A low-altitude patrol route planning method based on trajectory clustering

Abstract

The application discloses a kind of low-altitude patrol route planning method based on trajectory clustering, it is related to the technical field of aviation logistics, including: obtaining the historical flight trajectory data and low-altitude micro-meteorological field data in target low-altitude area, generate trajectory feature vector set containing multidimensional space-time and meteorological attribute;Generate multiple trajectory clusters that present different low-altitude passing characteristics;Obtain the target point set to be patrolled, determine the spatial matching degree between the target point set to be patrolled and each trajectory cluster, and distribute the target point set to be patrolled into corresponding trajectory cluster based on spatial matching degree;For each trajectory cluster that is assigned with the target point set to be patrolled, obtain the historical passing cost benchmark in trajectory cluster, and generate adaptive flight path planning constraint in combination with real-time environmental state parameter;Optimal patrol route covering the target points to be patrolled in each trajectory cluster is generated using multi-objective path optimization algorithm;The application realizes global full-automatic, dynamic adaptation low-altitude complex environment unmanned aerial vehicle intelligent patrol route planning.

Description

Technical Field

[0001] This invention relates to the field of aviation logistics technology, specifically to a low-altitude inspection route planning method based on trajectory clustering. Background Technology

[0002] With the booming development of the low-altitude economy, drones are increasingly being used in fields such as power line inspection, traffic monitoring, logistics distribution, and geographic mapping. Efficient and safe route planning is a key link in ensuring the smooth execution of low-altitude flight missions. Existing low-altitude inspection route planning technologies have the following limitations: lack of in-depth consideration of dynamic environmental factors: traditional methods are mostly based on static geographic information, such as building height and no-fly zones, for path searching, often ignoring the significant impact of low-altitude micro-meteorological fields, such as wind shear and thermal airflow, on the energy consumption, stability, and safety of drone flights. This leads to the theoretically planned routes potentially facing the risk of energy consumption surges or loss of control during actual flights.

[0003] Existing planning algorithms typically treat each task as an independent event, failing to fully utilize the massive amount of historical flight trajectory data. In reality, historical trajectories contain validated safe corridors and efficient paths; ignoring this information leads to repeated exploration of inefficient paths, reducing inspection efficiency. When faced with real-time changing environmental conditions, such as sudden strong winds or temporary no-fly zones, traditional algorithms often need to recalculate globally, lacking a mechanism for rapid adaptive adjustment based on historical travel cost benchmarks. This results in slow planning response speeds, and the generated routes often fail to balance energy consumption, time, and safety.

[0004] Therefore, there is an urgent need for a low-altitude inspection route planning method that can integrate historical trajectory experience with micro-meteorological environment characteristics to achieve adaptive and highly reliable results. Summary of the Invention

[0005] To address the aforementioned technical problems, a low-altitude inspection route planning method based on trajectory clustering is provided. This technical solution solves the problems mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A low-altitude inspection route planning method based on trajectory clustering includes: Historical flight trajectory data and low-altitude micro-meteorological field data within the target low-altitude area are acquired, and spatiotemporal alignment and feature extraction are performed on the historical flight trajectory data and low-altitude micro-meteorological field data to generate a trajectory feature vector set containing multi-dimensional spatiotemporal and meteorological attributes. Based on the trajectory feature vector set, a spatiotemporal and meteorological coupled clustering algorithm is used to cluster historical flight trajectory data to generate multiple trajectory clusters that exhibit different low-altitude travel characteristics. Obtain the set of target points to be inspected, determine the spatial matching degree between the set of target points to be inspected and each trajectory cluster, and assign the set of target points to be inspected to the corresponding trajectory clusters based on the spatial matching degree; For each trajectory cluster with assigned target points to be inspected, the historical travel cost benchmark within the trajectory cluster is obtained, and adaptive trajectory planning constraints are generated by combining real-time environmental state parameters. Based on adaptive trajectory planning constraints, an optimal inspection route covering the target points to be inspected within each trajectory cluster is generated using a multi-objective path optimization algorithm.

[0007] Furthermore, historical flight trajectory data and low-altitude micro-meteorological field data within the target low-altitude area are acquired, and spatiotemporal alignment and feature extraction are performed on the historical flight trajectory data and low-altitude micro-meteorological field data to generate a trajectory feature vector set containing multi-dimensional spatiotemporal and meteorological attributes, including: Obtain the latitude and longitude coordinate sequence, flight altitude sequence, and timestamp sequence from historical flight trajectory data; Determine the flight speed and acceleration between adjacent coordinate points based on the timestamp sequence; The latitude and longitude coordinate sequence, flight altitude sequence and timestamp sequence are mapped to the three-dimensional meteorological grid corresponding to the low-altitude micro-meteorological field data to obtain the wind shear index and thermal flow intensity at each adjacent coordinate point; By cascading flight speed, acceleration, wind shear index, and thermal airflow intensity, a set of trajectory feature vectors is generated.

[0008] Furthermore, based on the trajectory feature vector set, a spatiotemporal and meteorological coupled clustering algorithm is used to cluster historical flight trajectory data, generating multiple trajectory clusters exhibiting different low-altitude travel characteristics, including: The trajectory feature vector set is mapped onto a low-altitude three-dimensional spatial grid to determine the spatiotemporal and meteorological similarity matrix between each trajectory feature vector. Based on the spatiotemporal and meteorological similarity matrix, density clustering algorithm is used to identify high-density connected regions in historical flight trajectory data; Historical flight trajectories within a high-density connected region are divided into the same trajectory cluster, and the centroid coordinates and average passage cost of each trajectory cluster are determined.

[0009] Furthermore, the process involves acquiring the set of target points to be inspected, determining the spatial matching degree between the set of target points and each trajectory cluster, and assigning the set of target points to be inspected to the corresponding trajectory clusters based on the spatial matching degree, including: Obtain the set of three-dimensional coordinates of the target points to be inspected; Determine the Euclidean distance between each target point to be inspected and the centroid coordinates of each trajectory cluster; Based on Euclidean distance, spatial matching degree is determined by a monotonically decreasing mapping method; The target points to be inspected with a spatial matching degree not lower than the preset matching threshold are assigned to the trajectory cluster with the highest spatial matching degree, and the number of target points in each trajectory cluster is counted.

