A method and system for unmanned aerial vehicle path planning oriented to positioning drift accumulation

Abstract

This invention discloses a method and system for UAV path planning based on location drift accumulation, relating to the field of UAV path planning technology. A UAV path planning system for location drift accumulation includes: a semantic topology mapping module, a drift benefit evaluation module, a drift accumulation prediction module, a constraint risk requirement module, a path constraint verification module, a path active correction module, and an inspection path determination module. This invention comprehensively evaluates path length cost, energy consumption cost, obstacle avoidance cost, location drift accumulation cost, task object binding risk cost, semantic misbinding risk cost, location correction benefit, semantic evidence collection benefit, and counterfactual exclusion benefit, resulting in a comprehensive and reasonable inspection path selection.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) path planning technology, and in particular to a UAV path planning method and system for addressing positioning drift accumulation. Background Technology

[0002] With their high mobility, flexible deployment, and wide coverage, drones have been widely used in scenarios such as power facility inspection, industrial park equipment inspection, warehouse asset inventory, industrial site inspection, and urban infrastructure operation and maintenance. In these applications, drones typically need to fly along pre-set paths, acquiring images, recognizing conditions, or recording data from multiple objects, and accurately mapping the inspection results to specific objects. Therefore, the accuracy of path planning capabilities and object-specific mapping directly impacts the execution efficiency and reliability of the drone inspection system.

[0003] Existing UAV path planning methods primarily optimize path length, flight energy consumption, obstacle avoidance safety, or coverage efficiency. Some solutions combine satellite positioning, visual odometry, lidar, or anchor point relocation to reduce positioning errors during flight. However, in environments with numerous repetitive structures and similar task objects, such as warehouse shelves, substation array equipment, pipe gallery equipment areas, and parking facilities, UAVs may still experience accumulated positioning drift during continuous flight, causing deviations between the estimated and actual positions. Furthermore, the high similarity in appearance between adjacent task objects can easily lead to incorrect binding of collected data to neighboring objects. Existing technologies typically separate path planning, positioning correction, and target identification, lacking a collaborative mechanism for reliable binding of task objects, making it difficult to balance flight efficiency and binding accuracy. Summary of the Invention

[0004] This invention proposes a UAV path planning method oriented towards location drift accumulation, which is used to solve the problem of misbinding of inspection results caused by location drift accumulation and object semantic confusion in UAVs in repetitive structure environments and environments with dense similar task objects. This allows the UAV to comprehensively consider location drift control, object identity differentiation and path cost evaluation during the path planning stage, and output an inspection path that achieves reliable binding of task objects.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for UAV path planning oriented towards positioning drift accumulation includes: Obtain an environmental map of the area to be inspected, and construct a semantic topology map containing task objects, positioning and correction anchor points, repetitive structural regions, unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information; The location drift growth coefficient of path units is calculated based on semantic topology maps, and the location correction benefit of location correction anchors is evaluated. Based on the initial positioning uncertainty, the positioning drift growth coefficient, and the positioning correction benefit, the cumulative positioning drift of the candidate path is predicted. The trusted binding threshold is determined based on the spatial size of the task object, the distance between adjacent task objects, or the inspection accuracy requirements. The semantic confusion risk is determined based on the semantic similarity information of the object. The counterfactual exclusion requirement is determined by combining the cumulative location drift amount and the counterfactual exclusion feature description information. Determine whether the candidate path meets the trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirements; If the trusted binding threshold constraint is not met, a positioning correction anchor point is inserted; if the semantic confusion risk constraint is not met, a semantic evidence collection path is generated; if the counterfactual exclusion requirement is not met, a counterfactual exclusion observation pose is inserted. Evaluate candidate paths and determine the inspection path that simultaneously meets the requirements of trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion.

[0006] As a preferred embodiment of the present invention, the semantic topology map includes: task object nodes, positioning and correction anchor point nodes, repeating structure region nodes, and path units connecting adjacent nodes; the task object nodes are associated with unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information; the unique identifier feature description information includes the unique identifier feature type, the spatial position of the unique identifier feature in the task object ontology coordinate system, and the capture perspective conditions of the unique identifier feature; the object semantic similarity information includes the task object identifier adjacent to the task object and the corresponding semantic similarity; the counterfactual exclusion feature description information includes the difference feature type, the difference feature spatial position, the difference feature capture perspective conditions, and the corresponding exclusion object range.

[0007] As a preferred embodiment of the present invention, the calculation of the positioning drift growth coefficient includes: for any path unit traversed by the candidate path, obtaining the path length of the path unit, the overlap length with the region corresponding to the node of the repeating structure region, the number of observable positioning correction anchor points, the number of observable unique identifier features, and the positioning signal quality value within the path unit; determining the ratio of the overlap length to the path length as the structure repeatability; determining visual localization based on the weighted normalization result of the number of observable positioning correction anchor points and the number of observable unique identifier features; determining signal reliability based on the normalization result of the positioning signal quality value relative to a preset signal quality benchmark value; and, based on the preset benchmark drift growth coefficient, increasing the preset benchmark drift growth coefficient according to the structure repeatability, and decreasing the preset benchmark drift growth coefficient according to the visual localization and signal reliability, to obtain the positioning drift growth coefficient of the path unit.

[0008] As a preferred technical solution of the present invention, the evaluation of the positioning correction benefit includes: for any positioning correction anchor point corresponding to any positioning correction anchor point node in the semantic topology map, obtaining the anchor point type, anchor point spatial location, number of observable features, observable distance range, observable viewing angle range, and semantic distinguishability with adjacent task objects or adjacent positioning correction anchor points; determining semantic uniqueness based on semantic distinguishability, determining visual stability based on the number of observable features, observable distance range, and observable viewing angle range, and determining spatial stability based on the fixed attributes of anchor point type and anchor point spatial location in the semantic topology map; and performing weighted normalization processing on semantic uniqueness, visual stability, and spatial stability to obtain the positioning correction benefit corresponding to the positioning correction anchor point.

[0009] As a preferred embodiment of the present invention, the prediction of the cumulative positioning drift includes: taking the initial positioning uncertainty of the UAV as the current positioning drift at the starting point of the candidate path, and updating the current positioning drift by increasing it according to the connection order of the path units in the candidate path, based on the path length and positioning drift growth coefficient of each path unit; when the candidate path passes through a positioning correction anchor point, updating the updated positioning drift by decreasing it according to the positioning correction benefit of the positioning correction anchor point; repeating the increase update and decrease update to obtain the cumulative positioning drift when the UAV reaches the task object along the candidate path.

[0010] As a preferred technical solution of the present invention, the determination of the trusted binding threshold includes: for any task object, obtaining the spatial size of the task object, the minimum distance between the task object and adjacent task objects, and the inspection accuracy requirement of the task object; determining a first threshold based on the spatial size and a preset size ratio, determining a second threshold based on the minimum distance and a preset distance ratio, and determining a third threshold based on the allowable positioning error corresponding to the inspection accuracy requirement; and determining the minimum value among the first threshold, the second threshold, and the third threshold as the trusted binding threshold corresponding to the task object.

[0011] As a preferred embodiment of the present invention, the determination of the counterfactual exclusion requirement includes: determining the positioning uncertainty range corresponding to the task object based on the cumulative positioning drift when the UAV arrives at the task object; determining a set of similar task objects to be excluded from adjacent task objects based on the positioning uncertainty range and object semantic similarity information; determining the difference features, difference feature capture perspective conditions, and exclusion object range between the task object and each similar task object in the set of similar task objects based on the counterfactual exclusion feature description information; determining the minimum exclusion evidence strength corresponding to each similar task object based on object semantic similarity information, cumulative positioning drift, and the exclusion object range corresponding to the difference features; and constituting the counterfactual exclusion requirement by the set of similar task objects, the difference feature capture perspective conditions, and the minimum exclusion evidence strength.

