Cable path planning method based on deep learning and multi-objective optimization

Abstract

The invention relates to the technical field of path planning, in particular to a cable path planning method based on deep learning and multi-objective optimization, and the method comprises the following steps: based on remote sensing image data, employing a multi-scale semantic segmentation deep learning model, combining with an auxiliary accessibility tagging sample, and recognizing the type of each ground feature; constructing a ground feature accessible confidence field with continuous spatial distribution; constructing a confidence coupling path cost function, endowing a penalty weight to an area with low ground feature confidence, endowing a passage reward weight to an area with high confidence, and generating a confidence coupling path cost function for path search; and constructing an improved heuristic function, and executing path search by adopting a multi-objective optimization algorithm to obtain a cable path scheme. According to the method, the historical total cost of the current node is considered, and the average confidence risk from the current node to the target direction is predicted, so that prospective avoidance is realized.

Description

Technical Field

[0001] This invention relates to the field of route planning technology, and in particular to a cable route planning method based on deep learning and multi-objective optimization. Background Technology

[0002] With the continuous advancement of intelligent infrastructure construction in urban and rural areas, the demand for cable route planning is increasing due to power communication, power grid expansion, and intelligent upgrading of industrial parks. Especially in complex environments such as urban fringe, mountainous areas, farmland, and wetlands, how to scientifically plan cable routes while ensuring construction feasibility, reducing construction costs, and mitigating environmental risks has become one of the core issues in engineering practice.

[0003] Current cable route planning methods primarily rely on engineers manually designing plans based on topographic maps, satellite imagery, or field survey data, supplemented by simple shortest path algorithms for auxiliary calculations. Traditional route planning methods often prioritize the shortest geographical distance, neglecting factors such as accessibility, geological stability, and environmental sensitivity involved in construction. This results in routes that, while achieving the shortest distance, traverse areas unsuitable for construction or with extremely high construction costs, exhibiting the problem of being "geometrically optimal but technically infeasible." While some existing improved methods introduce slope factors or land cover type costs, they are mostly static weighted models that fail to dynamically reflect regional uncertainties or differences in risk levels, lacking adaptive adjustment capabilities and struggling to address the need for multi-objective trade-offs. Conventional heuristic search algorithms only consider geometric distance estimation and cannot predict the risk distribution trends ahead of the route. This can lead to routes mistakenly entering large areas unaccessible for construction during the search process, increasing computational redundancy and affecting the quality of results. Summary of the Invention

[0004] This invention provides a cable path planning method based on deep learning and multi-objective optimization, which is an intelligent cable path planning method that integrates remote sensing analysis and multi-objective optimization mechanism and has risk adaptive avoidance capabilities.

[0005] A cable path planning method based on deep learning and multi-objective optimization includes the following steps: S1, Ground Feature Accessibility Confidence Field Generation Step: Based on remote sensing image data, a multi-scale semantic segmentation deep learning model is adopted, combined with auxiliary accessibility annotation samples, to identify the types of ground features in each region, and to calculate the accessibility confidence of each pixel for cable construction, thereby constructing a ground feature accessibility confidence field with continuous spatial distribution. The accessibility confidence field is used to express the credibility of the accessibility of the ground area for construction in remote sensing identification and the uncertainty level of potential risk areas. S2, Construction of Confidence-Coupled Path Cost Function: Based on the accessibility confidence field of the terrain features, a confidence adjustment factor is defined and introduced into the multi-objective cost modeling process of path optimization to construct a confidence-coupled path cost function. The cost function integrates path length, construction cost, terrain slope, and confidence inverse factor, and assigns penalty weights to areas with low terrain feature confidence and passage reward weights to areas with high confidence, generating a confidence-coupled path cost function for path search. S3, Optimized path search based on the confidence-coupled path cost function: The confidence-coupled path cost function is used as the input for path search. An improved heuristic function is constructed, and a multi-objective optimization algorithm is used to perform path search to obtain a cable path scheme that meets the requirements of construction continuity, risk avoidance and path smoothness. After smoothing the path, the cable path vector result used for standard engineering deployment is output.

[0006] Optionally, S1 includes geometric correction, radiometric calibration, and image enhancement preprocessing of the acquired remote sensing image data; the multi-scale semantic segmentation deep learning model is constructed based on a convolutional neural network with an encoder-decoder structure, the encoder part extracts multi-scale land cover features through a pyramid pooling module, and the decoder part upsamples and fuses the features to output the initial land cover type classification and corresponding category probability for each pixel.

[0007] The auxiliary accessibility label samples are obtained corresponding to the remote sensing image regions. The auxiliary accessibility label samples include field survey accessibility scores for known land cover types. The accessibility scores are determined comprehensively based on the difficulty of passage for cable construction machinery, geological stability, and ecological sensitivity. During the training phase of the multi-scale semantic segmentation deep learning model, the auxiliary accessibility scores are used as additional supervision signals, which, together with the category labels of semantic segmentation, constitute a multi-task learning objective. The model is trained end-to-end, enabling the model to identify land cover types while associating its deep features with prior accessibility knowledge.

[0008] Optionally, the calculation of the accessibility confidence includes: for any pixel output by the trained model, obtaining the predicted probability distribution of its belonging to various land cover types; according to the "land cover type-standard accessibility score" mapping table established in the auxiliary accessibility annotation samples, weighting and fusing the probability distribution into the comprehensive accessibility expectation score of the pixel; and converting the comprehensive accessibility expectation score into an accessibility confidence between 0 and 1 through a preset monotonic mapping function, wherein a high score maps to high confidence, indicating that accessibility is certain and the risk is low, and a low score maps to low confidence, indicating that accessibility is poor or the risk is uncertain.

