Power inspection path optimization system based on dynamic planning algorithm
By optimizing the selection of candidate inspection points and path generation, the problem of inaccurate paths caused by false and omitted points in power inspection path planning was solved, achieving efficient and safe power inspection path optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-03-10
AI Technical Summary
Existing power line inspection route planning systems are prone to inaccurate route planning in complex environments due to false or missed locations, resulting in high computational complexity and impacting efficiency and safety.
By combining ant colony optimization and dynamic programming methods with ghost risk mapping, node compression, edge cost construction, backfilling selection, and open-path dynamic programming modules, the selection of candidate inspection points and path generation are optimized to ensure the accuracy and efficiency of the path.
It effectively reduces false locations, controls the scale of calculations, generates the shortest open path, and improves the safety and efficiency of power line inspection.
Smart Images
Figure CN121632151A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of path planning, more particularly, it relates to a dynamic programming algorithm in a power inspection path optimization system. BACKGROUND
[0002] With the popularization of high-resolution imaging technologies such as unmanned aerial vehicles and satellite remote sensing, power inspection has gradually transformed from traditional manual reconnaissance to automation and intelligence. Existing inspection path planning systems usually extract inspection nodes such as transmission towers and insulators from aerial or remote sensing images through image recognition algorithms, and then generate an optimal inspection path with the help of a dynamic programming algorithm. Dynamic programming algorithms are widely used in this field due to their global optimal solution characteristics. However, the performance of such algorithms is highly dependent on the accuracy of the input node set. Any errors or redundant nodes can cause state space expansion, resulting in exponential growth in computational complexity.
[0003] In actual inspection scenarios, the inspection area often faces complex environmental influences. Changes in light can cause shadows and clouds in images and high-reflectance areas. Terrain undulations and meteorological disturbances further exacerbate image brightness mutations. Meanwhile, real targets such as vegetation, ground objects, and line fittings exhibit high-frequency texture differences. Traditional inspection point identification algorithms mainly rely on brightness thresholds and edge gradients to determine targets, lacking effective capabilities to distinguish between brightness changes and real textures. When brightness changes dominate image regions, algorithms are prone to misjudging shadow boundaries as inspection targets, generating false point positions without actual significance. In areas with uneven lighting or strong reflections, edge signals can be obscured by background interference, preventing algorithms from accurately identifying real targets and resulting in missed point positions.
[0004] The presence of false and missed point positions directly undermines the input stability of dynamic programming algorithms. Excessive false point positions can form invalid path branches in the state transition matrix, while missed point positions cause path planning to deviate from actual inspection requirements. The combined effects of both significantly prolong path solution times and may generate infeasible paths that traverse terrain obstacles or exceed equipment operating ranges, severely impacting the efficiency and safety of power inspection. Existing technologies often filter nodes through empirical thresholds or position clustering, making it difficult to address node identification issues caused by image quality defects at the root. Therefore, there is an urgent need for targeted technical solutions to break through this problem. SUMMARY
[0005] The present application provides a dynamic programming algorithm in a power inspection path optimization system, which solves the technical problems in the background art.
[0006] The present application provides a dynamic programming algorithm in a power inspection path optimization system, which solves the technical problems in the background art. The ghost risk mapping module is used to acquire candidate inspection points in aerial images, decompose brightness information within a fixed window of each candidate inspection point, obtain the intensity of illuminance change and texture energy, and calculate the illuminance-texture decoupling ratio accordingly, which is then mapped to ghost risk through an empirical distribution function. The node compression module is used to calculate the nearest neighbor distance of each candidate inspection point, take the median as the edge radius to divide the candidate inspection point into several connected components, and retain only the representative point with the largest illumination texture decoupling ratio in each connected component to obtain the compressed node set. The edge cost construction module is used to calculate the average endpoint risk of any two points in the compressed node set, and obtain the shadow belt integral penalty through the ghost risk space integral, and construct the edge cost by combining the scale balancing coefficient. The backfill selection module is used to form a backfill pool from the candidate inspection points to be compressed and removed. Using the illumination-texture decoupling ratio as a heuristic value, it selects a backfill set of a limited size through the ant colony algorithm and merges it with the compressed node set to form the final node set. The open path dynamic programming module is used to generate the shortest open path containing all nodes by using a subset enumeration type dynamic programming method based on the final set of nodes and edge costs, following the open path Hamiltonian path rule, and constraining the specified start and end points. The total path duration calculation module is used to map the average risk of the endpoints to the speed contraction ratio through an adaptive scaling factor, and calculate the travel time of each side and summarize it into the total path duration.
[0007] Furthermore, a fixed window is set for each candidate inspection point to obtain the brightness channel of the aerial image. The brightness channel is then convolved and smoothed using a Gaussian low-pass filter. The gradient of the smoothed result is calculated, and the first norm of the gradient is calculated. All first norm gradients are summed within the fixed window, and the summation result is averaged using the number of pixels in the fixed window to obtain the intensity of the illumination change. The brightness channel is then convolved using a high-pass filter to extract high-frequency components. The square value of the high-frequency component is calculated, and the square values of all high-frequency components are summed within the fixed window. The summation result is averaged using the number of pixels in the fixed window to obtain the texture energy.
