Unmanned inspection system and method for oil taking and injection of converter station

Abstract

The invention discloses an unmanned inspection system and method for oil taking and injection of a converter station. The method comprises the following steps: firstly, carrying out initial exploration and positioning mapping, and constructing an initial three-dimensional environment map through a tight coupling SLAM algorithm; secondly, efficiently expanding a map coverage range by adopting a boundary-first exploration strategy, and synchronously constructing a hierarchical semantic map comprising a geometric terrain layer, a power equipment semantic layer and a feasible path layer; polling point autonomous calibration based on a knowledge base is carried out, an observation task is autonomously planned for electrical equipment on an equipment layer, and key polling points on the equipment are automatically calibrated; constructing a multi-target cost function fusing safety, energy consumption and timeliness by taking all calibrated inspection points as targets, and solving a global optimal inspection path; and meanwhile, environment change is continuously detected in inspection, and dynamic increment updating is carried out on the map and the path, so that the generated inspection path can adapt to dynamic change of the converter station environment.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of path planning, in particular to an unmanned inspection system and method for taking oil in a converter station. BACKGROUND

[0002] In view of the problems of the prior art and method, the application provides an unmanned inspection system and method for taking oil in a converter station, which can enable a robot to automatically complete environment exploration, hierarchical semantic map construction, power equipment identification, key inspection point calibration and global optimal path planning without any prior information, and has a closed-loop adaptive updating capability for environmental changes.

[0003] Currently, the implementation of autonomous inspection by inspection robots usually relies on the following technical paths: (1) navigation based on preset trajectory: by laying magnetic strips, two-dimensional codes or installing guide beacons on the ground, the robot travels along a fixed trajectory. It is found through retrieval that the invention patent "Converter station operation and maintenance path planning method, device, electronic equipment and readable medium", patent application number: CN202310946539.5, periodically acquires device data of public equipment through the positioning tags of work equipment and operation and maintenance personnel and locates faulty equipment. This scheme has problems such as complex deployment, high cost, rigid path, and inability to adapt to environmental changes. Once the device layout is adjusted or new inspection tasks are added, it needs to be reprogrammed and programmed. (2) navigation based on prior map: professional technicians first enter the converter station, use laser radar and other equipment to scan the environment with high precision, manually draw the environment map, and manually mark each device point (POI, Point of Interest) that needs to be inspected. It is found through retrieval that the invention patent "Unmanned aerial vehicle inspection path planning method for converter station equipment", patent application number: CN202411792351.0, constructs a device node set based on the location of the converter station equipment, and realizes the inspection path planning of the unmanned aerial vehicle. This process is time-consuming and laborious, and requires personnel with professional skills to operate. When the converter station replaces, adds or adjusts the layout of the equipment, the entire map and inspection points need to be re-mapped and calibrated. (3) traditional SLAM (simultaneous localization and mapping) technology: although the robot can explore and map autonomously, the generated is usually a pure geometric map (such as occupancy grid map), which lacks semantic information. The robot cannot distinguish specific device information and is difficult to form an operational inspection process. It is found through retrieval that the invention patent "Indoor mobile robot path planning method and system for converter station equipment inspection", patent application number: CN202510790622.7, constructs a work scene grid map, fuses the dielectric minimum strategy and improved DQN algorithm, and realizes collision-free and high-efficiency path planning; the invention patent "Converter station unmanned carrier device based on laser SLAM navigation", patent application number: CN202510878074.3, constructs a real-time three-dimensional map of the complex scene of the converter station based on laser SLAM. However, traditional SLAM is prone to positioning failure or map tearing in dynamic environments or when the environment changes significantly.

[0004] In summary, the existing converter station inspection robot path planning and deployment method has a series of defects such as long deployment cycle, high labor cost, poor environmental adaptability, and inability to make dynamic adaptive adjustment to changes in the station, which restricts the large-scale application of the inspection robot in the converter station. In order to overcome the above defects and realize "zero configuration" rapid deployment, self-cognition of the environment and long-term adaptive change of the environment, it is urgent to develop a new adaptive path planning method for the converter station inspection robot to solve the following key problems: (1) the robot realizes the full-automatic completion of map building and device identification in a strange environment without prior map; (2) the robot self-calibrates the key points to be checked by deeply integrating geometric navigation with business requirements of device inspection; and (3) a long-term maintenance mechanism of map and path is established to adapt to dynamic changes such as expansion and reconstruction of the converter station. SUMMARY

[0005] In view of the problems existing in the prior art and method, the present application provides an unmanned inspection system and method for oil taking and injection of a converter station, which can enable the robot to automatically complete environment exploration, hierarchical semantic map building, power device identification, key inspection point calibration and global optimal path planning without any prior information, and has a closed-loop adaptive updating capability for environmental changes.

[0006] The present application provides an unmanned inspection method for oil taking and injection of a converter station, which comprises initial exploration and positioning mapping, hierarchical semantic map incremental construction, autonomous calibration of inspection points based on a knowledge base, multi-target fusion global inspection path planning, and map and path closed-loop updating based on change detection, and comprises the following steps:

[0007] Step S100: initial exploration and positioning mapping, after the robot is started, the initial exploration action is performed through the tightly coupled radar-inertial SLAM algorithm to complete the high-precision positioning initialization of the robot and establish an initial three-dimensional environment map covering the starting area of the robot. The step S100 comprises:

[0008] Step S101: system initialization and multi-sensor data fusion. The system initializes the multi-sensor module, including: preheating the rotation speed of the three-dimensional laser radar (LiDAR) and establishing the data port; zero bias estimation and initial alignment of the inertial measurement unit (IMU); reading zero of the wheel speed encoder; automatic setting of white balance and exposure parameters of the vision camera.

[0009] Step S102: start the tightly coupled radar-inertial SLAM algorithm based on factor graph optimization to estimate the robot state by minimizing the total sum of residuals of all sensor measurement constraints. The objective function is shown in formula (1):

[0010] (1)

[0011] In the formula,

[0012] X*: Represents the optimal solution to the optimization problem defined by equation (1). It takes the form of a set of states or a state vector containing the optimal estimates xi* of all the state variables to be optimized, i.e., X* = {x1*, x2*,..., x...} n The optimal solution is the set of specific states among all possible X states that minimizes the weighted sum of squared residuals from all sensor measurements, representing the maximum posterior probability estimate of the robot's complete motion trajectory and internal states.

[0013] The set of robot state variables to be optimized;

[0014] : These represent the robot's position, orientation (quaternion), velocity, accelerometer bias, and gyroscope bias at time i, respectively;

[0015] : Prior factor residuals, global position constraints from GPS, and pose constraints generated by loop closure detection;

[0016] : LiDAR factor residual;

[0017] : The k-th frame of LiDAR point cloud measurement value. This function calculates the matching error between the current point cloud and the established map under the pose to be optimized X by point-to-point or point-to-edge matching.

[0018] IMU pre-integration factor residuals;

[0019] The IMU measurements (angular velocity and acceleration) between two laser frames are used to calculate the difference between the pose change predicted by the IMU measurements and the pose variable X to be optimized.

[0020] : Represents the index sets for lidar and IMU measurements, respectively;

[0021] The squared Mahalanobis norm represents a weighted error. P is a weighted matrix;

[0022] Prior factor covariance matrix The reverse, Includes GPS constraints and loop closure detection constraints;

[0023] LiDAR measurement covariance matrix The reverse, It is not a fixed value, but is dynamically calculated each time point cloud matching is performed, and it reflects the quality of the current matching.

[0024] IMU measurement covariance matrix The inverse is mainly determined by the Gaussian white noise in the IMU's measurement noise;

[0025] P: The inverse of the covariance matrix (information matrix), used to adjust the weight of each error in the total error according to the measurement uncertainty (noise). The higher the sensor accuracy, the greater its weight; the less accurate the sensor, the smaller its weight. The Mahalanobis norm can force the optimizer to prioritize more reliable (smaller noise variance) measurement constraints.

[0026] (2) Optimization solution

[0027] The entire optimization process is an iterative search. Starting with an initial value of X, the system continuously adjusts the values ​​of each variable in X using a nonlinear optimization algorithm (such as the Gauss-Newton method or the Levenberg-Marquardt method), causing the objective function f(X) to gradually decrease along the direction of gradient descent, eventually converging to a local or global minimum. The core of this iterative solution lies in linearizing the following incremental equation:

[0028] (2)

[0029] In the formula,

[0030] : The increment of the state variable to be solved;

[0031] r: All residual factors The total residual vector formed by splicing;

[0032] J: The Jacobian matrix obtained by differentiating the total residual vector r with respect to the state variable X, i.e. ;

[0033] H: Hessian matrix, a key structural element of the entire optimization problem, approximated in the Gauss-Newton method as: .

[0034] The Hessian matrix H is the second-order partial derivative matrix of the objective function with respect to the state variable X, describing the local curvature of the objective function near the current solution. The increment ΔX is obtained by solving equation (2), and the state X ← X + ΔX is updated. This process is repeated until ΔX converges to zero, finally yielding the optimal state estimate X*.

