Power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation

Abstract

This invention discloses a post-disaster power transmission inspection method based on 3D point cloud semantic segmentation and route generation. The specific steps include: S1, acquiring 3D point cloud data and performing semantic segmentation to generate a first scalar field; S2, generating a second scalar field based on the confidence level of the semantic segmentation; S3, fusing the first and second scalar fields to generate a unified utility field representing the comprehensive inspection value; S4, extracting local maxima points in the utility field to determine dynamic inspection target points; and S5, constructing a path cost field and generating the optimal inspection route. This invention constructs a unified inspection utility field, integrating physical risk, semantic entropy, and navigation integrity to build a dynamic risk-reward decision space. It can adaptively adjust the safety boundary according to the UAV's own state and autonomously generate the optimal observation point as a dynamic task target, achieving deep coupling closed-loop control of environmental perception, introspection, and route planning.

Description

Technical Field

[0001] This invention relates to the field of UAV 3D point cloud processing and autonomous navigation technology, specifically a post-disaster power transmission inspection method based on 3D point cloud semantic segmentation and flight path generation. Background Technology

[0002] The development of autonomous navigation and 3D environmental perception technologies is driving UAVs to play an increasingly important role in the intelligent operation and maintenance of critical infrastructure such as power systems. Improving the autonomous inspection capabilities of UAVs in unstructured and complex environments has become a core research direction in this field.

[0003] Existing technologies typically face the following challenges when dealing with such complex inspection tasks: Flight path planning often relies on incomplete, static 3D point cloud models acquired only once. When executing a predetermined route, an aircraft can only passively avoid local dynamic obstacles. Its path selection cannot dynamically adapt to the complexity and uncertainty of the post-disaster environment at a global level, resulting in room for improvement in flight reliability.

[0004] Inspection tasks are typically limited to confirming a set of pre-set observation points. This mode cannot dynamically adjust inspection priorities based on real-time perceived environmental uncertainties (such as cognitively ambiguous areas caused by occlusion or low confidence in semantic segmentation), which can easily lead to incomplete information acquisition on key hidden damage. Summary of the Invention

[0005] The purpose of this invention is to provide a post-disaster power transmission inspection method based on three-dimensional point cloud semantic segmentation and route generation, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A post-disaster power transmission inspection method based on 3D point cloud semantic segmentation and route generation includes the following steps: S1. Acquire three-dimensional point cloud data of power transmission facilities and perform semantic segmentation to generate a first scalar field to characterize the distribution of physical risks in the inspection space. S2. Based on the confidence level of semantic segmentation, generate a second scalar field to characterize the distribution of information uncertainty in the inspection space; S3. Through a preset utility function, the first scalar field and the second scalar field are nonlinearly fused to generate a utility field that uniformly represents the comprehensive inspection value of each point in the inspection space. S4. Extract local maxima points in the utility field to determine multiple dynamic inspection target points; S5. Invert the values ​​of the utility field to construct the path cost field, and generate the optimal inspection route connecting the dynamic inspection target points based on the path cost field.

[0007] Furthermore, S1 includes identifying point cloud clusters with risk semantic labels in the three-dimensional point cloud data, and calculating the spatial repulsive potential field of the point cloud based on the distribution density of the point cloud clusters, as the first scalar field.

[0008] Furthermore, S2 includes calculating the information entropy of the point cloud classification probability distribution within each voxel in the 3D point cloud data, generating a voxel semantic entropy field as a second scalar field.

[0009] Furthermore, the steps for generating the first scalar field are as follows: Input 3D point cloud data containing risk semantic labels and risk impact weight coefficients; A density-based spatial clustering algorithm with noise is used to extract point clouds with the same risk semantic label into independent target point cloud clusters; Calculate the Euclidean distance between the voxel center and the target cloud cluster center; divide the square of the Euclidean distance by twice the square of the Gaussian kernel width parameter to obtain the intermediate exponent; calculate the natural constant. The negative median exponent term raised to the power of the negative median exponent term; multiply the negative median exponent term by the risk impact weighting coefficient; The calculation results are then nonlinearly mapped using the Sigmoid function. We obtain a smooth and second-differentiable first scalar field.

[0010] Furthermore, the steps for generating the second scalar field are as follows: Input the posterior probability vector of each voxel for its class; Obtain the probability value of each category in the category posterior probability vector; calculate the base-2 logarithm for each probability value greater than 0; multiply each probability value by the corresponding logarithm and take the negative number; sum the product results of all categories to obtain the original information entropy of the voxel; Obtain the original information entropy of the 26 neighboring voxels in the current voxel's neighborhood; calculate the arithmetic mean of the original information entropy of the 26 neighboring voxels; and perform a weighted summation of the original information entropy of the current voxel and the average entropy of the neighborhood, with the weights determined by a preset spatial consistency correction factor. Divide the weighted summation result by the preset maximum theoretical entropy value; We obtain a second scalar field that reflects cognitive uncertainty.

[0011] Furthermore, in S3, if the value of the first scalar field at the target point exceeds the hard safety threshold, the value of the utility field at the target point is configured as a prohibited value; if the value of the first scalar field at the target point does not exceed the hard safety threshold, the value of the utility field at the target point is calculated based on the product of the exponential decay function of the value of the second scalar field at the target point and the value of the first scalar field at the target point.

[0012] Furthermore, the steps for generating the comprehensive inspection utility output value are as follows: Real-time acquisition of normalized physical risk intensity, normalized semantic entropy value, and navigation and positioning integrity index; Obtain the preset basic risk coefficient; The dynamic risk aversion coefficient is obtained by calculating the quotient of the basic risk coefficient and the navigation and positioning integrity index. Compare the magnitudes of normalized physical risk intensity and hard safety threshold. If the normalized physical risk intensity is greater than the hard safety threshold, the comprehensive inspection utility output value will be directly set to 0; and the area will be marked as an absolutely prohibited zone. If the normalized physical risk intensity is less than or equal to the hard safety threshold, then proceed to the next step of nonlinear fusion calculation; Obtain the dynamic risk aversion coefficient and the current normalized physical risk intensity, and perform the product calculation; With natural constant The risk decay factor is calculated by raising the negative product to the power of the base. The normalized semantic entropy value is multiplied by the risk decay factor to obtain the final comprehensive inspection utility output value. The output reflects the comprehensive inspection effectiveness value of spatial location.

[0013] Furthermore, during the generation of the optimal inspection route, the positioning reference point is corrected by utilizing the gradient distribution of the utility field, and navigation reliability monitoring is performed based on the magnitude relationship between the first scalar field and the preset hard safety threshold.

[0014] Furthermore, S5 includes the use of particle swarm optimization algorithm to search for a six-dimensional parameter sequence that minimizes the path integral cost in the path cost field, in order to generate the optimal inspection route that includes position and attitude information. The six-dimensional parameter sequence includes coordinate parameters used to characterize the three-dimensional spatial position of the UAV, and pitch angle, roll angle, and yaw angle parameters used to characterize the three-dimensional observation attitude of the UAV.

