Hierarchical path planning method, device and equipment of traction substation robot, storage medium and program product

Abstract

This application relates to a hierarchical path planning method, apparatus, computer equipment, readable storage medium, and program product for a traction substation robot, belonging to the field of robot path planning technology. This application can improve planning efficiency and address poor environmental adaptability. The method includes: constructing a two-dimensional grid map based on the physical environment information of the traction substation; determining the robot's starting point and a set of task points to be inspected on the two-dimensional grid map; encoding the task point set using a matrix encoding network based on a MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships; generating a global target inspection path based on the node pair feature matrix using a trained generative diffusion model; generating local navigation paths connecting each task point in the global target inspection path; and constructing a third-order Bézier curve to smoothly fit the key inflection points of the local navigation path to generate a final execution path that satisfies the robot's minimum turning radius and kinematic constraints.

Description

Technical Field

[0001] This application relates to the field of robot path planning technology, and in particular to a hierarchical path planning method, apparatus, computer equipment, computer-readable storage medium and computer program product for a traction substation robot. Background Technology

[0002] With the rapid development of high-speed rail and smart grid technologies, the operational safety and stability of traction substations, as the core hubs of railway power supply systems, are becoming increasingly important. Against this backdrop, inspection robots with autonomous movement and sensing capabilities are replacing or assisting manual labor, undertaking demanding tasks such as equipment status monitoring, infrared temperature measurement, instrument reading, and foreign object intrusion detection, becoming key equipment for ensuring the safe and reliable operation of traction power supply systems.

[0003] In terms of local navigation and obstacle avoidance, classic methods such as the A* algorithm, Dijkstra's algorithm, and artificial potential field method are relatively mature and can plan theoretical collision-free paths based on known static environment maps. However, most existing navigation algorithms are mainly designed for static environments and lack the ability to quickly adapt to changes in dynamic environments. When robots encounter sudden obstacles such as temporarily placed maintenance equipment or moving personnel during inspections, traditional algorithms such as the A* algorithm often need to abandon the original path and recalculate the entire map information from scratch. This results in high computational redundancy and long replanning time, leading to low planning efficiency and poor environmental adaptability, which seriously affects the robot's real-time response speed and inspection continuity. Summary of the Invention

[0004] Based on this, it is necessary to provide a hierarchical path planning method, device, computer equipment, computer-readable storage medium, and computer program product for traction substation robots to address the aforementioned technical problems.

[0005] Firstly, this application provides a hierarchical path planning method for a traction substation robot, including:

[0006] Based on the physical environment information of the traction substation, a two-dimensional grid map is constructed, which includes obstacle areas, electromagnetic high-risk areas, and passable areas. The starting point of the robot and the set of task points to be inspected are determined on the two-dimensional grid map.

[0007] The task point set is encoded using a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships. A global target inspection path is then generated based on the node pair feature matrix using a trained generative diffusion model.

[0008] In the global target inspection path, two adjacent task points are taken as local start and end points respectively. The D*Lite algorithm is used to perform a heuristic search in the two-dimensional grid map to generate a local navigation path connecting each task point.

[0009] Key inflection points in the local navigation path are extracted, and a third-order Bézier curve is constructed to smoothly fit the key inflection points, generating a final execution path that satisfies the robot's minimum turning radius and kinematic constraints; the final execution path is used to control the robot to complete the inspection.

[0010] In one embodiment, generating a global target inspection path based on the node pair feature matrix using a trained generative diffusion model includes:

[0011] The node pair feature matrix is ​​input into the generative diffusion model for multi-step iterative reverse denoising to generate a global probability distribution heatmap containing the task point set. The global probability distribution heatmap is decoded to obtain the global initial inspection path, and the global initial inspection path is validated for legality. If the global initial inspection path passes the validity validation, the 2-opt local search algorithm is used to eliminate cross connections in the global initial inspection path, thereby optimizing and obtaining the global target inspection path.

[0012] In one embodiment, the step of encoding the task point set using a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships includes:

[0013] Obtain the Euclidean distance between any two task points in the task point set, and construct a distance matrix based on the Euclidean distance; use the MatNet architecture to perform hybrid expert encoding on the distance matrix to generate the node pair feature matrix.

[0014] In one embodiment, the step of using the MatNet architecture to perform hybrid expert encoding on the distance matrix to generate the node pair feature matrix includes:

[0015] The MatNet architecture is used to perform hybrid expert coding through a row / column hybrid expert module, and the row and column features of the distance matrix are iteratively updated until a preset iteration condition is met, so as to obtain the node pair feature matrix, which is used as the condition input of the generative diffusion model.

[0016] In one embodiment, the method further includes:

[0017] The robot is simulated to perform inspections according to the local navigation path and perceive the spatial environment information on the local navigation path in real time through onboard sensors. If a temporary obstacle is detected on the local navigation path according to the spatial environment information, the incremental search characteristic of the D*Lite algorithm is used to update only the RHS value of the affected node in the local navigation path and add the updated RHS value to the priority queue to repair the local navigation path.