[0035] (3) Covariance evaluation of state estimation results

[0036] After the optimization converges to obtain the optimal state estimate X*, the system further quantifies the uncertainty of the estimation result. The posterior covariance matrix Σ_state of the final estimated state can be approximated by the inverse of the Hessian matrix H at the optimal solution X* of the entire factor graph optimization problem, as shown in equation (3):

[0037] (3)

[0038] In the formula,

[0039] : This refers to the final output pose covariance matrix. It is a large block matrix, and the block matrix Σi on its diagonal describes the uncertainty of a single state variable xi.

[0040] The inverse of the Hessian matrix H. The "size" of the Hessian matrix H (measured by its eigenvalues) reflects the strength of the constraints on the optimal solution, while its inverse matrix H⁻¹ directly quantifies the uncertainty of the optimal solution. For example, if the system has traversed a feature-rich scene and includes high-quality loop closure detection, the H matrix will be very "strong," and its inverse Σstate will be very "small," indicating that the final estimated robot trajectory has very high accuracy and confidence.

[0041] Step S103: In the SLAM system initialization phase, in order to quickly establish a stable local map, the robot performs simple exploration actions such as "walking along the wall" or "walking in a small range of spirals" at a safe speed. This provides sensor data with sufficient parallax and motion excitation for factor map optimization, thereby quickly converging to obtain a reliable initial pose (including 6 degrees of freedom, position: (x, y, z), attitude: Roll (roll angle), Pitch (pitch angle), Yaw (yaw angle)), and simultaneously generating a local geometric point cloud map.

[0042] Step S200: Incremental Construction of Layered Semantic Map: When the initial positioning accuracy reaches a threshold, i.e., the pose covariance output by the factor graph optimization process is less than a preset value, the system adopts a boundary-first exploration strategy to efficiently expand the map coverage and simultaneously construct a layered semantic map containing a geometric terrain layer, a power equipment semantic layer, and a feasible path layer. Step S200 includes:

[0043] Step S201: Boundary Cluster Detection and Clustering: Based on the geometric terrain layer map output by the SLAM algorithm, the system detects the boundaries (frontiers) between known free space and unknown space in real time. All detected boundary points are aggregated into multiple boundary clusters using the DBSCAN clustering algorithm to improve the stability and efficiency of the exploration.

[0044] Step S202: Select the next optimal exploration target point from the boundary cluster set F using the utility function shown in formula (2). To maximize the ratio of information gain to movement cost.

[0045] (4)

[0046] In the formula,

[0047] : Represents the exploration of boundary clusters The information gain is usually measured by the number of boundary points contained in the boundary cluster or its geometric length.

[0048] : Represents the robot's current position Go to The cost of the center point is determined by The shortest path length calculated by the algorithm on the current terrain layer map;

[0049] These are the weighting coefficients for information gain and movement cost, respectively, which can be adjusted according to the task strategy.

[0050] Step S203: Hierarchical Semantic Map Construction: As the inspection robot moves toward the target exploration point, it updates the hierarchical semantic map in parallel: the terrain layer. Equipment layer and path layer Among them, the topographic layer It is a 3D point cloud map continuously updated by the SLAM system, describing the physical structure and accessibility of the environment; device layer This involves feeding images captured by a visual camera into a pre-trained power equipment target detection network (such as YOLOv8), constructing a network using the identified equipment category and location information, and fusing images of the same target from different perspectives using a Kalman filter to improve positioning accuracy; path layer Record the robot's verified safe passage trajectories and abstract them into a topological path map. Terrain layer. Equipment layer and path layer The specific construction process is as follows:

[0051] (1) Topographic layer The three-dimensional point cloud map or its projection of a two-dimensional occupied grid map continuously output by the SLAM system in the S100 describes the physical structure of the environment and is used for basic navigation and obstacle avoidance.

[0052] (2) Equipment layer The visual camera continuously acquires images, which are then fed into a pre-trained power equipment target detection network (such as YOLOv8). For each detected device, the system extracts its category and 2D bounding box, and combines this with synchronized depth information or LiDAR point cloud data to calculate its 3D position in the world coordinate system. By fusing multiple observations of the same target using a Kalman filter, a more accurate pose estimate can be obtained. And assign a unique identifier (UUID) to each individual device;

[0053] (3) Path layer Record the trajectory that the robot has safely traversed and abstract it into a topological graph. , where node V is a key location point, and edge E is the path segment connecting the nodes and its travel cost.

[0054] Step S300: Knowledge-based autonomous inspection point calibration: For newly discovered power equipment at the equipment level, the system autonomously plans observation tasks and automatically calibrates key inspection points on the equipment through refined secondary identification. Step S300 includes:

[0055] Step S301: Triggering of the device focusing observation task: When the device layer When a new device object with a confidence level higher than the threshold is first identified, or when the location uncertainty of an existing device is still relatively large, the system automatically generates a "focused observation" subtask for that device.

[0056] Step S302: Next-Best-View Planning: To acquire high-quality images of the device details, the robot needs to move to an optimal viewing pose; the system calculates this viewing pose by optimizing a utility function. This function takes into account three factors: observation distance, observation angle, and avoiding occlusion by the device itself.

[0057] Step S303: Fine Identification and Calibration of Key Inspection Points (POIs): After the robot reaches the optimal observation pose, it uses a high-resolution camera to capture close-up images of the equipment. Simultaneously, the system queries the built-in "Equipment-Inspection Point Knowledge Base," which stores standard inspection locations and their visual features for various types of equipment in a structured format (e.g., "SF6 circuit breaker": ["pressure gauge", "opening / closing indicator"]). Subsequently, the system uses template matching, SIFT feature point matching, and a dedicated secondary recognition network on the close-up image to accurately locate the 2D image coordinates of the inspection points retrieved from the knowledge base. Finally, using camera intrinsics and depth data, the 2D coordinates are back-projected into three-dimensional coordinates in the world coordinate system. And set it as a mandatory inspection task point, and link it to the parent equipment object to which the inspection part belongs at the equipment layer. In the context of storage, associated data is stored.

[0058] Step S400: Global Inspection Path Planning with Multi-Objective Fusion: Using all marked inspection points as objectives, a multi-objective cost function integrating safety, energy consumption, and timeliness is constructed, and the globally optimal inspection path is solved to generate a comprehensively optimal inspection "work order" for the robot. Step S400 includes:

[0059] Step S401: Generation of inspection task point set: from the equipment layer Extract the three-dimensional coordinates of all marked inspection points to form a set of target points to be visited;

[0060] Step S402: Equivalent the path planning problem to a variant of the Traveling Salesman Problem (TSP), and construct a multi-objective optimization path cost function between any two inspection points, as shown in formula (3):

[0061] (3)

[0062] In the formula,

[0063] Safety cost: On the terrain layer map, based on the voltage level and location of high-voltage equipment in the equipment layer, a safety risk field composed of superimposed Gaussian functions is generated. This cost is the line integral of the path in the risk field, penalizing paths that are closer to the hazard source;

[0064] Energy consumption cost is a weighted sum of path length, cumulative turning angle, and climbing height, simulating the physical energy consumption of the robot.

[0065] Timeliness cost consists of the travel time (length / speed) of the route and the estimated dwell time (such as taking photos and measuring temperature) at each inspection point.

[0066] Step S403: Solving the optimal path using an improved genetic algorithm: For this NP-hard path solving problem, this invention uses an improved genetic algorithm to find an approximate optimal solution. By performing selection, crossover, and mutation operations on the inspection point sequence (chromosome), it iteratively converges to an inspection order that minimizes the total cost C. And the specific path that corresponds to it.

[0067] The specific operations for improving the genetic algorithm include:

[0068] Step (1): Parameter definition and optimization objective formalization

[0069] Define the parameters required for the algorithm, including: population size N_pop, maximum number of iterations G_max, crossover probability P_c, mutation probability P_m, and number of elites to be retained N_elite.

[0070] Define a candidate solution as a chromosome, which is the set of inspection points. An ordered arrangement The optimization objective of this method is to find an optimal permutation. This makes the total cost function Minimize the total cost function. The definition is as follows:

[0071] (6)

[0072] In the formula,

[0073] Represents arrangement The i-th inspection point in the process;

[0074] This is the robot's current starting position. The designated endpoint for the task;

[0075] Time(A,B) is the shortest travel distance for the robot from spatial location A to B. This value is determined by the path layer. Execution The path search algorithm calculates the result.

[0076] To transform the above minimization problem into the standard fitness maximization problem of genetic algorithms, a fitness function is defined. for:

[0077] (7)

[0078] Step (2): Population initialization based on heuristic algorithm

[0079] Generate an initial population of size N_pop. This step employs a hybrid strategy: one set of chromosomes is generated by randomly permuting the inspection point set P to ensure population diversity; the other set of chromosomes is generated using the nearest neighbor heuristic algorithm to introduce high-quality initial solutions into the population and accelerate algorithm convergence.

[0080] Step (3): Fitness assessment and elite retention selection

[0081] For each chromosome in the current population According to the fitness function defined in step (1) Calculate its fitness value. The selection operation is performed using a strategy that combines elitist retention and tournament selection. First, determine the N_elite individuals with the highest fitness values in the current population as the elite set E_g (where g is the current generation), and directly copy them to the next generation population. Second, generate the parental population for subsequent crossover operations through the tournament selection method.

