A snake robot path optimization method based on transformer internal inspection

Abstract

The present application relates to the technical field of internal detection of power equipment, and particularly relates to a serpentine robot path optimization method based on transformer internal detection, which comprises the following steps: constructing a first static three-dimensional model based on transformer internal point cloud data, dividing a passable space and obstacles, and adding rigid body or flexible category labels to the obstacles; constructing a robot forward kinematics model and determining pose information; when the robot performs a maintenance operation, recording a full joint angle sequence in real time to construct a reversible path memory bank, and determining a path difference degree according to image data and depth data during travel, and judging whether to start reverse backtracking planning; in response to the reverse backtracking planning, performing a reverse operation by reproducing joint angles in reverse chronological order, and retreating to a risk-free safety node. The present application improves the autonomous inspection reliability and operation efficiency of the robot in a complex environment inside the transformer.

Description

Technical Field

[0001] This invention relates to the field of internal inspection technology for power equipment, and in particular to a snake robot path optimization method based on transformer internal inspection. Background Technology

[0002] As a core component of the power system, the reliability and safety of the transformer's internal structure directly affect the stable operation of the power grid. The internal environment of a transformer contains various complex obstacles, including rigid components such as the core, windings, and clamps, as well as flexible materials such as insulating cardboard and insulating pads. Therefore, planning a safe and efficient robot path within the complex, dynamic, and confined internal environment of a transformer is crucial for achieving autonomous robot inspection.

[0003] Chinese Invention Patent Publication No. CN119779325A discloses a robot 3D path planning method based on deep reinforcement learning. This method combines an improved A* algorithm with a D3QN deep reinforcement learning model. In a known 3D environment, it performs a global path search using an energy consumption evaluation model (comprehensively considering path length, angle changes, and vertical movement) to generate obstacle avoidance paths. This method exhibits good path planning performance in structured environments or static obstacle scenarios. However, it does not distinguish between rigid and flexible obstacles and lacks an effective escape mechanism when the robot encounters blind holes or dead ends, making it difficult to directly apply to the complex and dynamic detection environment inside transformers.

[0004] However, existing methods still have the following significant shortcomings when applied to transformer internal inspection tasks: On the one hand, its approach to handling obstacles is relatively simple, failing to distinguish between rigid and flexible obstacles in terms of safety constraints. This results in insufficient avoidance of flexible materials such as insulating cardboard by the robot, which may cause insulation damage due to mechanical contact. On the other hand, many robots adopt forward planning and obstacle avoidance strategies. When they encounter blind holes, dead ends or other situations where they cannot move forward, they lack an effective escape mechanism. They often need manual intervention or global replanning, which results in long response times and high computational costs, making it difficult to meet the requirements of real-time performance and reliability.

[0005] Therefore, there is an urgent need for a robot path optimization method that can distinguish between rigid and flexible obstacles, has the ability to backtrack and escape from obstacles, and supports efficient local replanning, so as to improve the reliability of autonomous inspection inside transformers. Summary of the Invention

[0006] To address this, the present invention provides a snake robot path optimization method based on transformer internal inspection, which overcomes the technical problems in the prior art that result in insufficient reliability of the robot's autonomous inspection in the complex dynamic environment inside the transformer, caused by the lack of differentiation between rigid and flexible obstacles, the lack of an emergency response mechanism for being trapped in blind holes, and the low efficiency of global replanning after encountering obstacles.

[0007] To achieve the above objectives, the present invention provides a snake robot path optimization method based on transformer internal inspection, characterized in that it includes: Step S1: Construct a first static 3D model based on the point cloud data of the transformer's internal structure to obtain the passable space set and obstacle set within the first static 3D model, and add obstacle category labels to each obstacle, wherein the category labels include rigid obstacles and flexible obstacles; Step S2: Based on the set of maintenance points and the set of obstacles, a multi-path sampling algorithm is used to obtain a set of candidate maintenance paths for each maintenance point, so as to determine the maintenance path complexity of each maintenance path. The maintenance path complexity is determined according to the spatial density and number of rigid and flexible obstacles in the maintenance path, as well as the total path length. Step S3: Determine the maintenance path determination strategy for each maintenance point based on the maintenance path complexity, so as to determine the initial maintenance path for the corresponding maintenance point. The maintenance path determination strategy includes a fast search strategy based on rigid obstacle avoidance and a fine planning strategy for avoiding or contacting flexible obstacles. Step S4: The robot performs maintenance work based on the initial maintenance path, builds a reversible path memory, and determines the depth data ahead based on the image data during travel, so as to determine the start of reverse backtracking planning. Step S5: In response to the reverse backtracking planning, immediately stop the robot's forward movement, construct a reverse path control law based on the full joint angle sequence in the reversible path memory library to perform the backtracking operation until the robot retreats to a risk-free safe node; Step S6: In response to the current risk-free safe node where the robot is located, update the passable space set based on the obstacle change information collected during the backtracking process, and re-search for obstacle avoidance paths in the updated passable space set as the updated maintenance path; Step S7: Control the robot to continue the maintenance work along the replanned maintenance path, and repeat steps S4-S6 until the robot completes all maintenance tasks in the set of maintenance points in sequence.

[0008] Furthermore, in step S2, the process of determining the maintenance path complexity of each maintenance path includes: Step S21: Discretize the inspection path into several path segments, and construct an expansion channel space based on the robot outline for each path segment, wherein the radius of the expansion channel space is the sum of the robot's maximum cross-sectional half-width and a preset safety margin; Step S22: Based on the expansion channel space, the ratios of rigid obstacles and flexible obstacles to the expansion channel space are determined as the first density and the second density, respectively. Step S23: Count the first number of rigid obstacles and the second number of flexible obstacles in the spatial area traversed by the inspection path. Step S24: The ratio of the first density to the second density is determined as the obstacle category influence factor, wherein the obstacle category influence factor is used to characterize the degree of influence of the relative density of flexible obstacles in the expansion channel space on the difficulty of path passage; Step S25: The weighted sum of the obstacle category influence factor, the total number of obstacles, and the total path length is used to determine the maintenance path complexity.

[0009] Furthermore, in step S3, the process of determining the initial maintenance path for the corresponding maintenance point includes: The complexity of the maintenance path is compared and analyzed with the complexity of the preset maintenance path; Based on the result that the maintenance path complexity is greater than or equal to the preset maintenance path complexity, the maintenance path strategy is determined to be a fine-grained planning strategy. Alternatively, based on the result that the maintenance path complexity is less than the preset maintenance path complexity, the maintenance path strategy is determined to be a fast search strategy.