[0015] Furthermore, the steps for generating the optimal inspection route are as follows: Input the generated utility field and the corresponding path cost field, and the initial particle swarm size; In a six-dimensional normalized space, randomly generate the six-dimensional state vector and initial velocity vector of the initial position of 50 particles. For each path represented by a particle, the path integral cost is calculated using Riemannian and numerical integral logic. Check whether the first scalar field value at any point on the path exceeds the safety margin monitoring threshold; If the value exceeds the limit, the particle's fitness value will be set to the maximum value, and the navigation reliability monitoring obstacle avoidance logic correction will be triggered. Obtain the current individual optimal position and the group optimal position of the particle; Obtain the utility field gradient vector at the particle's current position, and multiply the utility field gradient vector by the environment guidance weight coefficient; Multiply the current velocity by the particle inertial weight, the individual optimal traction force, the group optimal traction force, and the environmental guiding momentum, and then sum them as a vector. The previous position and the updated velocity are vector-summed to obtain a new six-dimensional state vector. After a preset number of iterations, the six-dimensional parameter sequence represented by the particle with the smallest fitness value is the optimal inspection route.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a unified, multi-dimensional information fusion inspection task utility field. The inspection task utility field is not only a mathematical expression of physical space, but also a unified central hub for task decision-making. It integrates high-physical-risk regions representing physical obstacles with high-semantic-entropy regions representing key damage information. Physical risk applies a repulsive force to the field, ensuring safety boundaries, while semantic entropy gives the field an attractive force, guiding the detection direction. Route generation is no longer a simple path search, but a solution to a dynamic optimal solution, thus achieving deep coupling between global environmental perception and route planning.

[0017] This invention utilizes the utility field itself to dynamically generate task objectives; the local maxima in the utility field are the best observation candidate points under the risk-benefit trade-off in the current environment; thus, the inspection target is transformed from a set of static preset coordinates into a series of dynamic points of interest generated by real-time environmental data, thereby realizing intelligent and adaptive exploration of the inspection task and ensuring priority detailed investigation of high-value fuzzy areas.

[0018] This invention also introduces the UAV's own navigation and positioning integrity index as a key variable into the construction of the utility field; when the positioning signal is poor, it will automatically increase the sensitivity to physical risks, so that the safety boundary in the utility field will actively expand; the mechanism of linking the introspection state of the aircraft with the perception of the external environment constructs a higher-dimensional closed-loop control, ensuring high reliability under various complex conditions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram combining the isometric hidden line removal diagram and the technical roadmap of the present invention; Figure 2 This is a schematic diagram of the overall steps of the post-disaster power transmission inspection method of the present invention; Figure 3 This is a schematic diagram of the generation steps of the first scalar field of the present invention; Figure 4 This is a schematic diagram of the generation steps of the second scalar field of the present invention; Figure 5 This is a schematic diagram illustrating the steps for generating the comprehensive inspection utility output value of the present invention; Figure 6 This is a schematic diagram illustrating the steps for generating the optimal inspection route according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] Example 1 Please see Figures 1 to 6 This invention provides a technical solution: a post-disaster power transmission inspection method based on 3D point cloud semantic segmentation and route generation, comprising the following steps: S1. Acquire three-dimensional point cloud data of power transmission facilities and perform semantic segmentation to generate a first scalar field to characterize the distribution of physical risks in the inspection space.

[0023] S1 includes identifying point cloud clusters with risk semantic labels in 3D point cloud data, and calculating the spatial repulsive potential field of the point cloud based on the distribution density of the point cloud clusters, as the first scalar field. The semantic labels of point clouds are identified as obstacles, ground, or unknown. For each voxel in space, the spatial repulsive potential field of the point cloud is calculated as a weighted sum of the inverse of the distance to the risk point cloud and positively correlated with the point cloud density within the voxel. The higher the point cloud density or the closer to the risk point, the larger the calculated spatial repulsive potential field value of the point cloud. To calculate the spatial repulsion of point clouds for any voxel in space, perform the following calculation steps: From the global point cloud, select all point clouds with semantic labels of obstacle, ground, unknown, or other preset risk categories to form a risk point cloud set. ,in Index for risk points; for the voxel to be calculated. ,in For the classified voxel index, for voxels The number of risk point clouds contained within is counted to obtain the local point cloud density of the voxels. ; with voxels Using the center point coordinates as the calculation reference, the risk point cloud set is traversed. Each risk point ; For each risk point Calculate the voxel Single-point potential field contribution value The calculation logic is as follows: Calculate voxels Center and Risk Points Euclidean distance between ; Identify risk points The semantic tags are used to query the preset semantic risk level weight table to obtain the risk points. semantic risk weights ; Obtaining voxels Local point cloud density And based on this, the density influence weight is calculated. ; By using semantic risk weights Weighting of density influence Multiply by the product to obtain the combined weight; then divide the combined weight by the Euclidean distance between the risk point and the voxel center. The square of a fraction and small positive constants to prevent the denominator from being zero The sum of these values ​​ultimately generates the single-point potential field contribution value. ; All risk points to voxels Single-point potential field contribution value By summing the results, we obtain the final original value of the repulsive potential field of the voxel, which is specifically characterized as follows: ; The calculation logic is as follows: Obtain the set of coordinates of the center points of the target point cloud clusters extracted by the density-based spatial clustering algorithm; for each voxel to be calculated in space, extract the coordinates of the voxel center point; calculate the Euclidean distance between the voxel center point and the center points of all target point cloud clusters; for each target point cloud cluster, divide the square of the Euclidean distance by twice the square of the Gaussian kernel width parameter to obtain the intermediate exponent term; calculate the negative intermediate exponent term power of the natural constant, and multiply the result by the preset risk influence weight coefficient of the corresponding target point cloud cluster to obtain the single-cluster repulsive potential field contribution value of the point cloud cluster to the current voxel; sum up the single-cluster repulsive potential field contribution values ​​of all target point cloud clusters to the voxel to obtain the original repulsive potential field value of the voxel; perform nonlinear mapping on the original repulsive potential field value through the Sigmoid function to generate the final first scalar field with the value range strictly constrained in the interval (0.5,1).

[0024] Define semantic risk weights The risk level is assigned based on the inherent hazard level of the risk source, and objects with different semantic labels pose different threats to flight safety. In this embodiment, the danger of high-voltage power lines is much higher than that of the ground. A preset semantic risk level weight table is used to assign different fixed weight values ​​to different semantic labels. The specific allocation scheme is as follows: Semantic label = high-voltage conductor → semantic risk weight =1.5 (highest risk); Semantic label = tree branches → semantic risk weight =1.0 (Standard Risk); Semantic Label = Tower Structure → Semantic Risk Weight =0.8 (lower risk, due to its surface continuity, easy to avoid); Semantic label = ground → semantic risk weight =0.5 (lowest risk, used only to maintain a safe altitude); Semantic label = unknown → semantic risk weight =1.2 (due to its uncertainty, it is given a higher risk weight).