[0008] Furthermore, using texture energy as the numerator and the sum of illuminance variation intensity and numerical stability term as the denominator, the illuminance-texture decoupling ratio is obtained by dividing the numerator by the denominator. Logarithmic operation is performed on each illuminance-texture decoupling ratio to obtain a logarithmic normalization result. All logarithmic normalization results are collected to calculate the empirical distribution function. The empirical distribution function value corresponding to each logarithmic normalization result is subtracted from the numerical value to obtain the ghost risk.
[0009] Furthermore, for each candidate inspection point, the Euclidean distance between the candidate inspection point and all other candidate inspection points is calculated, and the minimum value is selected as the nearest neighbor distance of the candidate inspection point. The nearest neighbor distances of all candidate inspection points are collected, and the median is taken as the edge radius. A neighborhood graph is constructed using the set of candidate inspection points as the vertex set. When the Euclidean distance between two candidate inspection points is less than or equal to the edge radius, an edge is established between the two candidate inspection points, and the neighborhood graph is divided to obtain all connected components. In each connected component, the candidate inspection point with the largest illumination texture decoupling ratio is selected as the representative point of the connected component, and all representative points are collected to obtain a compressed node set.
[0010] Furthermore, the Euclidean distance between any two points in the compressed node set is calculated as the geometric distance, and the average ghost risk of any two points in the compressed node set is calculated as the endpoint average risk. On the straight line segment connecting any two points, multiple sampling points are selected at equal intervals, the arc length interval between adjacent sampling points is determined, and the trapezoidal method is used to calculate the shadow zone integral penalty. Specifically, the ghost risk field values of each adjacent sampling point are added together, multiplied by half of the arc length interval, and then all results are summed. The median of the geometric distance is divided by the sum of the median of the shadow zone integral penalty and the numerical stability term to obtain the scale balancing coefficient. The product of the geometric distance between the two nodes, the scale balancing coefficient, and the shadow zone integral penalty is added together to obtain the edge cost.
[0011] Furthermore, a heuristic value is assigned to each point in the replenishment pool. The heuristic value is equal to the illumination-texture decoupling ratio corresponding to that point. The pheromone is initialized for each point in the replenishment pool, and a scaling factor with a value between 0 and 1 is set. The upper limit of the scale is calculated by multiplying the scaling factor by the number of elements in the compressed node set and rounding up. The replenishment set is constructed by multiple rounds of sampling using the ant colony algorithm. In each round of sampling, a non-replacement selection method is used to select elements from the unselected points in the replenishment pool in turn until the number of elements in the replenishment set reaches the upper limit of the scale.
[0012] Furthermore, in each round of element selection, the selection probability is calculated. The selection probability is calculated as the power of the pheromone weight of the current candidate point multiplied by the power of the heuristic value of that point, then divided by the sum of the powers of the pheromone weights of all unselected points multiplied by the powers of their respective heuristic values; both weights are greater than zero. The set of nodes to be evaluated is determined as the union of the compressed node set and the current replenishment set. A subset enumeration dynamic programming method is called, using the edge cost set as the graph weight, following the open-path Hamiltonian rule, to calculate the total path cost. A value ranging from 0 to 1 is set. The volatile coefficient is used to update the pheromone for each point in the replenishment pool. The update method is to multiply the original pheromone by the difference between one and the volatile coefficient, and then add the volatile coefficient multiplied by the pheromone increment. The pheromone increment is determined as follows: if the point belongs to the current replenishment set, the increment is the reciprocal of the total path cost; if the point does not belong to the current replenishment set, the increment is zero. The operation of sampling and constructing replenishment sets, calculating the total path cost, and updating pheromones is repeated multiple times. The replenishment set with the minimum total path cost is selected as the optimal replenishment set. The final node set is the union of the compressed node set and the optimal replenishment set.
[0013] Furthermore, a start point and an end point are specified, both belonging to the final node set. The path visits each node only once and does not return to the start point. A minimum cost table is defined, where each entry corresponds to a subset of the final node set containing the start point and an end node in that subset. The entry value is the minimum cost of visiting all nodes in the subset once from the start point and ending at the end node. The minimum cost table is initialized, with only entries corresponding to the start point in the subset having the start point as the end node set to zero, and all other undefined minimum cost table entries set to maximum values. For all subsets and end node combinations containing the start point but whose end nodes are not the start point, the corresponding minimum cost table entry value is calculated. The calculation method is as follows: in the subset after removing the end node, each node is selected as the predecessor node, and the minimum cost table entry value corresponding to the predecessor node is added to the edge cost from the predecessor node to the current end node. The minimum value among all the sums is taken as the value of the current entry.
[0014] Furthermore, the predecessor node of each minimum cost table entry is recorded. The predecessor node is the node corresponding to the minimum value obtained when calculating the current minimum cost table entry value. The overall minimum cost is calculated by selecting each node in the final node set as a temporary terminal node, adding the minimum cost table entry value corresponding to that node to the edge cost from that node to the destination, and taking the minimum value among all the sums as the overall minimum cost. The optimal terminal node is determined. The optimal terminal node is the temporary terminal node corresponding to the minimum value obtained when calculating the overall minimum cost. Based on the predecessor node record, starting from the final node set and the entry corresponding to the optimal terminal node, backtracking backwards, all predecessor nodes are found in sequence to form a node sequence containing the shortest open path of all final nodes.