[0010] Furthermore, for each trajectory cluster assigned with a set of target points to be inspected, historical travel cost benchmarks within the trajectory cluster are obtained, and adaptive trajectory planning constraints are generated by combining real-time environmental state parameters, including: Historical average energy consumption and historical flight time are obtained from historical flight trajectory data within the trajectory cluster and used as historical travel cost benchmarks. Obtain real-time wind speed and direction from real-time environmental status parameters, and obtain real-time no-fly zone boundaries; Based on real-time wind speed and direction, the historical passage cost benchmark is corrected to generate the expected passage cost. By using the expected passage cost as a weight and mapping the real-time no-fly zone boundary as a spatial exclusion constraint, adaptive trajectory planning constraints are generated.

[0011] Furthermore, based on adaptive trajectory planning constraints, a multi-objective path optimization algorithm is used to generate the optimal inspection route covering the target points to be inspected within each trajectory cluster, including: Construct a directed graph model with the target points to be inspected as nodes and the expected passage cost as edge weights; In directed graph models, spatial exclusion constraints are incorporated as a penalty term into the cost function of multi-objective path optimization algorithms; The multi-objective path optimization algorithm is used to search for the traversal path in the directed graph model that connects all the target points to be inspected within the trajectory cluster and has the minimum total cost. After smoothing the traversal path, the optimal inspection route is generated.

[0012] Furthermore, after generating the optimal inspection route covering the target points to be inspected within each trajectory cluster, it also includes: Obtain the performance parameters of the UAV that executes the optimal inspection route, including maximum endurance and maximum payload; Determine the expected total travel cost of the optimal inspection route; When the expected total toll cost is not less than the energy consumption threshold corresponding to the maximum range, the route splitting strategy is triggered. Based on the centroid coordinates of the trajectory cluster, the optimal inspection route is divided into multiple sub-routes, and a collaborative inspection plan containing each sub-routes is generated.

[0013] Furthermore, after generating the optimal inspection route covering the target points to be inspected within each trajectory cluster, it also includes: Acquire real-time flight status data collected by the UAV during the execution of the optimal inspection route; By comparing flight status data with historical flight trajectory data, the deviation between the actual passage cost and the expected passage cost is determined. When the deviation value is not less than the preset deviation threshold, the flight status data is updated into the historical flight trajectory data as new historical flight trajectory data, and the trajectory feature vector set is regenerated based on the updated historical flight trajectory data.

[0014] Furthermore, before acquiring historical flight trajectory data and low-altitude micro-meteorological field data within the target low-altitude area, the following steps are also included: Acquire three-dimensional geographic information data and building outline data of the target low-altitude area; Based on building outline data, safe flight corridors and no-fly zones are identified; The safe flight corridor and no-fly zone are mapped onto a low-altitude 3D mesh model to generate a basic environmental constraint map; Historical flight trajectory data is filtered based on a basic environmental constraint map to remove abnormal trajectory data located within no-fly zones.

[0015] Furthermore, before applying the spatiotemporal and meteorological coupled clustering algorithm to low-altitude flight path planning, it also includes: Obtain a historical inspection sample set containing known optimal inspection paths; The historical inspection sample set is input into the spatiotemporal and meteorological coupled clustering algorithm to determine the overlap between the generated trajectory clusters and the known optimal inspection path; With the goal of maximizing the overlap, the spatiotemporal weight parameters in the spatiotemporal and meteorological similarity matrix are iteratively adjusted using the gradient descent algorithm to obtain the spatiotemporal and meteorological coupled clustering algorithm after parameter tuning.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention incorporates low-altitude micro-meteorological field data, aligning meteorological attributes such as wind shear index and thermal flow intensity with historical flight trajectories in both time and space for feature extraction. By constructing a trajectory feature vector set containing multi-dimensional temporal, spatial, and meteorological attributes, the generated flight path not only avoids static obstacles but also proactively evades high-risk areas with severe weather conditions. By combining historical travel cost benchmarks with real-time environmental parameters to correct expected travel costs, the planned flight path is ensured to be optimal in terms of both energy consumption and time, thereby improving the endurance and safety of UAV operations.

[0017] 2. This invention utilizes a spatiotemporal and meteorological coupled clustering algorithm to mine historical trajectories and identify trajectory clusters with different low-altitude travel characteristics. When planning new routes, the spatial matching degree between the target point to be inspected and the trajectory clusters is calculated, and mature travel areas are directly reused as basic constraints. This method avoids blind searching in a vast airspace, significantly reduces the optimization space, and enables the algorithm to quickly generate the optimal inspection route covering the target point, thus improving planning efficiency. 3. This invention constructs a closed-loop optimization mechanism. On the one hand, it can dynamically generate adaptive trajectory planning constraints based on real-time wind speed and direction and no-fly zone boundaries, ensuring the real-time response of flight routes to sudden environmental changes. On the other hand, by comparing the actual flight status data during execution with the expected cost, new data is backfilled into the historical database and the feature vector set is updated when the deviation is large. This online learning and continuous iteration mechanism makes the flight route planning system more and more accurate with the increase in usage, and can continuously adapt to the complex and ever-changing low-altitude environment. When the expected passage cost exceeds the UAV's endurance threshold, the system can automatically trigger a flight route splitting strategy based on the centroid of the trajectory cluster, generating a collaborative inspection plan for multiple sub-flight routes. This effectively solves the problem that the single-aircraft endurance is limited and cannot cover a large range of targets, and improves the scheduling capability of multi-aircraft collaborative operations. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the present invention. Detailed Implementation