[0012] As a preferred technical solution of the present invention, the insertion of the counterfactual exclusion observation pose includes: when the candidate path does not meet the counterfactual exclusion requirements, parsing the counterfactual exclusion requirements to obtain the similar task objects to be excluded, the difference feature capture viewpoint conditions, and the minimum exclusion evidence strength; determining the candidate observation poses that meet the difference feature capture viewpoint conditions and are located within the reachable flight space of the UAV in the semantic topology map; selecting the counterfactual exclusion observation pose from the candidate observation poses based on the difference feature visibility, deviation distance from the candidate path, obstacle avoidance constraints, and the expected exclusion evidence strength, and inserting the counterfactual exclusion observation pose into the candidate path.

[0013] As a preferred technical solution of the present invention, the determination of the inspection path includes: a comprehensive cost evaluation of candidate paths after inserting positioning correction anchor points, semantic evidence collection paths, or counterfactual exclusion observation poses; the comprehensive cost evaluation includes determining the path length cost, energy consumption cost, obstacle avoidance cost, positioning drift accumulation cost, task object binding risk cost, semantic misbinding risk cost, positioning correction benefit, semantic evidence collection benefit, and counterfactual exclusion benefit, and weighting them according to preset weights to obtain the comprehensive cost value of the candidate paths; the semantic evidence collection benefit is determined based on the degree of satisfaction of the capture viewpoint conditions of the unique identifier feature and the visibility of the unique identifier feature; the counterfactual exclusion benefit is determined based on the degree of satisfaction of the counterfactual exclusion requirements and the expected strength of the excluded evidence; the candidate paths that meet the credible binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirements, and whose comprehensive cost value meets the preset path selection conditions, are determined as inspection paths.

[0014] A UAV path planning system for addressing positioning drift accumulation includes: Semantic topology mapping module: acquires the environmental map of the area to be inspected, and constructs a semantic topology map containing task objects, positioning and correction anchor points, repetitive structural regions, unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information; Drift benefit assessment module: Calculates the positioning drift growth coefficient of path units based on semantic topology map, and evaluates the positioning correction benefit of positioning correction anchor points; Drift accumulation prediction module: Based on the initial positioning uncertainty, positioning drift growth coefficient, and positioning correction benefit, predict the cumulative positioning drift of candidate paths; Constraint Risk Requirements Module: Determine the trusted binding threshold based on the spatial size of the task object, the distance between adjacent task objects, or the inspection accuracy requirements; determine the semantic confusion risk based on the semantic similarity information of the object; and determine the counterfactual exclusion requirements by combining the cumulative location drift amount and the counterfactual exclusion feature description information. Path constraint verification module: Determines whether candidate paths meet the requirements of trusted binding threshold constraints, semantic confusion risk constraints, and counterfactual exclusion; Path active correction module: If the trusted binding threshold constraint is not met, a positioning correction anchor point is inserted; if the semantic confusion risk constraint is not met, a semantic evidence collection path is generated; if the counterfactual exclusion requirement is not met, a counterfactual exclusion observation pose is inserted. Inspection path determination module: Evaluate candidate paths and determine inspection paths that simultaneously meet the requirements of trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion.

[0015] The present invention has the following advantages: This invention calculates the positioning drift growth coefficient of path units and combines it with the positioning correction benefit of positioning correction anchors to predict the cumulative positioning drift of candidate paths, enabling UAVs to pre-identify high-drift-risk paths during the path planning stage. By determining the credible binding threshold based on the spatial size of the task object, the distance between adjacent task objects, and the inspection accuracy requirements, different task objects are subject to differentiated binding accuracy constraints, thereby improving the adaptability of path planning.

[0016] This invention determines the semantic confusion risk based on the semantic similarity information of objects and generates counterfactual exclusion requirements by combining counterfactual exclusion feature description information, so that the path planning process actively considers the identity differentiation requirements between similar task objects; by inserting positioning correction anchor points, semantic evidence collection paths and counterfactual exclusion observation poses when candidate paths do not meet the constraints, the path planning results simultaneously take into account positioning correction, identity confirmation and path efficiency.

[0017] This invention comprehensively evaluates the costs of path length, energy consumption, obstacle avoidance, cumulative positioning drift, task object binding risk, semantic misbinding risk, positioning correction benefits, semantic evidence collection benefits, and counterfactual exclusion benefits, thereby making the inspection path selection comprehensive and reasonable. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort. Figure 1 This is a schematic diagram of the structure of a UAV path planning system based on positioning drift accumulation used in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] Example 1: A UAV path planning method for addressing positioning drift accumulation, comprising the following steps: Step S1: Obtain the environmental map of the area to be inspected, and construct a semantic topology map containing task objects, positioning and correction anchor points, repeating structural regions, unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information; The environmental map of the area to be inspected is basic map data representing the spatial structure, access areas, obstacle distribution, and positional relationships of target objects in the inspection area. It originates from historical survey maps, manually modeled maps, laser scan maps, visual mapping results, or existing digital operation and maintenance maps acquired before the UAV performs the inspection mission. Taking a substation inspection scenario as an example, the environmental map includes the spatial relationships of the main transformer area, switchgear area, capacitor area, road access area, fence area, and building area, as well as the distance relationships and access boundaries between various devices.

[0021] The semantic topology map is a task map formed by structurally labeling the object categories, spatial connections, and identification information in the inspection area based on the environmental map. In the semantic topology map, "semantic" refers to the fact that nodes or regions in the map not only have geometric location attributes but also object categories, task attributes, identification attributes, and risk attributes; "topology" refers to the connectivity, adjacency, reachability, and order relationships between nodes and regions.

[0022] The semantic topology map includes: task object nodes, positioning and correction anchor point nodes, repeating structure region nodes, and path units connecting adjacent nodes.

[0023] The task object node represents a specific object that needs to be inspected, photographed, identified, or its status recorded, and records the spatial location and task attributes of the object in the inspection area. Taking a substation scenario as an example, the task object node includes a circuit breaker numbered CB-01, a disconnector numbered DS-02, and a main transformer numbered TR-01.

[0024] The positioning correction anchor nodes are used to represent stable reference objects for reducing the cumulative positioning drift of the UAV. These anchor nodes are derived from objects in the inspection area that are fixed in location, have stable characteristics, are easily identifiable, and exhibit minimal long-term changes. Taking a substation scenario as an example, the corners of the control room, equipment area number plates, and fence corners are used as positioning correction anchor nodes.

[0025] The repeating structure region nodes are used to represent areas within a region that have multiple similar appearances, similar arrangement patterns, or repeated distributions of structural units. These repeating structure region nodes are used to mark areas where UAVs are prone to visual confusion and accumulated positioning drift during flight. Taking a substation scenario as an example, continuously arranged disconnector switch areas serve as repeating structure region nodes.

[0026] The path unit is used to represent a flyable path segment connecting adjacent nodes. Each path unit has a starting node, an ending node, a path length, a travel height range, and obstacle avoidance boundary information. Taking a substation scenario as an example, a straight path segment within a height range of 3 to 8 meters above the main passage is used as a path unit.

[0027] The task object node is associated with unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information.

[0028] The unique identifier feature description information is used to represent the specific identification features that distinguish a task object from other task objects. This information includes the unique identifier feature type, the spatial position of the unique identifier feature in the task object's coordinate system, and the capture perspective conditions for the unique identifier feature. The unique identifier feature type includes equipment nameplates, QR codes, barcodes, serial numbers, specific color markings, dedicated shape structures, or unique interface layouts. The task object's coordinate system is a local coordinate reference system established using the task object's own center point, installation reference plane, or structural reference edge, used to uniformly describe the feature's position on the object. The capture perspective conditions limit the observation direction, distance range, pitch angle range, and obstruction conditions required for the UAV to acquire the unique identifier feature. Taking a circuit breaker with the serial number CB-01 as an example, its nameplate is located in the lower right corner of the front of the device, and the capture perspective conditions are met when the shooting distance is between 2 and 5 meters and the yaw angle is within ±25 degrees.