[0010] Further, in step S3, the process of the fast search strategy includes: Based on the target maintenance point in the set of maintenance points, a fast expanding random tree algorithm or A* algorithm is used for path search within the set of accessible spaces. During the search process, the rigid obstacles are used as collision constraints for path pruning, while the flexible obstacles are marked as accessible space and are not used as collision constraints for path search. When a feasible path is found that connects the current location to the target maintenance point, meets all rigid obstacle avoidance requirements, and has a path cost that meets a preset threshold, the feasible path is determined as the initial maintenance path.

[0011] Furthermore, in step S3, the process of refining the planning strategy includes: Based on the target maintenance point in the set of maintenance points, an initial path search is performed within the set of accessible spaces using a bidirectional fast expanding random tree algorithm or a potential field method, wherein the rigid obstacle is used as the first priority collision constraint. When the initial path intersects with the flexible obstacle, the path segment corresponding to the intersection is locally optimized. By introducing the upper limit constraint of the driving torque of each joint of the robot and the contact force threshold constraint between the robot body and the flexible obstacle, a smooth path is generated that maintains contact with or avoids the flexible obstacle within the contact force threshold range. This smooth path is determined as the initial maintenance path.

[0012] Further, in step S4, determining the forward depth data based on the in-journey image data to determine the process of initiating reverse backtracking planning includes: Real-time depth data of the robot's forward movement is obtained based on the image; The forward depth data is compared with the preset depth data extracted from the corresponding position on the initial maintenance path to determine the path difference. The path difference is compared and analyzed with the preset path difference. Based on the result that the path difference is greater than the preset path difference, it is determined to start the reverse backtracking planning.

[0013] Further, in step S5, the process of constructing the reverse path control law based on the full joint angle sequence in the reversible path memory includes: Extract the full joint angle sequence from the risk-free safe node to the backtracking starting point from the reversible path memory; The full joint angle sequence is arranged in reverse chronological order to generate a reverse joint angle sequence; Based on the reverse joint angle sequence, a reverse path control law is constructed to control each joint of the robot to perform actions in reverse order according to the reverse joint angle sequence, so that the robot can move backward accurately along the original path. The reverse path control law is a control rule that controls each joint of the robot to reproduce the full joint angle sequence in reverse order at the same time interval as when it was recorded.

[0014] Further, in step S6, the process of updating the passable space set based on the obstacle change information collected during the backtracking process includes: During the retrograde operation, the robot's sensors collect real-time information on changes in obstacles in the surrounding environment. The first static three-dimensional model is locally updated using the obstacle change information. The state of the flexible obstacles that have been displaced or deformed is updated to the current state, and the newly added obstacles are added to the obstacle set to form the updated passable space set.

[0015] Furthermore, in step S6, the updated process of re-searching for obstacle avoidance paths from the available space includes: Starting from the risk-free safety node and ending at the unfinished target maintenance point in the maintenance point set, the maintenance path determination strategy is re-executed in the updated passable space set to search for an obstacle avoidance path. The newly discovered obstacle avoidance path is then combined with the non-risk-free path segments that have not yet been executed in the initial maintenance path to form the updated maintenance path.

[0016] Compared with existing technologies, this invention adds rigid body category labels and flexible category labels to obstacles when constructing the first static three-dimensional model, thereby achieving differentiated processing of rigid and flexible obstacles. This allows path planning to set collision constraints according to the type of obstacle, strictly enforce obstacle avoidance for rigid obstacles, and allow contact with flexible obstacles within a controllable contact force range. As a result, while ensuring the insulation safety of the equipment, the accessibility of the path and the completion rate of the operation are improved.

[0017] Furthermore, this invention constructs a reversible path memory library to record the sequence of all joint angles and corresponding spatiotemporal nodes during the robot's movement in real time. When the path difference determined by the image data and depth data during movement exceeds a preset threshold, reverse backtracking planning is immediately initiated. The joint angles are reproduced in reverse order according to the timestamp to perform the reversing operation, enabling the robot to accurately retreat along the original path to a risk-free safe node. This effectively solves the problem that traditional methods are prone to getting stuck in blind holes, dead ends, and other scenarios, requiring manual intervention. It significantly improves the robot's ability to autonomously escape from the narrow and complex environment inside a transformer.

[0018] Furthermore, after backtracking to a risk-free safe node, the present invention only updates the passable space set based on the obstacle change information collected during the backtracking process, and re-searches for obstacle avoidance paths in the updated passable space set. These paths are then spliced ​​with risk-free path segments that have not yet been executed in the initial maintenance path, achieving local replanning rather than global replanning. This significantly reduces computational overhead and response time, while prioritizing the reuse of successfully passed path segments, thus improving the overall efficiency of maintenance operations.

[0019] Furthermore, in the determination of the inverse backtracking planning start, the present invention adopts the method of continuously sampling the path difference multiple times and comparing it with a preset threshold to effectively filter out the instantaneous noise interference of the sensor and avoid false triggering; at the same time, it combines the comparison of joint angle and joint limit angle for dual determination to ensure the executability and mechanical safety of the inverse backtracking operation and improve the robustness of the system in complex dynamic environments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the steps of the snake robot path optimization method based on transformer internal inspection in an embodiment of the present invention. Figure 2 A flowchart illustrating the steps for determining the maintenance path complexity of each maintenance path in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the steps of the fast search strategy in an embodiment of the present invention. Figure 4 The following is a logic diagram for determining whether to enable reverse backtracking planning in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] It should be noted that in the description of this invention, the terms such as up, down, left, right, inside, and outside indicate the direction or positional relationship based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0024] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] Example 1 This invention provides a method for optimizing the path of a snake robot based on transformer internal inspection. Please refer to [link / reference]. Figure 1 ,include: Step S1: Construct a first static 3D model based on the point cloud data of the transformer's internal structure to obtain the set of passable spaces and the set of obstacles within the first static 3D model, and add obstacle category labels to each obstacle, wherein the category labels include rigid obstacles and flexible obstacles. Step S2: Based on the set of maintenance points and the set of obstacles, a multi-path sampling algorithm is used to obtain a set of candidate maintenance paths for each maintenance point, so as to determine the maintenance path complexity of each maintenance path. The maintenance path complexity is determined according to the spatial density and number of rigid and flexible obstacles in the maintenance path, as well as the total path length. Step S3: Determine the maintenance path determination strategy for each maintenance point based on the maintenance path complexity, so as to determine the initial maintenance path for the corresponding maintenance point. The maintenance path determination strategy includes a fast search strategy based on rigid body obstacle avoidance and a fine planning strategy for avoiding or contacting flexible obstacles. Step S4: The robot performs maintenance work based on the initial maintenance path, builds a reversible path memory, and determines the depth data ahead based on the image data during travel, so as to determine the start of reverse backtracking planning. Step S5: In response to the reverse backtracking planning, immediately stop the robot's forward movement, construct the reverse path control law based on the full joint angle sequence in the reversible path memory bank to execute the backtracking operation until the robot retreats to a risk-free safe node; Step S6: In response to the current risk-free safe node where the robot is located, update the passable space set based on the obstacle change information collected during the backtracking process, and re-search for obstacle avoidance paths in the updated passable space set as the updated maintenance path; Step S7: Control the robot to continue the maintenance work along the replanned maintenance path, and repeat steps S4-S6 until the robot completes all maintenance tasks in the set of maintenance points in sequence.