[0025] Define the influence weight of density : Characterization is based on the spatial clustering pattern of risk sources; dense obstacles composed of a large number of point clouds have a higher probability of physical collision than sparse obstacles composed of a small number of point clouds; local point cloud density within weighted positive correlation voxels. ; Piecewise linear or logarithmic functions are used for calculation to avoid numerical explosion; the preferred implementation is: ;in, It is an adjustable density sensitivity coefficient, preferably 0.5 in this embodiment; ensuring that there is no local point cloud density within the voxel. At that time, density affects weight. A value of 1 indicates no enhancement effect; as the local point cloud density increases... The increase in density affects the weight. Logarithmic growth amplifies the risk level of the region while avoiding abrupt changes in potential field values ​​caused by individual extremely high-density points. Through the aforementioned dual-weighting mechanism, this embodiment calculates the spatial repulsive potential field of the point cloud, which not only reflects the distance relationship with obstacles but also encodes the type and density of obstacles more precisely, thereby providing unprecedented high-fidelity risk perception capabilities for subsequent route planning.

[0026] S2. Based on the confidence level of semantic segmentation, generate a second scalar field to characterize the distribution of information uncertainty in the inspection space.

[0027] S2 includes calculating the information entropy of the point cloud classification probability distribution within each voxel in the 3D point cloud data, generating a voxel semantic entropy field as the second scalar field.

[0028] Receive the classification probability vector of each point cloud output; for each voxel in space, the semantic entropy value is obtained by calculating the average of the Shannon entropy of the classification probability distribution of all point clouds in the voxel. The more dispersed the classification probability of a point is, the higher the information entropy, which in turn leads to a higher semantic entropy value of the voxel. For each voxel to be computed in space The number of point clouds contained within Perform a count; Conditional judgment and branch calculation: based on the number of point clouds The value is used to perform conditional judgment: If the number of point clouds If the voxel is not empty, then the calculation will proceed according to the original plan: traverse the voxel. For each point, obtain its classification probability vector, calculate the Shannon entropy, and finally average the Shannon entropies of all points to obtain a voxel. The semantic entropy value; If the number of point clouds If the voxel is empty, skip the calculation of the entropy average, directly read the unobserved spatial uncertainty constant from the preset parameter table, and assign the unobserved spatial uncertainty constant directly to the voxel. As its semantic entropy value; After the entropy values ​​of all voxels (including non-empty and empty voxels) have been determined, the entire second scalar field is uniformly normalized to ensure that all values ​​fall within the (0,1) interval. The value of the unobserved space uncertainty constant is subject to the following limitations: assigning a higher value (0.75 after normalization in this embodiment) is to mark the unobserved area as an area with high exploration value in the information entropy field; it will incentivize the route planning algorithm to generate paths that can pass through or enter these areas to obtain new observation data, thereby eliminating uncertainty and directly serving the goal of inspection without leaving any blind spots. Instead of directly assigning a theoretical maximum value (in this embodiment, the theoretical maximum value is 1.0), the highest uncertainty should be defined as the region that has been observed but has serious classification conflicts. In this embodiment, half of the point cloud within a voxel is identified as a wire and the other half as a tree branch. It is known that severe semantic conflicts represent the highest decision risk and information value. Setting the level of unobserved uncertainty slightly lower than the level of serious conflicts is a more refined and reasonable strategy, which allows the algorithm to prioritize solving the most difficult problems that have been discovered when resources are limited. In the preferred embodiment, the unobserved spatial uncertainty constant is set to 0.75.

[0029] In this embodiment, the key parameters involved are as follows: The parameter sign for the local point cloud density of voxels is set to... It represents the ratio of the number of original laser point clouds contained in a specific three-dimensional voxel space to the voxel volume, and is used to initially measure the solidity of obstacles. The original 3D point cloud is acquired using an airborne lidar; the entire inspection space is divided into a uniformly sized cubic voxel grid; in this embodiment, the voxel side length is set to 0.5 meters, balancing inspection accuracy and computational real-time performance; all original lidar point clouds are traversed, the voxel index of each point cloud is determined, and the total number of point clouds within each voxel is counted. ; Obtain the volume of voxels In this embodiment, it is 0.5. 3 =0.125 cubic meters; divide the total number of point clouds by the volume. The local point cloud density of voxels is obtained, i.e. .

[0030] The parameter sign of the risk impact weighting coefficient is set to The quantitative risk level assigned based on the label type identified by semantic segmentation (in this embodiment, it includes transmission towers, broken conductors, and vegetation) is used to adjust the repulsion intensity. The semantic risk lookup table is pre-stored in the system; calibration is based on offline experiments; the impact level of different objects on flight safety is determined by cluster analysis of historical inspection accident data; in this embodiment, the coefficient is set to 0.9 for the transmission tower label and 0.4 for the small branch label.

[0031] The parameter sign of the Gaussian kernel width parameter is set to When constructing the repulsive potential field, the influence range of the repulsive force generated by a single risk point and the smoothness of its decay with distance are controlled; essentially, the minimum thickness of the safety buffer zone around the UAV is defined; the system preset parameters are directly related to the physical dimensions of the UAV and the safety margin required by the mission; the determination steps are as follows: Obtain the maximum geometric dimensions of the drone in flight. In this embodiment, the maximum geometric dimension The length of the rotor diagonal; Determine the minimum safe distance based on the task scenario. ; Gaussian kernel width parameter Determined as the maximum geometric dimension and minimum safe distance The sum should be a multiple of the sum to ensure sufficient buffering; the preferred calculation formula is: , This represents the safety buffer factor, used to adjust the fault tolerance rate under environmental wind disturbances. Its preferred value range is [1.2, 2.0], and in this embodiment, it is preferably 1.5. This represents the maximum geometric dimensions of the drone; in this embodiment, the diagonal length of the drone's rotor is taken as 0.8 meters. This indicates the minimum safe distance set according to different inspection task scenarios. In this embodiment, 2.0 meters is preferred for the power transmission corridor scenario.

[0032] The sign of the parameter of the class posterior probability vector is set to The semantic segmentation model is a set of probability estimates for point clouds within voxels belonging to each preset semantic category; The output is from the Softmax layer of the preceding deep learning semantic segmentation network (PointNet++ in this embodiment); if the probability of a point cloud within a voxel being identified as a wire is 0.7, the probability of it being an insulator is 0.2, and the probability of it being background is 0.1, then the posterior probability vector of the category is... It consists of these three values; Table 1 below details the preset semantic categories and their corresponding risk impact weight coefficients, risk levels, and setting basis.

