A data processing method and system for low-altitude economical UAV based on large model

Abstract

The present invention belongs to the technical field of intelligent navigation of unmanned aerial vehicles (UAVs), and discloses a data processing method and system for low-altitude economic UAVs based on a large model, comprising: calculating the scene adaptation weights of multimodal data based on a dynamic attention mechanism based on a spatiotemporal alignment feature package, and fusing them to generate a multimodal joint feature matrix; constructing a three-dimensional topological model of the urban airspace based on the multimodal joint feature matrix, combining a spatiotemporal graph neural network to predict the trajectory of dynamic obstacles, and generating a hierarchical navigation instruction set; the UAV executes flight instructions according to the hierarchical navigation instruction set, and generates a flight status monitoring log by collecting data from the UAV in flight in real time; the present invention generates a multimodal joint feature matrix by driving weight distribution through environmental parameters, significantly improving the perception accuracy of complex scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle (UAV) intelligent navigation, and in particular relates to a data processing method and system for low-altitude economical UAV based on a large model. Background Art

[0002] The field of low-altitude economical drone navigation technology has developed rapidly in recent years, primarily relying on multi-sensor fusion and path planning algorithms to achieve autonomous flight. Conventional methods collect environmental data by integrating vision, lidar, and an inertial measurement unit (IMU), construct an airspace topology model through spatiotemporal registration, and generate a flight path based on static rules. For example, visual images identify obstacle semantic labels, lidar point clouds mark no-fly zone voxels, and real-time traffic control data are combined to generate a sequence of waypoints. In dynamic obstacle handling, traditional solutions use trajectory prediction models to estimate the motion path, and then use global planning algorithms to avoid collision risks. Static rule-based obstacle avoidance systems based on multi-source perception have achieved effective avoidance of static obstacles such as buildings and high-voltage towers, and have formed a standardized application framework for scenarios such as urban logistics and inspections.

[0003] However, existing methods have two limitations: First, the static weight mechanism of multimodal data fusion is difficult to adapt to changes in complex scenes. For example, under conditions of drastic lighting fluctuations or sparse point clouds, the fixed-weight vision-lidar fusion strategy is prone to missed detection of transparent obstacles, and is unable to dynamically adjust perception priority based on motion stability, increasing the risk of collision in low-visibility areas. Second, the separate processing of obstacle prediction and path planning weakens the real-time response of the system. Dynamic obstacle trajectory prediction and waypoint risk assessment are performed step by step, and an interactive feedback mechanism in the spatiotemporal graph network has not been established, resulting in a delayed response to sudden obstacles. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides a low-altitude economic UAV data processing method based on a large model to solve the problems of missed detection of transparent obstacles due to insufficient environmental adaptability and obstacle avoidance response delay caused by prediction-planning separation.

[0005] The technical solutions of the present invention are as follows:

[0006] In a first aspect, the present invention provides a method for processing data of a low-altitude economical UAV based on a large model, comprising:

[0007] S1. Collect multimodal data and preprocess them, perform time alignment and spatial calibration on the preprocessed multimodal data, and generate a spatiotemporal alignment feature package;

[0008] S2. Based on the spatiotemporal alignment feature package, the scene adaptation weights of multimodal data are calculated through the dynamic attention mechanism, and the multimodal joint feature matrix is ​​generated by fusion.

[0009] S3. Based on the multimodal joint feature matrix, a three-dimensional topological model of the urban airspace is constructed, and the dynamic obstacle trajectories are predicted by combining the spatiotemporal graph neural network to generate a hierarchical navigation instruction set.

[0010] S4. The UAV executes flight instructions according to the hierarchical navigation instruction set and generates a flight status monitoring log by collecting real-time data during the flight of the UAV;

[0011] S5. Based on the abnormal events in the flight status monitoring log, the root cause of the abnormal events is located in combination with the predefined expert rule base, and a parameter optimization strategy package containing dynamic attention parameter update rules and topological obstacle detection thresholds is generated.

[0012] Furthermore, in S1, the preprocessing includes data cleaning, format standardization, lidar point cloud denoising and camera dynamic illumination compensation.

[0013] Furthermore, the spatiotemporal alignment feature package is generated by temporally aligning and spatially calibrating the preprocessed multimodal data through a multi-sensor spatiotemporal registration algorithm.

[0014] Furthermore, in S2, the scene adaptation weights of the multimodal data are calculated based on the spatiotemporal alignment feature package through a dynamic attention mechanism, and are fused to generate a multimodal joint feature matrix, including:

[0015] Based on the spatiotemporal alignment feature package, a multi-sensor sliding window statistical method is used to synchronously obtain scene parameters such as light intensity, point cloud density, and motion stability, generating a multimodal feature set and scene perception parameter table.

[0016] Based on the multimodal feature set and scene perception parameter table, a dynamic attention mechanism is adopted to calculate the scene adaptation weight of multimodal data, generate a dynamic weight distribution table, and perform inter-modal interaction modeling and feature reorganization through the cross-attention mechanism to generate a multimodal joint feature matrix.

