Dynamic obstacle avoidance method and system for substation inspection robot

Abstract

The invention discloses a dynamic obstacle avoidance method and system for a substation inspection robot. The position of the robot is positioned based on laser point cloud data, and static and dynamic obstacle features are extracted at the same time. An environment association graph is constructed through perception information in the advancing process, and feature information of each node serves as an attribute of the environment association graph; calculating the distance, the relative speed and the relative angle between the edge weights of the nodes according to the nodes; on the basis of the environment association graph, aggregating all-graph obstacle node information to generate real-time feature representation for robot nodes, and correcting a calculation mode of attention weight according to edge weight between the nodes; and finally outputting a real-time feature vector of the robot node. And defining a reinforcement learning framework and executing path planning. According to the invention, through combination of GAT and dynamic planning, efficient obstacle avoidance and path planning of the robot in a substation environment are realized, and the decision and execution capability of the robot in a dynamic complex scene can be improved.

Description

Technical Field

[0001] This invention belongs to the field of mobile robot perception and obstacle avoidance technology, specifically relating to a dynamic obstacle avoidance method and system for substation inspection robots. Background Technology

[0002] With the rapid development of mobile robot technology, robots have demonstrated significant advantages in the field of inspection. Through autonomous localization and navigation technologies, robots can efficiently and safely plan paths and avoid obstacles, reducing human intervention and achieving the goal of independently completing tasks. These technologies not only improve the efficiency and accuracy of inspections but also effectively avoid human error and safety risks.

[0003] In terms of path planning and obstacle avoidance, several constructive methods have been proposed. For example, patent application CN120124969A analyzes inspection data to assess equipment status and dynamically adjusts the inspection items and times of subsequent tasks based on the status results (such as health or maintenance required), achieving intelligent optimization of inspection tasks and precise allocation of resources. Patent application CN114115286A detects obstacles ahead in real time and obtains their position coordinates. Combined with preset target positions, it uses particle swarm optimization algorithm to dynamically plan the optimal inspection path, enabling the robot to autonomously bypass obstacles and accurately reach the inspection point. Patent application CN109725233A deploys multiple specialized robots to handle ground and high-altitude areas respectively. Combining magnetic trajectory navigation, RFID positioning, and multi-sensor fusion, it achieves autonomous walking, obstacle crossing, slope climbing, and equipment inspection, improving inspection efficiency, coverage, and automation level.

[0004] While existing obstacle avoidance technologies each have their advantages, they still face challenges in the complex environment of substations, such as large obstacle positioning errors, low obstacle avoidance path calculation efficiency, and insufficient environmental adaptability. These issues make it difficult to meet the actual needs of substation inspection robots for efficient and accurate obstacle avoidance. As the density of equipment and the complexity of the environment in substations increase, the requirements for obstacle avoidance technology for inspection robots are becoming increasingly stringent. There is an urgent need for a new obstacle avoidance technology that can accurately locate obstacles, quickly plan obstacle avoidance paths, and adapt to complex environments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a dynamic obstacle avoidance method and system for substation inspection robots. The method involves extracting features from both static and dynamic obstacles and constructing an environmental association graph with the robot and obstacles as nodes, incorporating edge weights that integrate distance and motion relationships. A graph attention network is then used to aggregate information from the entire graph, generating real-time feature vectors for the robot nodes. These vectors are then used as state inputs to a reinforcement learning agent, which outputs control commands to complete path planning.

[0006] The first aspect of this application discloses a dynamic obstacle avoidance method for a substation inspection robot, which adopts the following technical solution: The robot loads a grid map of the substation to obtain a preliminary layout environment within the station, and calculates its real-time position in conjunction with laser point cloud data while moving; it acquires laser point cloud data in real time, and after preprocessing, identifies and extracts static and dynamic obstacle features in parallel; An environmental graph is constructed using the perception information during the movement. The nodes include robot nodes and obstacle nodes, and the feature information of each node is used as its attribute. The edge weights between nodes are calculated by combining the distance, relative speed and relative angle between the nodes. Based on the environmental graph, obstacle node information of the entire graph is aggregated to generate real-time feature representations for robot nodes; the feature information of each node is passed to neighboring nodes through the message passing mechanism of the graph attention network. During the passing process, the calculation method of attention weight is corrected according to the edge weight between nodes; finally, the real-time feature vector of the robot node is output. Define a reinforcement learning framework and perform path planning; convert the real-time feature vectors of robot nodes into reinforcement learning state vectors, select control commands based on the state vectors, and provide feedback guidance through a reward function; output the optimal path and train and validate it in a simulation environment.

[0007] Furthermore, the static obstacle features include the straight line features and circular features of the static obstacle, which are extracted by calling the Hough transform method.