[0033] Table 1: Semantic Categories and Their Corresponding Risk Impact Weight Coefficients, Risk Levels, and Definitions

[0034] For any point to be processed in the global point cloud, the following series of precise calculation steps will be performed to determine... : Risk points Centered on the target, the K-nearest neighbor algorithm is used for nearest neighbor search to select the surrounding points from the global point cloud. A set of neighboring points forms a local point cloud neighborhood. The neighborhood provides the model with insights into risk points. The necessary context of the local geometric environment in which it exists; For a local point cloud neighborhood set For each point in the matrix, extract multidimensional features to form the input feature matrix; the features include at least: 3D coordinates Laser reflection intensity Color information ; normal vector Online calculation or as a preprocessing step; The feature matrix of the constructed local point cloud neighborhood is used as input to the RandLA-Net model that has been loaded into memory; The input data propagates forward through the network; the data flows through a series of local feature aggregation modules and random downsampling layers; in this process, the network learns and extracts from low-order geometric primitives to high-order abstract combined features related to semantic categories layer by layer; The feature information compressed by the encoder will be upsampled layer by layer by the decoder, and the learned high-level semantic features will be gradually mapped back to the original point cloud resolution. The ultimate result of forward propagation is the identification of risk points in the input. The network outputs high-dimensional pointwise feature vectors containing rich semantic information. ; The generated point-by-point feature vector The data is input into the classification header at the end of the network; the classification header is usually composed of one or more fully connected layers or 1×1 convolutional layers. The role of fully connected layers is to transform high-dimensional pointwise feature vectors... A linear transformation is performed and mapped to a dimension equal to the total number of preset semantic categories. The output vector is called the Logits vector. ; ; Each element in the vector It is a real number (which can be positive or negative) representing the risk points that the RandLA-Net model considers to be risk points. Belongs to the The original confidence scores or strength of evidence for each category are not yet probabilities; In order to transform the original Logits vector Transform it into a valid probability distribution and apply the Softmax activation function to it; for the th logits vector... Each element corresponds to a final probability. The calculation formula is: ;in, A sequence represented as a semantic category; The new vector obtained after calculation using the Softmax function is the final class posterior probability vector. ; ; Class posterior probability vector It has two key characteristics: All elements The values ​​are all within the interval (0,1), and the sum of all elements is 1.

[0035] The parameter sign of the spatial consistency correction factor is set to The coefficients are used to weight and correct the current voxel information entropy based on the similarity of the semantic distribution of adjacent voxels; they are used to suppress isolated high-entropy noise caused by classification errors of a single point cloud, and improve the smoothness and reliability of the information entropy field; they are obtained by post-processing the second scalar field.

[0036] Specifically, the following calculation steps are defined: For each voxel, 26 neighboring voxels in three-dimensional space are defined, namely the mole neighborhood; Calculate the Jensen-Shannon divergence between the mean probability distribution of the current voxel and the mean probability distribution of each neighboring voxel; the Jensen-Shannon divergence is a symmetric, smooth measure of the difference between two probability distributions. The average divergence is obtained by averaging the Jensen-Shannon divergence values ​​of all neighborhoods; spatial consistency correction factor. In this embodiment, the function calculated as a decreasing function of the mean divergence is... ,in To adjust the hyperparameters of sensitivity, Expressed as the Jensen-Shannon divergence value, The average divergence is expressed as the Jensen-Shannon divergence value; the more semantically consistent adjacent voxels are, the better. The smaller, The closer it is to 1, the smaller the correction to the original entropy value.

[0037] The steps for generating the first scalar field are as follows: Input 3D point cloud data containing risk semantic labels and risk impact weight coefficients ; A density-based spatial clustering algorithm with noise is used to extract point clouds with the same risk semantic label into independent target point cloud clusters; Calculate the Euclidean distance between the voxel center and the target cloud center; divide the square of the Euclidean distance by the Gaussian kernel width parameter. Double the square of 1 to obtain the intermediate exponent; calculate the natural constant. The negative median exponent term raised to the power; multiply the negative median exponent term raised to the power by the risk impact weighting coefficient. ; The calculation results are nonlinearly mapped using the Sigmoid function to ensure that the output value range is within the (0.5, 1) interval; a smooth and twice-differentiable first scalar field is obtained. The computational steps of the spatial clustering algorithm are as follows: Traverse all point clouds that have completed semantic segmentation and group them according to semantic labels; subsequent clustering operations will be performed independently within each semantic group. Set two core parameters for the DBSCAN algorithm for each semantic category: The neighborhood range of a point is defined and is related to the point cloud density and the physical properties of the target object. For continuous, slender objects like conductors, the neighborhood range should be slightly larger than the average point spacing of the point cloud on the conductor to ensure that the points on the conductor can be connected into a cluster. If the average point spacing on the conductor is 0.2 meters, the neighborhood range can be set to 0.3 meters. For entities like transmission towers, the neighborhood range can be appropriately increased, such as to 0.5 meters. It defines the minimum number of points in the neighborhood required to form a core point; used to filter out sparse noise points. The value should be set based on the sensor's noise level and the expected minimum effective object size; a rule of thumb is to set MinPts to about twice the dimension of the point cloud data; in 3D space, a robust starting value is 5-10 points. Perform the DBSCAN algorithm on each semantic group: Traverse the point cloud group of the guide: using the settings configured for the guide The parameters are used to run DBSCAN, where, Represented as the neighborhood range of the point cloud group of the conductor. This represents the minimum number of points in the neighborhood of a traverse point cloud group; the algorithm automatically divides spatially connected traverse points into clusters; the output may be {traverse cluster 1, traverse cluster 2, ..., traverse cluster N} and isolated points marked as noise; traverse cluster 1 and traverse cluster 2 may represent two parallel traverses or traverses with different spans. Used for transmission towers The parameters are used to run DBSCAN, where, Represented as the neighborhood of the transmission tower. This represents the number of points within the smallest neighborhood of a transmission tower; if there are two towers in the scene, the algorithm will output {transmission tower cluster 1, transmission tower cluster 2}. Repeat this process for all other semantic categories; For each successfully extracted target point cloud cluster, calculate the coordinates of its geometric center point; this is the central source point for subsequent Gaussian kernel function calculations. Through the above steps, the original loosely organized point cloud by category is transformed into a list of independent target point cloud clusters with clear physical instances, laying the foundation for the subsequent construction of an object-centered potential field.

[0038] The steps for generating the second scalar field are as follows: Input the class posterior probability vector of each voxel Obtain the posterior probability vector of the class. The probability value of each category is calculated; for each probability value greater than 0, the logarithm to base 2 is calculated; each probability value is multiplied by the corresponding logarithm and the result is negative; the product of all categories is summed to obtain the original information entropy of the voxel. Obtain the original information entropy of the 26 neighboring voxels in the current voxel's neighborhood; calculate the arithmetic mean of the original information entropy of the 26 neighboring voxels; and perform a weighted summation of the original information entropy of the current voxel and the average entropy of the neighborhood, with the weights determined by the spatial consistency correction factor. Decide; Divide the summation result by the preset maximum theoretical entropy value, which is the preset total number of semantic categories. The base-2 logarithm of the logarithm ensures that the output range is in the interval (0,1); We obtain a second scalar field that reflects cognitive uncertainty.