[0082] Step (4): Recombination operation based on order crossover (OX)

[0083] To ensure that the offspring chromosomes are still legal permutations, perform the order crossover (OX) operator on the selected parental individuals according to the crossover probability P_c to generate new offspring individuals. This operator ensures the validity of the offspring chromosome permutation by retaining part of the gene sequence of the parent and filling the remaining genes in the order of the other parent.

[0084] Step (5): Perturbation operation based on inversion mutation

[0085] [[ID=1F]]To increase the population diversity and prevent the algorithm from falling into local optima, perform the inversion mutation operation on the newly generated individuals according to the mutation probability P_m. For a chromosome , randomly select two indices i and j (i < j), and generate the mutated chromosome , that is, reverse the gene subsequence between indices i and j.

[0086] Step (6): Iteration and termination condition judgment

[0087] Repeat steps (3) to (5). Let g be the current iteration generation, be the optimal individual in the g-th generation. When any of the following conditions is met, the algorithm terminates:

[0088] The current iteration generation g reaches the preset maximum iteration generation Gmax, that is, g ≥ Gmax.

[0089] ② The optimal fitness value has not been significantly improved for k consecutive generations, that is , where is the preset stagnation judgment threshold. After the algorithm terminates, the inspection order corresponding to the elite individual with the maximum fitness value is output as the result of the optimal path planning.

[0090] Step S500: Map and path closed-loop update based on change detection: Continuously detect environmental changes during the inspection process, and perform dynamic incremental updates on the map and path, so that the inspection path of the robot can adapt to the dynamic changes of the converter station environment. The step S500 includes:

[0091] Step S501: Environmental Change Detection: In subsequent routine inspections, the robot will use the currently acquired LiDAR scan data... With topographic layers Real-time registration is performed, and the registration error is calculated using the ICP algorithm. If the error exceeds a preset threshold... If this is detected, it is determined that a change in the physical environment has occurred in the area; simultaneously, by comparing the current visual recognition results with those at the device layer... Records of the addition, removal, or replacement of testing equipment;

[0092] Step S502: Incremental Map Update and Version Management: When a change is detected, the map management module archives the current map version (e.g., Map_v2.1) and triggers a local, small-scale exploration and calibration process in the changed area (i.e., re-execute S200 and S300); after the update is completed, a new version map with an update log is generated and enabled (e.g., Map_v2.2).

[0093] Step S503: Dynamic replanning of inspection paths: Once the map version is updated, the system will automatically discard the inspection paths generated based on the old map and re-execute the global path planning process of S400 to generate an optimal path that perfectly matches the new environment, ensuring that the inspection task is always carried out under the latest and most accurate environmental model.

[0094] Another aspect of the present invention provides an unmanned inspection system for oil intake and injection at a converter station, the system comprising:

[0095] An environment exploration module, used in the environmental cognition and initial exploration phase, is the process by which the robot progresses from a completely unknown state (cold start) to establishing initial localization and a local map. Its key feature is that it does not rely on any pre-set information and completes this entirely through its own interaction with the environment. The environment exploration module includes:

[0096] The initialization unit is used for system initialization and multi-sensor data fusion. The system initializes the onboard multi-sensor modules, including: preheating the rotational speed of the 3D LiDAR and establishing its data port; performing zero-bias estimation and initial alignment of the Inertial Measurement Unit (IMU); zeroing the wheel speed encoder readings; and automatically setting the white balance and exposure parameters for the vision camera.

[0097] The state estimation unit is used to initiate the LIO-SAM mapping method based on factor graph to tightly couple and fuse LiDAR, IMU and wheel speed encoder data to achieve high-precision and robust odometry. Its core algorithm is to maintain and optimize a factor graph, and its objective function is to minimize the sum of residuals of all measurement constraints to estimate the robot's state.

[0098] The point cloud generation unit is used in the initialization phase of the SLAM system to quickly establish a stable local map. The robot performs pre-set exploration actions such as "walking along the wall" or "walking in a spiral" in an open area at a safe speed. It provides sensor data with sufficient parallax and motion excitation for factor map optimization, thereby quickly converging to obtain a reliable initial pose (including 6 degrees of freedom, position: (x, y, z), attitude: Roll (roll angle), Pitch (pitch angle), Yaw (yaw angle)) and simultaneously generating a local geometric point cloud map.

[0099] The map building module is used for incremental construction of a hierarchical semantic map: when the initial positioning accuracy reaches a threshold, i.e., the pose covariance output by the factor graph optimization process is less than a preset value, the system adopts a boundary-first exploration strategy to efficiently expand the map coverage and simultaneously construct a hierarchical semantic world model containing a geometric terrain layer, a power equipment semantic layer, and a feasible path layer. The map building module includes:

[0100] Clustering units are used for boundary cluster detection and clustering: Based on the geometric terrain layer map output by the SLAM algorithm, the system detects the boundaries (frontiers) between known and unknown free spaces in real time. All detected boundary points are aggregated into multiple boundary clusters using the DBSCAN clustering algorithm. To improve the stability and efficiency of the exploration;

[0101] The exploration target decision unit is used for information gain-based exploration target decision-making: the system uses a utility function. Select the next optimal exploration target point from the set of boundary clusters F. ;

[0102] Layered map building units are used for layered semantic map construction: As the inspection robot moves towards the target exploration point, it updates the layered semantic map in parallel: the terrain layer. Equipment layer and path layer Among them, the topographic layer It is a 3D point cloud map continuously updated by the SLAM system, describing the physical structure and accessibility of the environment; device layer This involves feeding images captured by a visual camera into a pre-trained power equipment target detection network (such as YOLOv8), constructing a network using the identified equipment category and location information, and fusing images of the same target from different perspectives using a Kalman filter to improve positioning accuracy; path layer Record the robot's verified safe passage trajectories and abstract them into a topological path graph.

[0103] The inspection point calibration module is used for autonomous calibration of inspection points based on a knowledge base: for newly discovered power equipment at the equipment level, it autonomously plans observation tasks and automatically calibrates key inspection points on the equipment through refined secondary identification. The inspection point calibration module includes:

[0104] Subtask generation unit, used to trigger device-focused observation tasks: when the device layer... When a new device object with a confidence level higher than the threshold is first identified, or when the location uncertainty of an existing device is still relatively large, the system automatically generates a "focused observation" subtask for that device.

[0105] The pose planning unit is used for next-best-view planning: to obtain high-quality images of device details, the robot needs to move to an optimal view pose; the system calculates this view pose by optimizing a utility function. This function takes into account three factors: observation distance, observation angle, and avoiding occlusion by the device itself.

[0106] The inspection point marking unit is used for the precise identification and calibration of critical inspection points (POIs): After the robot reaches the optimal observation pose, it uses a high-resolution camera to capture close-up images of the equipment. Simultaneously, the system queries the built-in "Equipment-Inspection Point Knowledge Base," which stores the standard inspection locations and visual features of various types of equipment in a structured format (e.g., "SF6 circuit breaker": ["pressure gauge", "opening / closing indicator"]). Subsequently, the system uses template matching, SIFT feature point matching, and a dedicated secondary recognition network on the close-up image to accurately locate the 2D image coordinates of the inspection points retrieved from the knowledge base. Finally, using camera intrinsic parameters and depth data, the aforementioned 2D coordinates are back-projected into three-dimensional coordinates in the world coordinate system. And set it as a mandatory inspection task point, and link it to the parent equipment object to which the inspection part belongs at the equipment layer. In the context of storage, associated data is stored.

[0107] The path planning module is used for global inspection path planning with multi-objective fusion: Taking all marked inspection points as objectives, it constructs a multi-objective cost function that integrates safety, energy consumption, and timeliness, and solves for the globally optimal inspection path, generating a comprehensively optimal inspection "work order" for the robot. The path planning module includes:

[0108] The inspection point set generation unit is used to generate inspection task point sets: the system starts from the equipment layer. Extract the 3D coordinates of all successfully calibrated inspection points. This forms a set of target points to be visited.

[0109] The cost calculation unit is used to convert the path planning problem into a variant of the Traveling Salesman Problem (TSP) and construct a path cost function between any two inspection points for multi-objective optimization.

[0110] The path finding unit is used to find the optimal path using an improved genetic algorithm. For this NP-hard path finding problem, this invention employs an improved genetic algorithm to find an approximate optimal solution. Through selection, crossover, and mutation operations on the inspection point sequence (chromosomes), it iteratively converges to an inspection order that minimizes the total cost C. And the specific path that corresponds to it.

[0111] A map update module is used for closed-loop updating of the map and path based on change detection: During the inspection process, environmental changes are continuously detected, and the map and path are dynamically and incrementally updated, enabling the robot's inspection path to adapt to the dynamic changes in the converter station environment. The map update module includes:

[0112] The change detection unit is used to detect environmental changes: during subsequent routine inspections, the robot will use the currently acquired LiDAR scan data... With topographic layers Real-time registration is performed, and the registration error is calculated using the ICP algorithm. If the error exceeds a preset threshold... If this is detected, it is determined that a change in the physical environment has occurred in the area; simultaneously, by comparing the current visual recognition results with those at the device layer... Records of the addition, removal, or replacement of testing equipment;

[0113] The map update and version management unit is used for incremental map updates and version management: When a change is detected, the map management module archives the current map version (e.g., Map_v2.1) and triggers a local, small-scale exploration and calibration process in the changed area (i.e., re-executing S200 and S300). After the update is complete, a new version map with an update log (e.g., Map_v2.2) is generated and enabled.