[0015] Further, the median of the average endpoint risk is divided by one and the sum of the medians to obtain the adaptive scaling factor. The node sequence of the shortest open path is traversed to determine each edge formed by two adjacent nodes on the path. For each edge on the path, the average endpoint risk corresponding to that edge is calculated. The product of the adaptive scaling factor and the average endpoint risk is subtracted by one, and the result is multiplied by the nominal speed to obtain the travel speed of that edge. The edge cost of each edge is divided by the travel speed of that edge to obtain the travel time of that edge. The travel times of all edges on the path are summed to obtain the total time of the entire path.
[0016] The beneficial effects of this invention are as follows: This invention decomposes the brightness information within a fixed window of candidate inspection points to obtain the intensity of illuminance changes and texture energy, calculates the illuminance-texture decoupling ratio, and then maps it to ghost risk, accurately distinguishing between brightness interference and real texture, reducing the generation of false points; it calculates the nearest neighbor distance of candidate inspection points and takes the median as the edge radius, retaining high-value representative points after dividing connected components, controlling the computational scale; it forms a replenishment pool of candidate inspection points that have been compressed and removed, and selects high-value replenishment points using an ant colony algorithm with the illuminance-texture decoupling ratio as a heuristic value, ensuring a complete and high-quality node set. Simultaneously, it constructs edge costs by combining geometric distance, shadow zone integral penalty, and scale balancing coefficients to comprehensively reflect path risks; it generates the shortest open path containing all nodes through subset enumeration-type dynamic programming following the open-path Hamiltonian rule, avoiding invalid or deviated paths; and it adapts the travel speed based on the average risk of endpoints to accurately calculate the total path time. Ultimately, it solves the problems of invalid path planning, drastically increased solution time, and route deviation, improving the safety, feasibility, and efficiency of power line inspection. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of a dynamic programming algorithm of the present invention applied to a power line inspection path optimization system.
[0018] In the diagram: Ghost Risk Mapping Module 101, Node Compression Module 102, Edge Cost Construction Module 103, Backfill Selection Module 104, Open Path Dynamic Programming Module 105, Total Path Duration Calculation Module 106. Detailed Implementation
[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.
[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present 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 one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term 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.
[0021] like Figure 1 As shown, a dynamic programming algorithm is used in a power line inspection path optimization system, including: Ghost risk mapping module 101 is used to acquire candidate inspection points in aerial images, decompose brightness information within a fixed window of each candidate inspection point, obtain the intensity of illuminance change and texture energy, and calculate the illuminance-texture decoupling ratio accordingly, and map it to ghost risk through an empirical distribution function. The node compression module 102 is used to calculate the nearest neighbor distance of each candidate inspection point, take the median as the edge radius to divide the candidate inspection point into several connected components, and retain only the representative point with the largest illumination texture decoupling ratio in each connected component to obtain a compressed node set. The edge cost construction module 103 is used to calculate the average risk of endpoints between any two points in the compressed node set, and obtain the shadow zone integral penalty through the ghost risk space integral, and construct the edge cost by combining the scale balancing coefficient. The backfill selection module 104 is used to form a backfill pool from the candidate inspection points to be compressed and removed. Using the illumination-texture decoupling ratio as a heuristic value, it selects a backfill set of a limited size through the ant colony algorithm and merges it with the compressed node set to form the final node set. The open path dynamic programming module 105 is used to generate the shortest open path containing all nodes by using a subset enumeration type dynamic programming method based on the final set of nodes and edge costs, following the open path Hamiltonian path rule, and constraining the specified start and end points. The total path duration calculation module 106 is used to map the average risk of the endpoints to the speed contraction ratio through an adaptive scaling factor, and calculate the travel time of each side and summarize it into the total path duration.
[0022] It should be noted that the aerial images are acquired by drones equipped with high-definition digital cameras. Before flight, the flight path needs to be planned based on the terrain and distribution of power facilities in the power inspection area. The flight altitude is controlled between 50 and 100 meters to ensure that the images can clearly show the structural details of the power equipment while covering a sufficient inspection range. The flight speed is maintained at 3 to 8 meters per second, with an image overlap rate of no less than 60% for adjacent flight strips and no less than 30% for images within the same flight strip, to avoid inspection blind spots. The initial candidate inspection points are planned based on preset information such as power facility design drawings and geographic information system data. The first flight of the drone is an exploratory acquisition of these initial candidate points, that is, to verify whether each point is affected by shadows and texture interference and whether it can effectively obtain equipment details through aerial images. After effective candidate points are selected and the path is optimized through indicators such as illumination and texture decoupling ratio, the second flight of the drone is for precise inspection, and detailed inspection is carried out on the finally determined nodes. This phased operation not only avoids inspection blind spots caused by invalid candidate points in a single flight, but also improves the accuracy of the path through early verification.
[0023] In one embodiment of the present invention, a fixed window is set for each candidate inspection point to obtain the brightness channel of the aerial image. The brightness channel is convolved and smoothed using a Gaussian low-pass filter. The gradient of the smoothed result is calculated, and the first norm of the gradient is calculated. All first norms of the gradients are summed within the fixed window, and the summed result is averaged using the number of pixels in the fixed window to obtain the intensity of the illumination change. The brightness channel is convolved using a high-pass filter to extract high-frequency components. The square value of the high-frequency component is calculated, and the square values of all high-frequency components are summed within the fixed window. The summed result is averaged using the number of pixels in the fixed window to obtain the texture energy.
