Ring main unit inspection robot autonomous navigation method and system based on SLAM

Abstract

The invention discloses a ring main unit inspection robot autonomous navigation method and system based on SLAM, particularly relates to the technical field of robot autonomous navigation and intelligent inspection, and is used for solving the problem of positioning drift caused by repeated features of an existing ring main unit scene. Semantic feature analysis and topological constraints are introduced into an SLAM processing flow, acquired image data and point cloud data are processed through a deep learning model, objects such as an electrical cabinet, a corridor channel and a cable trench are identified, and a semantic feature set with category labels and spatial position information is generated; and constructing a topological graph containing node spacing, connectivity and directivity constraints based on the semantic features, adding the topological graph as a constraint factor into SLAM back-end optimization, and performing joint optimization in combination with vision, a laser odometer and inertial prior information, thereby avoiding only depending on repeated geometric feature positioning, and improving the positioning accuracy. The problems of loopback misjudgment and drifting caused by feature confusion are reduced, and the pose resolving stability in the ring main unit environment is improved.

Description

Technical Field

[0001] This invention relates to the field of robot autonomous navigation and intelligent inspection technology, specifically to an autonomous navigation method and system for a ring main unit inspection robot based on SLAM. Background Technology

[0002] With the continuous improvement of the intelligence level of power distribution networks, intelligent inspection methods based on autonomous mobile robots are gradually replacing traditional manual inspections and are widely used in scenarios such as ring main units and power distribution rooms to reduce maintenance costs and improve safety. Currently, inspection robots generally rely on Simultaneous Localization and Mapping (SLAM) technology, which uses sensors to collect environmental information and combines it with motion estimation algorithms to achieve autonomous navigation. Common SLAM methods include geometric feature mapping based on LiDAR and camera-based visual SLAM, which estimate pose by detecting environmental feature points and use loop closure detection and graph optimization to reduce cumulative errors. Some improved solutions adopt multi-sensor fusion strategies, introducing inertial measurement units (IMUs) to compensate for short-term sensor failures and improve positioning stability.

[0003] However, the inspection scenario of ring main unit has unique physical characteristics, posing significant challenges to existing SLAM technologies. Such scenarios typically involve long, narrow spaces with limited aisle width, numerous cabinets arranged at regular heights, resulting in highly repetitive geometric and textural features for localization. Traditional SLAM algorithms based on local feature matching are prone to feature confusion. Furthermore, the cabinet surfaces are primarily made of metal, exhibiting strong reflectivity, which can easily generate false points or reflection anomalies during laser ranging. Simultaneously, lighting conditions are affected by ambient light and cabinet occlusion, potentially creating localized highlights or shadows, reducing image quality for visual sensors and increasing recognition errors. These combined factors make existing SLAM systems prone to matching errors during loop closure detection and pose calculation, leading to a rapid increase in accumulated errors, resulting in localization drift or even pose loss.

[0004] To alleviate these problems, some existing solutions attempt to introduce multi-source sensing data fusion, improve feature extraction algorithms, or reduce drift through global closed-loop correction. However, due to the lack of global constraints specific to the layout characteristics of ring main units, it remains difficult to completely avoid the accumulation of positioning errors. Furthermore, traditional methods often rely on manual intervention or re-initialization when drift or loss of positioning occurs. This process is complex and prone to interruption, failing to meet the requirements of power operation and maintenance for autonomy, real-time performance, and continuity. Therefore, existing technologies still suffer from unstable positioning accuracy, path tracking deviation, and low relocation efficiency in ring main unit inspection scenarios, and these problems have not yet been effectively resolved in the industry. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides an autonomous navigation method and system for a ring network cabinet inspection robot based on SLAM. By introducing semantic feature parsing and topological constraints, and combining multi-source information to optimize SLAM, the problems mentioned in the background technology are solved.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution: an autonomous navigation method for a ring main unit inspection robot based on SLAM, characterized in that it includes: S1. Collect point cloud data, image data, and inertial data of the ring network cabinet inspection environment through lidar, vision sensor, and inertial measurement unit, and perform time synchronization processing. S2. Perform deep learning processing on the collected data to identify the outlines and spatial locations of electrical cabinets, corridors, walls and cable trenches, and generate a set of semantic features containing category labels and three-dimensional coordinates. S3. Construct a semantic topology map based on semantic features, abstract cabinet nodes and corridor paths into topological structures, and establish distance constraints and path connection relationships between nodes. S4. In the SLAM optimization process, the semantic topological constraints are introduced, and the pose calculation is optimized by combining the inertial measurement results to generate enhanced localization results. S5. During operation, the positioning health is calculated in real time. The positioning health is predicted to have a risk of deviation by combining the loop closure consistency index, feature matching confidence and inertial drift trend. S6. When the health level is lower than the threshold, the system triggers relocation, uses the cabinet QR code identification and key frame semantic features to correct the pose calculation results, and replans the inspection path based on the corrected pose.