[0008] Further, the dynamic obstacle features are extracted, including: Based on the extraction of static obstacle features, the robot uses LiDAR to perform continuous multi-frame scanning of the environment in the direction of robot movement; and performs Hough transform on the point cloud data of each frame according to the time series to extract the straight line features and circular features in each frame in the same way as the static obstacle extraction. By comparing the linear and circular features between consecutive frames, the temporal changes of these features are extracted as dynamic obstacle features.

[0009] Furthermore, the temporal variations of the linear features are extracted, including: For each frame of extracted straight line features, record the starting and ending coordinates of the line; if the straight line features change displacement between consecutive frames, they are identified as dynamic straight line obstacles. Based on the displacement changes between consecutive frames, the motion velocity of the straight line feature in the horizontal and vertical coordinate directions is calculated.

[0010] Furthermore, the temporal variations of the circular features are extracted, including: For each frame of extracted circular features, the positional changes of the circle in consecutive frames are tracked based on the center coordinates and radius; the motion velocity of the circular features is calculated by the displacement changes of the center coordinates in the horizontal and vertical directions.

[0011] Furthermore, for robot nodes, their attributes include pose, travel speed, and relative position to the target point; The pose refers to the position coordinates and orientation of the robot node in the global coordinate system; the travel speed refers to the linear velocity and angular velocity of the robot node in the horizontal and vertical coordinate directions; the relative position of the target point refers to the position coordinates of the inspection target point relative to the robot node.

[0012] Furthermore, for obstacle nodes, their attributes include geometry type, position, size, and movement speed; The geometric types include straight-line features and circular features; the position is the position coordinate in the global coordinate system; the size is the geometric dimensions of the obstacle, including the straight-line length of a straight obstacle and the radius of a circular obstacle; The movement speed is the same as the movement speed of the dynamic obstacle.

[0013] Furthermore, the edge weights between nodes are a combination of the distance, relative velocity, and relative angle between the nodes; including: For the edges between robot nodes and obstacle nodes, the edge weights are... It can be expressed as the product of the first distance influence term and the first association strength term; The first distance influence term is the reciprocal of the Euclidean distance between the robot node and the obstacle node; the first association strength term is an exponential function with the natural constant e as the base, and the exponent is the weighted sum of the absolute value of the first relative velocity and the cosine value of the first relative angle. The first relative velocity and the first relative angle are the relative velocity and relative angle between the robot node and the obstacle node.

[0014] Furthermore, for the edges between obstacle nodes, the edge weights... It is represented as the product of the second distance influence term and the second association strength term; The second distance influence term is the reciprocal of the Euclidean distance between obstacle nodes; the second correlation strength term is the weighted sum of the absolute values ​​of the second relative angle difference and the second relative velocity. The second relative velocity and the second relative angle are the relative velocity and relative angle between the obstacle nodes.

[0015] Furthermore, the calculation method for attention weights is modified based on the edge weights between nodes, including: Based on the fundamental formula for attention weights in graph attention networks, a bias term is introduced into the exponential function of the attention score; this bias term represents the edge weights. The product of a learnable hyperparameter; by Represents the edges between nodes, if Let be the edge between the robot node and the obstacle node. ;like If the edges are between obstacle nodes, then .

[0016] Furthermore, a reinforcement learning framework is defined, including: The state is defined as the feature vector output by the robot node after passing through the graph attention mechanism; the action is defined as the robot's linear velocity increment and angular velocity increment; the reward function is defined as the comprehensive weighted value of progress reward, collision penalty, safety penalty, smoothness reward and dynamic obstacle response reward. Among them, the dynamic obstacle response reward is based on the calculation of the relative speed between the robot and the dynamic obstacle.

[0017] The second aspect of this application discloses a dynamic obstacle avoidance system for a substation inspection robot, which operates the dynamic obstacle avoidance method described in the first aspect of this application. The system includes: Obstacle feature extraction module; used for the robot to load the substation grid map to obtain the preliminary layout environment of the station, and to calculate its real-time position with the help of laser point cloud during movement; real-time acquisition of laser point cloud data, and after preprocessing, parallel identification and extraction of static obstacle features and dynamic obstacle features; Graph structure building module; used to construct an environmental graph based on perception information during the movement process. Nodes include robot nodes and obstacle nodes, and the feature information of each node is used as its attribute; the edge weights between nodes are calculated by combining the distance, relative speed and relative angle between nodes. Attention weight correction module: Based on the environmental graph, it aggregates obstacle node information from the entire graph to generate real-time feature representations for robot nodes; the feature information of each node is passed to neighboring nodes through the message passing mechanism of the graph attention network. During the passing process, the calculation method of attention weight is corrected according to the edge weight between nodes; finally, the real-time feature vector of the robot node is output. The obstacle avoidance path planning module is used to define the reinforcement learning framework and perform path planning; it converts the real-time feature vectors of robot nodes into reinforcement learning state vectors, and the robot nodes select control commands based on the state vectors and provide feedback guidance through the reward function; it outputs the optimal path and trains and validates it in a simulation environment.