[0017] Furthermore, in S3, the three-dimensional topological model of the urban airspace is constructed based on the multimodal joint feature matrix, and the dynamic obstacle trajectory is predicted by combining the spatiotemporal graph neural network to generate a hierarchical navigation instruction set, including:

[0018] Based on the obstacle distribution and semantic information in the multimodal joint feature matrix, a three-dimensional topological model of the urban airspace is constructed, and the airspace connectivity map is generated by integrating real-time traffic control data.

[0019] Based on the three-dimensional topological model of the urban airspace and the airspace connectivity map, the spatiotemporal graph structure of dynamic obstacles is constructed through the spatiotemporal graph neural network. The obstacle trajectories are predicted by combining the multimodal joint feature matrix. Based on the obstacle trajectories, the spatiotemporal kernel density estimation is used to generate the dynamic obstacle trajectory heat map.

[0020] According to the dynamic interaction relationship between the spatiotemporal graph structure of dynamic obstacles and the dynamic obstacle trajectory heat map, global path planning is used to select a safe waypoint sequence, and risk assessment is performed through the edge attributes in the spatiotemporal graph structure. According to the risk assessment results, a hierarchical navigation instruction set is generated.

[0021] Furthermore, in S4, the UAV executes flight instructions according to the hierarchical navigation instruction set, and generates a flight status monitoring log by collecting data of the UAV in flight in real time, including:

[0022] After receiving the hierarchical navigation instruction set, the drone will fly, synchronously collecting environmental and status data during the flight, and using dynamic risk thresholds to identify sudden obstacles, and trigger the emergency obstacle avoidance protocol to dynamically adjust the hierarchical navigation instruction set;

[0023] The drone’s path deviation, obstacle avoidance events, power system overload alarm events, and the number of hierarchical navigation instruction set adjustments are recorded and marked as abnormal events to generate a flight status monitoring log.

[0024] Furthermore, in S5, the root cause of the abnormal event is located based on the abnormal event in the flight status monitoring log and combined with the predefined expert rule base, and a parameter optimization strategy package including dynamic attention parameter update rules and topological obstacle detection thresholds is generated, including:

[0025] Parse abnormal event records in flight status monitoring logs, classify them according to predefined expert rule bases, and generate abnormal event classification lists;

[0026] Construct a multi-level causal reasoning diagram based on the abnormal event classification list through forward reasoning chain to locate the root cause of the abnormal event;

[0027] Based on the root cause of the abnormal event, a parameter optimization strategy package including dynamic attention parameter update rules and topological obstacle detection thresholds is generated through JSON structured instruction encapsulation and simulation environment verification process.

[0028] In a second aspect, the present invention provides a low-altitude economic UAV data processing system based on a large model, comprising:

[0029] The data acquisition module is used to collect and preprocess multimodal data, perform time alignment and spatial calibration on the preprocessed multimodal data, and generate a spatiotemporal alignment feature package;

[0030] The scene fusion module is used to calculate the scene adaptation weights of multimodal data based on the spatiotemporal alignment feature package through a dynamic attention mechanism, and fuse them to generate a multimodal joint feature matrix;

[0031] The airspace collaborative navigation decision module is used to construct a three-dimensional topological model of the urban airspace based on a multimodal joint feature matrix, combine it with a spatiotemporal graph neural network to predict dynamic obstacle trajectories, and generate a hierarchical navigation instruction set;

[0032] The flight anomaly monitoring module is used for the UAV to execute flight instructions according to the hierarchical navigation instruction set, and to generate a flight status monitoring log by collecting real-time data during the UAV flight;

[0033] The autonomous optimization evolution module is used to locate the root cause of abnormal events based on the abnormal events in the flight status monitoring log and the predefined expert rule base, and generate a parameter optimization strategy package that includes dynamic attention parameter update rules and topological obstacle detection thresholds.

[0034] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the low-altitude economical UAV data processing method based on a large model as described in the first aspect of the present invention.

[0035] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the large model-based low-altitude economical UAV data processing method as described in the first aspect of the present invention.

[0036] The beneficial effects of the present invention are as follows: the present invention generates a multimodal joint feature matrix by driving weight distribution through environmental parameters, significantly improving the perception accuracy of complex scenes; through space-time graph risk assessment, hierarchical navigation instructions are generated according to the obstacle heat map and the spatial overlap rate of the flight segments, high-risk flight segments are split into fine instructions, and low-risk flight segments are merged into long-distance cruises, ensuring the timeliness of response to sudden obstacles, balancing safety and efficiency, and realizing the coordinated optimization of autonomous obstacle avoidance and route efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 This is a flow chart of the data processing method for low-altitude economic UAV based on large model;

[0039] Figure 2 This is a schematic diagram of the drone data processing architecture system;

[0040] Figure 3Flowchart of the dynamic attention mechanism;

[0041] Figure 4 Flowchart generated for a hierarchical navigation instruction set. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0043] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0045] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a low-altitude economic UAV data processing method based on a large model, comprising the following steps:

[0046] S1: Collect and preprocess multimodal data, perform time alignment and spatial calibration on the preprocessed multimodal data, and generate a spatiotemporal alignment feature package;

[0047] S1.1: Multimodal data includes visual images captured by the drone camera, 3D point clouds generated by the lidar, and attitude and motion parameters provided by the inertial measurement unit (IMU);