[0009] Optionally, the construction of the accessibility confidence field includes: performing an accessibility confidence calculation process on all pixels within the coverage area of ​​the remote sensing image to generate an initial discrete confidence map with the same spatial resolution as the original image, where each pixel corresponds to an accessibility confidence level; and processing the initial discrete confidence map using a spatial interpolation algorithm to eliminate local outliers or small-scale discontinuities, thereby generating a spatially continuous and smoothly varying ground feature accessibility confidence field.

[0010] Optionally, S2 includes the definition of the basic cost factor, specifically including: Path length cost factor: For any path segment on the path, its cost is proportional to the actual geometric length of the path segment; Construction cost factor: Based on the land cover types and historical construction data of the area traversed by the route, a standard construction cost coefficient per unit length is assigned to different land cover types. The total construction cost of the route segment is the product of its length and the corresponding regional cost coefficient. Terrain slope cost factor: The average slope of the area traversed by the path segment is calculated based on digital elevation model data. The slope value is mapped to a slope cost through a nonlinear function. The greater the slope, the higher the cost.

[0011] Optionally, the definition of the confidence adjustment factor includes extracting the accessibility confidence value of each pixel covered by the path segment from the ground accessibility confidence field generated by S1, and defining a confidence inverse factor based on the accessibility confidence value, wherein the confidence inverse factor is inversely proportional to the accessibility confidence.

[0012] Optionally, S2 also includes a penalty and reward weighting mechanism to differentiate the cost of different confidence regions in the path, thereby avoiding high-risk regions and incentivizing low-risk regions. Specifically, this includes setting an accessibility confidence threshold to divide path segments into two categories: When the accessibility confidence value is less than the accessibility confidence threshold, it is defined as a low-confidence region; A region is defined as having a high confidence level if its accessibility confidence level is greater than or equal to the accessibility confidence threshold. Based on the comparison results, different adjustment weights are assigned: Low confidence region: Multiply by a penalty weight to amplify its local path cost, thereby increasing this cost during path search and prompting the algorithm to automatically avoid the current region; High confidence region: Multiply by a reward weight to reduce its local path cost, thereby attracting the path to prioritize the current region.

[0013] Optionally, the construction of the confidence-coupled path cost function specifically includes: for any candidate path from the starting point to the ending point, discretizing it into a series of continuous path segments, calculating the local cost of each path segment separately, and accumulating the local costs to form the total cost of the entire path; wherein: The local cost of each path segment includes the sum of the basic costs and the accessibility adjustment factor; The sum of basic costs includes path length cost, construction cost cost, and slope cost, which are weighted and combined using a set basic cost factor weighting coefficient. Accessibility moderating factors include confidence inverse factor and weighting factor; The total cost of the current candidate path is obtained by summing up all the local costs of the candidate paths.

[0014] Optionally, S3 specifically includes modeling the planning area as a three-dimensional raster network including geographic coordinates and the location confidence attribute based on the digital elevation model and the accessibility confidence field of the ground features, as a basic search map; The improved heuristic function consists of two parts: the first part is the geometric cost estimated based on the Euclidean distance from the current node to the target point; the second part is the confidence cost estimated based on the average confidence field value in the direction from the current node to the target point, specifically the expected value of the confidence penalty that may be encountered in the remaining part of the predicted path. The improved heuristic function is used to guide the search direction, so that while moving towards the target, it avoids a large area of ​​low confidence and high risk ahead.

[0015] Optionally, S3 further includes employing multi-objectives. The search algorithm explores execution paths, using the constructed confidence-coupled path cost function as the core evaluation function; multi-objective When expanding the search nodes, the search algorithm not only accumulates the actual cost from the starting point to the current node, but also uses the constructed improved heuristic function to estimate the cost from the current node to the target point and calculates the evaluation value, which is the sum of the actual cost and the cost estimated by the improved heuristic function. The search process considers both constraints and optimization objectives simultaneously, specifically including: a) Construction continuity constraint: Set a minimum turning angle limit in the search space definition to ensure the mechanical construction operation of the path by constraining the directional changes of adjacent path segments; b) Risk avoidance optimization: When expanding nodes, the algorithm will prioritize exploring nodes with smaller evaluation values. Since the confidence coupling path cost function includes confidence penalty weights, it naturally guides the path to deviate from low-confidence units, thus achieving risk avoidance. c) Multi-objective trade-off: By adjusting the weight coefficients of the basic cost factors, multiple objectives such as path length, construction cost and terrain adaptability are optimized in a single search; finally, a cable path with the minimum total cost under the set weights is output.

[0016] The beneficial effects of this invention are: This invention fuses remote sensing imagery with auxiliary accessibility annotation samples and introduces a multi-task semantic segmentation model. This model not only identifies land cover types but also learns the construction accessibility scoring logic for different land covers, constructing a land cover accessibility confidence field with continuous spatial distribution characteristics. Compared to traditional path planning methods based on land cover classification, this invention can comprehensively identify high-risk areas such as water bodies, dense forests, and built-up areas, expressing them in quantitative confidence scores. The introduction of a scoring-confidence score conversion mechanism enables the model to perform risk classification at the pixel level, improving the path planning's ability to perceive on-site construction risks and providing a reliable spatial foundation for subsequent obstacle avoidance and cost modeling.