[0018] The beneficial effects of this invention are that, compared with the prior art, 1. This application enhances the intelligence and accuracy of dynamic obstacle avoidance. Specifically, by introducing a combination coefficient of multiple factors such as distance, relative velocity, and relative angle into the edge weight calculation between the robot and obstacle nodes, the robot can more accurately perceive and handle different types of obstacles in complex environments. The combination coefficient (e.g., the product of the distance influence term and the association strength term) can dynamically adjust the robot's reaction sensitivity to obstacles. Especially when dealing with dynamic obstacles, it can more flexibly adjust the obstacle avoidance strategy to avoid collisions caused by excessively close distances or rapidly changing obstacles.

[0019] 2. This application optimizes the decision-making process for path planning. Specifically, by combining edge weights to adjust attention weights and the relative speed between the robot and obstacles, path planning not only considers traditional factors such as progress and collisions but also flexibly responds to changes in the speed and position of dynamic obstacles. By defining appropriate weighting coefficients and learnable hyperparameters, the robot can more accurately adjust its response strategy to different situations, enabling it to adaptively adjust its path when performing inspection tasks, avoiding conflicts with dynamic obstacles and improving operational stability and safety. This approach comprehensively enhances the robot's ability to cope with complex, real-time changing environments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the execution flow of the dynamic obstacle avoidance method. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0022] As an embodiment of this application, a specific implementation method for a dynamic obstacle avoidance method for a substation inspection robot is disclosed. Please refer to the example flowchart for details. Figure 1 .

[0023] S1: As one implementation method of this embodiment, environmental information perception and processing are performed; this stage enables the robot to perceive and process the environmental information of the substation and accurately identify static and dynamic obstacles, ensuring the safety and efficiency of the inspection task.

[0024] 1.1: When the robot starts, it loads a pre-drawn substation grid map, which marks feasible areas (such as passageways) and static obstacles (such as electrical equipment, walls, etc.) within the substation. This map helps the robot understand the initial environmental layout.

[0025] After the robot initially acquires the static layout of the substation, it invokes localization algorithms such as laser SLAM to scan the laser point cloud data of the surrounding environment of the robot's current location in real time. This data is then matched with the loaded prior map (i.e., the pre-drawn substation grid map) to calculate the robot's accurate position within the substation, including its two-dimensional coordinates and orientation. In further explanation, the two-dimensional coordinates here are based on the global coordinate system of the substation plane, and the orientation is the angle between the robot's forward direction and the X-axis of the global coordinate system.

[0026] The acquired laser point cloud data is converted from polar coordinates to Cartesian coordinates in the robot coordinate system, and a filtering algorithm is called to remove noise in order to improve the accuracy of subsequent obstacle detection.

[0027] 1.2: Identify static and dynamic obstacles from preprocessed laser point cloud data and extract obstacle features within the substation.

[0028] For static obstacles: In a substation environment, static obstacles refer to stationary objects within the substation scene, such as walls, shelves, and pillars. The geometric features of these static obstacles can be directly extracted for environmental modeling and path planning. As an optional implementation method, the Hough transform method can be used to extract the straight-line and circular features of static obstacles.

[0029] For dynamic obstacles: In a substation environment, dynamic obstacles are simply objects that move in real time or can move at any time within the substation, such as inspection personnel and mobile equipment. Compared to static obstacles, the state of dynamic obstacles changes over time. Feature extraction not only needs to identify "what" and "where" an obstacle is, but also accurately determine "how it will move." Therefore, the Hough transform's capabilities are insufficient to meet the actual needs of substation scenarios.

[0030] Considering the above reasons, as one implementation method of this embodiment, this application constructs a dynamic processing mechanism based on continuous frames, which breaks down dynamic obstacle extraction into multiple consecutive static obstacle extraction methods to extract the shape and motion features of dynamic obstacles. Specifically: In a further implementation, based on the extraction of static obstacle features, the robot's onboard LiDAR performs multi-frame scans of the current environment in parallel. Each frame of point cloud data represents the environmental information "seen" by the robot at each moment. The data from consecutive frames includes spatial and temporal information about dynamic obstacles.