[0019] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0020] Example 1: Refer to Figure 1 As shown, a low-altitude inspection route planning method based on trajectory clustering includes: step 101, acquiring historical flight trajectory data and low-altitude micro-meteorological field data within the target low-altitude area, and performing spatiotemporal alignment and multi-dimensional feature extraction on the two types of data to generate a trajectory feature vector set; A 2km×3km urban power grid low-altitude inspection area was selected, with drones routinely inspecting at altitudes ranging from 20m to 120m. Historical flight trajectories were collected by multiple inspection drones operating continuously for 3 months, with a sampling frequency of 1Hz for each trajectory. Each sampling point on a single trajectory simultaneously recorded WGS84 longitude, WGS84 latitude, altitude, and millisecond-level timestamps. Low-altitude micro-meteorological fields were collected by a network of regionally distributed micro-meteorological stations, with a spatial grid resolution of 50m×50m×10m. Each grid simultaneously stored two types of meteorological values: real-time wind shear index and near-surface thermal airflow intensity. Complete spatiotemporal alignment operation: Using the trajectory's millisecond timestamp as the sole synchronization benchmark, all sampling points of each trajectory are traversed. Based on the three-dimensional coordinates of the point and the corresponding collection time, a three-dimensional meteorological grid is located. The trajectory point is bound to the meteorological value of the grid at the same time, achieving precise spatial and temporal matching. This can be understood as binding the drone flight data and turbulent meteorological data at the same time and spatial location with the turbulent meteorological data, eliminating the feature mismatch problem caused by collection time difference and spatial misalignment. For example, for the same tower point, the trajectory data at 10:02 AM will only match the meteorological data of the corresponding grid at 10:02 AM, and will not mix meteorological values ​​from other times. The complete feature extraction process involves: segmenting adjacent trajectory sampling points one by one; calculating the instantaneous velocity and average acceleration of the segment; reading the wind shear index and thermal flow intensity of the corresponding grid for that segment; and concatenating the "velocity, acceleration, corresponding wind shear, and corresponding thermal flow" in a fixed order to generate a four-dimensional feature vector for a single trajectory segment. All segment vectors of the entire trajectory are concatenated to form a complete trajectory, and all historical trajectories in the region are summarized to form a unified trajectory feature vector set. For example, a patrol trajectory with 500 sampling points will be split into 499 segment four-dimensional vectors, which are then concatenated to form the complete feature vector of the trajectory. The trajectory feature vector set is a four-dimensional motion + meteorological time series dataset of all historical inspection trajectories in the area; spatiotemporal alignment is a standardized preprocessing procedure to eliminate trajectory and meteorological collection time / space misalignment; Step 102: Based on the trajectory feature vector set, use spatiotemporal and meteorological coupled DBSCAN density clustering to complete trajectory grouping and divide independent trajectory clusters with consistent travel environments; All trajectory feature vectors are mapped to a 50m×50m×10m three-dimensional low-altitude grid, where X / Y represent planar latitude and longitude, and Z-axis represents ground elevation. The joint similarity is calculated for each pair of trajectories. The complete similarity calculation process is as follows: First, calculate the three-dimensional average straight-line distance and average operation time difference of all sampling points for both trajectories. Second, normalize the spatial distance, time difference, and four-dimensional meteorological feature difference from 0 to 1. Third, fix the spatiotemporal weight at 0.6 and the meteorological weight at 0.4, and calculate the joint similarity as 0.6 × (1 - normalized spatial distance) + 0.4 × (1 - normalized total meteorological difference). All pairwise trajectory similarities are summarized to generate a square matrix similarity matrix. DBSCAN core parameters are set: neighborhood similarity threshold 0.75, minimum number of core trajectories 3. All vectors are traversed, and those with similarity ≥ 0.75 and consecutively connected... All trajectories are grouped into the same high-density area; scattered and isolated trajectories are directly eliminated and do not participate in cluster division; each high-density area generates an independent trajectory cluster, and the cluster centroid is calculated by taking the arithmetic mean of the longitude, latitude, and altitude of all points within the area; the travel cost benchmark is calculated by statistically analyzing the total energy consumption and total flight time of each trajectory within the cluster, and taking the arithmetic mean of all trajectories as the static historical travel cost of that cluster; it can be understood that the closer two trajectories are in space, the similarity of their operation time, and the similarity of their turbulence conditions, the closer their similarity value is to 1, and the easier it is for them to be classified into the same trajectory cluster; for example, two inspection routes that both fly at a height of 100m along a river, operate at noon, and have similar wind shear values ​​can have a similarity of 0.88, which meets the 0.75 threshold and are classified into the same cluster; while a trajectory along a high-rise building and a trajectory along an open farmland have a similarity of only 0.32, and will be divided into two independent clusters; Spatiotemporal and meteorological coupled clustering is a density clustering algorithm that simultaneously integrates spatial, temporal, and turbulent multi-dimensional similarity; trajectory clusters represent long-term stable, low-energy mature inspection flight corridors; Step 103: Import the target point set to be inspected this time, calculate the spatial matching degree between each point and each trajectory cluster, and allocate them to the optimal trajectory cluster according to the matching threshold; A total of 126 targets (power poles and river sewage monitoring points) were to be inspected. All targets were uniformly converted to WGS84 3D coordinates, and each point was assigned a unique inspection number. For a single target point, the centroids of all trajectory clusters were traversed, and the 3D Euclidean straight-line distance was calculated group by group. The complete conversion logic for spatial matching degree was as follows: the maximum effective matching distance was set to 300m, and the matching degree value ranged from 0 to 1. At a distance of 0m, the matching degree = 1. For every 50m increase in distance, the matching degree decreased linearly by 0.05. When the distance was ≥300m, the matching degree was directly set to 0. The system preset the matching threshold to 0.6. For a single target, the matching degree of all clusters was traversed, and only the matching degree of each cluster was retained. Clusters with a matching degree ≥ 0.6 are considered candidates and ultimately assigned to the trajectory cluster with the highest matching degree. After assignment, the total number of points to be inspected within each trajectory cluster is counted. This can be understood as follows: the closer the target point is to the cluster centroid, the higher the matching degree, and the more suitable it is to reuse the mature, low-energy flight path of that cluster. For example, if the distance between the tower point and the cluster centroid is 100m, the matching degree = 1 - 100 ÷ 50 × 0.05 = 0.9, which is greater than 0.6, so it is assigned to that cluster. If the distance is 350m, the matching degree is set to 0 and it will not be assigned to that cluster. In densely populated urban high-rise buildings, the matching threshold can be increased to 0.7, while in open suburban areas it can be decreased to 0.5. Spatial matching measures the degree of fit between target locations and established inspection corridors; the preset matching threshold of 0.6 can be adjusted according to the building density of the inspection area, increasing to 0.7 in urban high-rise areas and decreasing to 0.5 in open suburban areas; Step 104: Retrieve the historical travel cost benchmark of the trajectory cluster and generate adaptive trajectory planning constraints by combining real-time wind speed and no-fly zone; The average single-segment energy consumption and average single-segment flight time stored in the current trajectory cluster are retrieved as the static traffic cost benchmark. Real-time environmental data is collected simultaneously from two sources: a micro weather station updates grid wind speed and direction every minute; the low-altitude airspace control platform pushes real-time information on temporary construction and event polygon no-fly zones for the day. The complete logic for traffic cost correction is as follows: real-time wind direction is decomposed into parallel and perpendicular flight path components, and only vertical crosswinds increase flight energy consumption; for every 1 m / s increase in crosswind, the expected energy consumption of that segment increases by 8% based on the benchmark; the no-fly zone polygon is mapped to a 3D grid, allowing arbitrary paths to be mapped. A penalty of 10,000 Wh is set when nodes or line segments enter the no-fly zone; the corrected segmented energy consumption is used as the edge weight for path optimization, and the no-fly zone penalty is used as a hard constraint. The two types of parameters together form the adaptive programming constraint. It can be understood that the stronger the crosswind, the higher the flight resistance, and the energy consumption increases accordingly; for example, the baseline energy consumption of this cluster is 50 Wh, the real-time crosswind is 3 m / s, and the expected passage cost is 50 × (1 + 3 × 0.08) = 62 Wh; if the flight path passes through the temporary event no-fly zone, a penalty of 10,000 Wh will be added, and the algorithm will automatically avoid the path; The expected cost of passage is the estimated energy consumption of a single flight segment after real-time weather correction; the spatial exclusion constraint is the path penalty rule corresponding to the no-fly zone; Step 105: Based on adaptive planning constraints, use multi-objective path optimization to solve for the optimal continuous inspection route covering all objectives; Using all inspection points within the trajectory cluster as graph nodes, and the expected travel cost of bidirectional flight between two points as the weight of directed edges, a fully connected directed graph is constructed. The total path cost = the sum of the expected travel costs of all segments + the sum of penalties for violations in no-fly zones. An improved ant colony multi-objective optimization algorithm is adopted, with optimization objectives including total energy consumption, total flight time, and the number of no-fly violations. The iterative convergence condition is: the optimal total cost does not decrease after 15 consecutive iterations. After iteration, a discrete point traversal sequence with no repetition and covering all points is output. The traversal sequence is smoothed using a third-order B-spline curve to eliminate sharp right-angle turns and output a continuous and smooth three-dimensional optimal inspection route. This can be understood as the algorithm automatically selecting the continuous inspection path with the lowest total energy consumption that does not enter no-fly zones. For example, for 126 tower points, the total energy consumption of the unoptimized random route is 820Wh, while the total energy consumption of the optimized route is only 630Wh, and the entire route does not pass through construction no-fly zones. In one implementation, when the number of monitoring points is less than 30, the dynamic programming TSP algorithm can be used instead of the ant colony algorithm, which is faster. When the number of monitoring points exceeds 200, a constrained genetic algorithm is used. For example, for only 22 river monitoring points, the TSP algorithm takes only 0.8 seconds to solve the problem, while the ant colony algorithm takes 3.2 seconds.