[0017] This invention constructs a multi-objective path cost function that incorporates fundamental factors such as length, construction cost, and slope. For the first time, it introduces an inverse accessibility confidence factor plus a penalty and reward weighting mechanism as an adjustment term, achieving a dynamic adaptive response of path cost to risk areas. Low-confidence areas automatically amplify the cost, guiding the path to avoid high-risk construction zones; high-confidence areas are rewarded, encouraging the path to gravitate towards areas with good accessibility and favorable construction conditions. By adjusting objective preferences through normalized weighting coefficients, it supports different scenarios such as optimal path length, optimal cost, or slope adaptation.

[0018] This invention employs 3D grid modeling and an improved heuristic function in the path search phase, incorporating accessibility confidence and spatial geometric information into the search logic. This improved mechanism not only considers the historical total cost of the current node but also predicts the average confidence risk from the current node to the target direction, thus achieving proactive risk avoidance. Simultaneously, construction continuity constraints and path post-processing are introduced to ultimately output a continuous vector path that meets the movement requirements of construction machinery. The search guidance is more intelligent, proactively avoiding large-scale risk areas. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the planning method flow according to an embodiment of the present invention; Figure 2This is a schematic diagram of the optimized path search in an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] like Figures 1-2 As shown, a cable path planning method based on deep learning and multi-objective optimization includes the following steps: S1, Ground Feature Accessibility Confidence Field Generation Step: Based on remote sensing image data, a multi-scale semantic segmentation deep learning model is used, combined with auxiliary accessibility annotation samples, to identify the types of ground features in each region, and to calculate the accessibility confidence of each pixel for cable construction, thereby constructing a ground feature accessibility confidence field with continuous spatial distribution. The accessibility confidence field is used to express the credibility of the surface area for construction accessibility in remote sensing identification and the uncertainty level of potential risk areas.

[0023] S11, Acquire remote sensing image data For the acquired remote sensing image data Perform the following preprocessing operations: Geometric correction: calibrating geographic coordinate distortion; Radiation calibration: standardizing the radiation intensity across different wavelength bands; Image enhancement: Improves feature clarity and edge contrast.

[0024] Construct a multi-scale semantic segmentation neural network based on an encoder-decoder structure Its structure is as follows: 1. Encoder Section: The input is preprocessed remote sensing imagery. A set of convolutional layers extracts primary spatial features, which are then fed into either a dilated convolution module or a pyramid pooling module to achieve joint modeling of image regions at different receptive field scales. Dilated convolution expands the receptive field of the convolution kernel without increasing computation, enabling the model to perceive a wider range of context. The pyramid pooling module divides the image into grids of different scales, pools them separately, and then fuses them uniformly, allowing the model to consider both global and local features at each layer. The output is the multi-scale fused ground feature data. It includes encoding information such as the shape, texture, and boundaries of ground features.

[0025] 2. Decoder section: Upsamples the high-dimensional feature map output by the encoder to restore it to a state similar to the original. Figure 1To achieve consistent resolution, shallow detail features from the encoder are gradually fused together, skip connections are made to preserve spatial boundary information, and convolutional layers are used to map to the number of land cover categories. Finally, the Softmax function is used to output the probability that each pixel belongs to each land cover category. ,in This represents the number of land cover categories. The final output is a semantic segmentation probability map, where each pixel corresponds to a vector representing the probability value of belonging to categories such as "arable land", "forest land", "water body", "road", and "bare land".

[0026] S12, Auxiliary Accessibility Label Sample Fusion: In remote sensing images, simply identifying "this area is woodland" or "that area is bare land" is insufficient, because cable laying depends not only on the type of terrain features but also on various factors such as on-site construction difficulty, geological stability, and ecological sensitivity. Therefore, a single terrain feature classification cannot comprehensively reflect which path is suitable for construction. To address this, auxiliary accessibility label samples are introduced as external prior information to supplement the construction difficulty features that cannot be directly perceived in remote sensing images, allowing the model to simultaneously learn "what terrain features are" and "whether construction is possible and how difficult it is."

[0027] These samples, derived from field surveys or historical engineering data, mainly include two elements: Land feature type labels: such as "arable land", "forest land", "built area", "tidal flat", etc.; Accessibility score: A value between 0 and 1, representing the construction accessibility of this type of surface. The higher the score, the easier it is to construct and the lower the risk. For example, flat bare land can be scored as 0.9, while swamps or ecological reserves are only 0.2 or lower. These data constitute a feature-accessibility correspondence database, that is, an auxiliary annotation sample set, which can be understood as assigning a construction friendliness score to each feature.

[0028] Each sample record in the auxiliary annotation sample set contains two fields: The land feature category labels are: "farmland", "grassland", "shrubland", "swamp", "road", "building", etc., which are consistent with the remote sensing classification standards.

[0029] The standard accessibility score for this type of land cover is a score that is formed by comprehensively evaluating three factors: the difficulty of passage for cable construction machinery, geological stability, and ecological sensitivity. This score is stored in the sample set as prior knowledge. Each factor is assigned a score, and the final score is obtained by weighted averaging.

[0030] Table 1. Partial Structure of the Auxiliary Annotation Sample Set Ground feature types Accessibility score (s) arable land 0.85 Woodland (dense) 0.3 Woodland (sparse) 0.55 tidal flats and wetlands 0.15 City roads 0.95 building 0.05 The process of constructing the auxiliary labeled sample set includes: 1. Collect accessibility data for ground features: cable engineering design manuals, construction logs, UAV survey data, project acceptance reports, etc. For each type of typical ground feature, extract comprehensive indicators such as the difficulty of mechanical access, path excavation difficulty, and construction period in actual construction.