[0031] Based on the time series, a Hough transform is performed on each frame of point cloud data to extract straight line and circular features in each frame, using a static obstacle extraction method. By comparing consecutive frames, the changes of the extracted features over time are analyzed. The steps are as follows: (1) For each frame of extracted straight line features, record the starting coordinates of the straight line. and endpoint coordinates If the straight line feature undergoes displacement changes between consecutive frames, it is identified as a dynamic straight line obstacle.

[0032] The velocity of the linear feature is calculated based on the displacement changes between consecutive frames, and is expressed as: ; In the above formula, and These are the displacement rates in the X and Y directions, respectively, i.e., the velocity of motion; and Let X represent the displacement changes in the X and Y directions. For time intervals.

[0033] (2) Similarly, for each frame of extracted circular features, based on the center coordinates and radius Track the position changes of a circle in consecutive frames, calculate its velocity and direction, and represent it as follows: ; In the above formula, and These are the center coordinates between consecutive frames. The rate of change of displacement in the X and Y directions, i.e., the velocity of motion; and Coordinates of the center of the circle The amount of displacement change in the X and Y directions. For time intervals.

[0034] As one implementation method of this embodiment, Kalman filtering or particle filtering algorithms are used to track the extracted dynamic obstacles and predict their position, velocity, and trajectory. In a further implementation method, for the dynamic obstacle features (including straight line and circular features) extracted in each frame, the position information of past frames and the data of the current frame are fused using methods such as Kalman filtering to estimate the current state of the obstacle, including its position, velocity, and acceleration.

[0035] S2: As one implementation method of this embodiment, an environmental association graph is constructed by sensing various information in the travel environment; the purpose of this graph structure data is to provide structured support for subsequent path planning, obstacle avoidance and task execution.

[0036] 2.1: In one specific implementation, the environment association graph consists of a set of nodes. Sum of edges The structure consists of nodes, which represent various entities in the substation environment. In this embodiment, they are specifically divided into robot nodes and obstacle nodes, with the characteristic information of each node serving as its attributes.

[0037] (1) A robot node represents the robot's position, motion state, and task objective. The attributes of a robot node include: Pose, i.e., the position coordinates and orientation obtained by the 1.1 positioning algorithm, is... The formal representation; Travel speed, linear velocity and angular velocity on two coordinate axes, is used to describe the robot's motion state and helps with subsequent trajectory planning and obstacle avoidance decisions. The relative position of the target point indicates the relative position of the inspection target point relative to the robot. This indicates that it is used for path planning and adjusting the travel trajectory.

[0038] (2) An obstacle node represents each obstacle in the environment. The type of obstacle can be static (e.g., walls, shelves) or dynamic (e.g., pedestrians, moving objects). For an obstacle node, its attributes include: Geometric type: The geometric shape of the obstacle, which may be a straight line (such as a wall or shelf edge) or a circle (such as a pillar or pedestrian). Location, the spatial position of an obstacle; static obstacles are represented by coordinates, while dynamic obstacles are represented by real-time coordinates; Size refers to the geometric dimensions of the obstacle (such as the length of a straight line, the radius of a circle, etc.). Size information helps calculate the minimum obstacle avoidance distance between the robot and the obstacle.

[0039] Movement speed, for dynamic obstacles, is used to describe the movement state of the obstacle.

[0040] In a further detailed description of the implementation, in order to solve the problem of different coordinate systems between the robot pose and the obstacle node, this application transforms the obstacle from the local coordinate system to the global coordinate system through coordinate system transformation, so that the nodes (especially obstacle nodes) in the environment association graph can be correctly located in the global coordinate system.

[0041] 2.2: In a further implementation, the edges in the environmental graph represent the spatial relationships between different nodes, and the edge weights reflect the relative influence or distance between the nodes. Edges can be categorized as follows: (1) The edge between the robot and the obstacle; For the edge between the robot node and the obstacle node, the edge weight can be expressed as the product of the first distance influence term and the first association strength term; The first distance influence term is the reciprocal of the Euclidean distance between the robot node and the obstacle node; the first association strength term is an exponential function with the natural constant e as the base, and the exponent is the weighted sum of the absolute value of the first relative velocity and the cosine value of the first relative angle. The first relative velocity and the first relative angle are the relative velocity and relative angle between the robot node and the obstacle node.

[0042] The mathematical form is as follows: Each robot node is connected to all obstacle nodes by an edge. The weight of each edge is inversely proportional to the distance and depends on the movement speed of both types of nodes. The weight of each edge of this type... It can be represented as: ; In the formula, This represents the Euclidean distance between the robot and the obstacle. The closer the distance, the larger the edge weight, indicating a stronger connection between the robot and the obstacle. The relative velocity between the robot node and the obstacle node; Indicates the relative angle between the robot and the obstacle; and These are adjustable weighting coefficients used to control the degree to which relative velocity and relative angle affect the weight of this type of edge.