[0021] Through the above technical solutions, mature inspection corridors are divided based on historical multi-dimensional trajectory clustering, inspection points are automatically allocated based on spatial matching degree, and path energy consumption constraints are dynamically corrected in conjunction with real-time weather. Multi-objective optimization is used to output low-energy, compliant and smooth inspection routes, which solves the defects of traditional routes that rely solely on planar distance partitioning, ignore low-altitude turbulence, resulting in high route energy consumption and easy entry into no-fly zones.

[0022] In some embodiments, historical flight trajectories and low-altitude micro-meteorological fields are acquired, and spatiotemporal alignment and feature extraction are performed to generate a trajectory feature vector set. The complete execution steps are as follows: Step 201: Read the complete historical flight trajectory's three-dimensional coordinate sequence, altitude sequence, and millisecond timestamp sequence; The sampling frequency for a single historical trajectory is 1Hz. Each sampling point on a single trajectory synchronously stores WGS84 longitude, WGS84 latitude, and altitude above the ground. The timestamp accuracy is down to milliseconds. All trajectories are stored as structured time-series tables. This can be understood as the millisecond-level timestamps ensuring accurate matching of meteorological data. For example, a sampling point at 10:02:15.320 will not match meteorological grid data at 10:02:15.500. Step 202: Based on the time difference between adjacent sampling points, calculate the instantaneous flight velocity and average acceleration segment by segment; Δt between two adjacent sampling points represents the number of seconds converted from milliseconds, and the three-dimensional straight-line distance represents the spatial distance between the two points. Instantaneous velocity = distance ÷ Δt. The average acceleration of the current segment is obtained by dividing the velocity difference between the current segment, the preceding segment, and the following segment by the total time of the three segments. For example, if the interval between two points is 0.2s and the spatial distance is 12m, the instantaneous velocity is 60m / s; the velocities of the preceding and following segments are 58, 60, and 62m / s respectively, and the average acceleration is 10m / s². 2 This can be understood as the acceleration value reflecting the acceleration and deceleration fluctuations of the UAV, serving as a trajectory motion characteristic to distinguish between two types of flight states: smooth cruise and variable speed circling. Step 203: Using the trajectory timestamp and three-dimensional coordinates as indexes, match the corresponding three-dimensional meteorological grid and read the wind shear index and thermal flow intensity; The low-altitude micro-meteorological field is stored in a 50m×50m×10m three-dimensional grid, with each grid bound to millisecond-level time-series meteorological data. The grid is located based on the latitude, longitude, and altitude of the point, and then synchronous meteorological parameters are retrieved with a millisecond timestamp. For example, at coordinates (116.32, 30.15, 60m) and 10:02:15.320, the wind shear value at the same time for that grid is directly retrieved. This can be understood as achieving one-to-one binding of single-sampling point flight data with synchronous turbulent meteorological data, without spatial or temporal misalignment. Step 204: Cascade the four types of values—velocity, acceleration, wind shear index, and thermal airflow intensity—in one dimension to generate a multi-dimensional feature vector for a single trajectory, and summarize them to form a trajectory feature vector set; The single-segment features are arranged in a fixed order: instantaneous velocity of this segment, average acceleration of this segment, wind shear of the corresponding grid, and corresponding thermal flow; all segment features of the entire trajectory are spliced ​​together in sequence; for example, a trajectory with 500 sampling points generates 499 four-dimensional segmented vectors that are concatenated into a long one-dimensional vector; it can be understood that each vector completely records the dual attributes of motion and turbulence of a segment of flight, which serves as the original input for clustering.