[0026] Specifically, the reverse path control law refers to a control method in which, when the robot's movement is obstructed, the control system reverses the time sequence of all joint angles according to the same time interval as when recording the sequence, generates a reverse joint angle sequence, and controls each joint to execute in sequence, so that the robot can accurately retreat along the original path.

[0027] In step S1 of this embodiment of the invention, the process of constructing the first static three-dimensional model includes: First, a 3D laser scanner is used to perform a full-range scan of the transformer's internal structure, acquiring high-precision point cloud data. This point cloud data contains the spatial coordinate information of all structural components inside the transformer. Subsequently, point cloud processing software (such as CloudCompare or the PCL library) is used to denoise, filter, and register the original point cloud, constructing an accurate first static 3D model.

[0028] Based on this, and using the first static 3D model, the space is divided using the voxel mesh method as follows: The entire 3D space is divided into cubic voxels with a side length of 1 mm. Each voxel is marked as an obstacle if it contains point cloud data points, and otherwise marked as passable.

[0029] Understandably, the set of all passable voxels constitutes the passable space set. For voxels labeled as obstacles, they are further classified according to their physical components (such as coils, cores, lead supports, etc.). For example, components with minimal or no deformation, such as metal supports and coil frames, are classified as rigid obstacles, while components that may deform under external forces, such as insulating cardboard and flexible cables, are classified as flexible obstacles. Each obstacle voxel is then labeled with a corresponding category to support differentiated obstacle avoidance strategies in subsequent path planning.

[0030] Please see Figure 2 This is a flowchart of the steps involved in determining the maintenance path complexity of each maintenance path according to an embodiment of the present invention.

[0031] Specifically, in step S2, the process of determining the maintenance path complexity of each maintenance path includes: Step S21: Discretize the inspection path into several path segments, and construct an expansion channel space based on the robot contour for each path segment. The radius of the expansion channel space is the sum of the robot's maximum cross-sectional half-width and the preset safety margin. Step S22: Based on the expansion channel space, the ratios of rigid obstacles and flexible obstacles to the expansion channel space are determined as the first density and the second density, respectively. Step S23: Count the first number of rigid obstacles and the second number of flexible obstacles in the spatial area traversed by the inspection path. Step S24: The ratio of the first density to the second density is determined as the obstacle category influence factor, wherein the obstacle category influence factor is used to characterize the degree of influence of the relative density of flexible obstacles in the expansion channel space on the difficulty of path passage. Step S25: The weighted sum of the obstacle category influence factor, the total number of obstacles, and the total path length is used to determine the maintenance path complexity.

[0032] In this embodiment of the invention, the candidate path is discretized into several path segments. The length of each path segment is A thickness of 10mm to 20mm is acceptable. Based on the robot's maximum cross-sectional profile, a cylindrical expansion channel space is generated along the path segment. The expansion radius is defined as the robot's half-width plus a safety margin (e.g., 5mm). This channel space defines the area the path occupies in the environment. The ratio of rigid obstacles to flexible obstacles and the expansion channel space is determined as the first density and the second density. Intersection determination is used to count the number of rigid obstacles and flexible obstacles intersecting with the expansion channel space within the spatial area traversed by the path. The ratio of the first density to the second density is determined as the obstacle category influence factor, which characterizes the degree to which the relative density of flexible obstacles within the expansion channel space affects the difficulty of path passage.

[0033] The formula for calculating path complexity is: , in, For path complexity, The obstacle category influence factor coefficient. Indicates the obstacle category influence factor. For the first density, For the second density, It is a very small correction factor. This is the coefficient for the number of rigid obstacles. The number of rigid obstacles, For the number of flexible obstacles, This is the coefficient for the total path length. This represents the total path length. , , The calibration process for the preset weight coefficients is as follows: In multiple typical transformer simulation scenarios containing different densities and distributions of rigid / flexible obstacles, the optimization objectives are path planning success rate and overall cost (such as the weighting of time and safety). A parameter optimization algorithm (such as grid search or Bayesian optimization) is used to iterate through the weight combinations. Finally, the set of weight values ​​(0.5, 0.3, 0.2) with the best overall performance in most scenarios is selected as the preset value. The resulting path complexity is dimensionless, and the larger the value, the higher the difficulty of path planning and control.

[0034] Understandably, the obstacle category influence factor reflects the proportion of flexible obstacles in the robot's path; a higher density means less space for the robot to adjust its posture, resulting in poorer path traversability. The number of obstacles directly affects the probability of collisions on the path, especially rigid obstacles, which pose a greater threat to the robot itself. The total path length directly affects operational efficiency. By weighted summation, these three factors are unified into a complexity index, which can quantify the overall traversability difficulty of the path, providing an objective basis for subsequent path selection. The weights can be adjusted according to the actual internal structural characteristics of the transformer. For example, when flexible insulation materials are dense, the weight of the density obstacle category influence factor can be appropriately increased to prioritize paths that avoid flexible materials.

[0035] Specifically, in step S3, the process of determining the initial maintenance path for the corresponding maintenance point includes: A comparative analysis was conducted based on the complexity of the maintenance path and the complexity of the preset maintenance path. Based on the result that the maintenance path complexity is greater than or equal to the preset maintenance path complexity, the maintenance path strategy is determined to be a fine-grained planning strategy. Alternatively, based on the result that the complexity of the maintenance path is less than the preset maintenance path complexity, the maintenance path strategy can be determined as a fast search strategy.

[0036] In this embodiment, a preset maintenance path complexity threshold is set. This threshold can be obtained through simulation or experimental calibration of typical scenarios inside the transformer (in this embodiment, it is taken as...). =2.5).

[0037] like A meticulous planning strategy is employed. This strategy is suitable for highly complex environments, requiring the simultaneous handling of strict avoidance of rigid obstacles and contact force control of flexible obstacles.