[0039] In the generation of the second scalar field, a spatial consistency smoothing step is introduced to eliminate isolated high-entropy noise points that may be caused by classification errors in a single point cloud. The core idea of ​​this step is that the final information entropy of a voxel should be determined by its own original information entropy and the average entropy of its neighborhood environment. A spatial consistency correction factor is also involved. This is the key to dynamically adjusting the weight between the two; The logic for determining the weights is as follows: If the semantic probability distribution of a voxel is highly consistent with the distribution of its neighbors, then there is reason to believe that the original information entropy of the voxel is reliable and should be retained; otherwise, if the distribution is inconsistent with the distribution of its neighbors, it is likely to be an outlier or noise due to misclassification, and the weight of the original information entropy should be reduced and corrected by referring more to the average situation of the neighborhood. Spatial consistency correction factor The range of values ​​is designed to be in the interval (0,1); When the spatial consistency correction factor →1 indicates a high degree of consistency; When the spatial consistency correction factor →0 indicates extreme inconsistency (possibly an outlier); Therefore, spatial consistency correction factor The weights are directly used as the weights of the original information entropy of the current voxel; correspondingly, the weights of the neighborhood average entropy are (1 ); The specific weight determination and weighted summation implementation are as follows: set up The original information entropy calculated for the current voxel; The arithmetic mean of the original information entropy of the 26 neighboring voxels; This is a spatial consistency correction factor calculated based on the semantic distribution differences between the current voxel and its neighborhood. The information entropy after spatial consistency smoothing correction The calculation formula is as follows: .

[0040] When a voxel is located on a face, edge, or corner of the entire 3D voxel mesh, the number of available neighboring voxels will be less than 26, and adaptive neighborhood calculation is used. This scheme does not create virtual data, but dynamically adapts to the available neighborhood during calculation. Specific implementation steps: For the voxel currently being processed, examine the standard 26 neighborhood locations. Only retain valid neighbors whose coordinates fall within the original grid area; obtain the set of valid neighbors and record their count. ; Calculate the arithmetic mean of the neighborhood At that time, only to Sum the original entropy values ​​of the effective neighbors, then divide by . ; ,in, Represented as a sequence of the original entropy values ​​of the effective neighbors; Computational Spatial Consistency Correction Factor At the same time, it is also based only on The average difference is calculated using the probability distribution of each effective neighbor. Using dynamically calculated arithmetic mean Spatial consistency correction factor It executes the same weighted summation formula as the internal voxels; The above calculation logic does not require additional memory to store the fill layer; and the smoothing effect at the boundary is slightly affected by the number of available neighbors; in this embodiment, the smoothing of the corner voxel will only be based on its 7 neighbors, and the statistical stability of the result will be slightly lower than that of the internal voxel with 26 neighbors.

[0041] S3. By using a preset utility function, the first scalar field and the second scalar field are nonlinearly fused to generate a utility field that uniformly represents the comprehensive inspection value of each point in the inspection space.

[0042] S3 includes: if the value of the first scalar field at the target point exceeds the hard safety threshold, then the value of the utility field at the target point is configured as the prohibition value; if the value of the first scalar field at the target point does not exceed the hard safety threshold, then the value of the utility field at the target point is calculated based on the product of the exponential decay function of the value of the second scalar field at the target point and the value of the first scalar field at the target point. The parameter symbols involved are set as follows: The parameter sign of the dynamic risk aversion coefficient is set to , which is a weighting parameter used to control the sensitivity of the utility function to physical risk.

[0043] The parameter sign of the hard safety threshold is set to Define the physical risk limit boundary that drones are allowed to enter; exceeding this value is considered impassable. In a controlled environment, the impact of different wind speeds on the drone's attitude is simulated, and the maximum radius of the drone's drift is recorded. This maximum radius is mapped onto the intensity curve of the first scalar field, and 0.85 times the corresponding field intensity value is selected as the safety threshold. In this embodiment, a hard safety threshold is used. The specific value is determined to be 0.75.

[0044] The parameter symbol for the comprehensive inspection utility output value is set to... The final generated scalar value, used to guide route generation, has a value range of 0 to 1, and is a comprehensive inspection utility output value. The larger the value, the more valuable and safer the point is for inspection.

[0045] The critical value of the repulsive potential field is used as the hard safety threshold; for any point in space, the calculation logic for the inspection task utility value is as follows: Determine whether the repulsive potential field value of the point cloud space of the target point is greater than the hard safety threshold. If so, the utility value of the target point is directly assigned to negative infinity. If not, the utility value is jointly determined by the voxel semantic entropy value of the target point and the repulsive potential field value of the point cloud space, and its preferred calculation formula is: ,in, A configurable, dynamic risk aversion coefficient. This is represented as the comprehensive inspection utility output value. This represents the normalized semantic entropy value. This is represented as a hard safety threshold.

[0046] In this embodiment, the key parameters involved are as follows: The parameter sign for the normalized physical risk intensity is set to , represents the physical collision risk level in the current voxel space output by the first scalar field, with a value range between 0 and 1, and is the direct output value of the first scalar field generated by S1.

[0047] The parameter sign of the normalized semantic entropy value is set to , represents the information value output by the second scalar field in the current voxel space, with a value range between 0 and 1, and is the direct output value of the second scalar field generated by S2.

[0048] The parameter sign of the navigation and positioning integrity index is set to... The reliability of the UAV's current positioning is characterized by both the number of visible satellites and the quality of its geometric distribution. The extended Kalman filter system of the UAV's onboard GNSS receiver and inertial measurement unit outputs in real time; the GDOP value of the current positioning system is obtained; 1 is divided by the GDOP value to obtain the original index; and it is scaled by a preset scaling factor. In this embodiment, the preset scaling factor is 0.2 to ensure that the generated navigation and positioning integrity index is between 0.8 and 1.0 under normal working conditions. In this embodiment, the scaling factor is preferably 0.2, with a reasonable range of 0.15 to 0.25. It is determined based on the positioning system performance baseline calibration experiment. The execution steps of the positioning system performance baseline calibration experiment are as follows: A UAV equipped with an onboard GNSS receiver and inertial measurement unit is placed in an unobstructed open area, and positioning data is continuously collected for at least 30 minutes. The numerical distribution of the geometrical precision factor (GDOP) during this period is recorded. The upper limit of the 95% confidence interval of the statistical distribution is defined as the GDOP reference value of the positioning system under normal, good working conditions. This value is then used to calculate the GDOP reference value from the extended Kalman filter system. The geometric precision factor (GDOP) value is read in real time; the smaller the GDOP value, the more accurate the positioning. A scaling factor is calculated that maps the reciprocal of the GDOP reference value to a preset integrity index lower limit. In this embodiment, the preset integrity index lower limit is set to 0.72. The specific calculation logic is as follows: obtain the preset integrity index lower limit value, multiply the integrity index lower limit value by the GDOP reference value, and thus obtain the final value of the scaling factor. This method ensures that positioning systems with different performance can adaptively determine the most suitable scaling factor through a one-time offline calibration, thereby ensuring universality and reproducibility. GDOP values ​​are typically categorized into several levels to assess positioning quality, as shown in Table 2.

[0049] Table 2: Numerical Classification of GDOP

[0050] The integrity availability lower bound is a critical decision threshold that is pre-calculated and embedded in the flight control system; When the real-time calculated navigation and positioning integrity index is lower than the integrity availability lower limit, it is determined that the reliability of the current positioning information is insufficient to support the autonomous inspection task that requires precise operation. At this time, the current route tracking will be interrupted immediately, and the preset navigation failure safety protocol will be automatically executed. In this embodiment, the system will immediately hover, switch to a more stable attitude control mode, return along a safe path, or land in an open area.