[0048] In this embodiment, the drone's camera continuously captures ambient optical information during flight, recording surface and obstacle features within its field of view in the form of RGB color images. The image resolution and frame rate are configured according to preset parameters. A laser radar (LiDAR) emits a pulsed laser beam at a set frequency, acquiring the three-dimensional spatial coordinates of the surrounding environment through time-of-flight measurement and multi-line scanning. This generates a point cloud dataset containing X, Y, and Z coordinate values ​​and reflection intensities, with a single-frame point cloud density reaching a specific level per square meter. The three-axis accelerometer embedded in the inertial measurement unit (IMU) measures the linear acceleration of the drone body in the X, Y, and Z axes at a high sampling rate. The three-axis gyroscope simultaneously acquires the angular velocity around the X, Y, and Z axes at corresponding sampling rates. The accelerometer range covers the multi-axis dynamic range, and the gyroscope range covers the multi-dimensional rotation range. Visual images are stored as JPEG format sequence files, the three-dimensional point cloud is packaged according to the E57 standard format, and the inertial measurement unit (IMU) data is recorded as a CSV format time series, with timestamp alignment accuracy maintained within millisecond deviations.

[0049] S1.2: Preprocessing includes data cleaning, format standardization, lidar point cloud denoising, and camera dynamic illumination compensation;

[0050] In this embodiment, data cleaning removes invalid data segments containing missing values ​​and abnormal values ​​outside the preset range through conditional filtering, such as performing sliding window mean correction on inertial measurement unit (IMU) data. Format standardization uniformly converts visual images into fixed resolution and color space, resamples the lidar point cloud into a uniform point distance distribution, and forces the inertial measurement unit (IMU) data to align with the time reference axis. LiDAR point cloud denoising uses a filtering method based on spatial distribution, such as using a statistical filtering algorithm to delete isolated points that deviate from the cluster center while retaining building edge feature points. Camera dynamic lighting compensation performs real-time brightness mapping adjustment, such as activating multi-region gamma correction in strong backlight scenes, dynamically enhancing the contrast of shadow areas based on the image histogram distribution, and suppressing overexposure of highlight areas. The preprocessed visual image maintains standard size and color balance, the three-dimensional point cloud eliminates noise caused by flight vibration, and the inertial measurement unit (IMU) data timestamps are aligned to a unified reference;

[0051] S1.3: Temporally align and spatially calibrate the pre-processed multimodal data using a multi-sensor spatiotemporal registration algorithm to generate a spatiotemporal alignment feature package.

[0052] The multi-sensor spatiotemporal registration algorithm first establishes a time reference axis for the preprocessed multimodal data. For example, it uses the lidar pulse trigger signal as the master clock source. Polynomial interpolation is then used to align the camera exposure time and the inertial measurement unit (IMU) sampling time to a unified time axis, with a time synchronization error within ±5 milliseconds. Spatial calibration performs a rigid transformation of the multi-sensor coordinate system. For example, based on calibration parameters, the lidar point cloud coordinate system and the camera pixel coordinate system are converted to the inertial measurement unit (IMU) body coordinate system using the rotation matrix R and translation vector T, achieving unified spatial coordinates in the X, Y, and Z axes. The spatiotemporal alignment feature package is encapsulated in a four-dimensional data structure. For example, each feature point contains a unified timestamp, the X, Y, and Z coordinates in the body coordinate system, as well as the associated RGB pixel values, laser reflection intensity, and angular velocity measurements. The data package is stored in a hierarchical HDF5 format.

[0053] S2: Based on the spatiotemporal alignment feature bag, the scene adaptation weights of multimodal data are calculated through the dynamic attention mechanism, and the multimodal joint feature matrix is ​​generated by fusion;

[0054] S2.1: Based on the spatiotemporal alignment feature package, a multi-sensor sliding window statistical method is used to synchronously obtain scene parameters such as light intensity, point cloud density, and motion stability, and generate a multimodal feature set and scene perception parameter table;

[0055] Based on spatiotemporally aligned feature packages, a multi-sensor sliding window statistical method captures time series data using a fixed-length window, for example, set to 500 milliseconds. Three calculations are performed simultaneously within the window: the Y-channel brightness mean is extracted from the RGB pixel values ​​in the spatiotemporally aligned feature package as a light intensity parameter; the number of 3D point clouds per unit volume within the window is counted as a point cloud density parameter; and the variance of the inertial measurement unit (IMU) linear acceleration along the X, Y, and Z axes is calculated as a motion stability parameter. The multimodal feature set encapsulates the processing results in a time series format. For example, each feature entry contains a normalized light intensity value, a point cloud density level, and an acceleration variance threshold. The scene perception parameter table records the timestamp, light intensity value, point cloud density value, and the X, Y, and Z acceleration variances in a structured table format. The row storage interval strictly matches the sliding window step size.

[0056] S2.2: Based on the multimodal feature set and scene perception parameter table, a dynamic attention mechanism is used to calculate the scene adaptation weights of the multimodal data, generate a dynamic weight distribution table, and perform inter-modal interaction modeling and feature reorganization through the cross-attention mechanism to generate a multimodal joint feature matrix.