[0009] Furthermore, S1 includes: The inspection robot performs omnidirectional or multi-plane scanning by using a lidar configured in the robot to obtain a three-dimensional point cloud that reflects the spatial geometry. Scene data, including color images and depth information, is collected by visual sensors installed on the inspection robot. The inertial measurement unit is used to collect acceleration and angular velocity information in real time to form inertial data describing the robot's motion state; The laser, vision, and inertial data are synchronized using a unified timestamp mechanism and a hardware clock to output a multi-source sensing data stream with consistent timing.

[0010] Furthermore, S2 includes: The time-synchronized image data and point cloud data are respectively input into the deep learning model. Convolutional feature extraction and target classification are performed on the image data, and spatial feature extraction and category recognition are performed on the point cloud data. Pixel-level and point-level semantic annotations are performed on the extracted results to identify the outlines and spatial locations of electrical cabinets, corridors, walls, cable trenches and other static structures; By combining depth information or laser point cloud computing, the three-dimensional spatial coordinates of the target object are generated to produce a semantic feature set containing category labeling and spatial location information.

[0011] Furthermore, S3 includes: Based on semantic features, electrical cabinets, corridors, walls, and cable trenches are identified and their three-dimensional center coordinates and category attributes are extracted as basic information of topology nodes. Based on scene constraints such as spatial distance between nodes, relative orientation, and corridor width, it is determined whether the connectivity conditions are met. If they are met, the passable paths between the nodes are abstracted as edges in the topology graph. When constructing the topological relationships, the connection relationships between adjacent nodes are determined based on the cabinet arrangement rules and scene geometric constraints, and the node category attributes and spatial orientation information are labeled. Construct a topology structure that includes node spacing constraints, connectivity constraints, and directional constraints to form a topology graph that can be used for constrained pose calculation.

[0012] Furthermore, S4 includes: The node positions, connections, and constraints in the topology graph are introduced into the SLAM backend optimization model, and the keyframe poses obtained from vision and laser odometry are used as optimization variables. When performing pose calculation, the keyframe poses generated by vision and laser odometry are used as optimization variables, and the topological constraints and motion prior information provided by the inertial measurement unit are superimposed to construct a joint optimization objective function. By iteratively solving the optimization model, the pose calculation results that satisfy semantic topological constraints and motion continuity requirements are output.

[0013] Furthermore, S5 includes: Based on the loop closure detection results generated during the pose calculation process, a loop closure consistency index is extracted. By combining the matching confidence of semantic feature points, the consistency between the current scene recognition and the map model is evaluated; Simultaneously, the acceleration and angular velocity information output by the inertial measurement unit are used to analyze the deviation trend between the current trajectory and the inertial-predicted trajectory; The above multidimensional indicators are weighted and fused to generate a comprehensive health index. The trend of this index is used to determine whether there is a potential risk of location deviation. When the risk exceeds the preset threshold, a drift warning signal is output for subsequent relocation module to call.

[0014] Furthermore, S6 includes: When the health indicators reach the trigger conditions, the repositioning process is initiated. The absolute position information corresponding to the QR code is obtained by identifying the QR code on the cabinet, and the position information is aligned with the current pose for calculation. During the relocation process, the cabinet QR code labels installed along the inspection path are first identified to extract the absolute position information corresponding to the QR code, and then the position information is compared with the current SLAM pose. Subsequently, pose correction was performed based on semantic feature matching of historical keyframes, combined with cabinet sequence features and corridor topology. After completing the pose correction, the inspection path is recalculated, and the navigation target point is updated based on the corrected pose.