[0029] The object semantic similarity information is used to represent the degree of similarity between a task object and its adjacent task objects in terms of appearance, structure, category, and spatial arrangement. The object semantic similarity information includes the identifiers of task objects adjacent to the task object and their corresponding semantic similarities. The semantic similarity is represented by a value between 0 and 1, with higher values ​​indicating greater similarity. Taking circuit breaker CB-01 and its adjacent circuit breaker CB-02 as an example, since they have the same appearance, the same installation method, and are close together, their corresponding semantic similarity is set to 0.92; however, since circuit breaker CB-01 and main transformer TR-01 have significantly different categories, their semantic similarity is set to 0.25.

[0030] The counterfactual exclusion feature description information is used to represent the basis for identifying the differences between a task object and similar task objects. "Counterfactual exclusion" refers to, during the path planning phase, not only confirming the characteristics of the target task object itself, but also pre-identifying which similar task objects need to be excluded, and using the difference features to complete the identity differentiation. The counterfactual exclusion feature description information includes the type of difference feature, the spatial location of the difference feature, the perspective conditions for capturing the difference feature, and the corresponding scope of the excluded objects.

[0031] The difference feature types include differences in nameplate position, color coding, handle direction, number of terminals, auxiliary components, or localized wear. The spatial location of the difference feature indicates its specific position on the task object. The difference feature capture perspective conditions indicate the observation direction and distance required to acquire the difference feature. The exclusion object range indicates the set of similar task objects excluded after acquiring the difference feature. Taking circuit breakers CB-01 and CB-02 as examples, CB-01's nameplate is located in the lower right corner, and CB-02's nameplate is located in the lower left corner. Therefore, "nameplate position difference" is considered a difference feature type, and the corresponding exclusion object range is the adjacent circuit breaker CB-02.

[0032] Step S2: Calculate the location drift growth coefficient of the path unit based on the semantic topology map, and evaluate the location correction benefit of the location correction anchor point; The positioning drift growth coefficient characterizes the rate at which positioning error accumulates with increasing flight distance or time when the UAV flies along a corresponding path unit. A larger positioning drift growth coefficient indicates that the UAV is more prone to accumulating positioning drift within that path unit due to environmental repetition, insufficient features, or weakened positioning signals; a smaller coefficient indicates that the UAV has better positioning constraints within that path unit. The positioning drift growth coefficient is calculated separately for each path unit and serves as the basis for subsequent prediction of accumulated positioning drift.

[0033] The calculation of the positioning drift growth coefficient includes: for any path unit traversed by the candidate path, obtaining the path length of the path unit, the overlap length with the region corresponding to the node of the repeating structure region, the number of observable positioning correction anchors, the number of observable unique identifiers, and the positioning signal quality value within the path unit; determining the ratio of the overlap length to the path length as the structure repeatability; determining visual localization based on the weighted normalization result of the number of observable positioning correction anchors and the number of observable unique identifiers; determining signal reliability based on the normalization result of the positioning signal quality value relative to a preset signal quality benchmark value; and, based on the preset benchmark drift growth coefficient, increasing the preset benchmark drift growth coefficient according to the structure repeatability, and decreasing the preset benchmark drift growth coefficient according to visual localization and signal reliability to obtain the positioning drift growth coefficient of the path unit.

[0034] The path length is the flight distance between the starting node and the ending node of the path unit, calculated directly from the node coordinates in the environmental map. Taking a substation inspection scenario as an example, the length of a straight path connecting the circuit breaker area and the disconnector area above the main passage is 18 meters. The overlap length with the corresponding area of ​​the repeating structure area node refers to the overlap length of the path unit with the repeating structure area after spatial projection, used to characterize the degree to which the UAV is continuously in a similar structural environment within the path unit. If 12 meters of the 18-meter path is located above a continuously arranged disconnector area, the overlap length is 12 meters, corresponding to a high degree of structural repetition.

[0035] The structural repeatability is used to represent the degree of interference of repetitive structures in the surrounding environment of a path unit on localization. When structural repeatability increases, environmental features used by visual odometry, feature matching, or map loop closure recognition are more prone to repeated matching, leading to a faster accumulation rate of localization errors. When calculating the localization drift growth coefficient, the preset baseline drift growth coefficient is correspondingly increased. The number of observable localization correction anchor points refers to the number of localization correction anchor points that the UAV can stably identify within a preset observation distance and field of view during its flight along the path unit. The number of observable unique identifier features refers to the number of unique identifier features of the task object that can be directly observed during its flight along the path unit. Both are used together to measure the visual localization of the path unit.

[0036] Visual localization refers to the ability of a UAV to constrain its own position through environmental visual features during flight. When there are many positioning correction anchor points, equipment nameplates, numbering markers, or structural corners around a path unit, visual localization is higher, correspondingly reducing the positioning drift growth coefficient. Taking a substation scenario as an example, if a path unit near the outer wall of the control room can observe wall corners, doorplates, and equipment numbering markers, then its visual localization is higher than in areas with dense equipment obstruction.

[0037] The positioning signal quality value indicates the stability of the positioning signal received by the UAV within the path unit. The positioning signal can be a satellite positioning signal, a ground-assisted base station positioning signal, an ultra-wideband positioning signal, a visual positioning-assisted signal, or a fused positioning signal. The positioning signal quality value is determined based on received signal strength, the number of available satellites, signal continuity, multipath interference, or positioning update frequency. Taking a substation scenario as an example, the satellite signal quality value in open areas is higher than that in areas with obstruction on the side of the control building. The signal reliability indicates the degree of support provided by external positioning sources within the path unit to the UAV's positioning results. Higher signal reliability results in a lower positioning drift growth coefficient; lower signal reliability results in a higher positioning drift growth coefficient.

[0038] The preset baseline drift growth coefficient is a reference drift growth level for the UAV under standard open environment conditions, normal feature distribution, and stable positioning signal. It is preset based on historical flight data, calibration test flight data, or aircraft parameters. For example, when the same model of UAV performs a station area inspection mission, the average drift growth level of the unobstructed area is used as the preset baseline drift growth coefficient. The positioning correction benefit is used to characterize the ability of the UAV to reduce the current positioning drift after passing a positioning correction anchor point and completing identification during flight. The higher the positioning correction benefit value, the stronger the positioning error correction effect of the positioning correction anchor point.

[0039] The evaluation of the positioning correction benefit includes: for any positioning correction anchor point in the semantic topology map, obtaining the anchor point type, anchor point spatial location, number of observable features, observable distance range, observable viewing angle range, and semantic distinguishability with adjacent task objects or adjacent positioning correction anchor points; determining semantic uniqueness based on semantic distinguishability, visual stability based on the number of observable features, observable distance range, and observable viewing angle range, and spatial stability based on the fixed attributes of anchor point type and anchor point spatial location in the semantic topology map; and performing weighted normalization on semantic uniqueness, visual stability, and spatial stability to obtain the positioning correction benefit corresponding to the positioning correction anchor point.

[0040] The anchor point type is used to distinguish the object category of the positioning and correction anchor point, such as building corners, fixed signs, fence corners, permanent QR code boards, or equipment foundation edges. Different anchor point types have different recognition stability and long-term retention. The anchor point spatial location is used to indicate the installation position of the positioning and correction anchor point in the inspection area and its surrounding occlusion relationship. For example, an anchor point located in the visible area of ​​the center of a passage has a higher correction benefit than an anchor point located in the occluded area behind equipment. The semantic distinguishability is used to indicate the degree of recognition difference between the positioning and correction anchor point and surrounding objects. When the anchor point has obvious appearance features and no similar objects around it, the semantic distinguishability is high. Taking the control room doorplate as an example, because the number is unique and the position is fixed, the semantic distinguishability is high. The visual stability is used to indicate the ability of the UAV to stably extract its recognition features when observing the anchor point from different distances and angles. Building corners, QR code signs, and large numbered signs usually have high visual stability. The spatial stability is used to indicate whether the position of the anchor point remains stable during long-term operation. Fixed building structures, permanent columns, and welded installation equipment possess high spatial stability; temporary storage or movable equipment are not considered high-level positioning correction anchor points. In actual evaluation, semantic uniqueness, visual stability, and spatial stability are uniformly quantified to form a positioning correction benefit level. Taking a substation inspection scenario as an example, control room exterior wall corners with clear outlines, no obstructions, and long-term fixation yield higher positioning correction benefits; ordinary fence posts with high visual repetition yield lower positioning correction benefits.