[0114] The path replanning unit is used for dynamic replanning of inspection paths: once the map version is updated, the system will automatically discard the inspection path generated based on the old map and re-execute the global path planning process of the path planning module to generate an optimal path that perfectly matches the new environment, ensuring that the inspection task is always carried out under the latest and most accurate environmental model.

[0115] Compared with the prior art, the present invention has the following advantages:

[0116] 1. Achieve zero configuration and rapid deployment: Eliminate the reliance on manual pre-map creation and inspection point calibration. The robot can automatically complete the deployment after startup, greatly shortening the deployment cycle and reducing labor and time costs.

[0117] 2. Achieving highly intelligent semantic cognition: By constructing a hierarchical semantic map and autonomously calibrating inspection points, the robot not only knows "how to go", but also deeply understands "where to look" and "what to look at", realizing a leap from geometric navigation to semantic navigation.

[0118] 3. Strong long-term environmental adaptability: The designed closed-loop incremental update mechanism enables the robot to autonomously adapt to dynamic changes such as the renovation and expansion of the converter station and equipment replacement, ensuring the long-term effectiveness of the map and path, and solving the pain point of the traditional solution of "one-time deployment and lifelong maintenance".

[0119] 4. Improve the overall efficiency of inspection operations: The multi-objective fusion path planning model takes into account both energy consumption and efficiency while ensuring safety, making the inspection path more scientific and economical. Attached Figure Description

[0120] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0121] Figure 1 This is an overall flowchart of an intelligent operation path planning method for oil intake and injection at a converter station according to the present invention.

[0122] Figure 2 This is a flowchart illustrating a specific implementation method of the present invention;

[0123] Figure 3 This is a schematic diagram of the adaptive mapping principle based on re-observation closed-loop correction in an embodiment of the present invention;

[0124] Figure 4 This is a schematic diagram illustrating the principle of incremental construction of hierarchical semantic maps in an embodiment of the present invention.

[0125] Figure 5 This is a schematic diagram of the map construction process based on the boundary-first exploration strategy in an embodiment of the present invention;

[0126] Figure 6 This is a schematic diagram of the autonomous calibration process for key inspection points (POIs) based on a knowledge base.

[0127] Figure 7 This is a schematic diagram of global inspection path planning for multi-target fusion in an embodiment of the present invention;

[0128] Figure 8 This is a schematic diagram of map and path closed-loop update based on change detection. Detailed Implementation

[0129] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0130] One embodiment of the present invention provides: an unmanned inspection method for oil intake and injection at converter stations, such as... Figure 1 As shown;

[0131] In this embodiment, we use a wheeled inspection robot platform. The core computing unit of this platform is an NVIDIA Jetson AGX Orin, running Ubuntu 20.04 and ROS Noetic. The robot is equipped with the following sensors: a Velodyne VLP-16 3D LiDAR, mounted on the top of the robot, providing a 360° horizontal field of view; an Intel RealSense D435i depth camera, including an RGB camera and depth sensor, and a built-in IMU; and a wheel speed encoder integrated into the wheeled chassis. Combined with... Figure 2 The specific implementation method flowchart, in this example, includes the following steps:

[0132] Step S100: Initial Exploration and Localization Mapping: In this embodiment, after the robot is powered on, the system first performs sensor initialization. For example... Figure 3As shown, the specific meanings are as follows: ① Initial State and Uncertainty: During the system initialization phase, all state variables are uncertain. This includes the robot's pose x1 at time t1, its pose x2 at time t2, and a fixed but unknown landmark L1 on the map. In the figure, this uncertainty is represented by the light-colored area around each variable (representing its covariance). ② Establishing Observation Constraints: When the robot is in pose x1, it performs a measurement (such as distance and angle) on landmark L1. This measurement itself contains noise (measurement covariance). Therefore, based on the uncertainty of x1 and the noise of the measurement, the system infers the first posterior probability distribution of the L1 position, as shown by the first light orange uncertainty ellipse in the figure. Subsequently, the robot moves to a new pose x2 and observes the same landmark L1 again. Similarly, this observation also establishes a new, independent constraint on the L1 position, as shown by the second light orange uncertainty ellipse parent circle in the figure. ③ Joint Optimization and Uncertainty Reduction:

[0133] This is the core of the diagram. Since landmark L1 is fixed, its true position must simultaneously satisfy all observation constraints. Therefore, the most likely position of L1 lies within the intersection of all these uncertain ellipses. The optimizer (as described in step S102) aims to find an optimal solution that minimizes the weighted sum of all observation residuals (i.e., the difference between prediction and measurement). By solving this joint optimization problem, the system obtains a new estimate of the L1 position that incorporates information from multiple observations. This new estimate has a much higher confidence level than any single observation, and its uncertainty (covariance) is significantly reduced, as shown by the dark orange ellipse in the diagram. ④ Backward correction of the state (information feedback): Once the system obtains a high-precision estimate of the location of landmark L1, this precise landmark acts as a "high-precision anchor point." The optimizer can use this anchor point to correct the robot's trajectory in reverse. The underlying logic is: "In order to observe this precise L1 at x1 and x2, the true positions of the robot's current pose x1 and x2 must also be adjusted accordingly." ⑤ Finally, through a global optimization, the system not only obtained an accurate estimate of landmark L1, but also simultaneously corrected the poses of historical trajectory points x1 and x2. Step S100 includes:

[0134] Step S101: The system loads the LIO-SAM (Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping) algorithm package. The Velodyne radar driver starts, and the rotation speed is set to 600 RPM. The RealSense camera driver starts, the RGB image resolution is set to 1920x1080, and depth stream and IMU data stream are enabled.

[0135] Step S102: The LIO-SAM algorithm node starts and begins receiving data from three ROS topics: / velodyne_points, / camera / imu, and / odom (from the wheel speed encoder). The backend for factor graph optimization uses the GTSAM 4.0 library. In Equation (1), the IMU pre-integration factor is generated between every two LiDAR scans;

[0136] (1)

[0137] Step S103: To quickly obtain a robust initial pose, the robot performs a pre-defined "wall-walking" behavior. The robot uses LiDAR data to detect the nearest wall within 2 meters and adjusts its posture to maintain a parallel distance of approximately 1 meter from it. Then, it moves forward at a constant linear velocity of 0.3 m / s for 30 seconds. This process is sufficient for the factor graph optimization of LIO-SAM to converge, establishing a stable local map.

[0138] Step S200: Incremental construction of hierarchical semantic map. For example... Figure 4 As shown, this illustrates the core loop of step S200: perception → mapping → decision → action. The robot is moving in an unknown gray environment. The black dotted line represents the path the robot has already traversed. Device A is a specific device identified by the robot through sensors; this is semantic information. The edge of the robot's field of vision is the boundary between the known and unknown areas. f1, f2, and f3 represent three different boundary regions, which are candidate targets for the next step of exploration. When the robot receives the same data stream from the sensors, it will simultaneously use it to build and update three maps for different purposes: the terrain layer, the device layer, and the path layer. When the information of all boundary points is sent to the decision module, the "gold mining" decision model will score each boundary point f and calculate its utility function value. This score comprehensively considers the reward (InfoGain, i.e., how much new area can be explored) and the cost (Cost, i.e., how far to travel). Decision output: The model calculates that f3 has the highest score (for example, it may connect to the largest unknown area, making it the most cost-effective). Therefore, the system makes the decision: "Go to f3!" This instruction (red arrow) is sent to the robot's motion control module, driving the robot to move towards f3. This process repeats continuously; the robot, like a clever explorer, constantly moves towards the area with the highest "information value" until the entire environment is explored and the three-layer map is completed. Step S200 includes:

[0139] Step S201: The map exploration node subscribes to the global point cloud map output by LIO-SAM and projects it onto a two-dimensional occupancy grid map with a resolution of 0.1 meters for boundary detection. The DBSCAN algorithm is used to cluster the boundary points, where the cluster radius eps is set to 0.5 meters and the minimum number of samples min_samples is set to 10.