[0057] Based on a multimodal feature set and a scene perception parameter table, the dynamic attention mechanism calculates the associated weights for visual imagery, LiDAR point cloud data, and inertial measurement unit (IMU) data based on real-time changes in illumination intensity, point cloud density, and motion stability parameters. For example, in low-light scenes, the dynamic attention mechanism assigns higher weights to visual imagery, medium weights to LiDAR point cloud data, and lower weights to IMU data, generating a dynamic weight assignment table that reflects inter-modal weight relationships. The cross-attention mechanism constructs the interactive relationships between the visual image feature space, LiDAR point cloud geometry, and IMU motion features. For example, multi-scale feature matching is used to calculate the spatial correlation between visual image texture and LiDAR point cloud curvature, establishing an inter-modal feature similarity map. The feature reconstruction stage integrates the weighting coefficients from the dynamic weight assignment table with the cross-attention feature map. For example, enhancing the high-frequency features of the visual image is then combined with the topological features of the LiDAR point cloud for multi-channel fusion. This results in a multimodal joint feature matrix containing multiple layers of abstract features. The time dimension of the matrix is ​​strictly synchronized with the spatiotemporally aligned feature package.

[0058] The dynamic weight allocation table is generated based on the illumination intensity parameters, point cloud density parameters, and motion stability parameters in the scene perception parameter table, and is implemented through the weight calculation rules of the dynamic attention mechanism. The specific process is as follows: First, the illumination intensity level, point cloud density classification, and motion stability status at the current moment in the scene perception parameter table are extracted as input features; then, according to the preset weight allocation strategy, for example, when the illumination intensity level is "low", the dynamic attention mechanism promotes the visual image weight coefficient to the primary level, adjusts the lidar point cloud weight coefficient to the secondary level, and sets the inertial measurement unit (IMU) data weight coefficient to the auxiliary level; finally, using the timestamp as the index, the visual image weight coefficient, lidar point cloud weight coefficient, inertial measurement unit (IMU) data weight coefficient, and the corresponding scene parameter status are written into a table in a structured format.

[0059] S3: Based on the multimodal joint feature matrix, a 3D topological model of the urban airspace is constructed. The dynamic obstacle trajectories are predicted by combining the spatiotemporal graph neural network and a hierarchical navigation instruction set is generated.

[0060] S3.1: Based on the obstacle distribution and semantic information in the multimodal joint feature matrix, a 3D topological model of the urban airspace is constructed and integrated with real-time traffic control data to generate an airspace connectivity map.

[0061] Based on the obstacle distribution and semantic information in the multimodal joint feature matrix, the team first extracts semantic labels for building outlines from visual images, obstacle 3D coordinates from LiDAR point clouds, and heading angle parameters from inertial measurement unit (IMU) data. For example, a convolutional neural network is used to identify the semantic category of "high-rise building" in the visual image and annotate its bounding box. The 3D topological model is constructed using a voxel grid spatial partitioning method. For example, the urban airspace is divided into cubic units with specific side lengths. Impassable voxels are labeled based on obstacle distribution data, and voxel attributes of building restricted flight zones are annotated based on semantic information. Real-time traffic control data is used to analyze altitude level assignments, temporary no-fly zone coordinates, and route priority parameters from air traffic control directives. For example, the vertex coordinates of flight corridors published in the "GTC1224" control directive are converted into topological model nodes. The airspace connectivity graph represents feasible paths through a graph structure. For example, the Dijkstra algorithm is used to calculate the connection weights of adjacent traversable voxel nodes in the 3D topological model. Dynamic constraints from the traffic control data are then overlaid to generate an adjacency matrix and reachability list containing node costs, edge connectivity status, and control constraint labels.

[0062] The obstacle distribution lidar point cloud annotates the 3D coordinates of building outline vertices. For example, a vertex coordinate sequence includes X-axis, Y-axis, and Z-axis coordinate values. Visual image recognition results provide semantic labels, such as "high-voltage tower" or "glass curtain wall building" labels, which are bound to corresponding point cloud coordinates. A safety extension defines a no-fly zone extending a specific distance beyond the obstacle surface. For example, all obstacles are extended by a specific size to create a safety buffer zone. The real-time motion envelope of dynamic obstacles annotates time-dependent no-fly zones, such as a cylindrical spatial range formed by the rotation of a construction crane. Obstacle distribution data is converted into impassable voxels in the 3D topological model. For example, a building coordinate set is labeled as a cubic no-fly zone, a tree safety extension is mapped as a cylindrical no-fly zone, and the dynamic obstacle envelope is associated with time-sliced ​​risk voxels. Semantic information defines passage rules. For example, the distance between two buildings must be greater than a specific distance to allow the creation of a flight corridor.

[0063] Among them, semantic information refers to the binding relationship between the obstacle attribute labels marked by visual image recognition and the point cloud coordinates, which includes two types of key data: one is the obstacle semantic label, and the other is the physical property extended label; this information is precisely associated with the lidar point cloud coordinates through spatial alignment, and together with the obstacle distribution in the three-dimensional topological model, it defines the passage rules.

[0064] Among them, real-time traffic control data refers to data used to dynamically update the distribution of obstacles, covering static annotations to ensure that the obstacle distribution reflects the current spatial constraints.