[0043] (2) The edges between obstacles; For the edges between obstacle nodes, the edge weight can be expressed as the product of the second distance influence term and the second association strength term; The second distance influence term is the reciprocal of the Euclidean distance between obstacle nodes; the second correlation strength term is the weighted sum of the absolute values ​​of the second relative angle difference and the second relative velocity. The second relative velocity and the second relative angle are the relative velocity and relative angle between the obstacle nodes.

[0044] The mathematical form is as follows: Edges also need to be established between adjacent obstacle nodes to model the structured information of the environment. In this embodiment, the weight of this type of edge is associated with the distance, relative position, and motion state (in a scenario where either end obstacle is a dynamic obstacle) between obstacles. This is represented as: ; In the formula, Represents the edges between obstacles; This represents the Euclidean distance between two obstacles; and These are the relative angles between obstacles and the reference angle, respectively. The relative velocity between obstacles can be calculated using a two-dimensional velocity synthesis method; and The adjustment coefficient controls the importance of relative angle and relative velocity in the edge weights.

[0045] It's important to further clarify that the goal of calculating the edge weights between robot nodes and obstacle nodes is to evaluate the relationship between the robot and obstacles, particularly how it affects the robot's obstacle avoidance behavior and path planning. Since robot behavior is dynamic, the edge weights between the robot and obstacles need to consider dynamic factors (such as relative speed and relative angle) and be more sensitive to these factors; therefore, an exponential calculation method is used. These factors directly affect the threat level between the robot and the obstacle at the current moment, and how the robot adjusts its path to avoid the obstacle.

[0046] The goal of edge weight calculation between obstacle nodes is to evaluate the relative relationships between obstacles, particularly how they affect the structural modeling of the environment. The relative positions, shapes, and motion states of obstacles (in the case of dynamic obstacles) have an indirect impact on robot path planning. In this case, the relationships between obstacles reflect more the spatial layout of the environment and the physical interactions between obstacles, rather than directly affecting how the robot interacts with the obstacles.

[0047] S3: As one implementation method of this embodiment, based on the environmental association graph, the feature information of surrounding obstacles is aggregated to generate new feature representations for robot nodes, enabling the robot to understand the environmental context and make more accurate decisions.

[0048] In S2, the environmental graph has been constructed. In this stage, the features of the robot node and the neighboring obstacle node will be used as input to the Graph Attention Network (GAT).

[0049] In a further implementation, feature aggregation is performed using message passing and attention mechanisms. Robot nodes need to share information with surrounding obstacle nodes and influence the robot nodes based on the characteristics of each obstacle (including the motion state of dynamic obstacles). The feature update of each node depends on its correlation with its neighboring nodes, and this correlation is dynamically adjusted through the attention mechanism. Specifically: Each node (including robot nodes and obstacle nodes) transmits its own characteristic information (i.e., attributes) to its neighboring nodes through the GAT message passing mechanism; Robot nodes update their own feature information based on the feature information of their neighbors, and the information of each neighbor node is weighted according to its importance to the robot; this importance is represented by edge weights. During aggregation, these edge weights are incorporated into the calculation of the GAT attention mechanism. Specifically: GAT's attention mechanism typically calculates the attention weights between nodes using the following formula: ; In the formula, Represents a node and nodes Attention weights between nodes (without distinguishing between robot nodes and obstacle nodes for now); and These are the feature representations of the two nodes, This is the transpose of an attention coefficient vector; It is a linear transformation matrix. For vector join operations, It is a non-linear activation function. Represents a node The set of neighboring nodes, Indicates the index of the neighboring node.

[0050] In this embodiment, when constructing the environmental relationship graph, the relationship between the robot and obstacles, and between obstacles themselves, is no longer simply based on changes in distance. Instead, the relative speed, distance, and angle between the robot and obstacles, and between obstacles, are considered. In other words, when calculating attention weights, this application considers not only the feature information of the nodes themselves, but also the movement of obstacles and the robot, and changes in the spatial structure, thereby enabling the robot to focus on the more pressing and threatening obstacles.

[0051] The attention weight formula in this embodiment is then modified as follows: ; In the formula, Let the edge weight be , if Let be the edge between the robot node and the obstacle node. ;like If the edge is between two obstacle nodes, then . This is a hyperparameter used to control the degree to which edge weights affect attention calculation.

[0052] S4: As one implementation method of this embodiment, the robot is trained to make optimal obstacle avoidance decisions in a dynamic and complex environment through reinforcement learning algorithms, so that it can efficiently and smoothly avoid obstacles and complete the task.