[0018] In one embodiment, the method further includes:

[0019] Obtain true optimal path samples, and gradually add Gaussian noise to the true optimal path samples until an isotropic standard Gaussian distribution is obtained. Based on the updated true optimal path samples, obtain a training set, a validation set, and a test set. Use the training set and the validation set to train the initial diffusion model to be trained using a deep learning algorithm and update the model parameters of the initial diffusion model until the performance index of the updated diffusion model on the test set meets the threshold condition, thus obtaining the generative diffusion model.

[0020] Secondly, this application also provides a hierarchical path planning device for a traction substation robot, comprising:

[0021] The map building module is used to construct a two-dimensional grid map containing obstacle areas, electromagnetic high-risk areas, and passable areas based on the physical environment information of the traction substation, and to determine the robot's starting point and the set of task points to be inspected on the two-dimensional grid map.

[0022] The network encoding module is used to encode the task point set through a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships, and to generate a global target inspection path based on the node pair feature matrix through a trained generative diffusion model.

[0023] The map search module is used to take two adjacent task points as local start and end points respectively in the global target inspection path, and use the D*Lite algorithm to perform heuristic search in the two-dimensional grid map to generate local navigation paths connecting each task point.

[0024] The path generation module is used to extract key inflection points in the local navigation path, construct a third-order Bézier curve to smoothly fit the key inflection points, and generate a final execution path that satisfies the robot's minimum turning radius and kinematic constraints; the final execution path is used to control the robot to complete the inspection.

[0025] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0026] A two-dimensional grid map containing obstacle areas, high-electromagnetic-risk areas, and passable areas is constructed based on the physical environment information of the traction substation. The robot's starting point and a set of task points to be inspected are determined on this two-dimensional grid map. The task point set is encoded using a matrix encoding network based on a MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships. A global target inspection path is generated based on this node pair feature matrix using a trained generative diffusion model. In this global target inspection path, two adjacent task points are used as local starting and ending points, respectively. A D*Lite algorithm is used to perform a heuristic search on the two-dimensional grid map to generate local navigation paths connecting the task points. Key inflection points in the local navigation paths are extracted, and a third-order Bézier curve is constructed to smoothly fit these key inflection points, generating a final execution path that satisfies the robot's minimum turning radius and kinematic constraints. This final execution path is used to control the robot to complete the inspection.

[0027] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0028] A two-dimensional grid map containing obstacle areas, high-electromagnetic-risk areas, and passable areas is constructed based on the physical environment information of the traction substation. The robot's starting point and a set of task points to be inspected are determined on this two-dimensional grid map. The task point set is encoded using a matrix encoding network based on a MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships. A global target inspection path is generated based on this node pair feature matrix using a trained generative diffusion model. In this global target inspection path, two adjacent task points are used as local starting and ending points, respectively. A D*Lite algorithm is used to perform a heuristic search on the two-dimensional grid map to generate local navigation paths connecting the task points. Key inflection points in the local navigation paths are extracted, and a third-order Bézier curve is constructed to smoothly fit these key inflection points, generating a final execution path that satisfies the robot's minimum turning radius and kinematic constraints. This final execution path is used to control the robot to complete the inspection.

[0029] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0030] A two-dimensional grid map containing obstacle areas, high-electromagnetic-risk areas, and passable areas is constructed based on the physical environment information of the traction substation. The robot's starting point and a set of task points to be inspected are determined on this grid map. The task point set is encoded using a matrix encoding network based on a MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships. A global target inspection path is generated based on this node pair feature matrix using a trained generative diffusion model. In this global target inspection path, two adjacent task points are used as local starting and ending points, respectively. A D*Lite algorithm is used to perform a heuristic search on the two-dimensional grid map to generate local navigation paths connecting the task points. Key inflection points in these local navigation paths are extracted, and a third-order Bézier curve is constructed to smoothly fit these key inflection points, generating a final execution path that satisfies the robot's minimum turning radius and kinematic constraints. This final execution path is used to control the robot to complete the inspection.

[0031] The aforementioned hierarchical path planning method, device, computer equipment, computer-readable storage medium, and computer program product for traction substation robots utilize a generative diffusion model based on matrix coding networks at the upper layer to achieve rapid generation and optimization of global traversal sequences for large-scale inspection task points. At the lower layer, the D*Lite algorithm is employed with an incremental cost update mechanism to generate specific navigation paths and perform online replanning for sudden obstacles. At the end layer, a third-order Bézier curve is used to smooth key inflection points of the path, thereby achieving smooth path generation that meets the robot's minimum turning radius and kinematic constraints. This application demonstrates high efficiency and quality in global planning. Leveraging the powerful data distribution modeling capabilities of the diffusion model, it can quickly escape local optima in large-scale task scenarios, generating a shorter and more efficient inspection sequence than traditional heuristic algorithms. Furthermore, this application exhibits strong adaptability to dynamic environments. Through the incremental search characteristics of the D*Lite algorithm, it only needs to update the node information of the affected area when environmental changes are detected, without requiring full graph recalculation, significantly reducing computational overhead and achieving millisecond-level dynamic obstacle avoidance response. The generated final execution path is smooth and highly executable, effectively solving the problems of numerous polylines and large turns in traditional grid paths, reducing robot motion wear, and ensuring the efficiency and safety of substation inspection operations. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is an application environment diagram of a hierarchical path planning method for a traction substation robot in one embodiment.