[0065] S3.2: Based on the 3D topological model of the urban airspace and the airspace connectivity map, a spatiotemporal graph structure of dynamic obstacles is constructed using a spatiotemporal graph neural network. The obstacle trajectories are predicted using a multimodal joint feature matrix. Based on the obstacle trajectories, a dynamic obstacle trajectory heat map is generated using spatiotemporal kernel density estimation.

[0066] Based on the 3D topological model of the urban airspace and its spatial connectivity map, a spatiotemporal graph neural network constructs a spatiotemporal graph structure of dynamic obstacles, using the voxel nodes of the 3D topological model as vertices and the paths of the spatial connectivity map as edges. For example, a construction crane is mapped into a spatiotemporal node containing position coordinates and motion vectors. Combining the obstacle motion features in the multimodal joint feature matrix, a gated graph convolutional layer is used to predict the future time series trajectory of the dynamic obstacle, for example, outputting a sequence of trajectory points of the construction crane at multiple consecutive moments. Spatiotemporal kernel density estimation uses the spatial kernel radius and time window parameters to estimate the probability density of the trajectory point sequence and generate a heat map of the dynamic obstacle trajectory. The heat map of the dynamic obstacle trajectory is represented as a 3D grid probability distribution. For example, labeling a specific voxel as a high-probability region indicates the likelihood of an obstacle existing at that location during the corresponding time period.

[0067] S3.3: Based on the dynamic interaction between the spatiotemporal graph structure of dynamic obstacles and the dynamic obstacle trajectory heat map, a safe waypoint sequence is selected using global path planning. Risk assessment is performed using edge attributes in the spatiotemporal graph structure. Based on the risk assessment results, a hierarchical navigation instruction set is generated.

[0068] Based on the dynamic interaction between the spatiotemporal graph structure of dynamic obstacles and the dynamic obstacle trajectory heat map, global path planning first matches high-probability areas in the heat map with the spatial locations of nodes in the spatiotemporal graph. For example, when the probability of a voxel in a construction boom trajectory heat map exceeds a certain threshold, the corresponding spatiotemporal graph node is marked as a temporary no-fly zone. A safe waypoint sequence is generated by searching within the 3D topological model using the A* algorithm. For example, from the starting point to the end point, voxel nodes that are away from high-risk areas in the heat map and meet the airspace connectivity map pass criteria are selected as waypoints. Risk assessment is performed using edge attributes in the spatiotemporal graph structure. For example, the spatial overlap ratio between the flight segment path and the dynamic obstacle trajectory heat map is calculated, and the risk level is output based on the edge connectivity status. A hierarchical navigation instruction set is generated based on the risk assessment results. For example, high-risk segments are split into low-speed fine navigation instructions, while low-risk segments are merged into long-distance cruise instructions (retaining only the coordinates of key turning points and altitude parameters). The instruction set stores the 3D coordinates of the waypoints, preset speed values, and risk control identifiers in a time series.

[0069] The risk assessment result contains the following fields: risk level, risk source, risk quantification, and control measure flag. For example, a flight segment's risk assessment result is recorded as "high" risk level, "construction crane trajectory heat map" risk source, "overlap area ratio exceeds a specific threshold" risk quantification, and "activate low-speed obstacle avoidance protocol" control measure. This result directly determines the hierarchical navigation instruction set generation strategy.

[0070] S4: The drone executes flight instructions according to the hierarchical navigation instruction set and generates a flight status monitoring log by collecting real-time data during the drone flight;

[0071] S4.1: After receiving the hierarchical navigation instruction set, the drone will conduct drone flight, synchronously collect environmental and status data during flight, use dynamic risk thresholds to identify sudden obstacles, and trigger the emergency obstacle avoidance protocol to dynamically adjust the hierarchical navigation instruction set;

[0072] After receiving the hierarchical navigation instruction set, the drone initiates flight, synchronously collecting camera visual images, lidar point clouds, and inertial measurement unit (IMU) attitude and motion parameters during flight. Dynamic risk threshold determination calculates the distance and relative speed of sudden obstacles in real time. For example, when the lidar detects that an unmodeled obstacle enters a specific distance threshold range and the relative speed exceeds the safety threshold, it is determined to be in a high-risk state. The emergency obstacle avoidance protocol is triggered to dynamically adjust the hierarchical navigation instruction set, such as deleting high-risk waypoints in the original instruction set and inserting emergency hover instructions (maintaining altitude, zero speed value) and obstacle avoidance maneuver instructions (adding three-dimensional coordinates of obstacle avoidance waypoints and decelerating to a specific speed value). The adjusted hierarchical navigation instruction set immediately overwrites the original instruction sequence, for example, replacing waypoint P5 with an obstacle avoidance path and updating the speed parameters. The instruction set version number and adjustment timestamp are also recorded.