[0031] 2. Expert experience scoring database construction: Invite engineering technicians or construction designers with on-site experience to assign relative scores to different land features based on their accumulated experience. The analytic hierarchy process (AHP) can be used to weight the scores of multiple factors.

[0032] 3. Convert data from different sources to a uniform range of 0 to 1 to ensure the comparability of the scoring system and ensure that the scores can serve as stable and generalizable supervision labels during the training phase of the remote sensing segmentation model.

[0033] In the training process of deep learning models, only one objective is used. For example, image segmentation tasks only focus on which category a pixel belongs to. However, this invention adopts a multi-task learning strategy: The first task is land cover classification: use the cross-entropy loss function to supervise each pixel and guide the model to identify which type of land cover the pixel belongs to.

[0034] The second task, accessibility score prediction, uses a regression loss function and requires the model to output a numerical value that approximates the accessibility score of the feature in the sample database.

[0035] Training the model together with these two tasks allows it to learn not only the morphological features of the terrain but also the relationship between these features and accessibility. This results in the output that tells you not only "what this is" but also "whether you can walk here and how difficult it is to walk here."

[0036] The specific plan is as follows: collect auxiliary labeled sample sets corresponding to the remote sensing areas. ;in, For land cover type labels, For the corresponding on-site accessibility score, and This reflects the difficulty of passage, geological stability, and ecological sensitivity during cable construction. During model training, auxiliary scoring will be used. As a second supervisory signal, it constitutes the multi-task loss function: ;in, The cross-entropy loss is used for semantic segmentation to ensure that the model can accurately identify the land cover type for each pixel. The regression loss for accessibility scoring guides the model to learn the relationship between different land features and accessibility. This is the loss weight.

[0037] S13 transforms the land cover classification probabilities output by the deep learning model into the construction accessibility confidence score for each pixel, i.e., a quantified expression. In simpler terms, the previous model output only the probability of classification such as farmland or water body, while construction is actually concerned with whether it's passable. This step transforms "what it is" into "whether it's passable," and into a probabilistic, uncertain expression. Specifically: 1. Obtain the probability distribution of land cover category for each pixel. After passing through the deep learning model, each pixel corresponds to a set of probabilities, such as 70% being cultivated land, 20% being forest land, and 10% being bare land. These probabilities reflect the model's confidence in the pixel belonging to different land cover types.

[0038] 2. Introduce accessibility scores as weights: Each land cover type has a pre-set accessibility score, which comes from the auxiliary annotation sample set. For example, cultivated land = 0.85, forest land = 0.4, bare land = 0.9. These accessibility scores are multiplied by the model prediction probability in a weighted manner to calculate the comprehensive accessibility expectation score of the pixel.

[0039] That is, for each pixel output by the model Obtain the probability distribution of its land cover categories. Then its accessibility expectation score is: ;in, For the pixel to belong to the first Predicted probability of class For the first The standard accessibility score for land features is provided by a mapping table constructed from auxiliary labeled samples. Since the expected score is merely a weighted average and lacks the ability to distinguish between reliable and unreliable scores, a monotonic mapping function is introduced to non-linearly convert the expected score into a confidence level. High scores become closer to 1 after the conversion, indicating high accessibility and reliability, while low scores are further compressed, approaching 0, indicating low accessibility or model uncertainty. The monotonic mapping function... Convert expected scores into accessibility confidence levels. This function can take the form of a logarithmic function, a sigmoid function, etc., to enhance the differentiation between high and low scores in terms of uncertainty.

[0040] S14, Continuous Confidence Field Construction: Perform the above calculations on all pixels within the remote sensing image coverage area. The output is the accessibility confidence value for each pixel. These values ​​constitute the initial discrete pixel map, called the pixel-level confidence map or initial confidence map, denoted as S14. Each pixel in this image has a confidence value ranging from 0 to 1. However, due to factors such as model prediction errors, image noise, or unclear ground feature boundaries, the image may exhibit significant local fluctuations, patches, and discontinuous edges. To improve the spatial continuity and practical usability of the data, spatial interpolation is used to transform the initial discrete pixel confidence map into a spatially continuous, visually smooth, and locally reasonable ground feature accessibility confidence field. The final output is a spatially continuous two-dimensional floating-point layer, denoted as [image source]. , is represented as: Spatial interpolation was performed on the map to obtain a spatially continuous and smooth ground feature accessibility confidence field. It serves as the basic input for path cost modeling.

[0041] S2, Construction of Confidence-Coupled Path Cost Function: Based on the accessibility confidence field of the ground features, a confidence adjustment factor is defined and introduced into the multi-objective cost modeling process of path optimization to construct a confidence-coupled path cost function. The cost function integrates path length, construction cost, terrain slope and confidence inverse factor, and assigns penalty weights to areas with low ground feature confidence and passage reward weights to areas with high confidence, generating a confidence-coupled path cost function for path search.

[0042] S21, Quantization of Basic Cost Factors: Define and quantify three basic path cost factors: S211, Path Length Cost Factor For any line segment on the path, its cost is related to its actual geometric length. Proportional: This factor measures the physical length of the path itself. Logically, the longer the cable path, the more materials, labor, and construction time are required, and therefore the higher the cost. For each segment in the path, its actual geometric length is directly used as the length cost of that segment. The length costs of all path segments will eventually be added together to form the total length cost, encouraging the path to be as short as possible and avoiding detours.