[0034] Figure 2 This is a flowchart illustrating a hierarchical path planning method for a traction substation robot in one embodiment.

[0035] Figure 3 This is a schematic plan view of a blank two-dimensional grid map of a traction substation in one embodiment;

[0036] Figure 4 This is a planar schematic diagram of the smoothing process of a third-order Bézier curve in one embodiment;

[0037] Figure 5 This is a planar schematic diagram of the global path planning results in one embodiment;

[0038] Figure 6 This is a flowchart illustrating the steps for generating a global target inspection path in one embodiment.

[0039] Figure 7 This is a schematic diagram of the MatNet architecture in one embodiment;

[0040] Figure 8 This is a flowchart illustrating a hierarchical path planning method for a traction substation robot in a specific embodiment.

[0041] Figure 9 This is a structural block diagram of a hierarchical path planning device for a traction substation robot in one embodiment.

[0042] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] The hierarchical path planning method for traction substation robots provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown illustrates this. In this environment, the terminal can communicate with the server via a network. The data storage system can store the data that the server needs to process. The data storage system can be integrated onto the server or located on the cloud or other network servers. In situations such as... Figure 1In the application environment shown, the terminal can be, but is not limited to, various personal computers, laptops, smartphones, and tablets. The server can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0045] In one embodiment, such as Figure 2 As shown, a hierarchical path planning method for a traction substation robot is provided, which can be applied to... Figure 1 In the terminal, the method may include the following steps:

[0046] Step S201: Construct a two-dimensional grid map containing obstacle areas, high-risk electromagnetic areas, and passable areas based on the physical environment information of the traction substation, and determine the robot's starting point and the set of task points to be inspected on the two-dimensional grid map.

[0047] Specifically, the core of this step lies in transforming the unstructured physical environment of the traction substation into a computer-processable digital semantic space, and establishing a refined cost field model for subsequent path planning algorithms. Traditional map building often only distinguishes between the binary states of "obstacles" and "empty land," which is far from sufficient in the complex substation environment. The core of this step not only utilizes LiDAR and SLAM technology to geometrically reconstruct physical boundaries such as walls and transformer bases, but more importantly, it introduces a semantic definition of "multi-level risk cost." In the construction... When creating a 2D raster map, the environment is divided into absolutely impassable physical obstacle areas, high-risk electromagnetic areas beneath high-voltage equipment, and passable areas. By assigning higher passability costs to high-risk electromagnetic areas, rather than simple binarization, the map incorporates potential energy information about the environment. This refined environmental modeling is the cornerstone of subsequent algorithm effectiveness: it ensures that the D*Lite algorithm naturally tends to choose safer paths when searching for routes, rather than simply the shortest path. Simultaneously, the choice of raster precision balances the granularity of environmental description with computational resource consumption, providing efficient data structure support for achieving millisecond-level dynamic replanning.

[0048] As an example, the terminal divides the traction substation environment into, for example, Figure 3 shown Blank 2D raster map, defining the raster state set ,in, Indicates a passable area. Indicates the area of ​​obstacles. This indicates a high-risk area for electromagnetic interference.

[0049] On the 2D grid map, the robot's starting point is set as P0, and the set of task points to be inspected is:

[0050]

[0051] in, Let p be the two-dimensional coordinates of the i-th task point. The higher-level objective of this application is to find an optimal permutation p that traverses all task points and returns to the starting point. i This makes the total path length Minimum:

[0052]

[0053] In the above formula, This represents the Euclidean distance between two task points.

[0054] Step S202: The task point set is encoded using a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships. Based on the node pair feature matrix, a global target inspection path is generated using a trained generative diffusion model.