[0041] Step S3: Based on the initial positioning uncertainty, positioning drift growth coefficient, and positioning correction benefit, predict the cumulative positioning drift of the candidate path; The initial positioning uncertainty represents the positioning error state of the UAV at the starting point of the candidate path. This initial positioning uncertainty originates from the UAV's pre-takeoff positioning state, the accuracy of the positioning equipment at the start of the mission, the initial map matching result, the visual positioning initialization result, or positioning error records from historical inspection paths. Taking a substation inspection scenario as an example, when the UAV enters the main channel from a fixed takeoff point, the initial positioning uncertainty at the starting point of the candidate path is obtained based on the accuracy of the takeoff point coordinate calibration, the satellite positioning state, and the pre-takeoff visual calibration result. The cumulative positioning drift represents the cumulative positioning deviation of the UAV relative to its actual position after flying along the candidate path and reaching a specific task object.

[0042] The prediction of the cumulative positioning drift includes: taking the initial positioning uncertainty of the UAV as the current positioning drift at the starting point of the candidate path, and updating the current positioning drift by increasing it according to the connection order of the path units in the candidate path, based on the path length and positioning drift growth coefficient of each path unit; when the candidate path passes through a positioning correction anchor point, updating the updated positioning drift by decreasing it according to the positioning correction benefit of that anchor point; repeating the increase and decrease updates to obtain the cumulative positioning drift when the UAV reaches the task object along the candidate path.

[0043] Specifically, the candidate path is formed by connecting multiple path units in flight sequence. Each path unit corresponds to a path length and a positioning drift growth coefficient. After the UAV departs from the starting point of the candidate path, the initial positioning uncertainty is used as the current positioning drift. When the UAV flies along the first path unit, the current positioning drift is increased and updated according to the path length and positioning drift growth coefficient of that path unit. When the UAV continues to enter the next path unit, the positioning drift updated after the previous path segment is used as the new current positioning drift, and it is increased and updated again.

[0044] The increased update is used to characterize the accumulation of errors caused by visual feature repetition, decreased positioning signal, increased path length, or environmental occlusion during UAV flight. The longer the path unit, the higher the positioning drift growth coefficient, and the more significant the effect of that path unit on increasing the current positioning drift. Taking a substation inspection scenario as an example, the candidate path passes sequentially from the takeoff point through the main channel, the continuous disconnector switch area, and the area where circuit breaker CB-01 is located. The path unit corresponding to the continuous disconnector switch area has a high degree of structural repetition, so its effect on increasing the current positioning drift is higher than that of the path unit in the open main channel.

[0045] When a candidate path passes a positioning correction anchor point, the updated positioning drift is reduced based on the positioning correction benefit of that anchor point. This reduction update characterizes how the UAV, after identifying the anchor point, corrects the current positioning state using the anchor point's known spatial location and stable identification features. Higher positioning correction benefits result in greater reductions in the current positioning drift; lower benefits result in smaller reductions.

[0046] Taking a substation inspection scenario as an example, when a drone flies along a candidate path to the vicinity of a corner of the control room's outer wall, if this corner is marked as a positioning correction anchor point and possesses high semantic uniqueness, visual stability, and spatial fixity, the drone will reduce and update its current positioning drift after identifying the anchor point. If the candidate path only passes through ordinary fence posts, the positioning correction benefit is lower due to the high visual repetition and low semantic distinguishability of these posts, resulting in a smaller reduction in the current positioning drift.

[0047] During the prediction process, if a candidate path does not pass through a positioning correction anchor point, an increase update is performed only according to the path unit sequence; if a candidate path passes through one or more positioning correction anchor points, a decrease update is performed once or multiple times at the corresponding location. For the same candidate path, the cumulative positioning drift when the UAV reaches different task objects is calculated separately. For example, the first cumulative positioning drift is obtained when the UAV reaches the circuit breaker CB-01, and the second cumulative positioning drift is obtained when it reaches the disconnector DS-02. These two are used for subsequent trusted binding judgments of the corresponding task objects.

[0048] The cumulative location drift is also related to the arrival order of the task object in the candidate path. If the target task object is located in the early part of the candidate path and has traversed fewer path units, its cumulative location drift is lower; if the target task object is located in the later part of the candidate path and has traversed more path units, and lacks location correction anchors in between, its cumulative location drift is higher. Therefore, in the same semantic topology map, different candidate paths correspond to different arrival orders of task objects, resulting in different cumulative location drift.

[0049] To maintain the prediction results in line with semantic topology Figure 1 In this system, the prediction of the cumulative positioning drift is updated using path units as the smallest update unit and positioning correction anchors as the trigger points for reduction updates. That is, each time a candidate path passes through a path unit, the current positioning drift is updated based on the positioning drift growth coefficient of that path unit; each time a candidate path passes through a positioning correction anchor, the current positioning drift is updated based on the positioning correction gain of that anchor. Thus, the cumulative positioning drift corresponds to the specific spatial regions traversed by the candidate path, the distribution of repeating structure regions, and the distribution of positioning correction anchors.

[0050] In a specific implementation process, the candidate path sequentially passes through the main channel path unit, the continuous disconnector switch area path unit, the control room exterior wall corner positioning correction anchor point, and the path unit in front of circuit breaker CB-01 from the UAV takeoff point. First, the initial positioning uncertainty at the takeoff point is taken as the current positioning drift. After passing through the main channel path unit, the current positioning drift increases slightly due to the shorter path and higher visual positioning accuracy. After passing through the continuous disconnector switch area path unit, the current positioning drift increases further due to the high structural repeatability. Upon reaching the control room exterior wall corner positioning correction anchor point, the current positioning drift is reduced based on the positioning correction benefit of this anchor point. Then, the UAV continues to fly along the path unit in front of circuit breaker CB-01, and the corresponding cumulative positioning drift is obtained upon reaching circuit breaker CB-01.

[0051] Step S4: Determine the trusted binding threshold based on the spatial size of the task object, the distance between adjacent task objects, or the inspection accuracy requirements; determine the semantic confusion risk based on the semantic similarity information of the object; and determine the counterfactual exclusion requirements by combining the cumulative location drift and the counterfactual exclusion feature description information. The trusted binding threshold represents the maximum positioning drift range within which inspection results are reliably bound to a task object when the UAV arrives at that task object. When the cumulative positioning drift upon arrival at the task object is not greater than the trusted binding threshold, the current positioning accuracy meets the task object binding requirements; when the cumulative positioning drift upon arrival at the task object is greater than the trusted binding threshold, the current positioning accuracy is insufficient to support accurate binding of inspection results to the task object.

[0052] The determination of the trusted binding threshold includes: for any task object, obtaining the spatial size of the task object, the minimum distance between the task object and adjacent task objects, and the inspection accuracy requirement of the task object; determining a first threshold based on the spatial size and a preset size ratio, determining a second threshold based on the minimum distance and a preset distance ratio, and determining a third threshold based on the allowable positioning error corresponding to the inspection accuracy requirement; and determining the minimum value among the first threshold, the second threshold, and the third threshold as the trusted binding threshold corresponding to the task object.

[0053] The spatial dimensions of the task object represent its geometric scale, including length, width, height, circumscribed rectangle size, main body diameter, or projected area. Smaller spatial dimensions increase the risk of the inspection result being bound to an incorrect object, corresponding to a lower trusted binding threshold. For example, in a substation inspection scenario, the main body size of circuit breaker CB-01 is larger than the nameplate area size; therefore, different trusted binding thresholds apply to overall equipment inspection tasks and nameplate recognition tasks.