[0043] S212, Construction Cost Factor Based on the terrain features of the area where the line segment is located, assign a standard construction cost coefficient per unit length. The construction cost of this line segment is: Different types of terrain features present different construction difficulties. For example, the cost of laying cables on hardened roads is much higher than that on bare land or farmland. Construction in water bodies, wetlands, and other areas may require additional structures or detours. Therefore, this invention assigns a "construction cost coefficient per unit length" to each type of terrain feature traversed by each segment of the path, based on a table. This means that for every 1 meter of terrain feature traversed, there is a corresponding construction cost. Then, the length of the path segment is multiplied by the unit cost coefficient of the terrain feature to obtain the construction cost of that segment.

[0044] S213, Terrain Slope Cost Factor Obtaining regional slope values ​​based on digital elevation models and through nonlinear functions Mapping to cost: To enhance the penalty for steep slopes, the slope reflects the inclination of the terrain. Generally, the steeper the terrain, the more difficult the construction, potentially involving more mechanical support and excavation stabilization, which not only increases costs but also safety risks. To quantify the impact of slope on path planning, the average slope of the area where the path is located is extracted from the digital elevation model. Then, the slope value is input into a nonlinear function, which maps the slope to a slope cost. To amplify the sensitivity to steep slopes, the cost does not change much at low slopes, but increases rapidly after exceeding a certain slope threshold, thus penalizing steep slopes and preventing the path from falling into high-slope areas. This invention uses a piecewise exponential mapping function to define the slope cost mapping rule: when the slope value is less than a set safety threshold, the cost increases slowly; once the slope exceeds the threshold, the cost increases exponentially, strongly suppressing paths traversing high-slope areas. The specific logic is as follows: If the slope is less than 5 degrees, the cost is the base value, for example, set to 1; If the slope is between 5 and 15 degrees, the cost is mapped according to a slowly increasing linear or exponential function, reflecting a mild risk. If the slope is greater than 15 degrees, the cost increases rapidly, and an exponential function or multiple penalty is used to make steep slope areas extremely uneconomical in path search.

[0045] It takes into account the slight slope fluctuations allowed by actual construction, and also effectively avoids the path falling into high slopes that are seriously unsuitable for construction.

[0046] The S22 confidence adjustment factor is defined as follows: The path cost is influenced not only by the path's inherent characteristics, such as length, slope, and construction cost, but also by the confidence level of the area's feasibility. Simply put, a path traversing a region with high confidence should have lower costs; conversely, entering regions with uncertain accessibility or high risk should automatically amplify costs, thus guiding the path to avoid these areas. Introducing confidence into the path cost function presents two challenges: First, confidence cannot be directly used as a weight: if a region has a confidence level of 0 and is completely unreliable, direct multiplication would result in a path cost of 0, violating logic. Second, it's necessary to emphasize the inverse effect of low confidence equaling high risk. Therefore, an inverse design is adopted: the lower the confidence level, the larger the adjustment factor; the higher the confidence level, the smaller the adjustment factor. This design effectively amplifies the path cost in low-confidence regions, guiding the optimization algorithm to avoid these areas. Specifically, each discrete path segment in the path is denoted as a unit. Perform the following operations: 1. Locate the spatial position of this unit in the remote sensing image, corresponding to the two-dimensional coordinates: ; 2. Ground features generated from S1 are accessible to the confidence field. Extract the confidence level of the pixels covered by the path segment. In other words, pixels provide data, which is then aggregated and calculated from line segments. 3. Define the confidence inverse factor: ;in, Represents path segment Accessibility confidence of the pixel in question. To prevent division by zero, take the smallest positive number. about.

[0047] It should be noted that the output of S1 is a spatial distribution layer, i.e., a ground feature accessibility confidence field. Each pixel corresponds to a confidence value between [0,1]. Essentially, it is a two-dimensional floating-point matrix with the same resolution as the original remote sensing image. The confidence extraction process described above is as follows: Path rasterization or discretization: Representing path segments as a series of continuous pixel coordinates, similar to tiling the path onto an image raster; Coordinate matching: for each path pixel The corresponding position can be found directly in the confidence field matrix; Read value: The confidence value corresponding to this pixel. It is the value at that position in the confidence field matrix; Assignment or interpolation: If the path pixels correspond one-to-one with the image pixels, they can be read directly; if the path is represented at the sub-pixel level, bilinear interpolation or nearest neighbor interpolation can be used to obtain its confidence estimate.

[0048] S23, Penalty and reward weighting mechanism: Set an accessibility confidence threshold. And define two adjustment weights: Low confidence areas (confidence level < threshold): These areas are considered to have high construction risk and strong uncertainty. High confidence areas (confidence level ≥ threshold): These areas are considered to have clear accessibility and low risk, making them suitable for construction.

[0049] Different weights are assigned based on the comparison results between their confidence level and the threshold: Low confidence regions: Multiply by a penalty weight to amplify the local path cost, thereby increasing this cost during path search and prompting the algorithm to automatically avoid these regions; High confidence regions: Multiply by a reward weight to reduce the local path cost, thereby attracting paths to prioritize these regions.

[0050] The formal representation is as follows: when Assign penalty weights: ; when Assign weight to passage rewards: This mechanism avoids paths falling into regions with low confidence, i.e., poor accessibility, by adjusting the cost function.