[0073] The formula for calculating the distance to sudden obstacles is:

[0074] ;

[0075] in, Indicates the distance between the drone and the sudden obstacle; Indicates the easting coordinate of the drone's position in the geodetic coordinate system; The north coordinate of the drone's position in the geodetic coordinate system; The elevation coordinate representing the position of the UAV in the geodetic coordinate system; It represents the position of the UAV in the geodetic coordinate system; Indicates the X-axis coordinate of the sudden obstacle in the lidar coordinate system; Indicates the Y-axis coordinate of the sudden obstacle in the lidar coordinate system; The Z-axis coordinate of the sudden obstacle in the lidar coordinate system; It represents the position of the obstacle in the LiDAR coordinate system;

[0076] Among them, the environment and status data include the following: environmental data: camera visual images, lidar point cloud, ambient light intensity; status data: inertial measurement unit (IMU) attitude parameters, inertial measurement unit (IMU) motion parameters, positioning data, and power system parameters.

[0077] The dynamic risk threshold judgment is based on the camera visual images, lidar point cloud, and inertial measurement unit (IMU) motion parameters collected during the flight, and performs three-layer judgment in real time: the spatial threat judgment layer calculates the obstacle distance value through the lidar point cloud, and the dynamic threshold adjusts the safety distance according to the current speed value of the drone; the time margin judgment layer uses the inertial measurement unit (IMU) data to calculate the relative speed value, and the dynamic threshold sets the collision time margin threshold; the semantic risk judgment layer identifies the semantic category of obstacles through visual images, and the dynamic threshold reduces the trigger distance for high-risk semantic categories (for example, transparent obstacles use a stricter judgment distance); when any judgment layer exceeds the limit, the risk level is output and the corresponding emergency obstacle avoidance protocol is triggered.

[0078] Among them, the emergency obstacle avoidance protocol refers to a hierarchical response mechanism triggered by the risk level determined by the dynamic risk threshold, which includes three levels of core operations: the first level risk triggers the waypoint fine-tuning protocol; the second level risk triggers the emergency hovering protocol; the third level risk triggers the full path replanning protocol. After the protocol is executed, the hierarchical navigation instruction set version number is immediately updated and the timestamp is recorded.

[0079] S4.2: Record the drone's path deviations, obstacle avoidance events, power system overload alarm events, and the number of hierarchical navigation instruction set adjustments, mark them as abnormal events, and generate a flight status monitoring log;

[0080] The flight status monitoring log recording process is as follows: The Euclidean distance deviation between the drone's actual flight trajectory and the path preset by the hierarchical navigation instruction set is calculated in real time and marked as a path deviation event. When the emergency obstacle avoidance protocol is activated, the obstacle avoidance event type and trigger timestamp are recorded. The power system overload alarm event captures motor speed or temperature exceeding limit conditions. The hierarchical navigation instruction set adjustment counts version number changes. All of these events are marked as abnormal events, and a flight status monitoring log is generated in a structured format. Each record includes a timestamp, event type, quantitative parameters, and protocol type. The log is stored as a CSV-formatted time series file.

[0081] Euclidean distance deviation calculation formula:

[0082] ;

[0083] in, represents the Euclidean distance deviation; Indicates time The drone's north coordinate; Indicates time The drone's north coordinate; Indicates time The elevation coordinates of the drone; Represents the time dimension, specifically the real-time timestamp; Indicates flight segment The easting coordinate of the projection point; Indicates flight segment The north coordinate of the projected point; Indicates flight segment The elevation coordinates of the projected point; Indicates the flight segment index; Indicates the total number of segments in the hierarchical navigation instruction set.

[0084] S5: Based on the abnormal events in the flight status monitoring log, the root cause of the abnormal events is located in combination with the predefined expert rule base, and a parameter optimization strategy package containing dynamic attention parameter update rules and topological obstacle detection thresholds is generated.

[0085] S5.1: Parse abnormal event records in the flight status monitoring log, classify them according to the predefined expert rule base, and generate an abnormal event classification list;

[0086] When parsing flight status monitoring logs, the timestamp, event type, quantified parameter, and protocol type fields of abnormal event records are extracted line by line. A predefined expert rule base matches event type with parameter thresholds to form classification rules: Path deviation events are categorized as "environmental perception anomaly" or "control response anomaly" based on the direction and magnitude of the deviation; obstacle avoidance events are categorized as "active obstacle avoidance - waypoint fine-tuning" or "emergency obstacle avoidance - full path replanning" based on protocol type; power system overload alarm events are categorized based on the type of exceeded parameters; and the number of hierarchical navigation instruction set adjustments is categorized based on frequency per unit time. Finally, a categorized list of abnormal events is generated, with each entry containing the original log timestamp, event classification code, and classification rule number.

[0087] Among them, the predefined expert rule library contains four core rules: path deviation rules classify anomalies according to the lateral or vertical continuous deviation value exceeding a specific distance threshold; obstacle avoidance event rules are classified according to protocol type and trigger frequency; power alarm rules are classified according to the type of over-limit parameters; instruction set adjustment rules are classified according to the number of adjustments per unit time or the mutation ratio of the waypoint sequence length; each rule is structured and stored with rule number, matching field (event type + parameter condition), output classification code, and threshold parameter. For example: when a power alarm event is parsed and the temperature value reaches a specific degree Celsius, the matching rule outputs the "cooling system abnormality" classification entry.