[0053] 4.1: In a further implementation, the basic framework of reinforcement learning is first established, including defining states, actions, and reward functions. Specifically: State: Defined as a feature vector of a robot node, representing the robot's current environmental perception. Specifically, it is the state vector output by S3's GAT, containing all key current environmental information; the state vector includes the robot's current position, the position and velocity of obstacles, relative positions, and other perception information (e.g., distance, angle, velocity, etc.). The robot understands its current state in the environment from this state information and selects appropriate actions based on this state.

[0054] Action: A control command that the robot can execute; in this embodiment, it is defined as a linear velocity increment. and angular velocity increment The action space defines all the possible actions a robot can take. It is usually a continuous value that represents the specific action a robot can perform at a given moment.

[0055] Reward function: Provides feedback to the robot, guiding it to learn the optimal strategy. The reward function comprehensively considers the robot's progress, collision situation, safety, smoothness, and especially its response to dynamic obstacles. It is expressed as: ; For the reward function; , , , and These are progress rewards, collision penalties, safety penalties, smoothness rewards, and dynamic obstacle response rewards. These are the weighting coefficients for each item.

[0056] In a further specific implementation, progress rewards are used to encourage the robot to move closer to the target point. The closer the robot is to the target, the greater the reward, thus encouraging the robot to get as close to the target as possible. This item is related to the current distance between the robot and the target, and is represented as: ; in, The coefficient for progress rewards adjusts the contribution of progress to the reward function; and These represent the current distance between the robot and the target point, and the initial distance between the robot and the target point, respectively.

[0057] In a further specific implementation, a collision penalty term is used to severely punish situations where the robot collides with obstacles, providing a very large negative reward. This is represented as: ; If a collision occurs, This indicates that this is unacceptable behavior and the punishment is very severe; if no collision occurs, There are no rewards or punishments.

[0058] In a further specific implementation, a safety penalty ensures that the robot maintains a sufficient safe distance from obstacles. The closer the distance, the greater the penalty. This is represented as: ; in, The coefficient representing the safety penalty adjusts the impact of the safety distance on the penalty. This indicates the minimum distance between the robot and the nearest obstacle. This penalty factor indicates that if the robot gets too close, the safety penalty will be greater. It can be adjusted... To control the severity of the punishment.

[0059] In a further specific implementation, the smoothness reward encourages the robot to adopt smooth control behavior, avoiding abrupt acceleration or turning. This is represented as: ; in, and These are the changes in linear velocity and angular velocity, respectively. and These are the coefficients for linear velocity smoothness and angular velocity smoothness, respectively. This formula encourages the robot to control its linear and angular velocities, resulting in smaller variations and smoother movements.

[0060] In a further specific implementation, the dynamic obstacle response reward is used to penalize the robot for slow response to dynamic obstacles and to reward the robot for avoiding dynamic obstacles as early as possible. This is represented as: ; in, The relative speed between the robot and dynamic obstacles, This is a coefficient for the dynamic obstacle response reward, adjusting the intensity of the reward. The meaning of the dynamic obstacle response reward formula is: if the robot approaches a dynamic obstacle and does not react in time, the reward will be lower to avoid a collision.

[0061] 4.2: Apply reinforcement learning methods to path planning and train and validate them in a simulation environment.

[0062] Based on the state, action, and reward functions defined in section 4.1 above, define all entities in the simulation environment (such as robots, obstacles, target locations, etc.). The environment needs to implement a feedback mechanism for robot actions and the calculation of state and reward functions.

[0063] In a further optional implementation step, a suitable reinforcement learning algorithm is selected, such as a policy gradient-based algorithm, PPO, or A3C. The neural network models are initialized for the policy network and the value network. The policy network is used to output actions, and the value network is used to estimate the value of the current state.

[0064] In a further implementation, the robot starts from its initial position and sets a target point according to the task requirements. The target point can be a fixed coordinate or a dynamically changing coordinate.

[0065] Based on the robot's current environmental information (robot's own pose, obstacle node information, relative velocity, etc.), the improved GAT in this application outputs the robot's current state vector. ; The robot selects an action (controlling its linear and angular velocities) to perform and interacts with the environment, which updates its state and reward based on the robot's actions. Action selection can be based on a greedy policy or a policy network.

[0066] Instant rewards are calculated based on the actions performed and environmental feedback. This is a combined value of target reward, collision penalty, safety, smoothness, and dynamic obstacle response reward. In a further optional implementation, experience replay and a target network are used to optimize the policy and value function. Experience replay stores historical interactions in a pool and breaks down data correlations through random sampling, thereby improving the stability of learning.