[0051] The core objective of S23 is to introduce a dynamic adjustment mechanism to differentiate the cost value of regions with different confidence levels in the path. This aims to avoid high-risk regions and incentivize low-risk regions, thereby improving the engineering feasibility and risk control capabilities of the entire path planning result. In the previous step, the accessibility confidence of the region was introduced into the cost function through the confidence inverse factor. However, this is only a continuous adjustment factor. To further strengthen the path's awareness of avoiding high-risk regions and its preference for low-risk regions, a penalty / reward weight mechanism based on threshold judgment is designed.

[0052] The accessibility confidence threshold is a real number between 0 and 1, representing the minimum acceptable accessibility confidence level. In the absence of special requirements, an empirical value of 0.5 to 0.7 is acceptable, indicating that if the model's confidence level in determining whether a region is suitable for construction is below 50%–70%, the region is considered unreliable.

[0053] If the semantic segmentation model is well-trained and the confidence field distribution is clear, the threshold can be appropriately increased to 0.7 to strengthen the avoidance of low-confidence regions. Construction risk tolerance: If the actual project has a low risk tolerance, such as passing through the city center or wetland protection area, a higher threshold is recommended; If the target area has diverse landforms and many transition zones, the threshold can be appropriately lowered to 0.5 to avoid the path being forced to detour too far.

[0054] S24, Construction of the confidence coupling path cost function: A total cost function integrating basic path cost and accessibility risk adjustment mechanisms is constructed as the core input for path search and optimization algorithms. This total cost function is not a simple shortest path algorithm, but a multi-objective weighted function with a risk-sensitive adjustment model. It comprehensively considers the actual path length, construction cost, terrain difficulty, and the reliability of regional construction accessibility, guiding the algorithm to generate more reasonable, safer, and more economical cable laying paths.

[0055] The construction logic is that the entire path, from the starting point to the end point, is discretized into multiple continuous path segments. Each segment has its local cost calculated separately, and finally these local costs are accumulated to form the total cost of the entire path.

[0056] 1. Local cost of each path unit: The cost of each path unit consists of two parts: Part 1: The sum of basic costs.

[0057] This includes path length cost, construction cost cost, and slope cost; The three are weighted and combined by setting weight coefficients; Different applications can adjust the weight distribution according to their focus, such as prioritizing construction costs or terrain adaptability.

[0058] Part Two: Accessibility Regulatory Factors.

[0059] This includes the confidence inverse factor and the weighting factor; It is a key design for path risk perception, used to dynamically amplify or reduce the previously calculated basic cost based on regional confidence. If the path traverses a region with low accessibility or uncertainty, the adjustment factor will amplify the cost of that segment, thus penalizing the path in terms of overall cost. If the path traverses a region with high confidence, a reward is given to reduce its cost.

[0060] 2. Total cost of the entire path: The total cost of the current path is obtained by summing the local costs of each segment on the path. The path optimization algorithm uses this total cost as the criterion for evaluation, searching for the optimal path solution with the lowest total cost in the entire search space.

[0061] Specifically, this means any candidate path from the starting point to the ending point. Discretized into a set of line segments The total cost of this path is calculated as follows: Each discrete unit The local cost is: ; in, These are the normalized weighting coefficients for the basic cost factor, used to adjust the importance of length, cost, and slope. This represents the cost adjustment factor obtained by inversely proportionalizing the confidence level. This indicates the penalty or reward weight assigned based on the level of confidence. path The total cost is: .

[0062] S3, Optimized path search based on the confidence-coupled path cost function: The confidence-coupled path cost function is used as the input for path search. An improved heuristic function is constructed, and a multi-objective optimization algorithm is used to perform path search to obtain a cable path scheme that meets the requirements of construction continuity, risk avoidance and path smoothness. After smoothing the path, the cable path vector result used for standard engineering deployment is output.

[0063] S31 models the entire planning area as a three-dimensional grid network structure: ,in, Represents two-dimensional plane coordinates. This represents the elevation value extracted from the digital elevation model at the corresponding location. This represents the accessibility confidence value for the corresponding location. This 3D raster map is used to define the topology of the path search space and the spatial attributes of each node.

[0064] Constructing improved heuristic functions Heuristic search, i.e. The algorithm relies on a heuristic function mechanism to predict the estimated cost from the current node to the target point. The more accurate the prediction, the more efficient the search and the more reasonable the path. Traditional heuristic functions typically only consider geometric distances, such as the straight-line distance from the current position to the target point. However, in this invention, path planning must not only consider taking the shortest route but also avoid high-risk routes. Therefore, an improved heuristic function that integrates terrain information and confidence information is designed, consisting of two parts: 1. Geometric cost term : Indicates starting from the current node To the target node Euclidean distance estimation: This is the standard part of a traditional heuristic, used to measure the straight-line distance from the current node to the target node, taking into account the actual physical distance in three-dimensional space, including elevation differences: The farther away from the target, the greater the value of the payment. This encourages the path to move in the direction of the goal.

[0065] 2. Confidence Cost Item : Represents the inverse expected value of the average confidence level along the path from the current node to the target node, and can be regarded as the potential confidence risk of the future path: ;in, Let be the mean confidence level along the straight line from the current node to the target point. This is a very small positive number; this is the key point of the invention. This term is used to predict the accessibility risks that may be encountered in the direction from the current node to the target node. First, samples are taken along a straight line between the current node and the target node; then, the confidence values ​​of all pixels on this expected path are calculated, and the average value is taken to obtain the average confidence in this path direction; finally, the inverse ratio is taken, that is, the lower the confidence, the higher the cost, to obtain the confidence cost term. This cost term guides the search algorithm to predict whether there are large areas of low-confidence risk regions ahead, and to deviate from these regions as early as possible, achieving a more robust path selection.