[0023] Through the above technical solution, four-dimensional features of flight motion and local turbulence are extracted simultaneously. The trajectory and meteorological conditions are accurately matched in time and space by relying on millisecond-level timestamps. The vectors completely reproduce each segment of the low-altitude flight environment, providing complete input data for coupled clustering.

[0024] In some embodiments, multiple trajectory clusters are generated based on the trajectory feature vector set using spatiotemporal meteorological coupling clustering. The complete execution steps are as follows: Step 301: Map all trajectory feature vectors to a 50m×50m×10m low-altitude 3D grid, calculate the joint similarity of each pair of trajectories, and generate a similarity matrix. The grid X / Y plane represents the inspection plane coordinates, and Z represents the flight altitude; the three-dimensional average distance, operation time difference, and meteorological feature difference between the two trajectories are calculated and normalized to 0~1 respectively; the joint similarity is 0.6×(1-normalized spatial distance)+0.4×(1-normalized meteorological difference); all trajectories are traversed and paired to generate a square matrix; for example, the similarity between two trajectories at noon along the river 100m is 0.88, and the similarity between the trajectories of tall buildings and farmland is 0.32; this can be understood as a spatiotemporal weight of 0.6 to prioritize the differentiation of spatial areas, and a meteorological weight of 0.4 to differentiate turbulent environments; Step 302: Identify high-density connected trajectory regions using DBSCAN density clustering based on the similarity matrix; The neighborhood similarity threshold is set to 0.75, and the minimum number of core trajectories is 3. Trajectories with similarity ≥ 0.75 and continuous connectivity are classified as the same area. Segments with only 1 or 2 similar trajectories are directly removed. For example, if the similarity of 5 consecutive river inspection trajectories in an area is higher than 0.75, they are classified as a complete trajectory cluster. Single scattered building-circling trajectories are directly discarded and do not participate in the zoning. This can be understood as removing scattered abnormal flight routes with no reference value and only reusing long-term mature continuous inspection corridors. Step 303: Divide the trajectories of the same high-density area into single trajectory clusters, and calculate the cluster centroid and the average passage cost of the cluster; The three-dimensional centroid is obtained by taking the arithmetic mean of the latitude, longitude and altitude of all points in the area; the average of the total energy consumption and total duration of all trajectories in the cluster is used as the reference for passage; for example, the average single-segment energy consumption of 15 historical trajectories in the cluster is 50Wh and the average flight time is 12s; it can be understood that the centroid of the cluster represents the geometric center of the area, and the average passage cost is used as the reference for calculating the energy consumption of the route. By using the above technical solutions, the similarity of flight space, operation time, and low-altitude turbulence is calculated in multiple dimensions. Density clustering is used to automatically identify long-term mature low-energy inspection corridors and output standardized trajectory clusters with centroid and energy consumption benchmarks, thereby realizing the automatic division of areas.

[0025] In some embodiments, the spatial matching degree between the target to be inspected and the trajectory cluster is matched, and the point allocation is completed. The complete execution steps are as follows: Step 401: Read the set of three-dimensional longitude, latitude, and altitude coordinates of all target points to be inspected in this inspection. Power poles and river monitoring points use the same WGS84 coordinate system, with each point having a unique number; for example, the coordinates of pole 1 are (116.321, 30.154, 55m). This unified coordinate system ensures that there are no systematic errors in distance calculations. Step 402: Calculate the three-dimensional Euclidean straight-line distance between each inspected target and the centroid of all trajectory clusters; The spatial straight-line distance is obtained by calculating the longitude difference, latitude difference, and altitude difference separately, summing the squares, and then taking the square root. For example, if the altitude difference between the target and the cluster centroid is 5m and the planar distance is 95m, the total three-dimensional distance is approximately 95.13m. It can be understood that the three-dimensional distance fully includes the altitude difference, and calculating only the planar distance will result in a matching deviation. Step 403: Calculate the spatial matching degree using a monotonically decreasing mapping based on the three-dimensional distance; The maximum effective matching distance is 300m. At a distance of 0m, the matching degree is 1. It decreases by 0.05 for every 50m increase, and is set to 0 for distances ≥300m. For example, the matching degree is 0.9 at a distance of 100m, 0.6 at a distance of 250m, and 0 at a distance of 320m. This can be understood as a linear monotonic mapping with no abrupt changes. The farther the distance, the lower the value of reusing mature routes. Step 404: Filter and assign points to the optimal trajectory cluster according to the matching threshold, and count the total number of targets in each cluster; With a preset matching threshold of 0.6, only candidate clusters with a matching degree ≥ 0.6 are retained and assigned to the cluster with the highest matching degree. For example, a certain tower has a matching degree of 0.9 for cluster A and a matching degree of 0.5 for cluster B, and is only assigned to cluster A. After all points are assigned, the number of towers in each cluster is counted. This can be understood as automatic zoning without the need for manual delineation of inspection areas, reducing human error.

[0026] The above technical solution uses a linear monotonically decreasing mapping to quantify the degree of alignment between the points and the mature inspection corridors. It automatically partitions the data based on a unified matching threshold, and the partitioning results match historical low-energy flight routes.