[0024] Specifically, the intensity of the illuminance change at the i-th candidate inspection point The calculation formula is as follows: ,in This indicates that the i-th candidate inspection point corresponds to a fixed window. The number of pixels, Represents the gradient operator, This represents a Gaussian low-pass filter. This represents the brightness channel of the aerial image. This indicates convolution smoothing.
[0025] Specifically, the texture energy of the i-th candidate inspection point The calculation formula is as follows: , where H represents a high-pass filter.
[0026] It should be noted that the brightness channel of the aerial image is an image layer that independently carries information about the brightness of light. The brightness channel extracts the numerical features of light intensity separately and does not involve the differentiation of color differences. The fixed window adopts a square pixel matrix. Preferably, the fixed window is set to 11×11 pixels, which can fully cover the effective feature area around the candidate inspection point without including too many irrelevant background pixels. The standard deviation of the Gaussian low-pass filter is a preset parameter. Preferably, the standard deviation of the Gaussian low-pass filter is set to 1.2. Aerial images of power inspection are easily affected by slight Gaussian noise. Setting the standard deviation of the Gaussian low-pass filter to 1.2 can effectively suppress noise while preserving the key details of the illumination gradient. The high-pass filter adopts a Gaussian high-pass filter. This type of filter has a smooth frequency response transition, which can avoid the edge effect caused by an ideal high-pass filter and adapt to the continuous distribution characteristics of the texture of power equipment.
[0027] It should be noted that illuminance variation intensity is the average level of light fluctuations within a fixed window, used to reflect the degree of slow light change in the window area. The larger the value, the more obvious the alternation of light brightness and darkness in the area, and the more unstable the lighting environment. Texture energy is the cumulative intensity of high-frequency components of the image within a fixed window, used to reflect the strength of rapidly changing features such as surface details and edges of objects within the window area. These features are usually generated by structural details of power equipment, such as tower material connections and insulator outlines. The larger the value, the richer the structural details of objects in the area, and the more effective equipment information can be provided for inspection.
[0028] In one embodiment of the present invention, texture energy is used as the numerator, and the sum of illuminance variation intensity and numerical stability term is used as the denominator. The illuminance-texture decoupling ratio is obtained by dividing the numerator by the denominator. Logarithmic operation is performed on each illuminance-texture decoupling ratio to obtain a logarithmic normalization result. All logarithmic normalization results are collected to calculate the empirical distribution function. The empirical distribution function value corresponding to each logarithmic normalization result is subtracted from the numerical value to obtain the ghost risk.
[0029] It should be noted that the numerical stability term is a preset parameter. Preferably, the numerical stability term is set to 10 to the power of -6 to avoid the denominator being zero. The specific calculation process of the empirical distribution function is as follows: collect the log-normalized results of all candidate inspection points and arrange them in ascending order. For each log-normalized result, count the number of samples less than or equal to that result in the total number of results, divide that number by the total number of samples, and the ratio is the empirical distribution function value corresponding to that log-normalized result.
[0030] It should be noted that the illuminance-texture decoupling ratio is used to separate the effects of illuminance changes and texture features. This ratio can clearly indicate whether the change in light or the structural details of the object dominates within the window area. That is, the larger the ratio, the more significant the influence of the structural details of the object within the window area, and the better it reflects the true structure of the device. The smaller the ratio, the more prominent the influence of the light change, and the higher the degree to which the structural details of the object are obscured by the light change. The ghost risk ranges from 0 to 1. The larger the value, the higher the probability that the corresponding area is dominated by the light change. It is likely to be a shadow or a pseudo-texture, or other area that is meaningless for inspection. These areas cannot provide information about the device structure and may also interfere with the inspection judgment.
[0031] In one embodiment of the present invention, for each candidate inspection point, the Euclidean distance between the candidate inspection point and all other candidate inspection points is calculated, and the minimum value among them is selected as the nearest neighbor distance of the candidate inspection point; the nearest neighbor distances of all candidate inspection points are collected, and the median is taken as the edge radius; a neighborhood graph is constructed using the set of candidate inspection points as the vertex set; when the Euclidean distance between two candidate inspection points is less than or equal to the edge radius, an edge is established between the two candidate inspection points, and the neighborhood graph is divided to obtain all connected components; in each connected component, the candidate inspection point with the largest illumination texture decoupling ratio is selected as the representative point of the connected component, and all representative points are collected to obtain a compressed node set.
[0032] It should be noted that the spatial density of candidate points in the inspection scenario varies with the aerial photography altitude and the distribution of inspection targets. For example, there are more points in densely populated areas of poles and fewer points in open areas. The median can reflect the overall distribution characteristics of all points, avoiding the problem of a fixed radius resulting in an excessively large connected component in dense areas and a dispersed connected component in sparse areas. Obtaining connected components by dividing the nearest neighbor graph can be achieved through a connected component search algorithm, such as the breadth-first search algorithm. The specific calculation process is as follows: mark all candidate inspection points as unvisited, select an unvisited point as the starting point, and sequentially search for all points that have an edge with that point, i.e., the Euclidean distance is less than or equal to the radius of the edge, forming an initial set. Continue to search for the associated points of each point in the initial set and add them to the set until no new associated points can be added. This set is a connected component. Mark all points in this connected component as visited. Repeat the above operation until all points are marked to obtain all connected components.