[0038] like A fast search strategy is employed. This strategy is suitable for low-complexity environments, using only rigid obstacles as collision constraints and excluding flexible obstacles as collision detection constraints to improve planning efficiency.

[0039] Specifically, the calibration and derivation process of the preset maintenance path complexity threshold is as follows: In this embodiment, in order to quantify and distinguish the applicable scenarios of fast search strategies and fine-grained planning strategies, a maintenance path complexity threshold is preset. =2.5. This threshold was obtained through the following simulation experiments and data calibration. Those skilled in the art can adaptively adjust this threshold using the same simulation method according to the actual internal structure of the transformer, the robot's motion capabilities, and safety requirements. The specific calibration process is as follows: 1. Simulation Scene Construction A typical transformer internal simulation environment is constructed. This environment is modeled based on the real point cloud data obtained in step S1, and includes: Rigid obstacles: non-deformable structures such as iron cores, clamps, and winding ends, totaling 32; Flexible obstacles: Deformable structures such as insulating cardboard, lead cables, and electrostatic rings, totaling 48.

[0040] Fifty different maintenance path samples were randomly generated in the environment. Each sample contained paths from different starting points to different maintenance points. The total path length ranged from 0.5m to 3.0m, and the obstacle space density ranged from 0.2 to 0.8.

[0041] 2. Two strategies were used for planning and simulating each path in the experiment, with different groups and evaluation indicators: Strategy A (Fast Search Strategy): Only rigid obstacles are used as collision constraints, while flexible obstacles are ignored; Strategy B (Fine Planning Strategy): Simultaneously considers rigid obstacle avoidance and flexible obstacle contact force constraints (contact force threshold set to 5N).

[0042] The following evaluation metrics are defined: Planning success rate The percentage of times a robot can complete the entire path without getting stuck or damaged; Average planning time : Average time required to complete a single path planning operation (in seconds); 3. Experimental Data Recording Table 1 shows the simulation experiment data. Table 1 4. Threshold Determination Analysis The experimental data show that: When the path complexity C < 2.5, the success rate of strategy A (fast search strategy) is higher than 85%, and the average planning time does not exceed 1.6 seconds, which is significantly lower than that of strategy B. This indicates that within this complexity range, ignoring flexible obstacles and performing fast planning can achieve a high success rate, and the efficiency advantage is obvious.

[0043] When path complexity When the value is ≥2.5, the success rate of strategy A drops to below 85%, and when C=2.8 it drops to 70%, posing a risk of the robot getting stuck or damaged due to flexible obstacles; Strategy B (detailed planning strategy) maintained a success rate of over 94% within this range. Although the planning time increased, it ensured the safety and reliability of the operation. Therefore, 2.5 was chosen as the threshold. This threshold balances planning efficiency and operational safety, enabling optimized switching between the two strategies.

[0044] 5. Adaptability Instructions It should be noted that, =2.5 is the value obtained through the simulation experiment calibrated using the transformer model, sensor configuration, and robot motion parameters employed in this embodiment. In practical applications, those skilled in the art can adaptively adjust this threshold based on the specific internal structural characteristics of the transformer, the robot's motion capabilities, and safety requirements, using similar simulation or measurement methods. Such adjustments require no creative effort and are all within the scope of protection of this technical solution.

[0045] Understandably, this solution achieves adaptive planning in both simple and complex environments by automatically switching strategies driven by complexity. It avoids overly conservative treatment of flexible obstacles, allowing the robot to interact with flexible objects within a safe contact force range, and significantly improves path accessibility and task completion rate.

[0046] Please see Figure 3 The flowchart illustrates the steps involved in the fast search strategy according to an embodiment of the present invention.

[0047] Specifically, in step S3, the process of the fast search strategy includes: Based on the target maintenance point in the maintenance point set, within the accessible space set, the fast expanding random tree algorithm or the A* algorithm is used for path search. During the search process, rigid obstacles are used as collision constraints for path pruning, while flexible obstacles are marked as accessible space and are not used as collision constraints for path search. When a feasible path is found that connects the current location to the target maintenance point, meets all rigid obstacle avoidance requirements, and has a path cost that meets a preset threshold, the feasible path is determined as the initial maintenance path.

[0048] The accessible space refers to the area in the collision detection module of the path search algorithm where flexible obstacles are not treated as hard collision constraints that must be avoided. The path planned by the algorithm is allowed to pass through this area, but the final path still needs to be ensured by the subsequent fine planning strategy or force control module to keep the contact force within the safety threshold. In the specific implementation of the Rapidly Expanding Random Tree (RRT) or A* algorithm, when performing a collision query on the search tree node or path point in the collision detection function, if it is only located within the voxel marked as accessible space, it is determined to be passable or given a minimal collision cost close to zero. If it is located within the voxel occupied by a rigid obstacle, it is considered a collision.

[0049] In this embodiment, within the accessible space set, a fast expanding random tree (RRT) or A* algorithm is used for path search. During the search process, only rigid obstacles are used as collision constraints for path pruning; flexible obstacles are marked as accessible space and are not used as constraints for collision detection. The initial maintenance path is determined as a path that connects the current location to the target maintenance point and satisfies all rigid obstacle avoidance requirements.

[0050] Understandably, this step significantly reduces computation by employing a fast search strategy for low-complexity paths, thereby greatly improving the real-time performance and response rate of path planning.

[0051] In step S3, the process of fine-tuning the strategy includes: Based on the target maintenance point in the maintenance point set, within the accessible space set, the initial path search is performed using a bidirectional fast expanding random tree algorithm or a potential field method, where rigid obstacles are used as the first priority collision constraints. When the initial path intersects with a flexible obstacle, the path segment corresponding to the intersection is locally optimized. By introducing the upper limit constraint of the driving torque of each joint of the robot and the contact force threshold constraint between the robot body and the flexible obstacle, a smooth path that maintains contact with or avoids the flexible obstacle within the contact force threshold range is generated, and this smooth path is determined as the initial maintenance path.

[0052] In this embodiment of the invention, for target maintenance points in the set of maintenance points, an initial path search is performed using a Bidirectional Rapid Expanding Random Tree (Bi-RRT) algorithm within the passable space set. The initial path is discretized and sampled, and each path point is checked to see if it is located within the occupied space or contact area of ​​the flexible obstacle. If all path points have no intersection with the flexible obstacle, the initial path can be directly determined as the initial maintenance path; if there is an intersection, all continuous path segments that intersect with the flexible obstacle or are less than a preset safety distance (set to 10mm in this embodiment) are extracted and marked as path segments to be optimized. .