[0055] Specifically, the core of this step lies in leveraging the unique row-column hybrid attention mechanism of the MatNet architecture to transform discrete task point coordinate information into a deep feature matrix containing global topological dependencies, thereby solving the problem that traditional encoders struggle to capture the duality of combinatorial optimization problems. In multi-objective point path planning, simply knowing "where the points are" is insufficient; the key is understanding "the connection tendencies between points." Traditional Transformer architectures often focus on the features of the nodes themselves, while the core innovation of the MatNet architecture lies in its "edge-centric" encoding architecture. It directly uses a normalized distance matrix as input and alternately aggregates features from the rows and columns of the matrix through multiple stacked hybrid expert attention modules. Row attention captures the potential associations between a node as a starting point and all other nodes, while column attention captures the probability of it being visited as a destination point. This alternating update mechanism simulates duality in graph theory, allowing the network to see through the geometric structure behind the distribution of task points. The final output context feature matrix is ​​no longer a simple physical distance, but a heatmap containing the optimal connectivity trend of the entire graph. It provides a highly guiding conditional input for the subsequent diffusion model, greatly reducing the search difficulty of the generative model in the huge solution space.

[0056] Furthermore, the core of this step lies in introducing the probabilistic distribution learning capability of generative AI. Through the inverse denoising process of "reconstructing order from noise," it breaks the constraint of traditional heuristic algorithms easily getting trapped in local optima, achieving the parallel generation of high-quality global inspection sequences. Traditional genetic algorithms or ant colony algorithms essentially search and trial and error in the solution space, which is inefficient and unstable. The core of this application is to reconstruct the path planning problem into a conditional data generation problem, using a denoising diffusion probability model to train the network to learn the data distribution characteristics of high-quality paths. In the inference phase, the core process starts from an isotropic standard Gaussian noise distribution. Guided by the strong prior features extracted by MatNet, it iterative inverse denoising gradually eliminates randomness, restoring a clear, optimally connected task point adjacency matrix. The essential advantage of this method lies in its powerful generalization and exploration capabilities: it does not rigidly search for a unique solution, but can sample multiple high-quality candidate path distributions in parallel, thus having a high probability of escaping local extremum traps. Finally, the 2-opt algorithm is used for fine-tuning, which is essentially a "geometric-level legality correction" of the generated results, ensuring that the generated solution is not only probabilistically optimal, but also geometrically non-intersecting and physically feasible.

[0057] Step S203: In the global target inspection path, two adjacent task points are taken as local start and end points respectively. The D*Lite algorithm is used to perform heuristic search in the two-dimensional grid map to generate local navigation paths connecting each task point.

[0058] Specifically, the core of this step lies in transforming the theoretical global target inspection path into actual executable motion commands for the robot, and solving the problems of real-time response and smooth control in dynamic environments through a dual mechanism of incremental calculation and kinematic fitting. First, the core value of the D*Lite algorithm lies in its incremental nature. In substation inspection, environmental changes are usually localized. Traditional A* algorithms require abandoning the entire path and recalculating when encountering obstacles, resulting in huge computational loads and robot lag. The D*Lite algorithm, however, only updates nodes whose RHS values ​​have changed and their affected downstream nodes, reusing most of the known path information, thus achieving millisecond-level replanning with extremely low computational power consumption.

[0059] Step S204: Extract the key inflection points of the local navigation path, construct a third-order Bézier curve to smoothly fit the key inflection points, and generate the final execution path that satisfies the robot's minimum turning radius and kinematic constraints; the final execution path is used to control the robot to complete the inspection.

[0060] Specifically, such as Figure 4As shown, the core of this step lies in solving the nonholonomic constraint problem of the grid path using a third-order Bézier curve. The path searched by the grid is essentially a polyline, containing many sharp 90-degree or 45-degree turns, which leads to frequent robot starts and stops and wear. By fitting the key inflection points with a third-order Bézier curve, the discrete polyline is transformed into a smooth trajectory with continuous curvature, ensuring a smooth transition of the robot's angular velocity. This improves inspection speed and ensures the safe transport of precision instruments, ultimately generating a path like... Figure 5 The final execution path shown satisfies the robot's minimum turning radius and kinematic constraints.

[0061] As an example, dynamic obstacle avoidance path planning and trajectory smoothing operations may include the following steps:

[0062] (1) D*Lite incremental search:

[0063] according to For adjacent task points For local navigation, D*Lite maintains two values ​​for each node s: the current cost g(s) and the look-ahead cost rhs(s). Its heuristic function uses Euclidean distance:

[0064]

[0065] Sort key value in priority queue :

[0066]

[0067] When the robot's sensors detect an obstacle that causes a change in the edge cost c(u,v), the edge cost c(u,v) is updated, and the RHS value of node u is also updated.

[0068]

[0069] like If u is not re-inserted into the priority queue, then ComputeShortestPath() is executed, recalculating only the affected nodes, not the entire graph, thus achieving... Rapid repair at the highest level.

[0070] (2) Smoothing of third-order Bézier curves:

[0071] To satisfy the nonholonomic constraints of the robot, the polyline path generated by D*Lite is... Smooth the curve. Select the midpoint of the path segment as the control point and construct a third-order Bézier curve. For three adjacent path points... Define control points .

[0072] The trajectory parametric equation B(u) is:

[0073]

[0074] By limiting the maximum curvature of the curve The final control command is generated based on the robot's maximum turning capability, driving the robot to smoothly complete the inspection.