[0051] Therefore, the integrity availability lower bound is the last line of defense to ensure that drones do not engage in catastrophic behavior when positioning signals deteriorate; The specific calculation steps are as follows: Starting from the task requirements themselves, we define how many meters the deviation between the estimated position of the UAV and the actual physical position should not exceed in the worst case, otherwise there will be a very high probability of collision with the inspection target or surrounding obstacles; this is determined by the task type, UAV size, and safety redundancy; it is calibrated by offline simulation of the inspection scenario or actual testing in a safe environment; for environments with dense obstacles such as power transmission corridors, considering the size of the UAV and the influence of gusts, this embodiment sets the maximum tolerable position error to 2.5 meters; Equivalent ranging error is a comprehensive parameter characterizing the inherent error level of a GNSS system, including satellite ephemeris error, ionospheric / tropospheric delay error, multipath effects, and receiver noise. It is usually provided by the GNSS receiver manufacturer or statistically evaluated through long-term static observation data. It can be calculated by consulting the technical specifications of the airborne GNSS receiver or using professional GNSS data analysis software. For the industrial-grade RTK-GNSS receiver used in this embodiment, the typical user equivalent ranging error is about 1.5 meters. The maximum acceptable geometrical precision factor (GDOP) is determined by an approximately linear relationship between position error, equivalent ranging error, and the GDOP reference value: position error ≈ GDOP × equivalent ranging error. Therefore, knowing the maximum tolerable position error and the system's basic ranging error, we can deduce the worst-case satellite geometrical distribution that allows the system to function normally, i.e., the maximum acceptable GDOP value. When the maximum acceptable GDOP value exceeds 1.67, even if the signal itself is not faulty, the position error caused solely by poor satellite distribution may exceed the safe range. Using the previously defined navigation and positioning integrity index The conversion relationship between the maximum acceptable GDOP value and the maximum acceptable GDOP value is used to convert the maximum acceptable GDOP value into the lower limit threshold of the navigation integrity index. To facilitate engineering applications and retain a small margin, the calculation results are rounded; in a preferred embodiment of the present invention, the final determined value of the lower limit of the integrity index is 0.72.

[0052] Ensured the integrity index of navigation and positioning The availability threshold is not an empirical value, but a traceable and verifiable scientific setting that is directly linked to security boundaries in the physical world.

[0053] In this embodiment, the comprehensive inspection utility output value The generation steps are as follows: Real-time acquisition of normalized physical risk intensity Normalized semantic entropy value and navigation and positioning integrity index ; A preset basic risk coefficient is obtained; in this embodiment, the basic risk coefficient is set to 2.0. Calculate the basic risk coefficient and navigation and positioning integrity index The business yields a dynamic risk aversion coefficient. When the navigation and positioning integrity index When it decreases, i.e., from 1.0 to 0.5, the dynamic risk aversion coefficient... It will double from 2.0 to 4.0, thereby amplifying the risk aversion tendency of the utility field; Comparison of normalized physical risk intensity With hard safety threshold The order of magnitude; If normalized physical risk intensity Greater than the hard safety threshold Then the comprehensive inspection utility output value will be directly displayed. The value is assigned to 0; this point is marked as an absolutely no-entry zone. If normalized physical risk intensity Less than or equal to the hard safety threshold Then proceed to the next step of nonlinear fusion calculation; Obtain dynamic risk aversion coefficient Compared with the current normalized physical risk intensity And perform product calculation; With natural constant The risk decay factor is calculated by raising the negative product to the power of the base. Normalized semantic entropy value Multiplying the result by the risk attenuation factor yields the final comprehensive inspection utility output value. ; Output value reflects the comprehensive inspection effectiveness based on spatial location. The value range is strictly distributed in the interval (0,1).

[0054] S4. Extract local maxima points in the utility field to determine multiple dynamic inspection target points.

[0055] S5. Invert the values ​​of the utility field to construct the path cost field, and generate the optimal inspection route connecting the dynamic inspection target points based on the path cost field.

[0056] In the process of generating the optimal inspection route, the positioning reference point is corrected by using the gradient distribution of the utility field, and navigation reliability monitoring is performed based on the magnitude relationship between the first scalar field and the preset hard safety threshold.

[0057] S5 includes the use of particle swarm optimization algorithm to search for a six-dimensional parameter sequence that minimizes the path integral cost in the path cost field, so as to generate the optimal inspection route including position and attitude information. The six-dimensional parameter sequence includes coordinate parameters used to characterize the three-dimensional spatial position of the UAV, and pitch angle, roll angle and yaw angle parameters used to characterize the three-dimensional observation attitude of the UAV. Using particle swarm optimization as the solver, each particle represents a candidate six-dimensional path. During the iterative optimization process, the fitness function of the particle swarm is determined by the integral of the path in the global path cost field. In each iteration, the utility field gradient of the particle's current position is calculated and used as a correction to compensate for the particle's position parameters to counteract potential positioning drift. The repulsive potential field value of the point cloud space on the path is continuously monitored. Once the repulsive potential field value of the point cloud space equals the hard safety threshold, a navigation reliability monitoring alarm is triggered.

[0058] In this embodiment, the key parameters involved are as follows: The parameter sign of the six-dimensional state vector is set to It includes the normalized coordinates of the UAV in three-dimensional space and three-dimensional attitude angles, including pitch, roll, and yaw, and is the basic unit of particle swarm search. Obtain the physical boundary of the inspection area; scale the physical coordinates and angles to the (0,1) interval through linear mapping to generate a dimensionless search space; Furthermore, in order to eliminate the inconsistency between the position and angle dimensions when performing vector summation of particle velocity and position updates, this embodiment introduces a dimensional balance matrix. The specific update logic is as follows: the particle at the... The six-dimensional state vector of the generation Determined by the following formula: ; Among them, symbols This represents the Hadamard product, which is the product of corresponding elements. The algebraic representation of particles; It is represented as a six-dimensional incremental vector containing three-dimensional position components and three-dimensional attitude components; It is represented as a dimensional balance matrix, specifically a 1×6 diagonal vector, which translates the change of angle into an equivalent change of length, thereby unifying the dimensions of all quantities; its first three elements are preferably 1.0; the last three elements are preferably equivalent rotation radius coefficients, and the attitude angle part is converted into position distance by multiplying it with the equivalent rotation radius coefficient; In this embodiment, the equivalent rotation radius coefficient is preferably set to 1.25; based on the typical field of view and observation distance of the onboard camera of the inspection drone, the 1 radian angle deviation is mapped to the projected displacement deviation generated on the surface of the target object; in this invention, the preferred range of the equivalent rotation radius coefficient is between 0.8 and 2.5.