[0015] Furthermore, including: After extracting image and point cloud features separately during the encoding stage, a feature alignment module is used to perform cross-modal fusion to form a unified feature representation, which is used to reduce registration errors caused by differences in viewpoints. When performing semantic annotation, a two-branch structure is adopted. One branch performs pixel-level classification for two-dimensional images, and the other branch performs point-level classification for three-dimensional point clouds. Then, pixel labels are mapped to point labels based on coordinate mapping. The generated semantic feature set contains object category identifiers, 3D center point coordinates, bounding boxes, and topological attribute labels in the form of a data structure, providing input for subsequent topological mapping.

[0016] Furthermore, including: After the QR code is decoded, the unique identifier obtained is queried against the preset map index table and bound to the predefined global coordinate system to obtain the absolute position coordinates; In the pose correction process, the position information provided by the QR code is first added to the graph optimization model as a strong constraint factor, and then the keyframe matching results based on semantic features are added to the optimization as a weak constraint factor to ensure computational stability. A priority matching strategy is adopted for cabinet sequence features and topological relationships. When the QR code is unavailable, the current position is inferred by combining the cabinet arrangement order with the known path topology, and the factor connection relationship in the SLAM graph is iteratively updated.

[0017] On the other hand, the present invention provides an autonomous navigation system for a ring main unit inspection robot based on SLAM, including: The multi-source data acquisition module, including lidar, vision sensor and inertial measurement unit, is used to acquire 3D point cloud, color image, depth information and inertial data, and complete time synchronization through a unified timestamp mechanism to output multi-source sensing data stream; The scene semantic parsing module is used to receive synchronous data, perform image feature and point cloud feature extraction, perform pixel-level semantic annotation on the image, perform point-level semantic annotation on the point cloud, and generate a semantic feature set containing object category and three-dimensional coordinates. The semantic topology modeling module is used to construct a topology graph based on semantic features, abstracting cabinets, corridors, etc. as nodes, paths as edges, and establishing distance constraints, connectivity constraints, and directional constraints between nodes. The topology constraint optimization module is used to introduce the topology graph as a constraint into the SLAM backend optimization model, and combine laser, visual odometry calculation and inertial prior to output the optimized pose calculation result. The positioning health assessment module is used to generate health indicators based on loop closure detection consistency, semantic feature matching confidence, and inertial trajectory deviation, and to trigger a relocation signal when the indicator is below a threshold. The relocation execution module is used to correct the pose, update the SLAM graph factor connections, and replan the inspection path when triggered, by recognizing the cabinet QR code and combining semantic keyframes and topological features.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing semantic feature parsing and topological constraints into the SLAM processing flow, the collected image data and point cloud data are processed by a deep learning model to identify objects such as electrical cabinets, corridors, and cable trenches, and generate a set of semantic features with category labels and spatial location information. Based on the semantic features, a topological graph containing node spacing, connectivity, and directional constraints is constructed, and this topological graph is added as a constraint factor to the SLAM backend optimization. Combined with vision, laser odometry, and inertial prior information, joint optimization is performed to avoid relying solely on repetitive geometric features for localization, reduce loop closure misjudgment and drift problems caused by feature confusion, and improve the stability of pose calculation in the ring network cabinet environment. 2. When the real-time calculated health index is lower than the preset threshold, relocation is triggered. The absolute position is obtained by recognizing the cabinet's QR code and binding it to the global coordinate system. This position is added to the optimization model as a strong constraint. At the same time, the semantic features and topological information of historical keyframes are combined as weak constraints to perform pose correction. When the QR code information is unavailable, the cabinet sequence rules and topological structure are used to calculate the position, realizing a multi-level relocation strategy. This reduces the impact of sensor interference or missing local features on positioning accuracy and ensures navigation continuity and trajectory reliability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the autonomous navigation method for a ring main unit inspection robot based on SLAM according to the present invention; Figure 2This is a schematic diagram of the autonomous navigation system for the ring main unit inspection robot based on SLAM of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: Figure 1 The present invention provides an autonomous navigation method for a ring main unit inspection robot based on SLAM, comprising: S1. Collect point cloud data, image data, and inertial data of the ring network cabinet inspection environment through lidar, vision sensor, and inertial measurement unit, and perform time synchronization processing. S2. Perform deep learning processing on the collected data to identify the outlines and spatial locations of electrical cabinets, corridors, walls and cable trenches, and generate a set of semantic features containing category labels and three-dimensional coordinates. S3. Construct a semantic topology map based on semantic features, abstract cabinet nodes and corridor paths into topological structures, and establish distance constraints and path connection relationships between nodes. S4. In the SLAM optimization process, the semantic topological constraints are introduced, and the pose calculation is optimized by combining the inertial measurement results to generate enhanced localization results. S5. During operation, the positioning health is calculated in real time. The positioning health is predicted to have a risk of deviation by combining the loop closure consistency index, feature matching confidence and inertial drift trend. S6. When the health level is lower than the threshold, the system triggers relocation, uses the cabinet QR code identification and key frame semantic features to correct the pose calculation results, and replans the inspection path based on the corrected pose.