[0054] The spacing between adjacent task objects represents the spatial distance between the current task object and its nearest neighboring task object, expressed as center-to-center distance, shortest edge distance, or mounting base spacing. The smaller the spacing between adjacent task objects, the higher the probability of misbinding to an adjacent object after positioning drift, and the lower the corresponding trusted binding threshold. Taking a continuously arranged disconnector switch area as an example, the installation spacing between adjacent disconnectors is small, and its trusted binding threshold is lower than that of an independently arranged main transformer area.

[0055] The inspection accuracy requirement refers to the spatial positioning accuracy standard required for a specific inspection task. If the task involves inspecting the appearance of the equipment, a larger positioning error range is allowed; if the task involves nameplate recognition, QR code reading, or local defect photography, a smaller positioning error range is allowed. Taking circuit breaker CB-01 as an example, if the task of photographing the overall status of the equipment is performed, the inspection accuracy requirement is lower than that of performing the task of recognizing nameplate characters.

[0056] The first threshold reflects the limitation of the task object's own size on the tolerance for binding errors; the second threshold reflects the limitation of the distance between adjacent task objects on the risk of misbinding; and the third threshold reflects the limitation of the task accuracy requirements on the positioning error. The minimum value among the three is used as the reliable binding threshold, so that the final threshold simultaneously satisfies the object size, proximity relationship, and task requirements.

[0057] The semantic confusion risk refers to the degree of risk that, after the UAV reaches the task object, even if the positioning error is within the allowable range, the target object may still be confused or confused due to its high similarity to adjacent objects in appearance, structure, or category. The semantic confusion risk is derived from the object semantic similarity information in step S1 and determined in combination with the adjacency relationship of the task object.

[0058] Specifically, if a task object is surrounded by multiple adjacent task objects that are similar in appearance, category, and arrangement, the semantic confusion risk of that task object is high. Conversely, if a task object is of an independent category, has distinct appearance features, and has no similar objects nearby, the semantic confusion risk of that task object is low. Taking a substation inspection scenario as an example, the semantic confusion risk between consecutively arranged circuit breakers CB-01 and CB-02 is higher than the semantic confusion risk between the main transformer TR-01 and the fence post.

[0059] In practice, semantic confusion risk is categorized based on the number of adjacent task objects, the semantic similarity between adjacent task objects, and their spatial proximity. For example, if there are two or more adjacent task objects with a semantic similarity higher than a preset similarity threshold, the task object is marked as a high-risk semantic confusion task object.

[0060] The counterfactual exclusion requirement indicates that, under the current location drift conditions, additional differential observation evidence needs to be collected to prevent inspection results from being incorrectly bound to similar task objects. "Counterfactual exclusion" means not only confirming that the current task object possesses the target characteristic, but also proving that adjacent similar task objects do not meet that target characteristic, thereby completing the object identity exclusion.

[0061] The determination of the counterfactual exclusion requirement includes: determining the positioning uncertainty range corresponding to the task object based on the cumulative positioning drift when the UAV arrives at the task object; determining the set of similar task objects to be excluded from adjacent task objects based on the positioning uncertainty range and object semantic similarity information; determining the difference features, difference feature capture perspective conditions, and exclusion object range between the task object and each similar task object in the set of similar task objects based on the counterfactual exclusion feature description information; determining the minimum exclusion evidence strength corresponding to each similar task object based on object semantic similarity information, cumulative positioning drift, and the exclusion object range corresponding to the difference features; and the counterfactual exclusion requirement is constituted by the set of similar task objects, the difference feature capture perspective conditions, and the minimum exclusion evidence strength.

[0062] The location uncertainty range refers to the area where the actual location of the UAV might be estimated based on the current cumulative location drift. A larger cumulative location drift results in a larger location uncertainty range; a smaller cumulative location drift results in a smaller location uncertainty range. If this location uncertainty range covers adjacent task objects, it indicates a possible mis-binding, requiring further determination of counterfactual exclusion requirements.

[0063] The set of similar task objects represents a set of task objects that are within the uncertain location range and have a high semantic similarity to the current task object. Taking circuit breaker CB-01 as an example, if the uncertain location range covers both CB-01 and CB-02, and the two have a high semantic similarity, then CB-02 is included in the set of similar task objects.

[0064] The distinguishing features are used to indicate characteristics that differentiate the current task object from similar task objects. These distinguishing features include differences in nameplate position, color markings, the number of terminals, handle orientation, auxiliary components, or localized wear. For example, CB-01's nameplate is located in the lower right corner, while CB-02's nameplate is located in the lower left corner; therefore, the difference in nameplate position is a distinguishing feature.

[0065] The difference feature capture viewpoint conditions are used to represent the flight observation direction, shooting distance, pitch angle, and unobstructed conditions required to acquire the difference features. For example, to observe the nameplate in the lower right corner of CB-01, the drone needs to be positioned slightly to the right of the device in front of it and maintain a preset shooting distance.

[0066] The minimum strength of exclusion evidence is used to represent the minimum level of evidence required to exclude an object. Higher semantic similarity, greater cumulative location drift, and closer adjacent objects correspond to a higher minimum strength of exclusion evidence; conversely, a lower minimum strength of exclusion evidence corresponds to a lower minimum strength of exclusion evidence. The minimum strength of exclusion evidence is categorized into low, medium, and high levels, or by numerical range.

[0067] Taking a substation inspection scenario as an example, when the drone arrives in front of circuit breaker CB-01, if the predicted cumulative positioning drift is close to the credible binding threshold, and CB-02 is located in an adjacent position and has a highly similar appearance, then firstly, the positioning uncertainty range of the current position is determined to cover CB-01 and CB-02; then CB-02 is included in the set of similar task objects; then, the difference features and corresponding capture perspective conditions are determined according to the "different nameplate positions"; finally, based on the high semantic similarity and small installation spacing between the two, the minimum exclusion evidence strength is determined to be high level, thus forming a counterfactual exclusion requirement for CB-01.

[0068] Step S5: Determine whether the candidate path meets the trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirements; The candidate path refers to a flight path to be evaluated, pre-generated based on the environmental map of the area to be inspected, the distribution of task objects, and reachable path units. The candidate path is either an initial inspection path generated according to the order of the task objects, or an alternative path after local adjustments. Each candidate path corresponds to several task object arrival nodes, as well as the cumulative positioning drift of the UAV when it arrives at each task object.

[0069] The requirement of satisfying the trusted binding threshold constraint means that when the UAV arrives at any task object along the candidate path, the corresponding cumulative positioning drift is not greater than the trusted binding threshold of that task object. If the cumulative positioning drift corresponding to a task object is greater than the trusted binding threshold, it means that although the UAV has arrived near the task object, its current position accuracy is insufficient to support the accurate binding of the inspection result to the task object, and the candidate path does not satisfy the trusted binding threshold constraint.

[0070] Taking a substation inspection scenario as an example, if the cumulative positioning drift of the drone when it reaches circuit breaker CB-01 along the candidate path exceeds the trusted binding threshold corresponding to CB-01, then the collected image data is at risk of being incorrectly bound to the adjacent circuit breaker CB-02. In this case, the candidate path does not meet the trusted binding threshold constraint.

[0071] The requirement to satisfy the semantic confusion risk constraint means that, under the observation conditions after the candidate path reaches the task object, the semantic confusion risk corresponding to the task object is not higher than a preset risk level or a preset risk threshold. If there are multiple highly similar objects around the current task object, and the current observation path does not provide sufficient distinguishing information, then the candidate path does not satisfy the semantic confusion risk constraint.

[0072] Taking a region of continuously arranged circuit breakers as an example, if a drone observes circuit breaker CB-01 from a distance directly in front along a candidate path, and the adjacent circuit breaker CB-02 has the same appearance as CB-01 and a small distance between them, then the two still have a high probability of confusion under the current path, and the candidate path does not meet the semantic confusion risk constraint.