[0053] Local replanning is performed on the path segment to be optimized. By introducing upper limit constraints on the driving torque of each joint of the robot and the contact force threshold constraints between the robot body and the flexible obstacle, a smooth and controllable path is generated.

[0054] Contact force constraint modeling: Establish an elastic contact model for the flexible obstacle, treating the flexible obstacle as having a stiffness coefficient. The elastomer (in this embodiment, the values ​​are specified according to different materials such as insulating paperboard and cable, with a range of 50-200 N / m); Define the contact force between the robot's torso and a flexible obstacle. ,in For penetration depth; set contact force threshold. (In this embodiment, we take) =5N), to ensure that the robot will not be damaged when it comes into contact with flexible obstacles.

[0055] Joint torque constraint modeling: Based on the robot's kinematics and dynamics model, calculate the driving torque of each joint of the robot during execution path tracking. ; Set the upper limit of joint torque (In this embodiment, the motor selection is based on...) = 2.5N m), to prevent joint overload caused by excessive path curvature.

[0056] Optimization Solution: A local path optimization problem is constructed with path smoothness as the objective function and contact force and joint moment constraints as conditions. A quadratic programming (QP) algorithm is used to solve the problem. The positions of the path points in the code are iteratively adjusted; when the conditions are met... and Under the premise of minimizing the rate of change of path curvature and the peak contact force with the flexible obstacle, the local path optimization problem is constructed as follows: Using the path point coordinate sequence P as the optimization variable, the objective function is to minimize the change in total path curvature Σ‖Δκ‖. 2 The constraints include: For all path points that come into contact with the flexible obstacle, the following conditions are met: ; For all joints, satisfy ; 3) The path must satisfy the robot's kinematic continuity constraint. The Sequence Quadratic Programming (SQP) algorithm is used to solve the above problem.

[0057] Path generation: After optimization, the adjusted path segment is spliced ​​with the unoptimized part of the original path; B-spline curves are used to smooth the spliced ​​path to generate a smooth path with continuous curvature that meets all constraints.

[0058] Understandably, by hierarchically processing path complexity, rigid obstacles are prioritized as collision constraints during the fine-grained planning process to ensure the robot avoids rigid collisions with non-deformable structures. For flexible obstacles, contact is allowed within a controllable contact force range, thus obtaining a better path in confined spaces. Computational complexity is reduced through local optimization, while contact force constraints and joint moment constraints are introduced to ensure operational safety from both the maintenance equipment and robot body perspectives, preventing equipment damage or robot overload due to improper path planning. Combined with a fast search strategy, this forms a complete path planning mechanism that considers both computational efficiency and robot operational safety, enabling it to efficiently and safely complete maintenance tasks in the complex environment inside transformers.

[0059] Specifically, in step S4, determining the forward depth data based on the in-journey image data to determine the process of initiating reverse backtracking planning includes: Real-time acquisition of depth data in front of the robot during its movement based on images; The depth data ahead is compared with the preset depth data extracted from the corresponding position on the initial maintenance path to determine the path difference. Compare and analyze the path difference with the preset path difference. Based on the result that the path difference is greater than the preset path difference, it is determined to start the reverse backtracking planning.

[0060] In this embodiment, a depth map of the foreground is acquired using a depth camera, and then compared point-by-point with preset depth data extracted from corresponding positions in the first static 3D model. The path difference is defined as the proportion of pixels whose absolute deviation from the current depth value to the preset depth value exceeds a safety threshold (set to 5mm in this embodiment) out of the total number of pixels in the entire region of interest. The specific calculation formula is as follows: in, This represents the depth value of the currently acquired i-th pixel. This refers to the preset depth value extracted from the corresponding location on the initial maintenance path. The threshold value is the safety deviation threshold, and N is the total number of pixels in the region of interest. This is an indicator function (it takes the value 1 if the condition is true, and 0 otherwise).

[0061] If the path difference D of three consecutive samples (sampling period 100ms) is greater than the preset path difference threshold. (In this embodiment) If the current angle of each joint of the robot is less than 0.2 and the current angle of each joint has not reached the joint limit angle, it is determined that an unexpected blockage or structural deformation has occurred on the path ahead, and the reverse backtracking planning is initiated.

[0062] Understandably, this path difference calculation method is based directly on depth data comparison, without the need to extract morphological changes or calculate path deviation distance, resulting in higher calculation efficiency. Through continuous sampling and judgment, it can effectively filter out single-frame noise from the depth sensor and avoid accidental triggering of the reverse backtracking function due to occasional data jumps.

[0063] Please see Figure 4 This is the logic diagram for determining whether to enable reverse backtracking planning in an embodiment of the present invention.

[0064] Specifically, in step S5, the process of determining to initiate reverse backtracking includes: Compare and analyze the path difference with the preset path difference. Based on the result that the path difference is greater than the preset path difference, If the joint angle of each joint is less than the joint limit angle, then reverse backtracking planning is initiated.

[0065] In this embodiment of the invention, the preset path difference reference value used to determine whether a path is passable is initially set to 0.6. This value is based on the safety distance requirements around typical obstacles inside the transformer (such as windings and insulating partitions), and is calibrated through Monte Carlo simulation experiments. The calibration method is as follows: In a simulation environment, the robot is controlled in different poses to simulate the scene where its depth camera captures images of obstacles within a predetermined distance in front of it. Through extensive sampling, the distribution of path difference degree D calculated under different actual obstacle distances was statistically analyzed. The maximum statistical upper bound of the D value was set as the 95th quantile in all scenarios where the distance between the robot's end effector and the obstacle could be greater than 5cm, and this was set as the preset path difference degree threshold Dt (0.6 in this embodiment). Experiments show that when D≤0.6, the distance between the robot's end effector and the nearest obstacle is greater than the safety threshold of 5cm, which can be considered as safe passage. That is, if the path difference in three consecutive samples (sampling period 100ms) all meet the requirements If the path is deemed impassable, this mechanism aims to filter out single-frame noise that may be generated by the depth sensor under complex lighting or texture conditions, and to avoid accidental triggering of the backtracking function due to occasional data jumps. When backtracking planning is enabled, the path difference is sampled three times consecutively with a sampling period of 100ms. If all three sampled values ​​are greater than the preset path difference and the joint angles of each joint are less than the joint limit angles, then backtracking planning is enabled. Simultaneously check the angles of each joint. Joint limiting angle The difference, where the joint limiting angle Determined by the mechanical structure of the robot joints, specifically by a software limit value that is 2° inward from the trigger angle of the joint's physical limit switch. If all joints meet this requirement... ( For safety margin, set to 5° (this value is determined based on the steady-state error of the joint controller ±2° and an additional 3° dynamic response margin), the joint has sufficient space to perform reverse backtracking. At this time, the reverse backtracking planning flag is set. If any joint is close to its limit, an alarm is issued first, and the operator intervenes or performs in-situ attitude adjustment.