[0140] Step S202: In the utility function of formula (2), information gain Directly take the boundary cluster Number of boundary points; movement cost Depend on The algorithm calculates the path length (in meters) on the current 2D raster map, as follows: Figure 5As shown, the specific process is as follows: ① Boundary recognition and candidate target generation: From its current pose, the robot identifies all reachable exploration boundaries by analyzing the global map. These boundary points are clustered into two main candidate boundary clusters (Frontier Clusters): F1 and F2, based on their spatial proximity. These two boundary clusters constitute the set of candidate targets for the robot's next exploration. ② The robot does not randomly select exploration targets, but rather uses a utility function to quantitatively evaluate each candidate boundary cluster to select the optimal target. This utility function integrates two key dimensions: information gain and exploration cost. Information Gain: Evaluates the size of the unknown area that can be observed after reaching a certain boundary. Generally, longer or more open boundaries (such as F1) indicate higher information gain. ② Exploration Cost: Evaluates the cost required for the robot to reach the boundary from its current position, mainly determined by the path length or estimated travel time. The diagram visually illustrates this decision-making process: Evaluation of F2 (alternative): The system calculates the optimal path (gray dashed line path) from the current pose® to F2, and determines its exploration cost to be 15.2. Evaluation of F1 (better solution): The system calculates the exploration cost to reach F1 to be 9.8. Optimal Decision: By comparison, the system finds that the cost of going to F1 is significantly lower than that of F2. After comprehensively considering information gain and cost (simplified here as cost priority), the system determines that F1 is the current optimal exploration direction. Therefore, the system selects a representative center point f* (asterisk * represents the optimal solution) on F1 as the optimal exploration target point for this exploration. ③ Once the optimal exploration target f* is determined, the robot's path planning module will generate a collision-free executable trajectory (blue solid arrow path) from the current pose® to f*. The robot will travel along this path. During the movement, the sensors will continuously transform unknown areas along the way into known areas, thereby updating the map. When the robot reaches the target area and completes the observation, the old boundary will disappear, and a new boundary will form in the deeper part of the map. The system will repeat steps 1 to 3 above, continuously identifying new boundaries, evaluating and selecting the optimal target, and navigating there, until all reachable boundaries have been explored, at which point the mapping process is completed automatically. During the exploration process, the weight coefficients of the cost function have been experimentally adjusted as follows: ;

[0141] (2)

[0142] Step S203: During the process of the robot moving towards the optimal boundary target, the device layer... RGB images are fed at 10Hz into a YOLOv8-L model pre-trained on the COCO dataset and fine-tuned on a private dataset containing 30,000 images of converter station equipment. A device (e.g., a "disconnect switch") is considered a valid discovery when its detection confidence is above 0.85 for three consecutive frames. The device's 3D position in world coordinates is calculated from the depth value corresponding to the center of its 2D bounding box and smoothed using a linear Kalman filter. Each device is assigned a UUID and stored in an SQLite database; path layer. The robot's trajectory points are recorded at a frequency of 1Hz and then sparsified using the Douglas-Puk algorithm (with a threshold of 0.2 meters) to form nodes in the topological graph.

[0143] Step S300: Autonomous calibration of inspection points based on the knowledge base, such as... Figure 6 As shown, step S300 includes:

[0144] Step S301: When at the device layer When a new device with high confidence is discovered, a focused observation task for that device is triggered;

[0145] Step S302: Plan an optimal viewing pose (Next-Best-View) for the device, moving the robot to the position most favorable for observing device details; such as when the device layer... When a new "SF6 circuit breaker" is first detected, the system retrieves information from an XML-formatted knowledge base file. The system identifies "SF6 pressure gauge" and "opening / closing status indicator" as associated inspection points. The system then plans an optimal observation position for the "SF6 pressure gauge," aiming to align the camera's optical axis directly with the center of the device and maintain an observation distance of 1.5 meters.

[0146] Step S303: After the robot reaches the pose, switch the RealSense camera to high-resolution mode. The system loads a 128x128 pixel grayscale image of an "SF6 pressure gauge" as a template and searches the currently captured 1920x1080 image using a template matching algorithm based on NCC (Normalized Cross-Correlation). Localization is considered successful when the matching score is higher than 0.9. The 2D pixel coordinates of the POI are combined with synchronized depth data and back-projected to obtain its 3D coordinates in the world coordinate system. It is then stored in the database and associated with the parent device "SF6 circuit breaker".

[0147] Step S400: Global inspection path planning with multi-target fusion, such as... Figure 7 As shown, step S400 includes:

[0148] Step S401: From the device layer Extract the three-dimensional coordinates of all marked inspection points to form a set of target points to be visited;

[0149] Step S402: Define the path cost function between any two inspection points, as shown in formula (3):

[0150] (3)

[0151] In the formula,

[0152] Safety costs are assessed on the topographic map by generating a safety risk field composed of superimposed Gaussian functions, based on the voltage levels and locations of high-voltage equipment within the equipment layer. This cost is the line integral of the path over the risk field, penalizing paths closer to the hazard source; for example, for equipment labeled as "220kV" in the knowledge base, a Gaussian risk field is generated around it, with a central peak penalty of 100 and a standard deviation of [missing value]. The standard deviation is 3 meters for "110kV" equipment. The distance is 1.5 meters, and the safety cost of the path is the integral of the risk values ​​of all points on the path.

[0153] Energy consumption cost is a weighted sum of path length, cumulative turning angle, and climbing height, simulating the physical energy consumption of the robot.

[0154] Timeliness cost consists of the travel time (length / speed) of the route and the estimated dwell time (such as taking photos and measuring temperature) at each inspection point.

[0155] The weighting coefficient is set to This indicates that in this embodiment, security is the primary consideration.

[0156] Step S403: We use an open-source genetic algorithm library (such as GALib) to solve the problem. The population size is set to 200, the number of iterations is 500 generations, the crossover probability is 0.8, and the mutation probability is 0.15. The algorithm takes about 15 seconds to complete one global planning cycle (approximately 50 inspection points) on an NVIDIA Jetson AGXOrin.

[0157] Step S500: Map and path closure update based on change detection, such as... Figure 8 As shown, step S500 includes:

[0158] Step S501: While the robot is inspecting along the planned path, the LiDAR scan data of the current frame is compared with the terrain layer in real time. The system performs ICP (Iterative Closest Point) registration on the global point cloud. If the root mean square error (RMSE) of 10 consecutive ICPs is greater than 0.25 meters, the system determines that a significant physical change has occurred in the region.

[0159] Step S502: Upon detecting a change, the system backs up the current map database (e.g., station_map_v1.2.db) and creates an "update area" with a radius of 8 meters around the point of change. The robot pauses its current inspection task and performs the local exploration and calibration procedures S200 and S300 on this area. After completion, a new map, station_map_v1.3.db, is generated.

[0160] Step S503: After the new map is generated, the system automatically calls the global path planning module of S400 to recalculate the optimal inspection path based on the new map, and then continues to execute the inspection task.

[0161] Through the above specific implementation methods, the method described in this invention can run stably on an actual robot hardware platform, realizing fully automated deployment from scratch and long-term adaptive intelligent inspection.

[0162] On the other hand, embodiments of this application provide an unmanned inspection system for oil intake and injection at a converter station, the system comprising:

[0163] An environment exploration module, used in the environmental cognition and initial exploration phase, is the process by which the robot progresses from a completely unknown state (cold start) to establishing initial localization and a local map. Its key feature is that it does not rely on any pre-set information and completes this entirely through its own interaction with the environment. The environment exploration module includes:

[0164] The initialization unit is used for system initialization and multi-sensor data fusion. The system initializes the onboard multi-sensor modules, including: preheating the rotational speed of the 3D LiDAR and establishing its data port; performing zero-bias estimation and initial alignment of the Inertial Measurement Unit (IMU); zeroing the wheel speed encoder readings; and automatically setting the white balance and exposure parameters for the vision camera.

[0165] The state estimation unit is used to initiate the LIO-SAM mapping method based on factor graph to tightly couple and fuse LiDAR, IMU and wheel speed encoder data to achieve high-precision and robust odometry. Its core algorithm is to maintain and optimize a factor graph, and its objective function is to minimize the sum of residuals of all measurement constraints to estimate the robot's state.

[0166] The point cloud generation unit is used in the initialization phase of the SLAM system to quickly establish a stable local map. The robot performs pre-set exploration actions such as "walking along the wall" or "walking in a spiral" in an open area at a safe speed. It provides sensor data with sufficient parallax and motion excitation for factor map optimization, thereby quickly converging to obtain a reliable initial pose (including 6 degrees of freedom, position: (x, y, z), attitude: Roll (roll angle), Pitch (pitch angle), Yaw (yaw angle)) and simultaneously generating a local geometric point cloud map.

[0167] The map building module is used for incremental construction of a hierarchical semantic map: when the initial positioning accuracy reaches a threshold, i.e., the pose covariance output by the factor graph optimization process is less than a preset value, the system adopts a boundary-first exploration strategy to efficiently expand the map coverage and simultaneously construct a hierarchical semantic world model containing a geometric terrain layer, a power equipment semantic layer, and a feasible path layer. The map building module includes:

[0168] Clustering units are used for boundary cluster detection and clustering: Based on the geometric terrain layer map output by the SLAM algorithm, the system detects the boundaries (frontiers) between known and unknown free spaces in real time. All detected boundary points are aggregated into multiple boundary clusters using the DBSCAN clustering algorithm. To improve the stability and efficiency of the exploration;

[0169] The exploration target decision unit is used for information gain-based exploration target decision-making: the system uses a utility function. Select the next optimal exploration target point from the set of boundary clusters F. ;

[0170] Layered map building units are used for layered semantic map construction: As the inspection robot moves towards the target exploration point, it updates the layered semantic map in parallel: the terrain layer. Equipment layer and path layer Among them, the topographic layer It is a 3D point cloud map continuously updated by the SLAM system, describing the physical structure and accessibility of the environment; device layer This involves feeding images captured by a visual camera into a pre-trained power equipment target detection network (such as YOLOv8), constructing a network using the identified equipment category and location information, and fusing images of the same target from different perspectives using a Kalman filter to improve positioning accuracy; path layer Record the robot's verified safe passage trajectories and abstract them into a topological path graph.