[0067] Use gradient descent or other optimization algorithms to update the weights of the policy network, enabling the robot to select actions that maximize cumulative rewards during training. Store the current state, action, reward, and next state in an experience replay pool for subsequent training. The robot learns through extensive trial and error in its environment, continuously adjusting its strategy and exploring and optimizing its path planning. During training, the cumulative reward is monitored; once the cumulative reward stabilizes, or after a predetermined number of training iterations (e.g., millions of interactions), the training can be considered converged. The resulting policy network is the optimal policy, capable of making optimal decisions based on the environmental state, thereby achieving optimal obstacle avoidance and path planning for the robot in complex and dynamic environments.

[0068] As an optional implementation, the trained strategy can be tested on a simulation platform to verify its effectiveness. Commonly used simulation platforms include Gazebo (integrated with ROS), V-REP, and Webots. These platforms simulate real-world environments and robot behavior, allowing for risk-free testing of the robot. The specific simulation process includes: Set up the robot model, configure the robot sensors (LiDAR, cameras, etc.), and the dynamic obstacles that need to be handled (such as other robots or pedestrians). Configure the substation environment, including static obstacles (such as walls and equipment) and dynamic obstacles (such as mobile equipment and personnel).

[0069] Deploy the trained optimal strategy into a simulation environment. Let the robot perform path planning tasks and observe how it avoids obstacles and reaches the target point in a dynamic environment.

[0070] By simulating and recording the robot's task execution process, the effectiveness of its path planning, such as obstacle avoidance, task completion time, and collision occurrence, can be analyzed. Robot path maps can be generated, marking obstacles, target locations, and the actual path. The robot's cumulative rewards can be recorded and analyzed to determine if overfitting or underfitting is occurring.

[0071] If the simulation results are unsatisfactory, the reward function, action space, or algorithm parameters can be adjusted for optimization. Through multiple simulations, the strategy can be gradually improved to enhance the accuracy and robustness of path planning. By training a path planning strategy using reinforcement learning algorithms and validating it in a simulation environment, robots can achieve intelligent obstacle avoidance and path planning in complex dynamic environments. The entire process includes initializing the environment and algorithm parameters, training the robot's interaction with the environment, updating the strategy, and finally, simulation verification. After training, the robot can autonomously plan paths based on the learned optimal strategy, thereby performing tasks in real-world scenarios.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A dynamic obstacle avoidance method for a substation inspection robot, characterized in that, include: The robot loads a grid map of the substation to obtain a preliminary layout environment within the station, and calculates its real-time position using laser point cloud data while moving. Real-time acquisition of laser point cloud data, followed by preprocessing and parallel identification and extraction of static and dynamic obstacle features; An environmental graph is constructed using the perception information during the movement process. The nodes include robot nodes and obstacle nodes, and the feature information of each node is used as its attribute. The edge weights between nodes are calculated by combining the distance, relative velocity, and relative angle between the nodes. Based on the environmental graph, obstacle node information of the entire graph is aggregated to generate real-time feature representations for robot nodes; the feature information of each node is passed to neighboring nodes through the message passing mechanism of the graph attention network. During the passing process, the calculation method of attention weight is corrected according to the edge weight between nodes; finally, the real-time feature vector of the robot node is output. Define a reinforcement learning framework and perform path planning; convert the real-time feature vectors of robot nodes into reinforcement learning state vectors, select control commands based on the state vectors, and provide feedback guidance through a reward function; output the optimal path and train and validate it in a simulation environment.

2. The dynamic obstacle avoidance method for a substation inspection robot according to claim 1, characterized in that, The static obstacle features include the straight line features and circular features of the static obstacle, which are extracted by calling the Hough transform method.

3. The dynamic obstacle avoidance method for a substation inspection robot according to claim 1, characterized in that, Extracting the dynamic obstacle features includes: Based on the extraction of static obstacle features, the robot uses LiDAR to perform continuous multi-frame scanning of the environment in the direction of robot movement; and performs Hough transform on the point cloud data of each frame according to the time series to extract the straight line features and circular features in each frame in the same way as the static obstacle extraction. By comparing the linear and circular features between consecutive frames, the temporal changes of these features are extracted as dynamic obstacle features.

4. The dynamic obstacle avoidance method for a substation inspection robot according to claim 3, characterized in that, Extracting temporal variations of linear features, including: For each frame of extracted straight line features, the starting and ending coordinates of the line are recorded; if the straight line features change displacement between consecutive frames, they are identified as dynamic straight line obstacles. Based on the displacement changes between consecutive frames, the motion velocity of the straight line feature in the horizontal and vertical coordinate directions is calculated.

5. A dynamic obstacle avoidance method for a substation inspection robot according to claim 3, characterized in that, Extracting temporal variations of circular features, including: For each frame of extracted circular features, the positional changes of the circle in consecutive frames are tracked based on the center coordinates and radius; the motion velocity of the circular features is calculated by the displacement changes of the center coordinates in the horizontal and vertical directions.