[0033] It should be noted that compressing the node set significantly reduces the number of candidate points without losing high-value inspection points, effectively controlling the computational scale of subsequent dynamic programming and other steps; and it can adapt to different aerial photography altitudes and different target distribution densities without adjusting parameters, avoiding insufficient robustness caused by threshold stacking, which will not be elaborated here.
[0034] In one embodiment of the present invention, the Euclidean distance between any two points in the compressed node set is calculated as the geometric distance, and the average value of the ghost risk of any two points in the compressed node set is calculated as the endpoint average risk. On the straight line segment connecting any two points, multiple sampling points are selected at equal intervals, the arc length interval between adjacent sampling points is determined, and the shadow zone integral penalty is calculated using the trapezoidal method. Specifically, the ghost risk field values of each adjacent sampling point are added together, multiplied by half of the arc length interval, and then all results are summed. The median of the geometric distance is divided by the sum of the median of the shadow zone integral penalty and the numerical stability term to obtain the scale balancing coefficient. The product of the geometric distance between the two nodes, the scale balancing coefficient, and the shadow zone integral penalty is added together to obtain the edge cost.
[0035] Specifically, the formula for calculating the edge cost between node i and node j is as follows: ,in and Let i and j represent the geometric distance and the shadow integral penalty, respectively. This represents the scalar balancing coefficient.
[0036] It should be noted that the number of equidistant sampling points is a preset parameter. Too few sampling points will result in insufficient accuracy in the calculation of the shadow belt integral penalty, failing to accurately reflect the risk distribution in the middle of the path. Too many sampling points will increase the computational load and reduce the overall execution efficiency of the process. Preferably, the number of equidistant sampling points is set to 30. The ghost risk field value of the sampling point can be obtained by interpolation. If the sampling point happens to fall on a candidate inspection point, the ghost risk of that point is directly used. If the sampling point is between two candidate inspection points, linear interpolation is used to calculate the risk field value of the sampling point based on the distance between the sampling point and the surrounding known risk points. This will not be elaborated here. The arc length interval is determined by the total length of the straight line segment connecting the two points and the number of sampling points. First, the total length of the straight line segment is calculated, and then the total length is divided by the number of sampling points minus one to obtain the arc length interval between adjacent sampling points, ensuring that the sampling points are equidistantly distributed.
[0037] It should be noted that the shadow belt integral penalty is used to quantify the overall risk of the path crossing the shadow-dominated area, extending the point risk between nodes to the line risk, avoiding the problem of only considering the risk at the endpoints and ignoring the problem of the path crossing a large area of shadow, so that the edge cost can fully reflect the risk level of the inspection path; the scale balancing coefficient is used to balance the difference in the dimensions of geometric distance and shadow belt integral penalty, so that the weights of the two in the edge cost calculation are adapted to the current scenario.
[0038] In one embodiment of the present invention, a heuristic value is set for each point in the replenishment pool. The heuristic value is equal to the illumination-texture decoupling ratio corresponding to that point. The pheromone is initialized for each point in the replenishment pool. A scaling factor with a value between 0 and 1 is set. The upper limit of the scale is calculated by multiplying the scaling factor by the number of elements in the compressed node set and rounding up. The replenishment set is constructed by multiple rounds of sampling using an ant colony algorithm. In each round of sampling, a non-replacement selection method is used to select elements from the unselected points in the replenishment pool in turn until the number of elements in the replenishment set reaches the upper limit of the scale.
[0039] It should be noted that the illumination-texture decoupling ratio is used as a heuristic value for the ant colony algorithm, causing the ant colony to prioritize points with high texture value during selection. This deeply integrates the physical characteristics of the inspection scene with the heuristic mechanism of the ant colony algorithm, ensuring that the points replenished are all real texture-dominated points with high inspection value, avoiding the replenishment of ghost points caused by shadows. At the same time, the collective intelligence of the ant colony is used to achieve accurate replenishment within a small range. The initial pheromone is set to 0.1. Too large an initial pheromone may cause the algorithm to get stuck in early local selection, while too small an initial pheromone will result in insufficient exploration. The scaling factor is used to control the size of the replenishment set. Preferably, the scaling factor is set to 0.2, which ensures sufficient flexibility of the path and controls the complexity of subsequent calculations, adapting to most inspection scenarios.
[0040] In one embodiment of the present invention, the selection probability is calculated in each round of element selection. The selection probability is calculated by multiplying the weight of the pheromone of the current candidate point by the weight of the heuristic value of that point by the power of the weight of the pheromone of all unselected points by the power of the weight of their respective heuristic values, where both weights are greater than zero. The set of nodes to be evaluated is determined as the union of the compressed node set and the current replenishment set. A subset enumeration dynamic programming method is called, using the edge cost set as the graph weight, and following the open-path Hamiltonian path rule, the total path cost is calculated. A value in the range of 0 is set. For each point in the replenishment pool, the pheromone is updated using a volatile coefficient between 1 and 1. The update method is to multiply the original pheromone by 1 and subtract the difference from the volatile coefficient, and then add the volatile coefficient multiplied by the pheromone increment. The pheromone increment is determined as follows: if the point belongs to the current replenishment set, the increment is the reciprocal of the total path cost; if the point does not belong to the current replenishment set, the increment is zero. The operation of sampling, constructing replenishment sets, calculating the total path cost, and updating pheromones is repeated multiple times. The replenishment set with the minimum total path cost is selected as the optimal replenishment set. The final node set is the union of the compressed node set and the optimal replenishment set.