[0022] S1. Point cloud data, image data, and inertial data of the ring main unit inspection environment are collected through lidar, vision sensors, and inertial measurement units, and time synchronization processing is performed. The specific implementation is as follows: The inspection robot utilizes a lidar system to perform omnidirectional or multi-planar scanning, acquiring a 3D point cloud reflecting the spatial geometry. The lidar, which can be configured as a multi-line laser supporting omnidirectional rotational scanning or a multi-planar arrangement for comprehensive measurement, ensures that the acquired point cloud covers the entire spatial area of ​​the ring network cabinet's passageway. During operation, the lidar continuously emits laser beams and receives reflected signals, calculating distances using time-of-flight or phase difference to obtain a 3D point cloud reflecting the scene's geometry. To improve the integrity of the point cloud, the system can stitch together multiple frames of data as the robot moves, eliminating data loss caused by single-frame occlusion.

[0023] The inspection robot collects scene data including color images and depth information using vision sensors mounted on it. These sensors can be RGB-D cameras, capable of capturing both color images and simultaneously acquiring depth information, or stereo cameras combined with computational methods to generate depth data. To ensure reliability in ring main unit environments, the camera's installation angle is adjustable to cover the front of the unit and both sides of the corridor, avoiding large blind spots. During data acquisition, the system triggers snapshots periodically, adjusting exposure based on lighting changes and using supplementary lighting when necessary to ensure clear and usable images under varying brightness levels.

[0024] An inertial measurement unit (IMU) is used to collect acceleration and angular velocity information in real time, forming inertial data describing the robot's motion state. The inspection robot is equipped with an IMU that outputs acceleration and angular velocity data in real time to reflect the robot's attitude changes and motion trends. The IMU may include a three-axis accelerometer and a three-axis gyroscope, mounted near the robot's center of gravity to reduce vibration interference. During robot movement, the IMU samples and records motion state information at a high frequency. This data not only serves as prior input for the SLAM system but also provides short-term pose estimation capabilities when visual or laser features are missing, reducing the risk of temporary positioning failures.

[0025] The laser, vision, and inertial data are synchronized using a unified timestamp mechanism and a hardware clock, resulting in a time-consistent multi-source sensing data stream. During data acquisition, the LiDAR, vision sensor, and inertial measurement unit output data at different frequencies. Inaccurate time alignment can lead to pose estimation errors. Therefore, this embodiment employs a hardware clock and a unified timestamp mechanism to strictly synchronize the multi-source data. Specifically, the system assigns a unified reference clock to all sensors, ensuring that the time difference between data acquisitions from each sensor remains within microseconds via a high-speed bus or time synchronization protocol (such as PTP or hardware trigger signals). A unified timestamp is appended to the data when it enters the buffer, ultimately outputting a time-consistent sensing data stream, providing accurate timing guarantees for subsequent semantic feature extraction and SLAM optimization.

[0026] S2. Perform deep learning processing on the collected data to identify the outlines and spatial locations of electrical cabinets, corridors, walls, and cable trenches, and generate a semantic feature set containing category labels and three-dimensional coordinates. The specific implementation is as follows: The inspection robot first synchronously acquires image and point cloud data from multiple sources, using a timestamp mechanism for unified synchronization. Then, the system inputs the time-synchronized image and point cloud data into the deep learning model. The image data undergoes convolutional neural network processing to extract color, texture, and edge features, while also performing target classification to identify electrical equipment outlines and background elements in the scene. The point cloud data is input into the spatial feature extraction network, focusing on analyzing 3D point distribution, surface structure, and spatial continuity, and performing category recognition. This dual-channel input method fully leverages the advantages of images in texture recognition and the strengths of point clouds in spatial morphology perception, achieving efficient extraction of complementary features.