[0075] In this embodiment, a generative diffusion model based on matrix coding network is used in the upper layer to realize the rapid generation and optimization of the global traversal sequence of large-scale inspection task points; in the lower layer, the D*Lite algorithm is used and an incremental cost update mechanism is introduced to realize the generation of specific navigation paths and online replanning for sudden obstacles; at the end, a third-order Bézier curve is used to smooth the key inflection points of the path to realize the smooth path generation that meets the robot's minimum turning radius and kinematic constraints. This application demonstrates high efficiency and quality in global planning. Leveraging the powerful data distribution modeling capabilities of the diffusion model, it can quickly escape local optima in large-scale task scenarios, generating a shorter and more efficient inspection sequence than traditional heuristic algorithms. Furthermore, this application exhibits strong adaptability to dynamic environments. Through the incremental search characteristics of the D*Lite algorithm, it only needs to update the node information of the affected area when environmental changes are detected, without requiring full graph recalculation, significantly reducing computational overhead and achieving millisecond-level dynamic obstacle avoidance response. The generated final execution path is smooth and highly executable, effectively solving the problems of numerous polylines and large turns in traditional grid paths, reducing robot motion wear, and ensuring the efficiency and safety of substation inspection operations.

[0076] In one embodiment, such as Figure 6 As shown, in step S202 above, generating a global target inspection path based on the node pair feature matrix using a trained generative diffusion model may include the following steps:

[0077] Step S601: Input the node pair feature matrix into the generative diffusion model for multi-step iterative reverse denoising to generate a global probability distribution heatmap containing the task point set.

[0078] Step S602: Decode the global probability distribution heatmap to obtain the global initial inspection path, and perform a validity check on the global initial inspection path.

[0079] Step S603: If the global initial inspection path passes the validity check, the 2-opt local search algorithm is used to eliminate cross connections in the global initial inspection path and optimize to obtain the global target inspection path.

[0080] Specifically, generating high-quality task point traversal sequences using a trained generative diffusion model can include the following steps:

[0081] (1) Forward diffusion process:

[0082] Gaussian noise is gradually added to the true optimal path solution. At step t, state X... t satisfy:

[0083]

[0084] In the above formula, β is a preset variance scheduling parameter. When T is sufficiently large, X T Approximately a standard Gaussian distribution .

[0085] (2) Reverse denoising and network training:

[0086] Build a prediction network Its input is the current noise. Figure X t Time step t and features H extracted by MatNet context The training objective is to minimize the mean square error between the predicted noise and the actual noise.

[0087]

[0088] (3) Path sampling and decoding:

[0089] In the reasoning stage, from Begin by performing a T-step reverse denoising process:

[0090]

[0091] The final generated X0 is a probability heatmap. A beam search strategy, combined with a masking mechanism to ensure no repeated node visits, is used to decode X0 into an initial path sequence. .

[0092] (4) 2-opt local optimization:

[0093] To address the potential local crossover issue in the generated sequence, a 2-opt optimization is performed, traversing all non-adjacent edge pairs in the path. and If the following conditions are met:

[0094]

[0095] Then perform a reversal operation, reversing the path segments from i+1 to j until no further optimization is possible, thus obtaining the globally optimal sequence. .

[0096] In one embodiment, step S202 above, encoding the task point set using a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships, may include the following steps:

[0097] Obtain the Euclidean distance between any two task points in the task point set, and construct a distance matrix based on the Euclidean distance. Using the MatNet architecture, perform hybrid expert coding through a row / column hybrid expert module, iteratively update the row and column features of the distance matrix until the preset iteration conditions are met, and obtain the node pair feature matrix, which serves as the conditional input of the generative diffusion model.

[0098] Specifically, the terminal constructs an encoder based on the MatNet architecture to extract topological relationships and global dependency features between task points. The MatNet architecture models the combinatorial optimization problem as a matrix, effectively capturing the dual relationships between nodes. The detailed steps are as follows:

[0099] (1) Construction of the distance matrix:

[0100] Calculate the Euclidean distance between any two points i and j in the task point set V, and construct the distance matrix:

[0101]

[0102] In the above formula, .

[0103] (2) Hybrid expert coding:

[0104] The MatNet architecture does not directly encode node coordinates; instead, it encodes the distance matrix. A row / column hybrid expert module iteratively updates the row features R and column features C. The update formula for the l-th layer is as follows:

[0105]

[0106]

[0107] In the above formula, MultiHeadAttn() represents the multi-head attention mechanism, FFN() represents the feedforward neural network, and Norm() represents layer normalization.

[0108] (3) Feature output:

[0109] After L iterations, a node pair feature matrix H that incorporates global topological information is obtained. ij This serves as the conditional input C for the subsequent generative diffusion model. The network structure of the MatNet architecture is as follows: Figure 7 As shown.