[0041] Specifically, the probability of selecting node u given the currently selected set S. The calculation formula is as follows: ,in and Let the pheromones of nodes u and v be represented respectively. and Let u and v represent the heuristic values of node u and node v, respectively. Indicates pheromone weights, Indicates heuristic value weights, It represents the set of remaining selectable nodes in the candidate node set U after excluding the already selected set S.
[0042] It should be noted that the linkage between pheromone updates and total path cost uses the reciprocal of the total path cost as the pheromone increment, allowing points in the replenishment set with smaller path costs to receive more pheromone. This reinforces the historical preference for high-quality selections. This linkage mechanism allows the algorithm to utilize pheromones from historical selections while continuously iterating and optimizing the total path cost through dynamic programming, preventing the ant colony algorithm from getting stuck in local optima and ultimately converging to the replenishment set that minimizes the path cost. The pheromone weight and heuristic value weight are preset parameters, controlling the influence of pheromones (historical selection preference) and heuristic values (illuminance-texture decoupling ratio) on ant colony selection, respectively. Preferably, the pheromone weight is set to 1 and the heuristic value weight is set to 2. The evaporation coefficient is used to control the pheromone forgetting rate. Preferably, the evaporation coefficient is set to 0.2, meaning that 80% of the pheromone is retained after each iteration. This allows the pheromone from historical high-quality selections to accumulate while preventing the algorithm from prematurely converging to local optima due to excessive accumulation of old pheromones.
[0043] It should be noted that the subset enumeration dynamic programming method (Held-Karp) solves the shortest path by enumerating all subsets of nodes that include the starting point and using dynamic programming recursion. It calculates the minimum cost of subsets of different sizes in turn and uses the results of the calculated smaller subsets to derive the results of the smaller subsets, thus avoiding brute-force enumeration of all paths and significantly reducing computational complexity. The open-path Hamiltonian rule represents a path that starts from a specified starting point, passes through all nodes to be inspected once and only once, and finally reaches the specified destination without returning to the starting point. This fits the actual work process of power inspection, which starts from the starting point, covers all nodes in one direction, and ends at the destination. The total cost of the path is the sum of the edge costs of all edges that make up the path. Finally, through dynamic programming recursion, the minimum total cost to reach the destination is selected from the states that include all nodes.
[0044] In one embodiment of the present invention, a start point and an end point are specified, both of which belong to the set of final nodes. The path visits each node only once and does not return to the start point. A minimum cost table is defined, where each entry in the minimum cost table corresponds to a subset of the set of final nodes containing the start point and an end node in that subset. The entry value is the minimum cost of visiting all nodes in the subset once from the start point and ending at the end node. The minimum cost table is initialized, with the entry values of only the subsets containing the start point and ending at the start point set to zero, and the values of all other undefined minimum cost table entries set to maximum values. For all subsets and end node combinations that contain the start point but whose end node is not the start point, the corresponding minimum cost table entry value is calculated. The calculation method is as follows: in the subset after removing the end node, each node is selected as the predecessor node, and the minimum cost table entry value corresponding to the predecessor node is added to the edge cost from the predecessor node to the current end node. The minimum value among all the sums is taken as the value of the current entry.
[0045] It should be noted that the minimum cost table is a structured table used to record the dynamic programming state. The value of each entry is the minimum cost of starting from the starting point, visiting all nodes of the subset once, and ending at the end node. Through this table, the optimal cost of subsets of different sizes can be derived step by step, and finally the shortest path cost covering all nodes can be obtained. The maximum value is a preset parameter. Preferably, the maximum value is set to 10 to the power of 9. The edge cost consists of geometric distance and risk penalty. In actual inspection scenarios, the total cost is much smaller than this value. Setting it to this value can ensure that unreached states will not be mistakenly selected in the recursion, while not affecting the stability of numerical calculation.
[0046] In one embodiment of the present invention, the predecessor node of each minimum cost table entry is recorded. The predecessor node is the node corresponding to the minimum value obtained when calculating the current minimum cost table entry value. The overall minimum cost is calculated by selecting each node in the final node set as a temporary terminal node, adding the minimum cost table entry value corresponding to the node to the edge cost from the node to the destination, and taking the minimum value among all the sums as the overall minimum cost. The optimal terminal node is determined. The optimal terminal node is the temporary terminal node corresponding to the minimum value obtained when calculating the overall minimum cost. Based on the predecessor node record, starting from the final node set and the entry corresponding to the optimal terminal node, backtracking is performed to find all predecessor nodes in sequence, forming a node sequence containing the shortest open path of all final nodes.
[0047] It should be noted that backtracking means starting from the entry corresponding to the optimal end node, recording that node first, then finding the previous node based on the predecessor node record, repeating this process until returning to the starting point, and finally reversing the node sequence to obtain the forward path sequence from the starting point to the end point.