[0066] The final heuristic function is a weighted combination of two terms: ;in, , which is a weighting coefficient used to balance the relationship between path convergence and future confidence risk prediction.

[0067] S32, Path Search Based on Multi-Objective Optimization: The goal is to find a cable path with the lowest total cost that meets construction requirements within the entire planning area, based on the previously constructed path cost function, which includes multiple factors such as distance, cost, gradient, and confidence level. Specifically, a multi-objective optimization approach is adopted. The algorithm performs path search, and during the search process, each candidate path node is assigned an evaluation value. : ;in, This represents the distance from the starting point to the current node. The actual cumulative total cost represents the total cost accumulated along the path already traversed, including path length cost, construction cost cost, slope cost, and weight-adjusted cost based on accessibility confidence. According to the confidence-coupled path total cost function defined in S2: ;in, For path unit Local cost, It is an improved heuristic function that estimates the potential total cost from the current node to the target node. It is a heuristic estimate that considers two aspects: the geometric distance from the current point to the destination; and the low-confidence area risk that the current point may encounter in the direction of the destination. This makes the path not only closer to the target, but also aware of the risks ahead and actively avoid them.

[0068] During the search process, The algorithm is based on the evaluation value Expand nodes from smallest to largest to minimize global path cost.

[0069] Synchronization constraints and optimization objectives: a) Construction continuity constraint: Set a minimum angle threshold for path turning changes in the search map to ensure smooth directional changes between adjacent path segments and meet the requirements for mechanical turning radius and continuous operation in cable construction.

[0070] b) Risk-averse optimization: due to the total path cost The algorithm has introduced an inverse confidence factor and a penalty weight, which significantly increases the cost of low-confidence regions. As a result, the search algorithm naturally deviates from high-risk regions when expanding nodes.

[0071] c) Multi-objective trade-off adjustment: By adjusting the parameter weights in S24 The system allows for flexible control over the relative importance of path length, construction cost, and terrain slope during path search, resulting in a more comprehensively optimized path that better reflects actual engineering needs. Weight adjustments can be made based on specific engineering requirements. If the construction budget is the most critical factor, then increase the weight of "construction cost"; If the mountainous terrain is complex, the impact of the "slope" cost will be amplified; If shorter distances are better, then the "length" factor is emphasized.

[0072] The final output is: the total cost under the current optimization objective. The smallest initial path The path satisfies: Do not cross dangerous areas; The construction path has smooth turns and strong continuity; The total cost, length, and slope are reasonably balanced.

[0073] S33 performs engineering smoothing on the discrete path node set obtained in S32. It uses Bézier curves or B-spline curve algorithms to fit the initial path, maintaining overall path consistency while smoothing the curvature at turning points and removing discontinuous polyline segments. During curve fitting, it ensures that minimum turning angles and maximum curvature radii are met to comply with engineering machinery construction continuity and cable bending specifications. The final smoothed curve is then combined with the GIS system coordinate reference and converted into a standard cable path vector format for use by engineering construction systems or automated deployment platforms.

[0074] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0075] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A cable path planning method based on deep learning and multi-objective optimization, characterized in that, Includes the following steps: S1, based on remote sensing image data, adopts a multi-scale semantic segmentation deep learning model, combined with auxiliary accessibility labeling samples, to identify the types of land cover in each region, and calculates the accessibility confidence of each pixel for cable construction, thus constructing a land cover accessibility confidence field with continuous spatial distribution. S2, based on the accessibility confidence field of the ground features, a confidence adjustment factor is defined and introduced into the multi-objective cost modeling process of path optimization to construct a confidence-coupled path cost function; the cost function integrates path length, construction cost, terrain slope and confidence inverse factor, and assigns penalty weights to areas with low ground feature confidence and passage reward weights to areas with high confidence, generating a confidence-coupled path cost function for path search; S3. Using the confidence-coupled path cost function as the path search input, an improved heuristic function is constructed, and a multi-objective optimization algorithm is used to perform the path search to obtain a cable path scheme that meets the requirements of construction continuity, risk avoidance, and path smoothness.

2. The cable path planning method based on deep learning and multi-objective optimization according to claim 1, characterized in that, S1 includes geometric correction, radiometric calibration and image enhancement preprocessing of the acquired remote sensing image data; the multi-scale semantic segmentation deep learning model is constructed based on a convolutional neural network with an encoder-decoder structure. The encoder part extracts multi-scale land cover features through a pyramid pooling module, and the decoder part upsamples and fuses the features to output the initial land cover type classification and the corresponding category probability for each pixel. The auxiliary accessibility label samples are obtained corresponding to the remote sensing image regions. The auxiliary accessibility label samples include field survey accessibility scores for known land cover types. The accessibility scores are determined comprehensively based on the difficulty of passage for cable construction machinery, geological stability, and ecological sensitivity. During the training phase of the multi-scale semantic segmentation deep learning model, the auxiliary accessibility scores are used as additional supervision signals, which, together with the category labels of semantic segmentation, constitute a multi-task learning objective. The model is trained end-to-end, enabling the model to identify land cover types while associating its deep features with prior accessibility knowledge.