[0171] The inspection point calibration module is used for autonomous calibration of inspection points based on a knowledge base: for newly discovered power equipment at the equipment level, it autonomously plans observation tasks and automatically calibrates key inspection points on the equipment through refined secondary identification. The inspection point calibration module includes:

[0172] Subtask generation unit, used to trigger device-focused observation tasks: when the device layer... When a new device object with a confidence level higher than the threshold is first identified, or when the location uncertainty of an existing device is still relatively large, the system automatically generates a "focused observation" subtask for that device.

[0173] The pose planning unit is used for next-best-view planning: to obtain high-quality images of device details, the robot needs to move to an optimal view pose; the system calculates this view pose by optimizing a utility function. This function takes into account three factors: observation distance, observation angle, and avoiding occlusion by the device itself.

[0174] The inspection point marking unit is used for the precise identification and calibration of critical inspection points (POIs): After the robot reaches the optimal observation pose, it uses a high-resolution camera to capture close-up images of the equipment. Simultaneously, the system queries the built-in "Equipment-Inspection Point Knowledge Base," which stores the standard inspection locations and visual features of various types of equipment in a structured format (e.g., "SF6 circuit breaker": ["pressure gauge", "opening / closing indicator"]). Subsequently, the system uses template matching, SIFT feature point matching, and a dedicated secondary recognition network on the close-up image to accurately locate the 2D image coordinates of the inspection points retrieved from the knowledge base. Finally, using camera intrinsic parameters and depth data, the aforementioned 2D coordinates are back-projected into three-dimensional coordinates in the world coordinate system. And set it as a mandatory inspection task point, and link it to the parent equipment object to which the inspection part belongs at the equipment layer. In the context of storage, associated data is stored.

[0175] The path planning module is used for global inspection path planning with multi-objective fusion: Taking all marked inspection points as objectives, it constructs a multi-objective cost function that integrates safety, energy consumption, and timeliness, and solves for the globally optimal inspection path, generating a comprehensively optimal inspection "work order" for the robot. The path planning module includes:

[0176] The inspection point set generation unit is used to generate inspection task point sets: the system starts from the equipment layer. Extract the 3D coordinates of all successfully calibrated inspection points. This forms a set of target points to be visited.

[0177] The cost calculation unit is used to convert the path planning problem into a variant of the Traveling Salesman Problem (TSP) and construct a path cost function between any two inspection points for multi-objective optimization.

[0178] The path finding unit is used to find the optimal path using an improved genetic algorithm. For this NP-hard path finding problem, this invention employs an improved genetic algorithm to find an approximate optimal solution. Through selection, crossover, and mutation operations on the inspection point sequence (chromosomes), it iteratively converges to an inspection order that minimizes the total cost C. And the specific path that corresponds to it.

[0179] A map update module is used for closed-loop updating of the map and path based on change detection: During the inspection process, environmental changes are continuously detected, and the map and path are dynamically and incrementally updated, enabling the robot's inspection path to adapt to the dynamic changes in the converter station environment. The map update module includes:

[0180] The change detection unit is used to detect environmental changes: during subsequent routine inspections, the robot will use the currently acquired LiDAR scan data... With topographic layers Real-time registration is performed, and the registration error is calculated using the ICP algorithm. If the error exceeds a preset threshold... If this is detected, it is determined that a change in the physical environment has occurred in the area; simultaneously, by comparing the current visual recognition results with those at the device layer... Records of the addition, removal, or replacement of testing equipment;

[0181] The map update and version management unit is used for incremental map updates and version management: When a change is detected, the map management module archives the current map version (e.g., Map_v2.1) and triggers a local, small-scale exploration and calibration process in the changed area (i.e., re-executing S200 and S300). After the update is complete, a new version map with an update log (e.g., Map_v2.2) is generated and enabled.

[0182] The path replanning unit is used for dynamic replanning of inspection paths: once the map version is updated, the system will automatically discard the inspection path generated based on the old map and re-execute the global path planning process of the path planning module to generate an optimal path that perfectly matches the new environment, ensuring that the inspection task is always carried out under the latest and most accurate environmental model.

[0183] The specific functions and effects of the unmanned inspection system for oil intake and injection at converter stations can be explained by referring to other embodiments in this manual, and will not be repeated here. Each module and / or unit in the unmanned inspection system for oil intake and injection at converter stations can be implemented entirely or partially through software, hardware, or a combination thereof. Each module 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.

[0184] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for unmanned inspection of oil taking of a converter station, characterized in that, Comprising the following steps: Step S100: environmental awareness and initial exploration stage, which is the process of robot from the cold start state of completely unknown state to establish the initial positioning and local map, characterized by not relying on any preset information, completely through the interaction of itself and the environment, the step S100 comprises: Step S101: system initialization and multi-sensor data fusion; the system initializes the multi-sensor module carried, including: preheating the rotation speed of three-dimensional laser radar (LiDAR) and establishing the data port; zero bias estimation and initial alignment of inertial measurement unit (IMU); reading zero of wheel speed encoder; automatic setting of white balance and exposure parameters of vision camera; Step S102: system starts the LIO-SAM mapping method based on factor graph, which is used to tightly coupled fusion of LiDAR, IMU and wheel speed encoder data, realizes high precision and high robustness odometer, the algorithm core is to maintain and optimize a factor graph, the objective function is to minimize the sum of residuals of all measurement constraints to estimate the state of robot; Step S103: in the initialization stage of SLAM system, in order to quickly establish a stable local map, the robot executes the preset "walking along the wall" or "spiral line walking" exploration action in the open area at a safe speed, provides sensor data with enough parallax and motion incentive for factor graph optimization, so as to quickly converge to get a reliable initial pose (including 6 degrees of freedom, position: (x, y, z), attitude: Roll (roll angle), Pitch (pitch angle), Yaw (yaw angle)), and a local geometric point cloud map is generated at the same time; Step S200: hierarchical semantic map incremental construction: when the initial positioning accuracy reaches the threshold value, that is, the pose covariance output by the factor graph optimization process is less than the preset value, the system adopts the boundary priority exploration strategy to efficiently expand the map coverage range, and a hierarchical semantic world model including geometric terrain layer, power equipment semantic layer and feasible path layer is constructed synchronously, comprising the following steps: Step S201: Frontier detection and clustering: On the basis of the geometric terrain layer map output by the SLAM algorithm, the system detects the frontiers of known free space and unknown space in real time; all detected frontier points are aggregated into multiple frontier clusters through the DBSCAN clustering algorithm to improve the stability and efficiency of exploration; Step S202: Exploration target decision based on information gain: The system selects the next exploration target point from the frontier set F by a utility function selecting the next optimal exploration target point from the frontier set F ; Step S203: Hierarchical semantic map construction: the inspection robot updates the hierarchical semantic map in parallel when moving towards the target exploration point: the terrain layer , the device layer and the path layer ; wherein the terrain layer is a three-dimensional point cloud map continuously updated by the SLAM system, describing the physical structure and passability of the environment; the device layer is constructed by feeding the images captured by the visual camera into a pre-trained power equipment target detection network (such as YOLOv8), using the identified device category and location information, and fusing different perspective observation images of the same target with the help of a Kalman filter to improve positioning accuracy; the path layer records the safe passable trajectories verified by the robot and abstracts them into a topological path graph; Step S300: knowledge base based self-calibration of inspection points: for the newly discovered power equipment in the device layer, the observation task is planned automatically, and the key inspection points on the equipment are automatically calibrated through fine secondary identification, comprising the following steps: Step S301: Triggering of the device focus observation task: When a new device object is first identified in the device layer with a confidence higher than a threshold, or when the position uncertainty of some existing device is still large, the system automatically generates a "focus observation" subtask for the device. Step S302: Next-Best-View planning: In order to acquire high quality images of the device, the robot needs to move to a Next-Best-View (NBV); the system computes the NBV by optimizing an utility function that considers the distance to the device, the angle of view, and the avoidance of occlusions by the device itself . Step S303: Fine identification and calibration of key POI: After the robot reaches the best observation pose, it uses a high-resolution camera to take a close-up shot of the equipment. The system simultaneously queries the built-in "equipment-POI knowledge base", which stores the standard POI of various types of equipment and their visual features in a structured form (for example, "SF6 circuit breaker": ["pressure gauge", "opening and closing indicator"]). Then, the system uses template matching, SIFT feature point matching, and a dedicated secondary recognition network to accurately locate the 2D image coordinates of the POI identified by the knowledge base on the close-up image. Finally, using the camera's intrinsic parameters and depth data, the above 2D coordinates are back-projected to three-dimensional coordinates in the world coordinate system , and stored in association with the parent equipment object to which the POI belongs in the equipment layer . Step S400: multi-objective fusion global inspection path planning: taking all calibrated inspection points as the target, a multi-objective cost function integrating safety, energy consumption and timeliness is constructed, and the global optimal inspection path is solved to generate a comprehensive optimal inspection "work order" for the robot, comprising the following steps: Step S401: Generation of the inspection task point set: the system extracts all successfully calibrated inspection point three-dimensional coordinates from the device layer , forming a target point set to be visited; , forming a target point set to be visited; Step S402: equivalent to the variant of traveling salesman problem (TSP), the path cost function between any two inspection points is constructed for multi-objective optimization; Step S403: solving the optimal path by using the improved genetic algorithm: for the NP-hard type path solving problem, the improved genetic algorithm is used to find the approximate optimal solution, and the selection, crossover and mutation operations are performed on the inspection point sequence (chromosome), and the iteration is converged to an inspection sequence which makes the total cost C minimum and the specific path corresponding thereto; Step S500: map and path closed loop update based on change detection: continuously detect the environmental changes in the inspection process, and dynamically update the map and path for incremental update, so that the inspection path of the robot can adapt to the dynamic changes of the converter station environment, comprising the following steps: Step S501: Environment change detection: in subsequent routine inspection, the robot will compare the current acquired LiDAR scan data with the previous scan data with the terrain layer to perform real-time registration, and calculate the registration error by using the ICP algorithm, if the error exceeds the preset threshold , it is determined that the physical environment change occurs in the region; at the same time, by comparing the current visual recognition result with the record of the device layer , the addition, removal or replacement of the device is detected; Step S502: Map incremental update and version management: After detecting the change, the map management module archives the current map version (e.g., Map_v2.1) and triggers a local, small-range exploration and calibration process (i.e., re-executes S200 and S300) in the changed area; after the update is completed, a new version of the map with an update log (e.g., Map_v2.2) is generated and enabled; Step S503: Dynamic re-planning of inspection path: Once the map version is updated, the system will automatically discard the inspection path generated based on the old map and re-execute the global path planning process of S400 to generate an optimal path that fully matches the new environment, ensuring that the inspection task is always performed under the latest and most accurate environmental model.