[0088] S5.2: Construct a multi-level causal reasoning diagram based on the abnormal event classification list through forward reasoning chain to locate the root cause of the abnormal event;

[0089] When processing a list of abnormal event classifications through a forward chain of reasoning, the system first extracts the abnormality classification codes and timestamp sequences from the list. Then, based on a predefined causal rule base, it conducts reasoning layer by layer: the first layer matches direct causal relationships, the second layer correlates the spatiotemporal correlations of multiple events, and the third layer integrates environmental data for verification. The final output is a multi-level causal reasoning graph, with the root node locating the root cause and each reasoning path annotated with a confidence score. For example, when a cooling system anomaly and a battery aging failure occur within a specific minute, the graph identifies the root cause as "power circuit voltage fluctuation" and generates a maintenance recommendation.

[0090] The root cause refers to the underlying fault source or design flaw located through a multi-level causal reasoning diagram. It must simultaneously meet three conditions: first, it can explain all associated anomalies in the abnormal event classification list; second, it must be located at the end of the causal chain and cannot be further decomposed; and third, it must be interventionable. The root cause description includes the specific component name, failure mode, environmental trigger, and comes with a confidence score and verification suggestions.

[0091] S5.3: Based on the root cause of the abnormal event, a parameter optimization strategy package containing dynamic attention parameter update rules and topological obstacle detection thresholds is generated through JSON structured instruction encapsulation and simulation environment verification process.

[0092] Based on the root cause of the anomaly, the parameter optimization strategy package generation process follows a three-step process: first, dynamic attention parameter update rules are encapsulated using JSON structured instructions; second, a strategy for adjusting the topological obstacle detection threshold is defined; and finally, a simulation environment verification process is performed, where a historical anomaly scenario is loaded, the optimized parameters are injected, and a hierarchical navigation instruction set is run. Verification metrics include increasing the obstacle avoidance success rate to a specific percentage and reducing the path deviation to a specific distance threshold. Successful strategy packages are stored as JSON key-value pairs; failures return the modified rule entry.

[0093] This embodiment also provides a low-altitude economic UAV data processing system based on a large model, including:

[0094] The data acquisition module is used to collect and preprocess multimodal data, perform time alignment and spatial calibration on the preprocessed multimodal data, and generate a spatiotemporal alignment feature package;

[0095] The scene fusion module is used to calculate the scene adaptation weights of multimodal data based on the spatiotemporal alignment feature package through a dynamic attention mechanism, and fuse them to generate a multimodal joint feature matrix;

[0096] The airspace collaborative navigation decision module is used to construct a three-dimensional topological model of the urban airspace based on a multimodal joint feature matrix, combine it with a spatiotemporal graph neural network to predict dynamic obstacle trajectories, and generate a hierarchical navigation instruction set;

[0097] The flight anomaly monitoring module is used for the UAV to execute flight instructions according to the hierarchical navigation instruction set, and to generate a flight status monitoring log by collecting real-time data during the UAV flight;

[0098] The autonomous optimization evolution module is used to locate the root cause of abnormal events based on the abnormal events in the flight status monitoring log and the predefined expert rule base, and generate a parameter optimization strategy package that includes dynamic attention parameter update rules and topological obstacle detection thresholds.

[0099] This embodiment also provides a computer device, which is suitable for the low-altitude economic drone data processing method based on a large model, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the low-altitude economic drone data processing method based on a large model proposed in the above embodiment.

[0100] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0101] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for processing low-altitude economical unmanned aerial vehicle data based on a large model as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0102] In summary, the present invention significantly improves the perception accuracy of complex scenarios by: generating a multimodal joint feature matrix driven by environmental parameter weight distribution. Through spatiotemporal risk assessment, hierarchical navigation instructions are generated based on obstacle heat maps and the spatial overlap rate of flight segments. High-risk segments are split into detailed instructions, while low-risk segments are combined into long-range cruises. This ensures timely response to sudden obstacles, balances safety and efficiency, and achieves coordinated optimization of autonomous obstacle avoidance and route efficiency.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, and all of these should be included in the scope of protection of the present invention.

Claims

1. A method for processing low-altitude economic UAV data based on a large model, characterized in that: include, Collect and preprocess multimodal data, perform time alignment and spatial calibration on the preprocessed multimodal data, and generate a spatiotemporal alignment feature package; Based on the spatiotemporal alignment feature package, the dynamic attention mechanism is used to calculate the scene adaptation weights of multimodal data and fuse them to generate a multimodal joint feature matrix. This includes: based on the spatiotemporal alignment feature package, a multi-sensor sliding window statistical method is used to synchronously obtain scene parameters such as light intensity, point cloud density, and motion stability to generate a multimodal feature set and a scene perception parameter table; based on the multimodal feature set and the scene perception parameter table, a dynamic attention mechanism is used to calculate the scene adaptation weights of the multimodal data and generate a dynamic weight distribution table. The cross-attention mechanism is then used to perform inter-modal interaction modeling and feature reorganization to generate a multimodal joint feature matrix. Based on the multimodal joint feature matrix, a three-dimensional topological model of the urban airspace is constructed. The dynamic obstacle trajectories are predicted by combining the spatiotemporal graph neural network and a hierarchical navigation instruction set is generated. The drone executes flight instructions according to the hierarchical navigation instruction set and generates a flight status monitoring log by collecting real-time data during the drone flight; According to the abnormal events in the flight status monitoring log, the root causes of the abnormal events are located in combination with the predefined expert rule base, and a parameter optimization strategy package including dynamic attention parameter update rules and topological obstacle detection thresholds is generated.