3. The cable path planning method based on deep learning and multi-objective optimization according to claim 1, characterized in that, The calculation of the accessibility confidence score includes: for any pixel output by the trained model, obtaining the predicted probability distribution of its belonging to various land cover types; according to the "land cover type-standard accessibility score" mapping table established in the auxiliary accessibility annotation samples, weighting and fusing the probability distribution into the comprehensive accessibility expectation score of the pixel; and converting the comprehensive accessibility expectation score into an accessibility confidence score between 0 and 1 through a preset monotonic mapping function, where a high score maps to high confidence, indicating that accessibility is certain and the risk is low, and a low score maps to low confidence, indicating that accessibility is poor or the risk is uncertain.

4. The cable path planning method based on deep learning and multi-objective optimization according to claim 3, characterized in that, The construction of the accessibility confidence field includes: performing an accessibility confidence calculation process on all pixels within the coverage area of ​​the remote sensing image to generate an initial discrete confidence map with the same spatial resolution as the original image, where each pixel corresponds to an accessibility confidence level; and using a spatial interpolation algorithm to process the initial discrete confidence map to eliminate local outliers or small-scale discontinuities, thereby generating a spatially continuous and smoothly varying ground feature accessibility confidence field.

5. The cable path planning method based on deep learning and multi-objective optimization according to claim 1, characterized in that, The S2 includes the definition of the basic cost factor, specifically including: Path length cost factor: For any path segment on the path, its cost is proportional to the actual geometric length of the path segment; Construction cost factor: Based on the land cover types and historical construction data of the area traversed by the route, a standard construction cost coefficient per unit length is assigned to different land cover types. The total construction cost of the route segment is the product of its length and the corresponding regional cost coefficient. Terrain slope cost factor: The average slope of the area traversed by the path segment is calculated based on digital elevation model data. The slope value is mapped to a slope cost through a nonlinear function. The greater the slope, the higher the cost.

6. The cable path planning method based on deep learning and multi-objective optimization according to claim 5, characterized in that, The definition of the confidence adjustment factor includes extracting the accessibility confidence value of each pixel covered by the path segment from the ground accessibility confidence field generated by S1, and defining a confidence inverse factor based on the accessibility confidence value. The confidence inverse factor is inversely proportional to the accessibility confidence.

7. The cable path planning method based on deep learning and multi-objective optimization according to claim 6, characterized in that, The S2 also includes a penalty and reward weighting mechanism to differentiate the cost of different confidence regions in the path, thereby avoiding high-risk regions and incentivizing low-risk regions. Specifically, this includes setting an accessibility confidence threshold to divide path segments into two categories: When the accessibility confidence value is less than the accessibility confidence threshold, it is defined as a low-confidence region; A region is defined as having a high confidence level if its accessibility confidence level is greater than or equal to the accessibility confidence threshold. Based on the comparison results, different adjustment weights are assigned: Low confidence region: Multiply by a penalty weight to amplify its local path cost, thereby increasing this cost during path search and prompting the algorithm to automatically avoid the current region; High confidence region: Multiply by a reward weight to reduce its local path cost, thereby attracting the path to prioritize the current region.

8. The cable path planning method based on deep learning and multi-objective optimization according to claim 7, characterized in that, The construction of the confidence-coupled path cost function specifically includes: for any candidate path from the starting point to the ending point, discretizing it into a series of continuous path segments, calculating the local cost of each path segment individually, and accumulating the local costs to form the total cost of the entire path; wherein: The local cost of each path segment includes the sum of the basic costs and the accessibility adjustment factor; The sum of basic costs includes path length cost, construction cost cost, and slope cost, which are weighted and combined using a set basic cost factor weighting coefficient. Accessibility moderating factors include confidence inverse factor and weighting factor; The total cost of the current candidate path is obtained by summing up all the local costs of the candidate paths.

9. The cable path planning method based on deep learning and multi-objective optimization according to claim 8, characterized in that, Specifically, S3 includes modeling the planning area as a three-dimensional raster network, including geographic coordinates and location confidence attributes, based on the digital elevation model and the accessibility confidence field of the ground features, as the basic search map; The improved heuristic function consists of two parts: the first part is the geometric cost estimated based on the Euclidean distance from the current node to the target point; the second part is the confidence cost estimated based on the average confidence field value in the direction from the current node to the target point, specifically the expected value of the confidence penalty that may be encountered in the remaining part of the predicted path. The improved heuristic function is used to guide the search direction, so that while moving towards the target, it avoids a large area of ​​low confidence and high risk ahead.

10. A cable path planning method based on deep learning and multi-objective optimization according to claim 9, characterized in that, S3 also includes employing multi-objectives. The search algorithm explores execution paths, using the constructed confidence-coupled path cost function as the core evaluation function; multi-objective When expanding the search nodes, the search algorithm not only accumulates the actual cost from the starting point to the current node, but also uses the constructed improved heuristic function to estimate the cost from the current node to the target point and calculates the evaluation value, which is the sum of the actual cost and the cost estimated by the improved heuristic function. The search process considers both constraints and optimization objectives simultaneously, specifically including: a) Construction continuity constraint: Set a minimum turning angle limit in the search space definition to ensure the mechanical construction operation of the path by constraining the directional changes of adjacent path segments; b) Risk avoidance optimization: When expanding nodes, the algorithm will prioritize exploring nodes with smaller evaluation values. Since the confidence coupling path cost function includes confidence penalty weights, it naturally guides the path to deviate from low-confidence units, thus achieving risk avoidance. c) Multi-objective trade-off: By adjusting the weight coefficients of the basic cost factors, multiple objectives such as path length, construction cost and terrain adaptability are optimized in a single search; finally, a cable path with the minimum total cost under the set weights is output.

Images