[0110] In one embodiment, the method of this application further includes the following steps:

[0111] The simulated robot performs inspections according to the local navigation path and perceives the spatial environment information on the local navigation path in real time through onboard sensors. If temporary obstacles are detected on the local navigation path according to the spatial environment information, the incremental search characteristics of the D*Lite algorithm are used to update only the RHS value of the affected nodes in the local navigation path and add the updated RHS value to the priority queue in order to repair the local navigation path.

[0112] Specifically, the terminal simulates a robot performing inspections according to a local navigation path. During the inspection process, the robot uses onboard sensors to perceive the spatial environment information along the local navigation path in real time. When a sudden obstacle is detected in the map that causes a change in the cost value of the current path node, the incremental search characteristic of the D*Lite algorithm is used to update only the RHS value of the affected node and add it to a priority queue to quickly repair the local path and avoid re-searching the entire map.

[0113] In one embodiment, the method of this application further includes the following steps:

[0114] Obtain true optimal path samples, and gradually add Gaussian noise to the true optimal path samples until an isotropic standard Gaussian distribution is obtained. Based on the updated true optimal path samples, obtain training set, validation set, and test set. Use the training set and validation set to train the initial diffusion model to be trained using a deep learning algorithm and update the model parameters of the initial diffusion model until the performance index of the updated diffusion model on the test set meets the threshold condition, thus obtaining the generative diffusion model.

[0115] In diffusion models, model parameters (such as the weights and biases of the neural network) are crucial components, determining the connection strength between neurons. During training, these parameters are continuously updated using optimization algorithms (such as adaptive optimization algorithms) to enable the model to better simulate the addition and subtraction of noise to generate data.

[0116] Specifically, during the training phase, the terminal gradually adds Gaussian noise to the real optimal path samples until they become an isotropic standard Gaussian distribution. The real optimal path samples are split into training, validation, and test sets according to a certain ratio. Then, the initial diffusion model to be trained is trained using deep learning algorithms on the training and validation sets. This allows the model to learn different types of noise data, helping to adjust model parameters during training to ensure the accuracy of path generation. Finally, the trained diffusion model is tested using the test set to ensure its noise removal capability. Testing helps confirm the actual denoising effect of the model and promptly identify potential problems.

[0117] In one embodiment, such as Figure 8 As shown, a hierarchical path planning method for a traction substation robot is provided in a specific embodiment, which includes the following steps:

[0118] Step S801: Construct a two-dimensional grid map containing obstacle areas, high-risk electromagnetic areas, and passable areas based on the physical environment information of the traction substation. Determine the robot's starting point and the set of task points to be inspected on the two-dimensional grid map.

[0119] Step S802: Obtain the Euclidean distance between any two task points in the task point set, construct a distance matrix based on the Euclidean distance, and use the MatNet architecture to perform hybrid expert coding through a row / column hybrid expert module to iteratively update the row and column features of the distance matrix until the preset iteration conditions are met, thereby obtaining the node pair feature matrix, which serves as the conditional input for the generative diffusion model.

[0120] Step S803: Input the node pair feature matrix into the generative diffusion model for multi-step iterative reverse denoising to generate a global probability distribution heatmap containing the task point set; decode the global probability distribution heatmap to obtain the global initial inspection path, and perform a validity check on the global initial inspection path; if the global initial inspection path passes the validity check, use the 2-opt local search algorithm to eliminate cross connections in the global initial inspection path and optimize to obtain the global target inspection path.

[0121] Step S804: In the global target inspection path, take two adjacent task points as local start and end points respectively, and use the D*Lite algorithm to perform heuristic search in the two-dimensional grid map to generate local navigation paths connecting each task point.

[0122] Step S805: Extract the key inflection points in the local navigation path, construct a third-order Bézier curve to smoothly fit the key inflection points, and generate the final execution path that satisfies the robot's minimum turning radius and kinematic constraints.

[0123] The beneficial effects of the above embodiments are as follows:

[0124] 1) Significantly Improved Global Planning Efficiency and Solution Quality. Addressing the combinatorial optimization challenges posed by the large scale of inspection tasks in traction substations, this application innovatively introduces a generative diffusion model based on the MatNet architecture. Compared to traditional genetic algorithms or ant colony algorithms, this application no longer relies on time-consuming iterative searches. Instead, it learns the probability distribution of high-quality paths and directly generates high-quality initial solutions in parallel from noise. The unique row-column hybrid attention mechanism of the MatNet architecture can capture complex topological dependencies between task points, giving the model extremely strong generalization capabilities. Experiments show that on large-scale TSP problems, the solution speed of this application is approximately 40% faster than traditional heuristic algorithms, and the total length of generated paths is shortened by an average of more than 15%, effectively reducing the robot's ineffective travel and energy consumption, and significantly improving the efficiency of substation inspection operations. Strong Real-Time Obstacle Avoidance Response in Dynamic Environments. For sudden dynamic obstacles such as personnel movement and temporary maintenance equipment placement within the substation, this application employs the D*Lite incremental search algorithm. Unlike the traditional A* algorithm, which requires a full map recalculation upon encountering obstacles, this application utilizes a local update mechanism for RHS values, replanning only the nodes affected by obstacles and their affected areas, reusing most of the historical path information. This incremental calculation method reduces the replanning computation time from seconds to milliseconds, ensuring that the robot can generate a new collision-free path the instant it detects an obstacle, achieving "seamless obstacle avoidance" during movement and greatly ensuring the continuity and safety of inspection operations. The path exhibits high smoothness, conforming to robot kinematic constraints. Addressing the problem of numerous polylines and sharp corners in paths generated by traditional grid map searches, leading to robot jerking and severe wear, this application introduces a third-order Bézier curve smoothing technique at the end effector. By continuously fitting the curvature of key inflection points on the path generated by the D*Lite algorithm, a smooth trajectory conforming to the robot's minimum turning radius constraint is generated. This not only eliminates hard inflections in the path, making the robot's angular velocity changes more continuous and smooth, reducing the mechanical impact of sudden stops and turns on motors and onboard precision testing instruments (such as infrared thermal imagers), but also further improves the robot's average travel speed and extends the equipment's maintenance cycle and service life.

[0125] 2) Excellent adaptability to complex environments and strong system robustness. The layered architecture design of this application achieves a perfect decoupling between "global planning" and "local adaptation". The upper-layer diffusion model focuses on macroscopic task ranking optimization, while the lower-layer D*Lite focuses on microscopic navigation and obstacle avoidance. This loosely coupled structure allows the system to adapt to changes in substation equipment layout or the addition or removal of task points without large-scale reconstruction of the overall algorithm; only local parameters need to be updated. Simultaneously, the introduced 2-opt local optimization strategy, as a post-processing module, further ensures the geometric legitimacy and superiority of the generated path, enabling the application to maintain extremely high stability and robustness even in complex operating conditions such as strong electromagnetic interference in traction substations and unstructured roads.

[0126] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0127] Based on the same inventive concept, this application also provides a hierarchical path planning device for a traction substation robot to implement the hierarchical path planning method for the traction substation robot described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the hierarchical path planning device for a traction substation robot provided below can be found in the limitations of the hierarchical path planning method for the traction substation robot described above, and will not be repeated here.

[0128] In one exemplary embodiment, such as Figure 9 As shown, a hierarchical path planning device for a traction substation robot is provided, which may include:

[0129] The map building module 901 is used to build a two-dimensional grid map containing obstacle areas, electromagnetic high-risk areas and passable areas based on the physical environment information of the traction substation, and to determine the robot's starting point and the set of task points to be inspected on the two-dimensional grid map.

[0130] The network coding module 902 is used to encode the task point set through a matrix coding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships. Based on the node pair feature matrix, a global target inspection path is generated through a trained generative diffusion model.

[0131] The map search module 903 is used to take two adjacent task points as local start and end points respectively in the global target inspection path, and use the D*Lite algorithm to perform heuristic search in the two-dimensional grid map to generate local navigation paths connecting each task point.

[0132] The path generation module 904 is used to extract key inflection points in the local navigation path, construct a third-order Bézier curve to smoothly fit the key inflection points, and generate a final execution path that meets the robot's minimum turning radius and kinematic constraints; the final execution path is used to control the robot to complete the inspection.

[0133] In one embodiment, the network encoding module 902 is further configured to input the node pair feature matrix into a generative diffusion model for multi-step iterative reverse denoising to generate a global probability distribution heatmap containing a set of task points; decode the global probability distribution heatmap to obtain a global initial inspection path; perform a validity check on the global initial inspection path; and, if the global initial inspection path passes the validity check, use a 2-opt local search algorithm to eliminate cross-connections in the global initial inspection path to optimize and obtain a global target inspection path.

[0134] In one embodiment, the network encoding module 902 is further configured to obtain the Euclidean distance between any two task points in the task point set, construct a distance matrix based on the Euclidean distance, and use the MatNet architecture to perform hybrid expert encoding on the distance matrix to generate a node pair feature matrix.

[0135] In one embodiment, the network coding module 902 is further configured to use the MatNet architecture to perform hybrid expert coding through a row / column hybrid expert module, iteratively updating the row and column features of the distance matrix until a preset iteration condition is met, to obtain a node pair feature matrix, which serves as the conditional input of the generative diffusion model.

[0136] In one embodiment, the device may further include: a path repair module, used to simulate a robot performing an inspection according to a local navigation path and to perceive spatial environment information on the local navigation path in real time through onboard sensors; if a temporary obstacle is detected on the local navigation path according to the spatial environment information, the incremental search characteristic of the D*Lite algorithm is used to update only the RHS value of the affected node in the local navigation path and add the updated RHS value to a priority queue to repair the local navigation path.