[0048] In one embodiment of the present invention, the median of the average endpoint risk is divided by a sum of the medians to obtain an adaptive scaling factor. The node sequence of the shortest open path is traversed to determine each edge formed by two adjacent nodes on the path. For each edge on the path, the average endpoint risk corresponding to the edge is calculated. The product of the adaptive scaling factor and the average endpoint risk is subtracted by a, and the result is multiplied by the nominal speed to obtain the travel speed of the edge. The edge cost of each edge is divided by the travel speed of the edge to obtain the travel time of the edge. The travel times of all edges on the path are summed to obtain the total time of the entire path.
[0049] It should be noted that the adaptive scaling factor is calculated using the median of the average risk at all endpoints across the entire map. This is used to quantitatively convert risk intensity into a scaling ratio for travel speed, ensuring that the impact of risk on speed remains consistent across different inspection scenarios. This avoids over- or under-adjustment of speed due to scenario differences, achieving a proper match between risk and speed. The median of the average risk at endpoints must be calculated based on the average risk of all node pairs across the entire map, not just node pairs on the path. Similarly, the median of all node pairs reflects the overall risk level of the current inspection scenario, ensuring that the adaptive scaling factor adapts to the entire scenario and that speed adjustments on each side of the path conform to the global risk characteristics. To avoid adjustment deviations caused by local risks, the nominal speed is the standard travel speed of the inspection equipment under ideal environmental conditions with no risk. It is a preset parameter, and the preferred nominal speed is set to 5 meters per second. The system traverses the path edges and calculates the travel speed to match the speed of each edge with its own risk. The higher the risk, the slower the speed, so as to reserve sufficient inspection time for high-risk areas. In addition, the starting time is set to zero, and the travel time of the previous edge is added to the starting time to obtain the timestamp of the next node. This process is repeated until the end point. Finally, the combination of the node sequence and the corresponding timestamp is output, which makes it convenient for operators or equipment to execute in sequence and ensures that the inspection process is orderly and controllable.
[0050] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0051] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.
Claims
1. A dynamic programming algorithm in a power inspection path optimization system, characterized in that, The method comprises the following steps: A ghost risk mapping module is used to obtain candidate inspection points in the aerial image, decompose the luminance information in a fixed window of each candidate inspection point, obtain the illumination change intensity and the texture energy, and calculate the illumination-texture decoupling ratio according to the same, and map the ghost risk through an empirical distribution function; A node compression module is used to calculate the nearest neighbor distance of each candidate inspection point, take the median as the edge radius, divide the candidate inspection points into several connected components, and only keep the representative point with the largest illumination-texture decoupling ratio in each connected component to obtain a compressed node set; An edge cost construction module is used to calculate the end-point average risk of any two points in the compressed node set, and obtain the film integral penalty through the ghost risk space integral, and construct the edge cost in combination with the scale balance coefficient; A backfill selection module is used to form a backfill pool by the candidate inspection points removed by compression, take the illumination-texture decoupling ratio as the heuristic value, select a limited scale of backfill set through the ant colony algorithm, and merge the backfill set with the compressed node set to obtain a final node set; An open-loop dynamic programming module is used to generate a shortest open-loop path containing all nodes according to the final node set and the edge cost through a subset enumeration type dynamic programming method, follow the open-loop Hamilton path rule, and constrain the specified start and end points; A total path total time calculation module is used to map the end-point average risk to a speed contraction ratio through an adaptive scaling coefficient, calculate the travel time of each edge according to the same, and sum up the total path total time.
2. The dynamic programming algorithm in the power inspection path optimization system according to claim 1, wherein, A fixed window is set for each candidate inspection point, the luminance channel of the aerial image is obtained, a Gaussian low-pass filter is used for convolution smoothing processing on the luminance channel, the gradient of the smoothed result is solved, the norm of the gradient is calculated, the sum of all gradient norms in the fixed window is calculated, and the sum result is averaged by the number of pixels in the fixed window to obtain the illumination change intensity; a high-pass filter is used for convolution processing on the luminance channel to extract high-frequency components, the square value of the high-frequency components is calculated, the sum of all high-frequency component square values in the fixed window is calculated, and the sum result is averaged by the number of pixels in the fixed window to obtain the texture energy.
3. The dynamic programming algorithm in the power inspection path optimization system according to claim 1, wherein, The texture energy is taken as the numerator, and the sum of the illumination change intensity and the numerical stability term is taken as the denominator, and the illumination-texture decoupling ratio is obtained by dividing the numerator by the denominator. The logarithmic operation is performed on each illumination-texture decoupling ratio to obtain a logarithmic normalized result, the empirical distribution function is calculated by collecting all the logarithmic normalized results, and the numerical value is subtracted from the empirical distribution function value corresponding to each logarithmic normalized result to obtain the ghost risk.
4. The dynamic programming algorithm in the power inspection path optimization system according to claim 1, wherein, For each candidate inspection point, the Euclidean distance between the candidate inspection point and all other candidate inspection points is calculated, and the minimum value is selected as the nearest neighbor distance of the candidate inspection point; The nearest neighbor distances of all candidate inspection points are collected, and the median is taken as the edge radius; The candidate inspection point set is taken as the vertex set to construct a neighborhood graph, an edge is established between two candidate inspection points when the Euclidean distance between the two candidate inspection points is less than or equal to the edge radius, and the neighborhood graph is divided to obtain all connected components; in each connected component, the candidate inspection point with the largest illumination-texture decoupling ratio is selected as the representative point of the connected component, and all representative points are collected to obtain a compressed node set.