[0066] Understandably, three consecutive checks prevent false triggers caused by instantaneous sensor noise or accidental disturbances, thus improving the system's robustness. Joint limit angle determination ensures the feasibility of reverse backtracking actions, preventing the robot from getting stuck because joints cannot reproduce the angles from the memory library. This dual determination mechanism guarantees both the timeliness of reverse backtracking planning and mechanical safety.

[0067] Specifically, in step S5, the process of constructing the reverse path control law based on the full joint angle sequence in the reversible path memory bank includes: Extract the full joint angle sequence from the risk-free safe node to the backtracking starting point from the reversible path memory; The full joint angle sequence is arranged in reverse chronological order to generate a reverse joint angle sequence; A reverse path control law is constructed based on the reverse joint angle sequence. This law controls each joint of the robot to perform actions in reverse order according to the reverse joint angle sequence, so that the robot can move backward accurately along the original path. The reverse path control law is a control rule that controls each joint of the robot to reproduce the full joint angle sequence in reverse order at the same time interval as when it was recorded.

[0068] Specifically, the reverse path control law refers to the control rule that controls each joint of the robot to reproduce the entire joint angle sequence in reverse order at the same time interval (10Hz) as when it was recorded.

[0069] In this embodiment, the reversible path memory stores the sequence of all joint angles and the robot's end-effector pose at each time point, tagged with timestamps. When reverse backtracking is initiated, the control system reproduces the joint angles frame by frame at the same frequency as during recording (10Hz in this embodiment), enabling the robot to precisely retreat along the original path.

[0070] Specifically, in step S5, the process of determining that the robot has retreated to a risk-free safe node includes: In response to initiating reverse backtracking planning, the robot's current position is recorded as the backtracking starting point; During the backtracking operation, the node that is closest to the backtracking starting point and has no obstacles within the set range is obtained based on the reversible path memory. This node is identified as a risk-free and safe node, and the backtracking operation is stopped.

[0071] In this embodiment of the invention, each node stored in the reversible path memory is associated with a set range (e.g., 50mm) around that node, determined based on the robot's body length, joint range of motion, and safe buffer distance. This ensures that if there are no obstacles within this range, the robot has sufficient space to initiate a new local path planning. When the robot retreats along the reverse path, it checks the obstacle status information associated with each preceding node. If there are no obstacles within the set range around the node, it is identified as a candidate safe node, and the node closest to the backtracking starting point is selected as the risk-free safe node. The reverse backtracking operation is immediately stopped, the robot remains in the pose of that node, and this node is marked as the new starting point.

[0072] Understandably, using the closest point to the backtracking starting point and the absence of obstacles as the stopping criterion allows the robot to accurately backtrack to the last safe node, rather than blindly backtracking to the path start point or a fixed distance. This on-demand backtracking method preserves successfully traversed path segments to the greatest extent, reduces unnecessary backtracking distance, improves operational efficiency after overcoming obstacles, and ensures that the replanning starting point is located in a truly safe area. Specifically, in step S6, the process of updating the passable spatial set based on obstacle change information collected during the backtracking process includes: During the retrograde operation, the robot's onboard sensors collect real-time information on changes in obstacles in the surrounding environment. The first static 3D model is locally updated using obstacle change information. The state of flexible obstacles that have been displaced or deformed is updated to the current state, and newly added obstacles are added to the obstacle set to form an updated passable space set.

[0073] Specifically, in step S6, the process of re-searching for obstacle avoidance paths in the updated obstacle avoidance space set includes: Starting from a risk-free safe node and ending at an unfinished target maintenance point in the maintenance point set, the maintenance path determination strategy is re-executed in the updated passable space set to search for an obstacle avoidance path. The newly discovered obstacle avoidance path is concatenated with the risk-free path segments that have not yet been executed in the initial maintenance path to form the updated maintenance path.

[0074] In this embodiment, a risk-free safe node is used as the new starting point, and the currently incomplete maintenance point is used as the new target point. The maintenance path determination strategy is re-executed within the updated passable space set. Specifically, based on the updated environmental complexity, a fast search strategy or a fine-grained planning strategy is adaptively selected for obstacle avoidance path search. During the search, the grid size and heuristic function from the initial path planning are used, and the portion overlapping with successfully traversed path segments is preferentially retained to reduce redundant exploration. This preferential retention is implemented in the cost function of the path search algorithm. For spatial voxels located on successfully traversed path segments, their passage cost is multiplied by a discount factor less than 1 (e.g., 0.3). This makes the algorithm more inclined to select paths containing these low-cost voxels when searching for new paths, thus achieving priority reuse. If the new path search is successful, it is concatenated with the risk-free path segments that have not yet been executed in the initial maintenance path to form the updated maintenance path.

[0075] Understandably, dividing step S6 into two independent sub-steps—spatial update and path replanning—makes the technical solution more hierarchical. Replanning only for local areas significantly reduces computational complexity, enabling the robot to complete escape and path update within seconds, meeting real-time operational requirements. Prioritizing the reuse of successful trajectories during local updates effectively utilizes the robot's environmental adaptation information from previous movements, avoiding repeated traversal of known safe areas and improving the efficiency and adaptability of path planning. This mechanism, in conjunction with the backtracking function, forms a complete closed loop of being trapped—retreating—replanning—continuing execution, significantly enhancing the robot's autonomous operation capability in the complex environment inside the transformer.

[0076] Example 2 This embodiment provides a snake robot path optimization method based on transformer internal inspection, which can be applied to rapid search and reverse backtracking escape scenarios in low-complexity environments.

[0077] Taking the task of inspecting foreign objects in the bottom oil duct of a 110kV oil-immersed transformer as an example, the inspection point is P3 (foreign object point in the bottom oil duct). Step S1 is executed, using a 3D laser scanner to acquire point cloud data inside the transformer, constructing the first static 3D model, dividing the passable space set into the obstacle set, adding rigid body labels to rigid components such as the core and clamps, and adding flexible labels to insulating cardboard, lead cables, etc.