[0137] In one embodiment, the apparatus may further include: a model training module, configured to acquire true optimal path samples, progressively add Gaussian noise to the true optimal path samples until an isotropic standard Gaussian distribution is obtained, and obtain a training set, a validation set, and a test set based on the updated true optimal path samples; use the training set and validation set to train the initial diffusion model to be trained using a deep learning algorithm and update the model parameters of the initial diffusion model until the performance index of the updated diffusion model on the test set meets a threshold condition, thereby obtaining a generative diffusion model.

[0138] Each module in the hierarchical path planning device of the traction substation robot described above can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0139] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a hierarchical path planning method for a traction substation robot. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0140] Those skilled in the art will understand that Figure 10The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0141] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0143] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0144] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0145] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0146] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0147] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A hierarchical path planning method for a traction substation robot, characterized in that, The method includes: Based on the physical environment information of the traction substation, a two-dimensional grid map is constructed, which includes obstacle areas, electromagnetic high-risk areas, and passable areas. The starting point of the robot and the set of task points to be inspected are determined on the two-dimensional grid map. The task point set is encoded using a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships. A global target inspection path is then generated based on the node pair feature matrix using a trained generative diffusion model. In the global target inspection path, two adjacent task points are taken as local start and end points respectively. The D*Lite algorithm is used to perform a heuristic search in the two-dimensional grid map to generate a local navigation path connecting each task point. Key inflection points in the local navigation path are extracted, and a third-order Bézier curve is constructed to smoothly fit the key inflection points, generating a final execution path that satisfies the robot's minimum turning radius and kinematic constraints; the final execution path is used to control the robot to complete the inspection.

2. The method according to claim 1, characterized in that, The generation of a global target inspection path based on the node pair feature matrix using a trained generative diffusion model includes: The node pair feature matrix is ​​input into the generative diffusion model for multi-step iterative reverse denoising to generate a global probability distribution heatmap containing the task point set. The global probability distribution heatmap is decoded to obtain the global initial inspection path, and the global initial inspection path is validated for legality. If the global initial inspection path passes the validity check, the 2-opt local search algorithm is used to eliminate cross connections in the global initial inspection path and optimize the global target inspection path.

3. The method according to claim 1, characterized in that, The step of encoding the task point set using a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships includes: Obtain the Euclidean distance between any two task points in the set of task points, and construct a distance matrix based on the Euclidean distance; The distance matrix is ​​subjected to hybrid expert encoding using the MatNet architecture to generate the node pair feature matrix.

4. The method according to claim 3, characterized in that, The step of using the MatNet architecture to perform hybrid expert encoding on the distance matrix to generate the node pair feature matrix includes: The MatNet architecture is used to perform hybrid expert coding through a row / column hybrid expert module, and the row and column features of the distance matrix are iteratively updated until the preset iteration conditions are met, so as to obtain the node pair feature matrix, which is used as the condition input of the generative diffusion model.

5. The method according to claim 1, characterized in that, The method further includes: The robot is simulated to perform inspections according to the local navigation path, and to perceive the spatial environment information on the local navigation path in real time through onboard sensors. If a temporary obstacle is detected on the local navigation path based on the spatial environment information, the incremental search characteristic of the D*Lite algorithm is used to update only the RHS value of the affected nodes in the local navigation path and add the updated RHS value to the priority queue in order to repair the local navigation path.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Obtain the true optimal path sample, and gradually add Gaussian noise to the true optimal path sample until an isotropic standard Gaussian distribution is obtained. Based on the updated true optimal path sample, obtain the training set, validation set and test set. The initial diffusion model to be trained is trained using a deep learning algorithm on the training set and the validation set, and the model parameters of the initial diffusion model are updated until the performance index of the updated diffusion model on the test set meets the threshold condition, thus obtaining the generative diffusion model.

7. A hierarchical path planning device for a traction substation robot, characterized in that, The device includes: The map building module is used to construct a two-dimensional grid map containing obstacle areas, electromagnetic high-risk areas, and passable areas based on the physical environment information of the traction substation, and to determine the robot's starting point and the set of task points to be inspected on the two-dimensional grid map. The network encoding module is used to encode the task point set through a matrix encoding network based on the MatNet architecture to obtain a node pair feature matrix that integrates global topological relationships, and to generate a global target inspection path based on the node pair feature matrix through a trained generative diffusion model. The map search module is used to take two adjacent task points as local start and end points respectively in the global target inspection path, and use the D*Lite algorithm to perform heuristic search in the two-dimensional grid map to generate local navigation paths connecting each task point. The path generation module is used to extract key inflection points in the local navigation path, construct a third-order Bézier curve to smoothly fit the key inflection points, and generate a final execution path that satisfies the robot's minimum turning radius and kinematic constraints; the final execution path is used to control the robot to complete the inspection.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images