[0073] Meeting the counterfactual exclusion requirement means that the candidate path already contains sufficient observation conditions to exclude similar task objects, or already has path segments and observation poses that meet the counterfactual exclusion requirement. If the candidate path does not pass through a location where the difference characteristics can be observed, or if the expected strength of evidence collected is lower than the minimum strength of exclusion evidence, then the candidate path does not meet the counterfactual exclusion requirement.

[0074] Taking circuit breakers CB-01 and CB-02 as examples, if the difference in the nameplate position used to distinguish the two is located at the lower right corner of the equipment, and the candidate path only passes through the high-altitude position directly in front of the equipment, it cannot meet the difference feature capture perspective condition, then the candidate path does not meet the counterfactual exclusion requirement.

[0075] In the specific judgment process, constraints are verified one by one for each task object. For each task object in the candidate path, its trusted binding threshold, semantic confusion risk, and counterfactual exclusion requirements are read sequentially, and the judgment is made in combination with the cumulative positioning drift when the UAV reaches the task object along the candidate path and the current path observation conditions. If any task object in the candidate path fails any constraint, the candidate path proceeds to step S6 to perform path correction.

[0076] When all task objects in the candidate path meet the trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirement, the candidate path is retained as a feasible candidate path and proceeds to step S7 for comprehensive cost evaluation.

[0077] Step S6: If the trusted binding threshold constraint is not met, insert a positioning correction anchor point; if the semantic confusion risk constraint is not met, generate a semantic evidence collection path; if the counterfactual exclusion requirement is not met, insert a counterfactual exclusion observation pose. When a candidate path does not meet the trusted binding threshold constraint, it indicates that the cumulative positioning drift of the UAV before reaching the target task object is too large, and it is necessary to reduce the cumulative positioning error before reaching the task object. In this case, a positioning correction anchor point is inserted into the candidate path.

[0078] The insertion of positioning correction anchors refers to adding a path segment passing through one or more positioning correction anchors before the original candidate path reaches the target mission object. This allows the UAV to correct its current position using the positioning correction anchors before reaching the mission object. The inserted positioning correction anchors are preferentially selected based on their high positioning correction benefit, small deviation from the original path, and flight reachability.

[0079] Taking a substation inspection scenario as an example, if the cumulative positioning drift of the drone exceeds the limit before it reaches CB-01 along the original candidate path, a positioning correction anchor point at the corner of the control room's outer wall is inserted into the original path so that the drone can first pass through the positioning correction anchor point to complete the positioning correction before heading to CB-01.

[0080] When a candidate path does not meet the semantic confusion risk constraint, it indicates that although the positioning accuracy meets the requirements, the current observation path cannot adequately distinguish between the target task object and similar task objects. In this case, a semantic evidence collection path is generated.

[0081] The semantic evidence acquisition path refers to a supplementary path that guides the UAV to a location that meets the viewing angle conditions for capturing unique identifier features, in order to collect evidence of the specificity of the task object. The unique identifier features include nameplates, serial numbers, QR codes, color markings, or proprietary structural features.

[0082] Taking CB-01 as an example, if the current path can only observe the overall outline of the device and cannot read the nameplate number, then an additional path segment close to the right side of the front of the device is added to the candidate path so that the drone can reach a position where the nameplate is clearly visible and complete the semantic evidence collection.

[0083] When a candidate path does not meet the counterfactual exclusion requirement, it means that although the current path has certain identification conditions, it has not yet obtained sufficient evidence of difference to exclude similar task objects. At this time, a counterfactual exclusion observation pose is inserted.

[0084] The insertion of the counterfactual exclusion observation pose includes: when the candidate path does not meet the counterfactual exclusion requirements, parsing the counterfactual exclusion requirements to obtain the similar task objects to be excluded, the difference feature capture viewpoint conditions, and the minimum exclusion evidence strength; determining the candidate observation poses in the semantic topology map that meet the difference feature capture viewpoint conditions and are located within the reachable flight space of the UAV; selecting the counterfactual exclusion observation pose from the candidate observation poses based on the difference feature visibility, deviation distance from the candidate path, obstacle avoidance constraints, and the expected exclusion evidence strength, and inserting the counterfactual exclusion observation pose into the candidate path.

[0085] The counterfactual exclusion observation pose refers to a combination of spatial flight positions and attitudes that simultaneously meet the requirements for observing differences in features and for path feasibility. The attitude includes the UAV's heading angle, pitch angle, and shooting direction. This counterfactual exclusion observation pose differs from ordinary observation points; its goal is to generate evidence excluding similar task objects.

[0086] Taking CB-01 and CB-02 as examples, if the only difference between the two is the position of the nameplate, then the counterfactual exclusion observation pose is set to the lower right angle in front of the equipment, so that the UAV can observe the nameplate in the lower right corner of CB-01, and at the same time confirm that CB-02 does not have this position feature.

[0087] In practice, if a candidate path does not meet two or three constraints simultaneously, a combined correction is performed on the candidate path. The combined correction is performed sequentially according to the priority of the trusted binding threshold constraint, the semantic confusion risk constraint, and the counterfactual exclusion requirement, or by merging path actions within the same region for joint correction.

[0088] For example, if the cumulative positioning drift exceeds the limit before the UAV reaches CB-01 and it is impossible to distinguish between CB-01 and CB-02, then first insert the positioning correction anchor point at the corner of the control room's outer wall, and then insert the semantic evidence collection path or counterfactual exclusion observation pose in front of CB-01, so that positioning correction and identity differentiation are completed simultaneously.

[0089] Step S7: Evaluate candidate paths and determine the inspection path that simultaneously meets the trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirements.

[0090] The determination of the inspection path includes: a comprehensive cost evaluation of candidate paths after inserting positioning correction anchor points, semantic evidence collection paths, or counterfactual exclusion of observation poses; the comprehensive cost evaluation includes determining the path length cost, energy consumption cost, obstacle avoidance cost, positioning drift accumulation cost, task object binding risk cost, semantic misbinding risk cost, positioning correction benefit, semantic evidence collection benefit, and counterfactual exclusion benefit, and weighting them according to preset weights to obtain the comprehensive cost value of the candidate path; the semantic evidence collection benefit is determined based on the degree to which the capture viewpoint conditions of the unique identifier feature are met and the visibility of the unique identifier feature; the counterfactual exclusion benefit is determined based on the degree to which the counterfactual exclusion requirements are met and the expected strength of the excluded evidence; the candidate path that meets the credible binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirements, and whose comprehensive cost value meets the preset path selection conditions, is determined as the inspection path.

[0091] The comprehensive cost assessment is used to transform the advantages and disadvantages of different paths into a unified comparison standard. For multiple candidate paths, their comprehensive cost value is calculated separately, and the candidate paths are ranked according to their comprehensive cost value. The candidate path with the highest priority is selected as the inspection path.

[0092] The path length cost is used to characterize the execution cost corresponding to the total flight distance of the candidate path. The longer the candidate path, the longer the flight time, the lower the task execution efficiency, and the higher the corresponding path length cost. The path length cost is obtained by accumulating the lengths of each path unit. Taking a substation inspection scenario as an example, a candidate path that bypasses multiple positioning correction anchor points has a higher path length cost than a candidate path that directly traverses the main passage.

[0093] The energy consumption cost is used to characterize the power consumption cost of the UAV when executing candidate paths. In addition to being affected by path length, the energy consumption cost is also related to the number of accelerations and decelerations, the number of hovering operations, changes in climb and descent altitudes, and the distribution of headwind areas. If a candidate path contains multiple stopping points for semantic evidence collection, the hovering action increases, corresponding to a rise in energy consumption cost.