6. A dynamic obstacle avoidance method for a substation inspection robot according to claim 1, characterized in that, For a robot node, its attributes include pose, travel speed, and relative position to the target point; The pose refers to the position coordinates and orientation of the robot node in the global coordinate system; the travel speed refers to the linear velocity and angular velocity of the robot node in the horizontal and vertical coordinate directions; the relative position of the target point refers to the position coordinates of the inspection target point relative to the robot node.

7. A dynamic obstacle avoidance method for a substation inspection robot according to claim 1, characterized in that, For obstacle nodes, their attributes include geometry type, position, size, and movement speed; The geometric types include straight line features and circular features; The position refers to the position coordinates in the global coordinate system; the size refers to the geometric dimensions of the obstacle, including the straight length of a straight obstacle and the radius of a circular obstacle. The movement speed is the same as the movement speed of the dynamic obstacle.

8. A dynamic obstacle avoidance method for a substation inspection robot according to claim 1, characterized in that, The edge weights between nodes are a combination of the distance, relative velocity, and relative angle between the nodes; including: For the edges between robot nodes and obstacle nodes, the edge weights are... It is represented as the product of the first distance influence term and the first association strength term; The first distance influence term is the reciprocal of the Euclidean distance between the robot node and the obstacle node; the first association strength term is an exponential function with the natural constant e as the base, and the exponent is the weighted sum of the absolute value of the first relative velocity and the cosine value of the first relative angle. The first relative velocity and the first relative angle are the relative velocity and relative angle between the robot node and the obstacle node.

9. A dynamic obstacle avoidance method for a substation inspection robot according to claim 8, characterized in that, Also includes: For edges between obstacle nodes, the edge weight... It is represented as the product of the second distance influence term and the second association strength term; The second distance influence term is the reciprocal of the Euclidean distance between obstacle nodes; the second correlation strength term is the weighted sum of the absolute values ​​of the second relative angle difference and the second relative velocity. The second relative velocity and the second relative angle are the relative velocity and relative angle between the obstacle nodes.

10. A dynamic obstacle avoidance method for a substation inspection robot according to claim 1, characterized in that, The calculation method for adjusting attention weights based on edge weights between nodes includes: Based on the fundamental formula for attention weights in graph attention networks, a bias term is introduced into the exponential function of the attention score; this bias term represents the edge weights. The product of a learnable hyperparameter; by Represents the edges between nodes, if Let be the edge between the robot node and the obstacle node. ;like If the edges are between obstacle nodes, then .

11. A dynamic obstacle avoidance method for a substation inspection robot according to claim 1, characterized in that, Define a reinforcement learning framework; including: The state is defined as the feature vector output by the robot node after passing through the graph attention mechanism; the action is defined as the robot's linear velocity increment and angular velocity increment; the reward function is defined as the comprehensive weighted value of progress reward, collision penalty, safety penalty, smoothness reward and dynamic obstacle response reward. Among them, the dynamic obstacle response reward is based on the calculation of the relative speed between the robot and the dynamic obstacle.

12. A dynamic obstacle avoidance system for a substation inspection robot, operating the dynamic obstacle avoidance method as described in any one of claims 1-11, characterized in that, The system includes: Obstacle feature extraction module; used for the robot to load the substation grid map to obtain the preliminary layout environment of the station, and to calculate its real-time position with the help of laser point cloud during movement; real-time acquisition of laser point cloud data, and after preprocessing, parallel identification and extraction of static obstacle features and dynamic obstacle features; Graph structure building module; used to construct an environmental graph based on perception information during the movement process. Nodes include robot nodes and obstacle nodes, and the feature information of each node is used as its attribute; the edge weights between nodes are calculated by combining the distance, relative speed and relative angle between nodes. Attention weight correction module: Based on the environmental graph, it aggregates obstacle node information from the entire graph to generate real-time feature representations for robot nodes; the feature information of each node is passed to neighboring nodes through the message passing mechanism of the graph attention network. During the passing process, the calculation method of attention weight is corrected according to the edge weight between nodes; finally, the real-time feature vector of the robot node is output. The obstacle avoidance path planning module is used to define the reinforcement learning framework and perform path planning; it converts the real-time feature vectors of robot nodes into reinforcement learning state vectors, and the robot nodes select control commands based on the state vectors and provide feedback guidance through the reward function; it outputs the optimal path and trains and validates it in a simulation environment.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the dynamic obstacle avoidance method according to any one of claims 1-11.

14. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the dynamic obstacle avoidance method according to any one of claims 1-11.

Images