5. The dynamic programming algorithm in the power inspection path optimization system according to claim 1, wherein, The Euclidean distance between any two points in the compressed node set is calculated as the geometric distance, and the average of the ghost risk between any two points in the compressed node set is calculated as the endpoint average risk; On the straight line segment connecting any two points, a plurality of sampling points are selected with equidistant distribution, the arc length interval between adjacent sampling points is determined, and the trapezoidal method is used to calculate the band integral penalty. Specifically, the values of the ghost risk field of each adjacent sampling point are added, multiplied by half of the arc length interval, and then all the results are summed up. The median of the geometric distance is divided by the sum of the median of the band integral penalty and the numerical stability term to obtain the scale balancing coefficient; The product of the geometric distance of the two nodes and the product of the scale balancing coefficient and the band integral penalty is added to obtain the edge cost.
6. The dynamic programming algorithm in power inspection path optimization system of claim 1, wherein, Set a heuristic value for each point in the backfill pool, which is equal to the corresponding illumination texture decoupling ratio of the point. Initialize the pheromone of each point in the backfill pool. Set a proportion coefficient with a value ranging from 0 to 1. The upper limit of the calculation scale is the product of the proportion coefficient and the number of elements in the compressed node set, rounded up. Use the ant colony algorithm to construct the backfill set through multiple rounds of sampling. In each round of sampling, elements are selected from the unselected points in the backfill pool in turn without replacement until the number of elements in the backfill set reaches the upper limit of the scale.
7. The dynamic programming algorithm in the power inspection path optimization system according to claim 6, wherein, Calculate the selection probability when selecting elements in each round. The calculation method of the selection probability is to divide the weight power of the pheromone of the current candidate point by the weight power of the heuristic value of the point, and then divide by the sum of the weight power of the pheromone of all unselected points and the weight power of their respective heuristic values. Both weights are greater than zero. Determine the evaluation node set as the union of the compressed node set and the current backfill set. Call the subset enumeration type dynamic programming method to calculate the total path cost with the edge cost set as the graph weight and follow the open road Hamilton path rule. Set a volatility coefficient with a value ranging from 0 to 1. Update the pheromone of each point in the backfill pool. The update method is to multiply the original pheromone by one minus the difference of the volatility coefficient, and then add the product of the volatility coefficient and the pheromone increment. The pheromone increment is determined as follows: if the point belongs to the current backfill set, the increment is the reciprocal of the total path cost, and if the point does not belong to the current backfill set, the increment is zero. Repeat the operations of constructing the backfill set through multiple rounds of sampling, calculating the total path cost, and updating the pheromone. Select the backfill set with the smallest total path cost as the optimal backfill set, and the final node set is the union of the compressed node set and the optimal backfill set.
8. The dynamic programming algorithm in the power inspection path optimization system according to claim 1, wherein, Specify the starting point and the end point, both of which belong to the final node set. The path only visits each node once and does not return to the starting point. Define the minimum cost table. Each entry of the minimum cost table corresponds to a subset of the final node set containing the starting point and an end node in the subset, and the entry value is the minimum cost of visiting all nodes in the subset once from the starting point and ending with the end node. Initialize the minimum cost table, set the entry value of the subset containing the start point and ending in the start point to zero, and set the rest of the undefined minimum cost table entry values to the maximum value; for all subsets containing the start point and the end node is not the start point, calculate the corresponding minimum cost table entry value, the calculation method is to select each node in the subset after removing the end node as the predecessor node, add the minimum cost table entry value corresponding to the predecessor node and the edge cost from the predecessor node to the current end node, and take the minimum value of all addition results as the value of the current entry.
9. The dynamic programming algorithm in the power inspection path optimization system according to claim 8, wherein, Record the predecessor node of each minimum cost table entry, which is the node corresponding to the minimum value when calculating the current minimum cost table entry value; Calculate the overall minimum cost, the calculation method is to select each node in the final node set as a temporary end node, add the minimum cost table entry value corresponding to the node and the edge cost from the node to the end point, and take the minimum value of all addition results as the overall minimum cost; determine the optimal end node, which is the temporary end node corresponding to the minimum value when calculating the overall minimum cost; According to the predecessor node record, start from the entry corresponding to the final node set and the optimal end node, and backtrack in reverse to find all the predecessor nodes in turn to form the node sequence of the shortest open circuit path containing all the final nodes.
10. The dynamic programming algorithm in the power inspection path optimization system according to claim 1, wherein, Divide the median of the end point average risk by the sum of one plus the median to obtain an adaptive scaling coefficient, traverse the node sequence of the shortest open circuit path, determine the travel speed of each edge composed of two adjacent nodes on the path, calculate the end point average risk corresponding to each edge on the path, subtract the product of the adaptive scaling coefficient and the end point average risk from one, and then multiply the result by the nominal speed to obtain the travel speed of the edge; divide the edge cost of each edge by the travel speed of the edge to obtain the travel time of the edge; accumulate the travel time of all edges on the path to obtain the total path time.