[0094] The obstacle avoidance cost is used to characterize the safe avoidance cost of a candidate path traversing a complex environment. The obstacle avoidance cost is determined based on the obstacle density near the path unit, safety margin, turning frequency, and the proportion of low-altitude narrow areas. Taking a path in a densely populated equipment area as an example, if a candidate path needs to traverse a narrow passage between a circuit breaker bracket and a fence, its obstacle avoidance cost is higher than that of a path through an open passage.

[0095] The cumulative positioning drift cost is used to characterize the risk level of overall positioning stability during the execution of the candidate path. The larger the cumulative positioning drift amount corresponding to each task object along the candidate path, the higher the cumulative positioning drift cost. If there are many highly repetitive path units in the candidate path and few positioning correction anchor points, the cumulative positioning drift cost will increase.

[0096] The task object binding risk cost is used to characterize the risk level of inspection results being bound to incorrect task objects under candidate paths. This cost is related to the remaining margin of the trusted binding threshold for task objects, the distance between task objects, and the cumulative location drift. If the cumulative location drift is close to the trusted binding threshold, the task object binding risk cost is high.

[0097] The semantic misbinding risk cost is used to characterize the risk of identity confusion between the target task object and similar task objects due to similar appearance, close arrangement, or limited viewpoint. If the candidate path only provides long-distance frontal view observation and there are multiple highly similar objects around the target task object, the semantic misbinding risk cost is high.

[0098] The positioning correction benefit is used to characterize the contribution of positioning correction anchors in candidate paths to reducing the cumulative positioning drift. Candidate paths with a large number of positioning correction anchors, high benefit levels, and reasonable distribution have higher positioning correction benefits. In the overall cost assessment, the positioning correction benefit is used as a favorable factor to offset part of the cost burden.

[0099] The semantic evidence acquisition benefit is determined based on the degree to which the capture viewpoint conditions of the unique identifier feature are met and the visibility of the unique identifier feature. Specifically, if the candidate path enables the UAV to reach the optimal observation position of the unique identifier feature, and there is no obstruction and the resolution meets the recognition requirements, the semantic evidence acquisition benefit is high; if only the observation angle is partially met or there is obstruction, the semantic evidence acquisition benefit is low.

[0100] Taking the CB-01 circuit breaker as an example, if a candidate path includes a close-range observation point slightly to the right of the device in front of it, and the nameplate number can be clearly read, then the semantic evidence collection benefit of this candidate path is higher than that of a candidate path that only passes through a long-distance path above the device.

[0101] The counterfactual exclusion benefit is determined based on the degree to which the counterfactual exclusion requirement is met and the expected strength of the exclusion evidence. Specifically, if the candidate path covers the observation poses of key difference features used to exclude similar task objects, and the expected strength of the collected evidence reaches or exceeds the minimum strength of exclusion evidence, the counterfactual exclusion benefit is high; if only some difference features can be observed, the counterfactual exclusion benefit is low.

[0102] Taking CB-01 and CB-02 as examples, if the candidate path includes a position slightly to the lower right of the front of the device, and the nameplate in the lower right corner of CB-01 is clearly observed and it is confirmed that CB-02 does not have the same positional characteristics, then the counterfactual exclusion benefit of this candidate path is higher.

[0103] The preset weights are used to reflect the importance of different evaluation indicators in the current task. For tasks with limited battery life, the weight of energy consumption cost is increased; for high-precision equipment inspection tasks, the weights of positioning drift accumulation cost, task object binding risk cost, and semantic misbinding risk cost are increased; for tasks in areas with dense similar equipment, the weight of counterfactual exclusion benefit is increased.

[0104] In the specific implementation process, multiple candidate paths that meet the constraints are generated simultaneously. For example: the first candidate path is the shortest, but has fewer positioning correction anchor points and a higher cumulative cost of positioning drift; the second candidate path adds a positioning correction anchor point at the corner of the control room's outer wall, making the path slightly longer, but with higher positioning correction benefits; the third candidate path adds the counterfactual exclusion observation pose in front of CB-01 on the basis of the second candidate path, making the path the longest, but with the highest semantic evidence collection and counterfactual exclusion benefits.

[0105] After a comprehensive cost evaluation of the above candidate paths, if the second candidate path achieves a better balance between path length cost, positioning stability, and task object binding risk, then the second candidate path will be determined as the inspection path; if the third candidate path is longer but has significant advantages in high-risk misbinding scenarios, then the third candidate path will be determined as the inspection path.

[0106] The candidate path's comprehensive cost value meeting the preset path selection criteria means that the candidate path's comprehensive cost value is at the optimal level among all candidate paths, or is lower than the preset cost threshold, or ranks first among multiple candidate paths. If multiple candidate paths have the same comprehensive cost value, a secondary selection is performed based on the principles of shorter path length, shorter total execution time, fewer positioning and correction anchor points, or better task object access order.

[0107] Example 2, a UAV path planning system for addressing positioning drift accumulation, see [link to example]. Figure 1 As shown, it includes the following modules: The semantic topology mapping module is used to acquire an environmental map of the area to be inspected and construct a semantic topology map that includes task objects, positioning and correction anchor points, repeating structural regions, unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information.

[0108] The drift benefit assessment module, connected to the semantic topology mapping module, is used to calculate the positioning drift growth coefficient of the path unit based on the semantic topology map and to assess the positioning correction benefit of the positioning correction anchor point.

[0109] The drift accumulation prediction module is connected to the drift benefit evaluation module and the initial positioning uncertainty data source, respectively, and is used to predict the positioning drift accumulation of the candidate path based on the initial positioning uncertainty, the positioning drift growth coefficient and the positioning correction benefit.

[0110] The constraint risk requirement module is connected to the drift accumulation prediction module and the task object attribute data source. It is used to determine the reliable binding threshold based on the spatial size of the task object, the distance between adjacent task objects or the inspection accuracy requirements, determine the semantic confusion risk based on the semantic similarity information of the object, and determine the counterfactual exclusion requirement by combining the positioning drift accumulation amount and the counterfactual exclusion feature description information.

[0111] The path constraint verification module is connected to the constraint risk requirement module and the candidate path data source, respectively, and is used to determine whether the candidate path simultaneously meets the trusted binding threshold constraint, semantic confusion risk constraint and counterfactual exclusion requirement.

[0112] The path active correction module, connected to the path constraint verification module, is used to perform correction operations when the candidate path does not meet the corresponding constraints: if the trusted binding threshold constraint is not met, a positioning correction anchor point is inserted; if the semantic confusion risk constraint is not met, a semantic evidence collection path is generated; if the counterfactual exclusion requirement is not met, a counterfactual exclusion observation pose is inserted, and the corrected path is re-sent to the path constraint verification module for verification.

[0113] The inspection path determination module, connected to the path constraint verification module, is used to comprehensively evaluate path length, energy consumption, obstacle avoidance, positioning drift, binding risk, semantic misbinding risk, positioning correction benefit, semantic evidence collection benefit, and counterfactual exclusion benefit when the candidate path simultaneously meets the three constraints, and determine the final inspection path.

[0114] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for unmanned aerial vehicle path planning for positioning drift accumulation, characterized in that, include: Obtain an environmental map of the area to be inspected, and construct a semantic topology map containing task objects, positioning and correction anchor points, repetitive structural regions, unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information; The location drift growth coefficient of path units is calculated based on semantic topology maps, and the location correction benefit of location correction anchors is evaluated. Based on the initial positioning uncertainty, the positioning drift growth coefficient, and the positioning correction benefit, the cumulative positioning drift of the candidate path is predicted. The trusted binding threshold is determined based on the spatial size of the task object, the distance between adjacent task objects, or the inspection accuracy requirements. The semantic confusion risk is determined based on the semantic similarity information of the object. The counterfactual exclusion requirement is determined by combining the cumulative location drift amount and the counterfactual exclusion feature description information. Determine whether the candidate path meets the trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirements; If the trusted binding threshold constraint is not met, a positioning correction anchor point is inserted; if the semantic confusion risk constraint is not met, a semantic evidence collection path is generated; if the counterfactual exclusion requirement is not met, a counterfactual exclusion observation pose is inserted. Evaluate candidate paths and determine the inspection path that simultaneously meets the requirements of trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion.

