the connection edge, including the azimuth angle information and the height difference information, and encoding the azimuth angle information and the height difference information into edge attribute parameters; combining the distance parameters and edge attribute parameters to construct a spatial relationship graph network containing a node attribute matrix and an edge attribute matrix, the row dimension of the node attribute matrix corresponds to the node number, and the column dimension corresponds to the attribute feature dimension; performing structure optimization on the spatial relationship graph network, removing redundant connection edges with a conflict risk prediction contribution degree lower than a preset threshold, retaining key connection relationships, and generating a simplified spatial relationship graph network; associating and binding the simplified spatial relationship graph network with the airspace state implicit representation, so that the graph network nodes can call the implicit representation parameters of the corresponding airspace position in real time.

10. A computer system, characterized by comprise: a processor; and a memory, wherein the memory has stored therein computer readable code which, when executed by the processor, causes the processor to carry out the method of any one of claims 1 to 9.

Images