[0059] The parameter sign of the path integral cost value is set to , which represents the cumulative risk and cost of the entire candidate route, is the fitness function value of particle swarm optimization; Decompose the candidate path into equally spaced micro-segments; obtain the value of the endpoint of each micro-segment in the path cost field; sum up the cost values ​​of all micro-segments; multiply by the length of the micro-segment to obtain the total cost of the entire route; Specifically, path integral cost value In order to accurately approximate the cost of a continuous path in a discrete three-dimensional voxel field, this embodiment adopts a discrete accumulation model based on sampling step size compensation; its mathematical expression is as follows: ; in, Represented as a physical definition on the path, the first The three-dimensional spatial coordinates of a discrete sampling point; The physical definition is a sampling point. The original path cost value of the voxel is obtained by inverting the utility field. It is a dimensionless quantity, and its preferred range is (0,1). This is expressed as the physical definition of the Euclidean distance step size between adjacent sampling points; in this embodiment, It is preferable to set it to 0.5 times the voxel side length to ensure that no risk features are missed when the path passes through the voxel; The physical definition is the attitude transformation penalty term; in this embodiment, the attitude transformation penalty term is preferably the absolute value of the second-order difference of the angle component in the six-dimensional state vector between adjacent sampling points, which is used to suppress violent attitude swings in the flight path. Physically defined as the total number of sampling points along the path, calculated by dividing the total path length by the step size. Then round up to confirm; By introducing a sampling step size anchored to voxel resolution Couple the attitude penalty term during accumulation. This avoids the obstacle omission mechanism caused by discrete sampling, thereby improving the physical feasibility and flight stability of the generated route in complex post-disaster environments.

[0060] The parameter sign of the utility field gradient vector is set to The direction of the vector with the fastest rate of change of the utility field at a certain point in space is used to characterize the guiding force of security and value. Obtain the utility values ​​of adjacent voxels of the target point along the three-dimensional axes within the utility field; calculate the difference between the utility values ​​of positive and negative adjacent points along each axis; divide the difference by twice the voxel spacing to obtain the gradient components of the axes; synthesize the three-dimensional vector.

[0061] The parameter sign of the environment-guided weight coefficient is set to The scaling factor used to adjust the degree of influence of the utility field gradient on the particle trajectory is a fixed value, determined through offline simulation experiments by balancing convergence speed and quality search. In this embodiment, The preferred value is 1.5.

[0062] The parameter sign of the particle inertia weight is set to A coefficient that controls the ability of particles to maintain their original state of motion, used to balance global search and local exploitation; A strategy of linearly decreasing with the number of iterations is adopted to encourage global exploration in the early stage of optimization and promote local convergence in the later stage.

[0063] Calculation formula: ,in This represents the current iteration number. This represents the total number of iterations. For maximum inertia weight, The minimum inertia weight; in this embodiment, the maximum inertia weight. Set to 0.9, minimum inertia weight Set to 0.4.

[0064] The parameter symbol for the safety margin monitoring threshold is set to A physical risk threshold used to trigger navigation reliability monitoring alerts; Safety margin monitoring threshold With hard safety threshold Correlation serves as an early warning line; Calculation formula: ,in This is a preset safety margin coefficient used to set a buffer space for early warning, with a value within the range of (0,1). In this embodiment, it is designed for complex environments such as post-disaster power grids. The preferred value is 0.9, combined with the previously determined hard safety threshold. The value is 0.75. Substituting this value into the formula, we can obtain the safety margin monitoring threshold in this embodiment. The final calculated value is 0.675, that is... During the route optimization process, once the value of the first scalar field at the current position of a particle reaches 0.675, a navigation reliability monitoring alarm is immediately triggered.

[0065] In this embodiment, the steps for generating the optimal inspection route are as follows: Input the generated utility field and the corresponding path cost field, the initial particle swarm size (in this embodiment, it is set to 50 particles), and the preset number of iterations; The six-dimensional state vectors of the initial positions of 50 particles are randomly generated in a six-dimensional normalized space. and the initial velocity vector; For each path represented by a particle, the path integral cost is calculated using Riemannian sum and numerical integral logic. ; Check whether the first scalar field value at any point on the path exceeds the safety margin monitoring threshold. ; If the value exceeds the limit, the fitness value of the particle is set to the maximum value, and the navigation reliability monitoring obstacle avoidance logic correction is triggered. Obtain the current individual optimal position and the group optimal position of the particle; Obtain the utility field gradient vector at the current position of the particle. The gradient vector of the utility field Multiplied by the environmental guidance weighting factor ; Multiply the current velocity by the particle inertia weight. The optimal traction force of individuals, the optimal traction force of the group, and the environmental guiding momentum are vector summed. The previous position and the updated velocity are vector-summed to obtain a new six-dimensional state vector. ; After a preset number of iterations, the six-dimensional parameter sequence represented by the particle with the smallest fitness value is the optimal inspection route.

[0066] To verify the superiority of the gradient-guided adaptive particle swarm optimization algorithm proposed in this invention over existing technologies, a MATLAB-based simulation environment was built in this embodiment. A typical scenario was simulated: a power transmission tower partially collapsed after a disaster, accompanied by complex obstacles (broken conductors, fallen trees). The method of this invention was compared with a traditional particle swarm optimization algorithm using a fixed risk coefficient. The simulation data are shown in Table 3 below.

[0067] Table 3: Comparison of Route Planning Performance under Different Positioning Integrity Scenarios

[0068] The Reliable Inspection Efficiency Index is calculated by combining three core output indicators. The calculation logic is as follows: obtain the coverage rate of key information points; obtain the path integral cost value and the reciprocal of the minimum safe distance; finally, calculate the product of the key information point coverage rate and the latter two. It is used to comprehensively quantify the information acquisition efficiency of inspection tasks under the constraints of high safety and low cost. The higher the value, the better the overall performance of route planning.

[0069] Scenario 1: Under these ideal conditions, the method of the present invention reduces the path integral cost by 12.7% compared to the control group, while increasing the minimum safe distance by 14.3%; indicating that the gradient guidance mechanism of the present invention can find a safe path more efficiently; the reliability inspection efficiency index of the present invention is 1.63, which is 41.7% higher than that of the control group (1.15).

[0070] Scenario 2: This is a key scenario for verifying the core inventiveness of the invention. When the positioning integrity index drops to 0.5, the invention adaptively increases the dynamic risk aversion coefficient to 4.00, demonstrating strong environmental adaptability. The minimum safe distance remains at 2.9 meters, compared to the control group's dangerous distance of 1.6 meters due to the inability to perceive the deterioration in positioning accuracy, representing an 81.3% improvement in safety. More importantly, this safety guarantee does not come at the expense of task efficiency. Data shows that the key information point coverage rate of the invention's method (91.5%) is still higher than that of the control group (82.3%), achieving a dual improvement in safety and efficiency. The invention's reliable inspection efficiency index is as high as 1.41, while the control group's is only 0.53, with a comprehensive efficiency 2.66 times higher, fully demonstrating the technical superiority of the invention in complex and harsh environments.

[0071] This not only proves the technical feasibility of the present invention, but more importantly, it quantitatively reveals the significant progress of the present invention in flight safety and overall mission effectiveness compared to existing technologies in harsh environments through positioning-planning control.

[0072] Figure 1The main view on the left depicts a complex post-disaster inspection scenario, in which drones are operating around damaged power transmission facilities. The abstract field lines and contour lines superimposed in the image are a visual representation of the core technical concepts of this invention. Specifically, the sparse field lines surrounding ground obstacles and damaged tower structures intuitively represent the first scalar field (environmental risk field) constructed according to S1 to characterize physical risk. The densely distributed field lines at key connection points or obstructed areas of the power transmission facilities represent the second scalar field constructed according to S2 to characterize information uncertainty. The smooth optimal trajectory followed by the drone is calculated by nonlinearly fusing the two fields according to S3 to generate a unified utility field, and finally calculated in S5 based on the path cost field constructed by negating the utility field.