[0078] Step S2 is executed, using a multi-path sampling algorithm to obtain a set of candidate maintenance paths. Taking path P3 as an example, the total path length is 0.25m, discretized into 17 path segments (each approximately 15mm), constructing an expanded channel space (radius = robot half-width 25mm + safety margin 5mm = 30mm). The first density ρ_rigid = 0.05, the second density ρ_flex = 0.03, the number of rigid obstacles N_rigid = 1, and the number of flexible obstacles N_flex = 2. Substituting into the complexity formula (weights w1 = 0.5, w2 = 0.3, w3 = 0.2, ε = 0.01): C_path = 0.5×(0.05 / (0.03+0.01)) + 0.3×(1+2) + 0.2×0.25 = 0.5×1.25 +0.9 + 0.05 = 0.625+0.95=1.575. With a preset threshold C_th = 2.5, and since C_path < 2.5, the fast search strategy in step S3 is executed: the A* algorithm is used to search within the accessible space, with rigid obstacles only considered as collision constraints and flexible obstacles marked as accessible space. A feasible path from the starting point to P3 is obtained and determined as the initial maintenance path.

[0079] The robot executes step S4 along the initial inspection path, constructing a reversible path memory database in real time (recording the full joint angle sequence and corresponding spatiotemporal nodes at a frequency of 10Hz). During movement, the depth camera acquires a depth map of the robot's path every 100ms. When the robot reaches a distance of approximately 0.15m from P3, a piece of insulating cardboard inside the bottom oil channel shifts (changing from being attached to the wall to crossing the oil channel), causing a significant deviation between the forward depth data and the preset depth data. The path difference degree D is calculated: the total number of pixels in the region of interest N=6400, the number of pixels with a depth deviation exceeding 5mm is 5120, and D=0.8. Three consecutive samplings yielded D values ​​of 0.78, 0.82, and 0.80, all greater than the preset threshold D_th=0.6. Simultaneously, the joint angles are detected; the current maximum joint angle is 45°, the joint limit angle is 80°, and the safety margin is 5°, satisfying 45°<75°, therefore, reverse backtracking planning is initiated.

[0080] Step S5: Immediately stop moving forward and record the current position as the backtracking starting point. Extract the full joint angle sequence (12 frames) from the reversible path memory from the risk-free safe node (0.12m behind the backtracking starting point, with no obstacles within 50mm) to the backtracking starting point. Arrange this sequence in reverse chronological order to generate a reverse joint angle sequence, and control each joint to execute sequentially at the same time interval (100ms / frame). The robot precisely backtracks along the original path to the safe node and stops backtracking.

[0081] Step S6: During the backtracking process, the sensors on the robot collect real-time displacement information of the insulating cardboard. This information is used to locally update the first static 3D model, moving the cardboard to a new position and updating the passable space set. Starting from the safe node and ending at P3, the fast search strategy is re-executed in the updated passable space set, finding an obstacle avoidance path around the displaced cardboard with a length of 0.18m. This path is then combined with the unexecuted risk-free path segment from the initial maintenance path to form the updated maintenance point path.

[0082] Step S7: Control the robot to continue along the new path and successfully reach P3 to complete the foreign object inspection operation.

[0083] Example 3 This embodiment provides a snake robot path optimization method based on transformer internal inspection, which is applied to fine planning and contact force constraint scenarios in highly complex environments.

[0084] Taking the inspection of the fastening bolts of the core clamps inside a 500kV large oil-immersed transformer as an example, the inspection point is the bolt point of the core clamp. The path needs to pass through the narrow and winding oil channel between the winding and the core, with a large number of insulating cardboard (flexible) and metal clamps (rigid) distributed on both sides. Step S1 is executed to construct a static three-dimensional model, with a total of 120 obstacles (75 rigid and 45 flexible).

[0085] Step S2 calculates the complexity of the candidate path: The total path length is 0.85m, discretized into 56 path segments. The radius of the expanded channel space = robot half-width 30mm + safety margin 8mm = 38mm. Calculations show ρ_rigid = 0.28, ρ_flex = 0.35, N_rigid = 18, and N_flex = 27. Substituting into the formula: C_path = 0.5×(0.28 / (0.35+0.01)) + 0.3×(18+27) + 0.2×0.85 = 0.5×0.7778 + 0.3×45 + 0.17 = 0.3889+13.5+0.17=14.0589, which is greater than the threshold of 2.5, so a fine-grained planning strategy is executed.

[0086] Step S3: Fine-tuning: An initial path search is performed using a bidirectional RRT algorithm, with rigid obstacles as the first priority collision constraint, resulting in a preliminary path. This path intersects with three locations of insulating cardboard; the intersection path segment (0.12m in length) is extracted for local optimization. An elastic model for the flexible obstacle is established: the stiffness coefficient of the insulating cardboard is k_f = 120N / m, and the contact force threshold F_max = 5N; the upper limit of the robot joint torque τ_max = 2.5N·m. With path smoothness as the objective, a quadratic programming problem is constructed to obtain an adjusted path point sequence, ensuring that the maximum contact force between the robot and the cardboard is 4.2N (less than 5N), and the maximum joint torque is 2.1N·m (less than 2.5N·m). An initial maintenance path is generated after smoothing using a B-spline curve.

[0087] The robot executes step S4, proceeding along the path. Upon reaching the depth of the winding, a previously fixed insulating pad falls off and obstructs the path. The depth camera collects depth data ahead and calculates the path difference D: the total number of pixels within the ROI is 12800, and the number of pixels with a depth deviation exceeding 5mm is 10240, so D=0.8. Three consecutive sampling values ​​are greater than 0.6, and none of the joint angles are close to their limits (current maximum joint angle 52°, limit 85°), confirming the initiation of reverse backtracking planning.

[0088] Step S5: The reversible path memory records the sequence of all joint angles (20 frames) from the safe node (0.2m behind, with no obstacles within 60mm) to the current point. Reproduce this sequence in reverse order, and the robot moves back 0.2m precisely at a frequency of 10Hz to reach the safe node.

[0089] Step S6: During the backtracking process, the stereo vision sensor collects the position and size information of the detached pad in real time, adding it as a new obstacle to the obstacle set and updating the passable space set. Starting from the safety node and ending at the iron core clamp bolt point, the fine planning strategy is re-executed in the updated space, finding an obstacle avoidance path around the detached pad with a length of 0.33m. This path is then spliced ​​with the path segment (0.65m) before the safety node in the initial maintenance path to form the updated maintenance point path.

[0090] Step S7: The robot continues along the new path and successfully reaches the bolt point of the iron core clamp, completing the tightening torque detection. No rigid collisions occurred throughout the process, and the contact force of the flexible obstacle was controlled within 5N.