2. The unmanned inspection method for oil taking of a converter station according to claim 1, characterized in that, The specific operations in step S102 include: (1) The objective function of the factor graph is to minimize the residual synthesis of all measurement constraints, as shown in equation (1): （1） In the formula, X* : represents the optimal solution of the optimization problem defined by equation (1); it is a state set or state vector containing the optimal estimates of all the state variables x i i i *, i.e. X* = {x1*, x2*,...,x n *} ; the optimal solution is a specific state set that minimizes the weighted sum of squared sensor measurement residuals among all possible X, representing the maximum a posteriori probability estimate of the complete motion trajectory and internal state of the robot; : set of robot state variables to be optimized; : represent the robot's position, pose (quaternion), velocity, accelerometer bias and gyroscope bias at time i, respectively; : prior factor residuals, e.g. global position constraints from GPS and pose constraints resulting from loop closure detection; : lidar factor residual; : the k-th frame lidar point cloud measurement, this function calculates the matching error of the current point cloud with the built map under the to-be-optimized pose X through point-to-point or point-to-edge matching; : IMU pre-integration factor residual; : IMU measurements (angular velocity and acceleration) between two laser frames, the function is based on the difference between the predicted pose change from IMU measurements and the pose variable X to be optimized; : sets of indices representing lidar and IMU measurements, respectively; : the square of the Mahalanobis norm, representing a weighted error, P is a weighting matrix. : inverse of the prior factor covariance matrix , and comprise GPS constraints and loop closure constraints; : laser radar measurement covariance matrix inverse of the laser radar measurement covariance matrix, is not a fixed value, but is dynamically calculated each time the point cloud is matched, and it reflects the quality of the current matching; : IMU measurement covariance matrix the inverse of which is determined mainly by the Gaussian white noise in the IMU's measurement noise; P: inverse of the covariance matrix (information matrix), used to adjust the weight of each error in the total error according to the uncertainty (noise) of the measurement; the higher the sensor accuracy, the greater the weight; the less accurate the sensor, the smaller the weight; the Mahalanobis norm can force the optimizer to preferentially satisfy the measurement constraints with higher credibility (smaller noise variance); (2) Optimization solution The entire optimization process is an iterative search process; the system starts with an initial X value, continuously adjusts the values of each variable in X through a nonlinear optimization algorithm (such as the Gauss-Newton method or the Levenberg-Marquardt method), so that the value of the objective function f(X) gradually decreases in the direction of the "gradient", and finally converges to a local or global minimum; the core of this iterative solution is to solve the following incremental equation by linearization: （2） In the formula, : state variable increment to be solved; r: all residual factors the total residual vector resulting from the stitching J: Jacobian matrix obtained by deriving the total residual vector r with respect to the state variable X, i.e. ; H: Hessian matrix, is the key structure of the whole optimization problem, in the Gauss-Newton method is approximated as: ; The Hessian matrix H is the second-order partial derivative matrix of the objective function with respect to the state variable X, which describes the local curvature of the objective function near the current solution; by solving equation (2), the increment ΔX is obtained, and the state X is updated as X ← X + ΔX, and the iteration is repeated until ΔX converges to zero, and finally the optimal state estimate X* is obtained; (3) Covariance evaluation of state estimation results After the optimization converges to the optimal state estimate X*, the system further quantifies the uncertainty of the estimation result; the posterior covariance matrix Σ_state of the final estimation state can be approximated by the inverse of the Hessian matrix H of the entire factor graph optimization problem at the optimal solution X*, as shown in equation (3): (3) In the formula, : i.e. the final output pose covariance matrix; it is a large block-diagonal matrix with block matrices i : i.e. the final output pose covariance matrix; it is a large block-diagonal matrix with block matrices i : i.e. the final output pose covariance matrix; it is a large block-diagonal matrix with block matrices : inverse of the hessian matrix; the "size" of the hessian matrix H (measured by its eigenvalues) reflects the strength of the constraints on the optimal solution, and its inverse H-1 directly quantifies the uncertainty of the optimal solution; for example, if the system has been through a feature-rich scene and contains high-quality loop closures, the H matrix will be very "strong" and its inverse state will be very "small", indicating that the final estimated robot trajectory has very high precision and confidence.

3. The unmanned inspection method and device for oil taking of a converter station according to claim 1, characterized in that, The decision function in step S202 aims to maximize information acquisition efficiency, which can be represented as equation (4): （4） In the formula, : represents the gain in information brought by exploring the cluster of boundaries is usually measured by the number of boundary points it contains or its geometric length; : Represents the robot's current position Go to The cost of the center point is usually determined by The shortest path length calculated by the algorithm on the current terrain map is represented by the shortest path length. : are the weight coefficients of information gain and moving cost respectively, which can be adjusted according to the task strategy.

4. The unmanned inspection method for oil taking of a converter station according to claim 1, characterized in that, The terrain layer in the step S203 , the device layer and the path layer are respectively: (1) Terrain layer : A 3D point cloud map or its projected 2D occupancy grid map continuously output by the SLAM system in S100, describing the physical structure of the environment for basic navigation and obstacle avoidance; (2) Device layer : Visual camera continuously captures images, which are fed into a pre-trained power equipment target detection network (e.g., YOLOv8); for each detected equipment, the system extracts its category, 2D bounding box, and combines the synchronous depth information or LiDAR point cloud to calculate its three-dimensional position in the world coordinate system ; Fusing multiple observations of the same target through a Kalman filter to obtain a more accurate pose estimate and assigning a unique identifier (UUID) to each individual device; (3) Path layer : Record the trajectories that the robot has safely passed through and abstract them into a topological graph , where nodes V are key position points and edges E are path segments connecting nodes and their passing costs.

5. The unmanned inspection method for oil taking of a converter station according to claim 1, characterized in that: In the step S402, the multi-objective optimization cost function is a weighted sum of the path cost between any two points ​​ （5） In the formula, : safety cost, on the topographic map, according to the voltage level and position of high-voltage equipment in the equipment layer, a safety risk field is generated by superimposing Gaussian functions This cost is the line integral of the path on the risk field, which punishes the path close to the dangerous source; : Energy cost is the weighted sum of path length, cumulative turn angle and climbing height, which simulates the physical energy consumption of the robot. : timeliness cost, composed of travel time (length / speed) of the path and estimated stay operation time (e.g. taking pictures, measuring temperature) at each inspection point.