[0027] In some embodiments, according to the method of claim 1, historical travel references for trajectory clusters are extracted, and adaptive trajectory planning constraints are generated in combination with the real-time environment. The complete execution steps are as follows: Step 501: Read the historical average energy consumption and average flight time stored in the current trajectory cluster as the benchmark for passage cost; Each cluster permanently stores the average energy consumption and average flight time of a single segment; for example, the baseline energy consumption of the river inspection cluster is 50Wh, and the average flight time of a single segment is 12s; this can be understood as using this value as the path weight when there is no real-time wind. Step 502: Synchronously collect two types of real-time environmental parameters: real-time wind speed and direction across the entire region, and the boundary of the temporary no-fly zone for the day; Miniature weather stations refresh grid wind speed and angle every minute; low-altitude platforms push closed polygonal no-fly zones for construction and events; for example, the set of vertex coordinates of the no-fly zone next to a riverbank construction site on a given day; this can be understood as real-time dynamic airspace and meteorological data constraining daily automatic updates. Step 503: Calculate the expected passage cost for each segment based on the real-time crosswind component correction benchmark; Only crosswinds on the vertical flight path increase energy consumption, with an 8% increase for every 1m / s crosswind; the baseline is 50Wh, the expected energy consumption is 62Wh at 3m / s crosswind, and it remains at 50Wh in the windless state; this can be understood as tailwinds do not increase energy consumption, only lateral crosswinds bring additional flight drag, which is consistent with the aerodynamic characteristics of real drones. Step 504: Set the expected cost as the path edge weight, and map the no-fly zone as a spatial exclusion penalty to form an adaptive planning constraint; Any point on the flight path that falls into a no-fly zone incurs a massive penalty of 10,000 Wh; the optimization algorithm prioritizes avoiding paths with high total cost; this can be understood as the massive penalty value ensuring that the algorithm will not choose flight paths that intrude into controlled airspace; Through the above technical solutions, the flight energy consumption weight is dynamically adjusted in real time by crosswinds, rigid path penalties are set in temporary no-fly zones, and constraints are automatically updated according to the low-altitude environment of the day. The generated routes take into account both energy economy and airspace compliance.

[0028] In some embodiments, the optimal inspection route is generated based on adaptive constraint multi-objective optimization. The complete execution steps are as follows: Step 601: Construct a topological directed graph with the points to be inspected as nodes and the expected passage cost as the weight of the directed edges; For any two points, the two paths to and from each other are assigned independent weights; for example, the expected energy consumption from tower A to B is 62Wh, and the expected energy consumption from B to A is 58Wh, so the weights of the two directed edges are different; this can be understood as the crosswind conditions for the round trip being different, resulting in different energy consumption, and the directed graph closely reflects the actual flight conditions. Step 602: Incorporate the no-fly zone spatial exclusion penalty into the total path cost function; Total cost = Segmented energy consumption + No-fly zone penalty; As long as the path crosses the no-fly zone, the full penalty of 10,000Wh will be included; This can be understood as the total cost of the illegal route increasing significantly, and the algorithm will automatically discard it; Step 603: Optimize the multi-objective path by traversing the graph and finding the traversal path with the minimum total cost that connects all points. The improved ant colony algorithm optimizes objectives including total energy consumption, duration, and number of violations; it stops iterating if there is no optimal solution after 15 consecutive rounds; for example, after 126 towers, the total energy consumption is 630Wh; it can be understood as balancing energy consumption and compliance with multiple objectives, rather than simply pursuing the shortest distance. Step 604: Use B-spline to smooth the discrete points and generate a continuous and executable optimal inspection route; Third-order B-spline interpolation eliminates right-angle turns and outputs continuous three-dimensional flight coordinates; this can be understood as preventing the UAV from making sharp right-angle turns and smoothing the flight path to adapt to flight control execution commands. In one implementation, when the number of monitoring points is less than 30, a dynamic programming TSP algorithm can be used instead of an ant colony, while a genetic algorithm is used for a large number of monitoring points; for example, the TSP calculation speed is improved by 70% for 20 monitoring points. Through the above technical solution, a directed graph model with real-time energy consumption and no-fly zone penalties is built. Multi-objective optimization is used to comprehensively balance energy consumption, duration, and airspace compliance. The smoothed flight path is adapted to the continuous flight commands of the UAV flight control system.

[0029] In some embodiments, after generating the optimal inspection route, the following steps are performed to split the remaining range and generate a multi-aircraft collaborative plan: Step 701: Read and execute the drone's hardware performance parameters: maximum flight time and maximum payload; Example of parameters for an industrial inspection drone: maximum flight time of 28 minutes, full payload of 1.2kg; parameters are fixed in the system configuration file; this can be understood as the parameters being written to the actual inspection camera and radar equipment being carried, without the need for repeated input; Step 702: Calculate the total expected toll cost for the entire optimal route and convert it into the corresponding total flight time; The total energy consumption is obtained by summing up the expected energy consumption of all segments, and then the flight time is calculated by combining the average power consumption of 25Wh / min. For example, if the total energy consumption is 650Wh, the total flight time = 650 ÷ 25 = 26 minutes. This can be understood as accurately calculating the flight time through energy consumption, without relying on rough distance estimation. Step 703: Compare the total flight time with the 90% safety threshold of maximum range; if the threshold is exceeded, trigger the split strategy. Safety threshold = 28 × 0.9 = 25.2 min; Example total duration 26 min > 25.2 min, trigger split; This can be understood as reserving 10% battery life redundancy to avoid running out of power and having to make a forced landing midway; Step 704: Based on the centroid of each trajectory cluster, cut the complete route, generate multiple sub-routes, and generate a collaborative inspection plan containing each sub-routes; Based on the natural division of cluster centroids, each sub-route has a duration of no more than 25.2 minutes; the collaborative plan marks the takeoff site, drone number, and location list; for example, the original route is split into 3 sub-routes, each assigned to 3 drones for simultaneous operation; this can be understood as splitting based on clustered areas, without the need for manual re-division of the inspection scope; Through the above technical solutions, the flight path endurance load is automatically verified, and ultra-long flight paths are split based on the original trajectory clusters to automatically output multi-aircraft collaborative inspection plans, which are suitable for large-scale continuous low-altitude inspection scenarios.