[0073] Figure 1 The technology roadmap on the right clearly reveals the core logical steps for realizing intelligent inspection routes. Among them, the construction of the environmental risk field corresponds to S1; the construction of the information uncertainty field corresponds to S2; the fusion of the risk-benefit utility field corresponds to S3; and the generation of dynamic targets and routes corresponds to the combination process of determining dynamic inspection target points in S4 and generating the optimal inspection route in S5. The entire diagram illustrates how, starting from environmental perception, through the construction and fusion of multi-dimensional information fields, an intelligent inspection route with both high reliability and high information acquisition efficiency can be generated.

[0074] It should be noted that all calculation formulas in this invention employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected relevant parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas of this invention, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, Min-Max Normalization and Z-Score standardization. The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of this invention.

[0075] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for post-disaster inspection of power transmission lines based on 3D point cloud semantic segmentation and route generation, characterized in that, The specific steps include: S1. Acquire three-dimensional point cloud data of power transmission facilities and perform semantic segmentation to generate a first scalar field to characterize the distribution of physical risks in the inspection space. S2. Based on the confidence level of semantic segmentation, generate a second scalar field to characterize the distribution of information uncertainty in the inspection space; S3. Through a preset utility function, the first scalar field and the second scalar field are nonlinearly fused to generate a utility field that uniformly represents the comprehensive inspection value of each point in the inspection space. S4. Extract local maxima points in the utility field to determine multiple dynamic inspection target points; S5. Invert the values ​​of the utility field to construct the path cost field, and generate the optimal inspection route connecting the dynamic inspection target points based on the path cost field.

2. The power transmission post-disaster inspection method based on 3D point cloud semantic segmentation and route generation according to claim 1, characterized in that, S1 includes identifying point cloud clusters with risk semantic labels in 3D point cloud data, and calculating the spatial repulsive potential field of the point cloud based on the distribution density of the point cloud clusters, as the first scalar field.

3. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 2, characterized in that, S2 includes calculating the information entropy of the point cloud classification probability distribution within each voxel in the 3D point cloud data, generating a voxel semantic entropy field as the second scalar field.

4. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 3, characterized in that, The steps for generating the first scalar field are as follows: Input 3D point cloud data containing risk semantic labels and risk impact weight coefficients; A density-based spatial clustering algorithm with noise is used to extract point clouds with the same risk semantic label into independent target point cloud clusters; Calculate the Euclidean distance between the voxel center and the target cloud cluster center; divide the square of the Euclidean distance by twice the square of the Gaussian kernel width parameter to obtain the intermediate exponent; calculate the natural constant. The negative intermediate exponent term raised to the power of; Multiply the negative intermediate index term by the risk impact weighting coefficient; The calculation results are then nonlinearly mapped using the Sigmoid function. We obtain a smooth and second-differentiable first scalar field.

5. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 4, characterized in that, The steps for generating the second scalar field are as follows: Input the posterior probability vector of each voxel for its class; Obtain the probability value of each category in the category posterior probability vector; calculate the base-2 logarithm for each probability value greater than 0; multiply each probability value by the corresponding logarithm and take the negative number; sum the product results of all categories to obtain the original information entropy of the voxel; Obtain the original information entropy of the 26 neighboring voxels in the current voxel neighborhood; Calculate the arithmetic mean of the original information entropy of 26 neighboring voxels; then sum the original information entropy of the current voxel with the average entropy of its neighborhood, with the weights determined by a preset spatial consistency correction factor. Divide the weighted summation result by the preset maximum theoretical entropy value; We obtain a second scalar field that reflects cognitive uncertainty.

6. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 5, characterized in that, S3 includes a first scalar field value at the target point exceeding a hard safety threshold, then the utility field value at the target point is configured as a prohibited value; if the first scalar field value at the target point does not exceed the hard safety threshold, then the utility field value at the target point is calculated based on the product of the second scalar field value at the target point and the first scalar field value at the target point using an exponential decay function.

7. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 6, characterized in that, The steps for generating the comprehensive inspection utility output value are as follows: Real-time acquisition of normalized physical risk intensity, normalized semantic entropy value, and navigation and positioning integrity index; Obtain the preset basic risk coefficient; The dynamic risk aversion coefficient is obtained by calculating the quotient of the basic risk coefficient and the navigation and positioning integrity index. Compare the magnitudes of normalized physical risk intensity and hard safety threshold. If the normalized physical risk intensity is greater than the hard safety threshold, the comprehensive inspection utility output value will be directly set to 0; and the area will be marked as an absolutely prohibited zone. If the normalized physical risk intensity is less than or equal to the hard safety threshold, then proceed to the next step of nonlinear fusion calculation; Obtain the dynamic risk aversion coefficient and the current normalized physical risk intensity, and perform the product calculation; With natural constant The risk decay factor is calculated by raising the negative product to the power of the base. The normalized semantic entropy value is multiplied by the risk decay factor to obtain the final comprehensive inspection utility output value. The output reflects the comprehensive inspection effectiveness value of spatial location.

8. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 7, characterized in that, During the generation of the optimal inspection route, the positioning reference point is corrected by the gradient distribution of the utility field, and navigation reliability monitoring is performed based on the magnitude relationship between the first scalar field and the preset hard safety threshold.

9. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 8, characterized in that, S5 includes the use of particle swarm optimization algorithm to search for a six-dimensional parameter sequence that minimizes the path integral cost in the path cost field, so as to generate the optimal inspection route including position and attitude information. The six-dimensional parameter sequence includes coordinate parameters used to characterize the three-dimensional spatial position of the UAV, and pitch angle, roll angle, and yaw angle parameters used to characterize the three-dimensional observation attitude of the UAV.

10. The power transmission post-disaster inspection method based on three-dimensional point cloud semantic segmentation and route generation according to claim 9, characterized in that, The steps for generating the optimal inspection route are as follows: Input the generated utility field and the corresponding path cost field, and the initial particle swarm size; In a six-dimensional normalized space, randomly generate the six-dimensional state vector and initial velocity vector of the initial position of 50 particles. For each path represented by a particle, the path integral cost is calculated using Riemannian and numerical integral logic. Check whether the first scalar field value at any point on the path exceeds the safety margin monitoring threshold; If the value exceeds the limit, the particle's fitness value will be set to the maximum value, and the navigation reliability monitoring obstacle avoidance logic correction will be triggered. Obtain the current individual optimal position and the group optimal position of the particle; Obtain the utility field gradient vector at the particle's current position, and multiply the utility field gradient vector by the environment guidance weight coefficient; Multiply the current velocity by the particle inertial weight, the individual optimal traction force, the group optimal traction force, and the environmental guiding momentum, and then sum them as a vector. The previous position and the updated velocity are vector-summed to obtain a new six-dimensional state vector. After a preset number of iterations, the six-dimensional parameter sequence represented by the particle with the smallest fitness value is the optimal inspection route.

Images