2. The method for processing data of a low-altitude economic UAV based on a large model according to claim 1 is characterized in that: The preprocessing includes data cleaning, format standardization, lidar point cloud denoising and camera dynamic illumination compensation.

3. The method for processing data of a low-altitude economic UAV based on a large model according to claim 2 is characterized in that: The spatiotemporal alignment feature package is generated by temporally aligning and spatially calibrating the preprocessed multimodal data through a multi-sensor spatiotemporal registration algorithm.

4. The method for processing data of a low-altitude economic UAV based on a large model according to claim 3 is characterized in that: The method constructs a three-dimensional topological model of the urban airspace based on the multimodal joint feature matrix, combines the spatiotemporal graph neural network to predict the trajectory of dynamic obstacles, and generates a hierarchical navigation instruction set, including: Based on the obstacle distribution and semantic information in the multimodal joint feature matrix, a three-dimensional topological model of the urban airspace is constructed, and the airspace connectivity map is generated by integrating real-time traffic control data. Based on the three-dimensional topological model of the urban airspace and the airspace connectivity map, the spatiotemporal graph structure of dynamic obstacles is constructed through the spatiotemporal graph neural network. The obstacle trajectories are predicted by combining the multimodal joint feature matrix. Based on the obstacle trajectories, the spatiotemporal kernel density estimation is used to generate the dynamic obstacle trajectory heat map. According to the dynamic interaction relationship between the spatiotemporal graph structure of dynamic obstacles and the dynamic obstacle trajectory heat map, global path planning is used to select a safe waypoint sequence, and risk assessment is performed through the edge attributes in the spatiotemporal graph structure. According to the risk assessment results, a hierarchical navigation instruction set is generated.

5. The method for processing data of a low-altitude economic UAV based on a large model according to claim 4 is characterized in that: The UAV executes flight instructions according to the hierarchical navigation instruction set and generates a flight status monitoring log by collecting real-time data during the flight of the UAV, including: After receiving the hierarchical navigation instruction set, the drone will fly, synchronously collecting environmental and status data during the flight, and using dynamic risk thresholds to identify sudden obstacles, and trigger the emergency obstacle avoidance protocol to dynamically adjust the hierarchical navigation instruction set; The drone’s path deviation, obstacle avoidance events, power system overload alarm events, and the number of hierarchical navigation instruction set adjustments are recorded and marked as abnormal events to generate a flight status monitoring log.

6. The method for processing data of a low-altitude economic UAV based on a large model according to claim 5 is characterized in that: The method locates the root cause of abnormal events in the flight status monitoring log and combines it with a predefined expert rule base to generate a parameter optimization strategy package containing dynamic attention parameter update rules and topological obstacle detection thresholds, including: Parse abnormal event records in flight status monitoring logs, classify them according to predefined expert rule bases, and generate abnormal event classification lists; Construct a multi-level causal reasoning diagram based on the abnormal event classification list through forward reasoning chain to locate the root cause of the abnormal event; Based on the root cause of the abnormal event, a parameter optimization strategy package including dynamic attention parameter update rules and topological obstacle detection thresholds is generated through JSON structured instruction encapsulation and simulation environment verification process.

7. A large-model-based low-altitude economic UAV data processing system, based on a large-model-based low-altitude economic UAV data processing method according to any one of claims 1 to 6, characterized in that: include, The data acquisition module is used to collect and preprocess multimodal data, perform time alignment and spatial calibration on the preprocessed multimodal data, and generate a spatiotemporal alignment feature package; The scene fusion module is used to calculate the scene adaptation weights of multimodal data based on the spatiotemporal alignment feature package and the dynamic attention mechanism, and fuse them to generate a multimodal joint feature matrix. This includes: based on the spatiotemporal alignment feature package, a multi-sensor sliding window statistical method is used to synchronously obtain scene parameters such as light intensity, point cloud density, and motion stability to generate a multimodal feature set and a scene perception parameter table; based on the multimodal feature set and the scene perception parameter table, a dynamic attention mechanism is used to calculate the scene adaptation weights of the multimodal data, generate a dynamic weight distribution table, and perform inter-modal interaction modeling and feature reorganization through the cross-attention mechanism to generate a multimodal joint feature matrix. The airspace collaborative navigation decision module is used to construct a three-dimensional topological model of the urban airspace based on a multimodal joint feature matrix, combine it with a spatiotemporal graph neural network to predict dynamic obstacle trajectories, and generate a hierarchical navigation instruction set; The flight anomaly monitoring module is used for the UAV to execute flight instructions according to the hierarchical navigation instruction set, and to generate a flight status monitoring log by collecting real-time data during the UAV flight; The autonomous optimization evolution module is used to locate the root cause of abnormal events based on the abnormal events in the flight status monitoring log and the predefined expert rule base, and generate a parameter optimization strategy package that includes dynamic attention parameter update rules and topological obstacle detection thresholds.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for processing low-altitude economical drone data based on a large model according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for processing low-altitude economical UAV data based on a large model according to any one of claims 1 to 6 is implemented.

Images