[0091] Understandably, dividing step S6 into two independent sub-steps—spatial update and path replanning—makes the technical solution more hierarchical. Replanning only for local areas significantly reduces computational complexity, enabling the robot to complete escape and path update within seconds, meeting real-time operational requirements. Prioritizing the reuse of successful trajectories during local updates effectively utilizes the robot's environmental adaptation information from previous movements, avoiding repeated traversal of known safe areas and improving the efficiency and adaptability of path planning. This mechanism, in conjunction with the backtracking function, forms a complete closed loop of being trapped—retreating—replanning—continuing execution, significantly enhancing the robot's autonomous operation capability in the complex environment inside the transformer.

[0092] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for optimizing the path of a snake-like robot based on transformer internal inspection, characterized in that, Includes the following steps: Construct the first static 3D model of the transformer to obtain the passable space set and obstacle set, and match flexible obstacle labels for the obstacles; Obtain the candidate maintenance path set for each maintenance point and determine the maintenance path complexity of each maintenance path to determine the maintenance path determination strategy for each maintenance point, and determine the initial maintenance path for the corresponding maintenance point according to the maintenance path determination strategy. The robot performs maintenance work based on the initial maintenance path and determines the depth data ahead based on the image data during travel to determine whether to initiate reverse backtracking planning; In response to the reverse backtracking planning, a reverse path control law is constructed based on the full joint angle sequence in the reversible path memory to perform the backtracking operation; in response to the risk-free safe node where the robot is currently located, the obstacle space set is updated based on the obstacle change information collected during the backtracking process, and the obstacle avoidance path is re-searched in the updated obstacle space set as the updated maintenance path.

2. The snake robot path optimization method based on transformer internal inspection according to claim 1, characterized in that, The process of determining the maintenance path complexity of each maintenance path includes: The maintenance path is discretized into several path segments, and an expansion channel space based on the robot outline is constructed for each path segment. The radius of the expansion channel space is the sum of the robot's maximum cross-sectional half-width and a preset safety margin. Based on the expansion channel space, the ratios of rigid obstacles and flexible obstacles to the expansion channel space are respectively determined as the first density and the second density. The number of rigid obstacles and the number of flexible obstacles within the spatial area traversed by the inspection path are counted separately. The ratio of the first density to the second density is determined as the obstacle category influence factor, wherein the obstacle category influence factor is used to characterize the degree of influence of the relative density of flexible obstacles in the expansion channel space on the difficulty of path passage; The maintenance path complexity is determined by combining the weighted sum of the obstacle category influence factor, the total number of obstacles, and the total path length.

3. The snake robot path optimization method based on transformer internal inspection according to claim 1, characterized in that, The process of determining the initial maintenance path for the corresponding maintenance point includes: The complexity of the maintenance path is compared and analyzed with the complexity of the preset maintenance path; Based on the result that the maintenance path complexity is greater than or equal to the preset maintenance path complexity, the maintenance path strategy is determined to be a fine-grained planning strategy. Alternatively, based on the result that the maintenance path complexity is less than the preset maintenance path complexity, the maintenance path strategy is determined to be a fast search strategy.

4. The snake robot path optimization method based on transformer internal inspection according to claim 3, characterized in that, The process of the fast search strategy includes: Based on the target maintenance point in the maintenance point set, a fast expanding random tree algorithm or A* algorithm is used for path search within the accessible space set. During the search process, the rigid obstacles are used as collision constraints for path pruning, while the flexible obstacles are marked as accessible space and are not used as collision constraints for path search. When a feasible path is found that connects the current location to the target maintenance point, meets all rigid obstacle avoidance requirements, and has a path cost that meets a preset threshold, the feasible path is determined as the initial maintenance path.

5. The snake robot path optimization method based on transformer internal inspection according to claim 3, characterized in that, The process of the refined planning strategy includes: Based on the target maintenance point in the set of maintenance points, an initial path search is performed within the set of accessible spaces using a bidirectional fast expanding random tree algorithm or a potential field method, wherein the rigid obstacle is used as the first priority collision constraint. When the initial path intersects with the flexible obstacle, the path segment corresponding to the intersection is locally optimized. By introducing the upper limit constraint of the driving torque of each joint of the robot and the contact force threshold constraint between the robot body and the flexible obstacle, a smooth path is generated that maintains contact with or avoids the flexible obstacle within the contact force threshold range. This smooth path is determined as the initial maintenance path.

6. The snake robot path optimization method based on transformer internal inspection according to claim 1, characterized in that, The process of determining the forward depth data based on in-journey image data to determine the initiation of reverse backtracking planning includes: Real-time depth data of the robot's forward movement is obtained based on the image; The forward depth data is compared with the preset depth data extracted from the corresponding position on the initial maintenance path to determine the path difference. The path difference is compared and analyzed with the preset path difference. Based on the result that the path difference is greater than the preset path difference, it is determined to start the reverse backtracking planning.

7. The snake robot path optimization method based on transformer internal inspection according to claim 1, characterized in that, The process of determining the risk-free safe node includes: In response to initiating the reverse backtracking plan, the robot's current position is recorded as the backtracking starting point; Based on the reversible path memory, the node that is closest to the backtracking starting point and has no obstacles within a set range is identified as the risk-free safe node.

8. The snake robot path optimization method based on transformer internal inspection according to claim 1, characterized in that, The process of constructing a reverse path control law based on the full joint angle sequence in the reversible path memory includes: Extract the full joint angle sequence from the risk-free safe node to the backtracking starting point from the reversible path memory; The full joint angle sequence is arranged in reverse chronological order to generate a reverse joint angle sequence; Based on the reverse joint angle sequence, a reverse path control law is constructed to control each joint of the robot to perform actions in reverse order according to the reverse joint angle sequence, so that the robot can move backward accurately along the original path. The reverse path control law is a control rule that controls each joint of the robot to reproduce the full joint angle sequence in reverse order at the same time interval as when it was recorded.

9. The snake robot path optimization method based on transformer internal inspection according to claim 1, characterized in that, The process of updating the passable space set based on obstacle change information collected during the backtracking process includes: During the retrograde operation, the robot's sensors collect real-time information on changes in obstacles in the surrounding environment. The first static three-dimensional model is locally updated using the obstacle change information. The state of the flexible obstacles that have been displaced or deformed is updated to the current state, and the newly added obstacles are added to the obstacle set to form the updated passable space set.

10. The snake robot path optimization method based on transformer internal inspection according to claim 1, characterized in that, The process of re-searching for obstacle avoidance paths in the updated available space set includes: Starting from the risk-free safety node and ending at the unfinished target maintenance point in the maintenance point set, the maintenance path determination strategy is re-executed in the updated passable space set to search for an obstacle avoidance path. The newly discovered obstacle avoidance path is then combined with the risk-free path segments that have not yet been executed in the initial maintenance path to form the updated maintenance path.

Images