2. The UAV path planning method based on positioning drift accumulation according to claim 1, characterized in that, The semantic topology map includes: task object nodes, positioning and correction anchor point nodes, repeating structure region nodes, and path units connecting adjacent nodes; the task object nodes are associated with unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information; the unique identifier feature description information includes the unique identifier feature type, the spatial position of the unique identifier feature in the task object ontology coordinate system, and the capture perspective conditions of the unique identifier feature; the object semantic similarity information includes the task object identifiers adjacent to the task object and their corresponding semantic similarity; the counterfactual exclusion feature description information includes the difference feature type, the difference feature spatial position, the difference feature capture perspective conditions, and the corresponding exclusion object range.

3. The UAV path planning method based on positioning drift accumulation according to claim 2, characterized in that, The calculation of the positioning drift growth coefficient includes: for any path unit traversed by the candidate path, obtaining the path length of the path unit, the overlap length with the region corresponding to the node of the repeating structure region, the number of observable positioning correction anchors, the number of observable unique identifiers, and the positioning signal quality value within the path unit; determining the ratio of the overlap length to the path length as the structure repeatability; determining visual localization based on the weighted normalization result of the number of observable positioning correction anchors and the number of observable unique identifiers; determining signal reliability based on the normalization result of the positioning signal quality value relative to a preset signal quality benchmark value; and, based on the preset benchmark drift growth coefficient, increasing the preset benchmark drift growth coefficient according to the structure repeatability, and decreasing the preset benchmark drift growth coefficient according to visual localization and signal reliability to obtain the positioning drift growth coefficient of the path unit.

4. The UAV path planning method based on positioning drift accumulation according to claim 2, characterized in that, The evaluation of the positioning correction benefit includes: for any positioning correction anchor point in the semantic topology map, obtaining the anchor point type, anchor point spatial location, number of observable features, observable distance range, observable viewing angle range, and semantic distinguishability with adjacent task objects or adjacent positioning correction anchor points; determining semantic uniqueness based on semantic distinguishability, visual stability based on the number of observable features, observable distance range, and observable viewing angle range, and spatial stability based on the fixed attributes of anchor point type and anchor point spatial location in the semantic topology map; and performing weighted normalization on semantic uniqueness, visual stability, and spatial stability to obtain the positioning correction benefit corresponding to the positioning correction anchor point.

5. The UAV path planning method for positioning drift accumulation according to claim 1, characterized in that, The prediction of the cumulative positioning drift includes: taking the initial positioning uncertainty of the UAV as the current positioning drift at the starting point of the candidate path, and updating the current positioning drift by increasing it according to the connection order of the path units in the candidate path, based on the path length and positioning drift growth coefficient of each path unit; when the candidate path passes through a positioning correction anchor point, updating the updated positioning drift by decreasing it according to the positioning correction benefit of that anchor point; repeating the increase and decrease updates to obtain the cumulative positioning drift when the UAV reaches the task object along the candidate path.

6. The UAV path planning method based on positioning drift accumulation according to claim 1, characterized in that, The determination of the trusted binding threshold includes: for any task object, obtaining the spatial size of the task object, the minimum distance between the task object and adjacent task objects, and the inspection accuracy requirement of the task object; determining a first threshold based on the spatial size and a preset size ratio, determining a second threshold based on the minimum distance and a preset distance ratio, and determining a third threshold based on the allowable positioning error corresponding to the inspection accuracy requirement; and determining the minimum value among the first threshold, the second threshold, and the third threshold as the trusted binding threshold corresponding to the task object.

7. The UAV path planning method for positioning drift accumulation according to claim 2, characterized in that, The determination of the counterfactual exclusion requirement includes: determining the positioning uncertainty range corresponding to the task object based on the cumulative positioning drift when the UAV arrives at the task object; determining the set of similar task objects to be excluded from adjacent task objects based on the positioning uncertainty range and object semantic similarity information; determining the difference features, difference feature capture perspective conditions, and exclusion object range between the task object and each similar task object in the set of similar task objects based on the counterfactual exclusion feature description information; determining the minimum exclusion evidence strength corresponding to each similar task object based on object semantic similarity information, cumulative positioning drift, and the exclusion object range corresponding to the difference features; and the counterfactual exclusion requirement is constituted by the set of similar task objects, the difference feature capture perspective conditions, and the minimum exclusion evidence strength.

8. The UAV path planning method for positioning drift accumulation according to claim 7, characterized in that, The insertion of the counterfactual exclusion observation pose includes: when the candidate path does not meet the counterfactual exclusion requirements, parsing the counterfactual exclusion requirements to obtain the similar task objects to be excluded, the difference feature capture viewpoint conditions, and the minimum exclusion evidence strength; determining the candidate observation poses in the semantic topology map that meet the difference feature capture viewpoint conditions and are located within the reachable flight space of the UAV; selecting the counterfactual exclusion observation pose from the candidate observation poses based on the difference feature visibility, deviation distance from the candidate path, obstacle avoidance constraints, and the expected exclusion evidence strength, and inserting the counterfactual exclusion observation pose into the candidate path.

9. The UAV path planning method for positioning drift accumulation according to claim 8, characterized in that, The determination of the inspection path includes: a comprehensive cost evaluation of candidate paths after inserting positioning correction anchor points, semantic evidence collection paths, or counterfactual exclusion of observation poses; the comprehensive cost evaluation includes determining the path length cost, energy consumption cost, obstacle avoidance cost, positioning drift accumulation cost, task object binding risk cost, semantic misbinding risk cost, positioning correction benefit, semantic evidence collection benefit, and counterfactual exclusion benefit, and weighting them according to preset weights to obtain the comprehensive cost value of the candidate path; the semantic evidence collection benefit is determined based on the degree to which the capture viewpoint conditions of the unique identifier feature are met and the visibility of the unique identifier feature; the counterfactual exclusion benefit is determined based on the degree to which the counterfactual exclusion requirements are met and the expected strength of the excluded evidence; the candidate path that meets the credible binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion requirements, and whose comprehensive cost value meets the preset path selection conditions, is determined as the inspection path.

10. A UAV path planning system for addressing positioning drift accumulation, characterized in that, The system applies the UAV path planning method for positioning drift accumulation as described in any one of claims 1 to 9, including: Semantic topology mapping module: acquires the environmental map of the area to be inspected, and constructs a semantic topology map containing task objects, positioning and correction anchor points, repetitive structural regions, unique identifier feature description information, object semantic similarity information, and counterfactual exclusion feature description information; Drift benefit assessment module: Calculates the positioning drift growth coefficient of path units based on semantic topology map, and evaluates the positioning correction benefit of positioning correction anchor points; Drift accumulation prediction module: Based on the initial positioning uncertainty, positioning drift growth coefficient, and positioning correction benefit, predict the cumulative positioning drift of candidate paths; Constraint Risk Requirements Module: Determine the trusted binding threshold based on the spatial size of the task object, the distance between adjacent task objects, or the inspection accuracy requirements; determine the semantic confusion risk based on the semantic similarity information of the object; and determine the counterfactual exclusion requirements by combining the cumulative location drift amount and the counterfactual exclusion feature description information. Path constraint verification module: Determines whether candidate paths meet the requirements of trusted binding threshold constraints, semantic confusion risk constraints, and counterfactual exclusion; Path active correction module: If the trusted binding threshold constraint is not met, a positioning correction anchor point is inserted; if the semantic confusion risk constraint is not met, a semantic evidence collection path is generated; if the counterfactual exclusion requirement is not met, a counterfactual exclusion observation pose is inserted. Inspection path determination module: Evaluate candidate paths and determine inspection paths that simultaneously meet the requirements of trusted binding threshold constraint, semantic confusion risk constraint, and counterfactual exclusion.

Images