6. The unmanned inspection method for oil taking of a converter station according to claim 1, characterized in that, In step S403, the specific operations of the improved genetic algorithm include: Step (1): Parameter definition and optimization objective formalization Define the parameters required by the algorithm, including: population size N_pop, maximum iteration G_max, crossover probability P_c, mutation probability P_m, and elite retention number N_elite; A candidate solution is defined as a chromosome, which is an ordered permutation of the set of inspection points The optimization goal of the method is to find an optimal permutation such that the total cost function is minimized; the total cost function is defined as follows:​ （6） In the formula, represents the i-th polling point in the arrangement represents the i-th polling point in the arrangement a current start position for the robot, an end position prepared for the task; Time(A, B) is the shortest travel distance of the robot from the spatial position A to B, which is calculated by executing the path search algorithm on the path layer above calculated by the path search algorithm To convert the above minimization problem into a standard fitness maximization problem of genetic algorithm, a fitness function is defined as f(x) = 1 / (1 + x) （7） Step (2): Population initialization based on heuristic algorithm An initial population of size N pop is generated; this step adopts a hybrid strategy, in which a part of chromosomes are generated by randomly arranging the set of inspection points P to ensure the diversity of the population, and the other part of chromosomes are generated by using the nearest neighbor heuristic algorithm to introduce high-quality initial solutions into the population and accelerate the convergence of the algorithm; Step (3): fitness evaluation and elite reservation selection for each chromosome in the current population its fitness value is calculated according to the fitness function defined in step (1); the selection operation is performed by combining the strategy of elite reservation with that of tournament selection; first, the N_elite individuals with the highest fitness in the current population are determined as the elite set E_g (g is the current generation number) and are directly copied into the next generation population; second, the parent population for the subsequent crossover operation is generated by the tournament selection method; its fitness value is calculated according to the fitness function defined in step (1); the selection operation is performed by combining the strategy of elite reservation with that of tournament selection; first, the N_elite individuals with the highest fitness in the current population are determined as the elite set E_g (g is the current generation number) and are directly copied into the next generation population; second, the parent population for the subsequent crossover operation is generated by the tournament selection method; Step (4): recombination operation based on order crossover (OX) To ensure that the offspring chromosomes are still legal arrangements, the selected parent individuals are subjected to the order crossover (OX) operator according to the crossover probability P c to generate new offspring individuals; the operator ensures the validity of the arrangement of the offspring chromosomes by retaining part of the gene sequence of the parent and filling the remaining genes in the order of the other parent; Step (5): perturbation operation based on reverse mutation To increase population diversity and avoid the algorithm falling into local optimum, the newly generated individual is subjected to inversion mutation operation according to mutation probability P_m; for a chromosome , two indexes i and j (i < j) are randomly selected to generate the mutated chromosome , that is, the gene sequence between indexes i and j is arranged in reverse order; Step (6): iteration and termination condition judgment Steps (3) to (5) are repeatedly performed; let g be the current iteration index, is the optimal individual of the gth generation; the algorithm terminates when any of the following conditions is met: ① the current iteration algebra g reaches a preset maximum iteration algebra G max , i.e. g ≥ G max ; ② The optimal fitness value of the continuous k generations has no significant improvement, that is, wherein, is a preset stagnation judgment threshold value; After the algorithm terminates, the elite individual with the largest fitness value corresponds to the inspection order Output as a result of optimal path planning.

7. An unmanned inspection system for oil taking of a converter station, characterized in that, The system comprises: An environment exploration module for environment cognition and initial exploration stage, which is the process of the robot from the cold start state of complete unknown state to the establishment of initial positioning and local map, characterized by not relying on any preset information, but completing through the interaction of itself and the environment, the environment exploration module comprising: An initialization unit for system initialization and multi-sensor data fusion; the system initializes the multi-sensor module carried, including: preheating the rotation speed of the three-dimensional laser radar (LiDAR) and establishing the data port; zero bias estimation and initial alignment of the inertial measurement unit (IMU); reading zero of the wheel speed encoder; automatic setting of white balance and exposure parameters of the vision camera; A state estimation unit for starting the LIO-SAM mapping method based on factor graph, which is used to tightly coupled fusion of LiDAR, IMU and wheel speed encoder data to realize high-precision and high-robustness odometer, the core of the algorithm is to maintain and optimize a factor graph, and the objective function is to minimize the sum of residuals of all measurement constraints to estimate the state of the robot; A point cloud generation unit for quickly establishing a stable local map in the initialization stage of the SLAM system, the robot performs a preset "walking along the wall" or a small range "spiral line walking" exploration action in an open area at a safe speed, to provide sensor data with sufficient parallax and motion incentive for factor graph optimization, so as to quickly converge to obtain a reliable initial pose (including 6 degrees of freedom, position: (x, y, z), attitude: Roll (roll angle), Pitch (pitch angle), Yaw (yaw angle)), and simultaneously generate a local geometric point cloud map; A map construction module for hierarchical semantic map incremental construction: when the initial positioning accuracy reaches the threshold, i.e. the pose covariance output by the factor graph optimization process is less than the preset value, the system adopts a boundary priority exploration strategy to efficiently expand the map coverage range and simultaneously construct a hierarchical semantic world model containing a geometric terrain layer, a power equipment semantic layer and a feasible path layer, the map construction module comprising: A clustering unit is configured to frontier cluster detection and clustering. Based on the geometric terrain layer map output by the SLAM algorithm, the system detects the frontier between the known free space and the unknown space in real time. All detected frontier points are aggregated into multiple frontier clusters through the DBSCAN clustering algorithm. to improve the stability and efficiency of exploration. exploration target decision unit for exploration target decision based on information gain: the system selects the next optimal exploration target point from the boundary cluster set F ; and ; and A hierarchical map construction unit is configured to construct a hierarchical semantic map, wherein the inspection robot updates the hierarchical semantic map in parallel when moving towards a target exploration point, including a terrain layer , a device layer , and a path layer ; the terrain layer is a three-dimensional point cloud map describing the physical structure and passability of the environment, which is continuously updated by a SLAM system; the device layer is constructed by feeding the images captured by a visual camera into a pre-trained power equipment target detection network (such as YOLOv8), using the identified device category and location information, and fusing different perspective observation images of the same target with a Kalman filter to improve positioning accuracy; the path layer records the safe passable trajectories verified by the robot and abstracts them into a topological path graph; The inspection point calibration module is configured to autonomously calibrate the inspection points based on the knowledge base. For newly discovered power equipment in the device layer, an observation task is autonomously planned, and through refined secondary identification, key inspection points on the equipment are automatically calibrated. The inspection point calibration module includes: A sub-task generation unit is used to trigger a device focus observation task: when a new device object with a confidence higher than a threshold is first identified in the device layer or when the position uncertainty of a certain existing device is still large, the system automatically generates a "focus observation" sub-task for the device; a pose planning unit for Next-Best-View (NBV) planning: in order to acquire high-quality images of the device, the robot needs to move to a pose that provides the best view of the device; the system computes this pose by optimizing an utility function that takes into account the distance to the device, the angle of view, and the avoidance of occlusions by the device itself . The patrol point marking unit is used for fine identification and calibration of key patrol points (POIs). After the robot reaches the best observation pose, a high-resolution camera is used to take a close-up shot of the equipment. The system simultaneously queries the built-in "equipment-patrol point knowledge base", which stores the standard patrol points of various equipment and their visual features in a structured form (for example, "SF6 circuit breaker": ["pressure gauge", "opening and closing indicator"]). Subsequently, the system uses template matching, SIFT feature point matching, and a special secondary recognition network to accurately locate the 2D image coordinates of the patrol points queried from the knowledge base on the close-up image. Finally, the above 2D coordinates are back-projected into three-dimensional coordinates in the world coordinate system using the camera intrinsic parameters and depth data , and stored in association with the parent equipment object to which the patrol point belongs in the equipment layer . The path planning module is configured to plan a global inspection path based on multi-objective fusion. With all the calibrated inspection points as targets, a multi-objective cost function that fuses safety, energy consumption, and timeliness is constructed, and a globally optimal inspection path is solved to generate a comprehensive optimal inspection "work order" for the robot. The path planning module includes: The inspection point set generating unit is configured to generate an inspection task point set. The system extracts all successfully calibrated inspection point three-dimensional coordinates from the device layer , and forms a target point set to be visited. ​ The cost calculation unit is configured to equivalently convert the path planning problem into a variant of the traveling salesman problem (TSP) and construct a multi-objective optimization path cost function between any two inspection points. The path solving unit is used for solving the optimal path by using the improved genetic algorithm; for the NP-hard type path solving problem, the improved genetic algorithm (Genetic Algorithm) is used to find the approximate optimal solution, and through the selection, crossover and mutation operation on the inspection point sequence (chromosome), the iteration converges to an inspection sequence which makes the total cost C minimum and the specific path corresponding thereto; The map updating module is configured to update the map and path based on change detection. During the inspection process, environmental changes are continuously detected, and the map and path are dynamically updated incrementally, so that the inspection path of the robot can adapt to the dynamic changes of the converter station environment. The map updating module includes: Change detection unit for environment change detection: in subsequent routine inspection, the robot will compare the current acquired LiDAR scan data with the terrain layer to perform real-time registration, calculate the registration error by using ICP algorithm, and if the error exceeds the preset threshold , it is determined that the region has undergone physical environment change; at the same time, by comparing the current visual recognition result with the record of the device layer , the addition, removal or replacement of the device is detected; The map updating and version management unit is configured to perform incremental updating and version management of the map. When a change is detected, the map management module archives the current map version (e.g., Map_v2.1) and triggers a local and small-range exploration and calibration process (i.e., re-executes S200 and S300) in the changed area. After updating, a new version of the map with an update log (e.g., Map_v2.2) is generated and enabled. The path re-planning unit is configured to dynamically re-plan the inspection path. Once the map version is updated, the system will automatically discard the inspection path generated based on the old map and re-execute the global path planning process of the path planning module to generate an optimal path that fully matches the new environment, ensuring that the inspection task is always performed under the latest and most accurate environmental model.

Images