[0030] In some embodiments, the steps for completing the closed-loop iterative update of flight data after route execution are as follows: Step 801: The UAV transmits a complete set of flight status timing data back in real time during the flight process; It transmits three-dimensional coordinates, real-time power consumption, and airborne wind speed every second; it stores the complete historical trajectory; this can be understood as adding real working condition data with each operation, continuously expanding the trajectory library. Step 802: Compare the flight status data with the planned expected cost to determine the deviation between the actual and expected passage costs; Single-segment deviation = |actual power consumption - expected power consumption|; for example, if the planned power consumption is 50Wh and the actual power consumption is 57Wh, the single-segment deviation is 7Wh; this can be understood as the deviation reflecting the energy consumption prediction error caused by changes in weather and buildings. Step 803: When the deviation value is not less than the preset 12% baseline energy consumption deviation threshold, update the trajectory feature vector set; Threshold = 50 × 12% = 6Wh; 7Wh ≥ 6, the deviation is judged to be excessive, the current trajectory is entered into the database, and all feature vectors are regenerated; this can be understood as the error exceeding the reasonable range indicating a change in the spatial environment, and the clustering benchmark needs to be updated; By using the above technical solutions, a closed loop for route execution feedback is constructed, and real flight data is used to update the trajectory database in reverse. This adapts to the slow changes in the low-altitude environment caused by seasonal airflow and new buildings, thus maintaining the accuracy of route planning in the long term.

[0031] In some embodiments, the pre-trajectory environment filtering step before trajectory acquisition is as follows: Step 901: Obtain three-dimensional geographic information data and building outline data of the target low-altitude area; The 3D geographic model includes the elevation of the entire region; the accuracy of the closed polygonal outlines of buildings is 1 meter; for example, the outlines of high-rise buildings and power transmission towers in urban areas are all recorded; it can be understood as a complete obstacle model that accurately delineates no-fly boundaries. Step 902: Based on building outline data, obtain safe flight corridors and no-fly zones; A 15m buffer zone extending outward from a building is a no-fly zone; an unobstructed area more than 20m above the ground is a safe corridor; for example, the entire 15m radius outside a 30m high building is designated as a no-fly zone; this can be understood as reserving a safe avoidance distance to prevent the risk of collision with the building. Step 903: Map the safe flight corridor and no-fly zone onto the low-altitude 3D mesh model to generate a basic environmental constraint map; Each cell in a 50m×50m×10m grid is marked "Safe / No-Fly"; this can be understood as a gridded map enabling rapid trajectory determination and filtering. Step 904: Filter historical flight trajectory data based on the basic environmental constraint map to remove abnormal trajectory data located in no-fly zones; If any sampling point on a trajectory falls within a no-fly zone, the entire trajectory is discarded. For example, if a trajectory passes through a building buffer zone, it is directly discarded and does not participate in clustering. This can be understood as removing illegal flight data from the source to ensure that clustering is based entirely on legal flight routes. The above technical solution filters out abnormal trajectories based on three-dimensional building meshes in advance, and completes area clustering and flight route planning using only compliant historical data, thus avoiding the risk of trespassing into buildings from the source.

[0032] In some embodiments, before applying the spatiotemporal and meteorological coupled clustering algorithm to low-altitude flight path planning, the method further includes: Step 101: Obtain a historical inspection sample set containing known optimal inspection paths; The optimal routes that have been verified manually in the field over the past year are collected; each route is fully annotated with three-dimensional coordinates; for example, 10 sets of manually optimized river and power grid inspection routes are used as optimization samples; this can be understood as using real high-quality routes as optimization reference benchmarks; Step 102: Input the historical inspection sample set into the spatiotemporal and meteorological coupled clustering to determine the overlap between the trajectory clusters generated by the clustering and the known optimal inspection path; Overlap ratio = Length of the cluster grid in which the optimal route falls ÷ Total length of the optimal route; value ranges from 0 to 1; for example, if 80% of the manually selected optimal routes fall within the corresponding trajectory clusters, the overlap ratio is 0.8; it can be understood that the higher the overlap ratio, the more closely the clustering is aligned with the mature routes. Step 103: With the goal of maximizing the overlap, the spatiotemporal and meteorological weights are iteratively adjusted using the gradient descent algorithm to obtain the coupled clustering algorithm after parameter tuning; The initial spatiotemporal weight is 0.6 and the meteorological weight is 0.4; the single adjustment range is ±0.02; the iteration limit is 50 rounds, and the algorithm stops after 8 consecutive rounds of no improvement in overlap; after the example iteration, the spatiotemporal weight is 0.68 and the meteorological weight is 0.32, and the overlap increases from 0.72 to 0.91; this can be understood as eliminating the need for manual weight setting based on experience, with the algorithm automatically finding the optimal ratio and reducing the overall flight energy consumption of the route; Through the above technical solution, the cluster similarity weight is automatically iteratively optimized based on the optimal flight route by real human intervention. The clustered areas are highly matched with mature low-energy inspection corridors, reducing the total flight energy consumption of subsequent routes.

[0033] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A low-altitude inspection route planning method based on trajectory clustering, characterized in that, include: Historical flight trajectory data and low-altitude micro-meteorological field data within the target low-altitude area are acquired, and spatiotemporal alignment and feature extraction are performed on the historical flight trajectory data and low-altitude micro-meteorological field data to generate a trajectory feature vector set containing multi-dimensional spatiotemporal and meteorological attributes. Based on the trajectory feature vector set, a spatiotemporal and meteorological coupled clustering algorithm is used to cluster historical flight trajectory data to generate multiple trajectory clusters that exhibit different low-altitude passage characteristics. Obtain the set of target points to be inspected, determine the spatial matching degree between the set of target points to be inspected and each trajectory cluster, and assign the set of target points to be inspected to the corresponding trajectory clusters based on the spatial matching degree; For each trajectory cluster with assigned target points to be inspected, the historical travel cost benchmark within the trajectory cluster is obtained, and adaptive trajectory planning constraints are generated by combining real-time environmental state parameters. Based on adaptive trajectory planning constraints, an optimal inspection route covering the target points to be inspected within each trajectory cluster is generated using a multi-objective path optimization algorithm.

2. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, Historical flight trajectory data and low-altitude micro-meteorological field data within the target low-altitude area are acquired, and spatiotemporal alignment and feature extraction are performed on the historical flight trajectory data and low-altitude micro-meteorological field data to generate a trajectory feature vector set containing multi-dimensional spatiotemporal and meteorological attributes, including: Obtain the latitude and longitude coordinate sequence, flight altitude sequence, and timestamp sequence from historical flight trajectory data; Determine the flight speed and acceleration between adjacent coordinate points based on the timestamp sequence; The latitude and longitude coordinate sequence, flight altitude sequence and timestamp sequence are mapped to the three-dimensional meteorological grid corresponding to the low-altitude micro-meteorological field data to obtain the wind shear index and thermal flow intensity at each adjacent coordinate point; By cascading flight speed, acceleration, wind shear index, and thermal airflow intensity, a set of trajectory feature vectors is generated.

3. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, Based on trajectory feature vector sets, a spatiotemporal and meteorological coupled clustering algorithm is used to cluster historical flight trajectory data, generating multiple trajectory clusters exhibiting different low-altitude passage characteristics, including: The trajectory feature vector set is mapped onto a low-altitude three-dimensional spatial grid to determine the spatiotemporal and meteorological similarity matrix between each trajectory feature vector. Based on the spatiotemporal and meteorological similarity matrix, density clustering algorithm is used to identify high-density connected regions in historical flight trajectory data; Historical flight trajectories within a high-density connected region are divided into the same trajectory cluster, and the centroid coordinates and average passage cost of each trajectory cluster are determined.

4. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, Obtain the set of target points to be inspected, determine the spatial matching degree between the set of target points and each trajectory cluster, and assign the set of target points to be inspected to the corresponding trajectory clusters based on the spatial matching degree, including: Obtain the set of three-dimensional coordinates of the target points to be inspected; Determine the Euclidean distance between each target point to be inspected and the centroid coordinates of each trajectory cluster; Based on Euclidean distance, spatial matching degree is determined by a monotonically decreasing mapping method; The target points to be inspected with a spatial matching degree not lower than the preset matching threshold are assigned to the trajectory cluster with the highest spatial matching degree, and the number of target points in each trajectory cluster is counted.

5. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, For each trajectory cluster assigned with a set of target points to be inspected, the historical travel cost benchmark within the trajectory cluster is obtained, and adaptive trajectory planning constraints are generated by combining real-time environmental state parameters, including: Historical average energy consumption and historical flight time are obtained from historical flight trajectory data within the trajectory cluster and used as historical travel cost benchmarks. Obtain real-time wind speed and direction from real-time environmental status parameters, and obtain real-time no-fly zone boundaries; Based on real-time wind speed and direction, the historical passage cost benchmark is corrected to generate the expected passage cost. By using the expected passage cost as a weight and mapping the real-time no-fly zone boundary as a spatial exclusion constraint, adaptive trajectory planning constraints are generated.

6. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, Based on adaptive trajectory planning constraints, a multi-objective path optimization algorithm is used to generate the optimal inspection route covering the target points to be inspected within each trajectory cluster, including: Construct a directed graph model with the target points to be inspected as nodes and the expected passage cost as edge weights; In directed graph models, spatial exclusion constraints are incorporated as a penalty term into the cost function of multi-objective path optimization algorithms; The multi-objective path optimization algorithm is used to search for the traversal path in the directed graph model that connects all the target points to be inspected within the trajectory cluster and has the minimum total cost. After smoothing the traversal path, the optimal inspection route is generated.

7. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, After generating the optimal inspection route covering the target points to be inspected within each trajectory cluster, the process also includes: Obtain the performance parameters of the UAV that executes the optimal inspection route, including maximum endurance and maximum payload; Determine the expected total travel cost of the optimal inspection route; When the expected total toll cost is not less than the energy consumption threshold corresponding to the maximum range, the route splitting strategy is triggered. Based on the centroid coordinates of the trajectory cluster, the optimal inspection route is divided into multiple sub-routes, and a collaborative inspection plan containing each sub-routes is generated.

8. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, After generating the optimal inspection route covering the target points to be inspected within each trajectory cluster, the process also includes: Acquire real-time flight status data collected by the UAV during the execution of the optimal inspection route; By comparing flight status data with historical flight trajectory data, the deviation between the actual passage cost and the expected passage cost is determined. When the deviation value is not less than the preset deviation threshold, the flight status data is updated into the historical flight trajectory data as new historical flight trajectory data, and the trajectory feature vector set is regenerated based on the updated historical flight trajectory data.

9. The low-altitude inspection route planning method based on trajectory clustering according to claim 1, characterized in that, Before acquiring historical flight trajectory data and low-altitude micro-meteorological field data within the target low-altitude area, the following steps are also included: Acquire three-dimensional geographic information data and building outline data of the target low-altitude area; Based on building outline data, safe flight corridors and no-fly zones are identified; The safe flight corridor and no-fly zone are mapped onto a low-altitude three-dimensional mesh model to generate a basic environmental constraint map; Historical flight trajectory data is filtered based on a basic environmental constraint map to remove abnormal trajectory data located within no-fly zones.

10. The low-altitude inspection route planning method based on trajectory clustering according to claim 3, characterized in that, Before applying spatiotemporal and meteorological coupled clustering algorithms to low-altitude flight path planning, the following were also included: Obtain a historical inspection sample set containing known optimal inspection paths; The historical inspection sample set is input into the spatiotemporal and meteorological coupled clustering algorithm to determine the overlap between the generated trajectory clusters and the known optimal inspection path; With the goal of maximizing the overlap, the spatiotemporal weight parameters in the spatiotemporal and meteorological similarity matrix are iteratively adjusted using the gradient descent algorithm to obtain the spatiotemporal and meteorological coupled clustering algorithm after parameter tuning.

Images