[0027] The extracted results are subjected to pixel-level and point-level semantic annotation to identify the contours and spatial locations of electrical cabinets, corridors, walls, cable trenches, and other static structures. After feature extraction, the system performs pixel-level and point-level semantic annotation on the output results. Specifically, the image channel annotation uses each pixel as a processing unit, assigning a corresponding semantic category to each pixel, giving the image clear region classification information; the point cloud channel annotation uses each 3D point as a basic unit, combining spatial neighborhood and feature distribution for label assignment. In this way, accurate identification of multiple objects in the inspection environment can be achieved, such as electrical cabinets, corridors, walls, cable trenches, and other fixed facilities. During the annotation process, the system also further subdivides the cabinet contours and corridor boundaries to ensure that key structural features are preserved in subsequent topology modeling.

[0028] By combining depth information or laser point cloud data to calculate the 3D spatial coordinates of the target object, a semantic feature set containing category labels and spatial location information is generated. After semantic labeling, the system calculates the 3D spatial coordinates of each target object using depth information or laser point cloud data. This process includes not only extracting the center point coordinates but also generating a corresponding 3D bounding box to describe the object's spatial occupancy. Finally, the system integrates the category labels and spatial location information to form a complete set of semantic features. This feature set is not only used for subsequent semantic topology map construction but also provides a constraint basis for pose optimization, enabling the inspection robot to achieve more robust localization and navigation in complex ring network cabinet scenarios.

[0029] S3. Construct a semantic topology map based on semantic features, abstract cabinet nodes and corridor paths into a topological structure, and establish distance constraints and path connection relationships between nodes. The specific implementation is as follows: Based on semantic features, electrical cabinets, corridors, walls, and cable trenches are identified and their 3D center coordinates and category attributes are extracted as basic information for topology nodes. Based on scene constraints such as spatial distance between nodes, relative orientation, and corridor width, connectivity conditions are determined. If met, the traversable paths between the nodes are abstracted as edges in the topology graph. Image features and point cloud features are first mapped to coordinates through extrinsic parameter calibration. Then, a cross-modal feature alignment module is used to map image semantic labels to point cloud 3D coordinates, forming a unified semantic feature set. This set includes node categories, center coordinates, boundary information, and topology attributes, serving as the foundational data for constructing the topology graph. After extracting semantic features from images and point clouds, the system abstracts structural elements such as electrical cabinets, corridors, walls, and cable trenches into topology nodes based on the identified scene objects. Here, "nodes" are not simply geometric coordinate points, but data entities with semantic attributes. Each node includes a category identifier (e.g., "electrical cabinet"), spatial coordinates, and connection attributes to other nodes. Simultaneously, the system abstracts these traversable paths in the inspection corridors into "edges" connecting the nodes. In this way, the entire inspection environment is transformed from a point cloud model into a topological structure composed of nodes and edges, which can more intuitively describe the logical relationships of the environment, rather than just geometric features. This approach allows subsequent path planning and SLAM optimization to incorporate more structured information, rather than relying solely on unordered point cloud matching.

[0030] When constructing the topology, the connection relationships between adjacent nodes are determined based on the cabinet arrangement rules and scene geometric constraints, and node category attributes and spatial orientation information are labeled. In generating the topology, this embodiment does not simply connect nodes based on proximity, but rather makes judgments based on the cabinet arrangement rules and scene geometric constraints. The arrangement rules refer to the fact that ring network cabinets are typically arranged in a standardized linear or semi-closed ring shape, with roughly consistent spacing between adjacent cabinets and parallel to the corridor direction. Therefore, the system analyzes the orientation information in the semantic features to determine which nodes are on the same straight line or in adjacent sequences, and establishes connections between these nodes. Simultaneously, the system labels each connection edge with directional information (e.g., the direction of travel from cabinet A to cabinet B) and adds its category attributes to the nodes for subsequent navigation strategies. When encountering corners or branch passages, the system also creates new edges and labels the branch attributes to ensure the topology map fully reflects the actual spatial orientation.

[0031] Construct a topology structure that includes node spacing constraints, connectivity constraints, and directional constraints to form a topology graph that can be used for constrained pose calculation.

[0032] After constructing the nodes and edges, this embodiment also establishes spatial constraints for the topology to improve its effectiveness in SLAM optimization. These constraints include node spacing constraints, path connectivity constraints, and directional constraints. Node spacing constraints ensure that the abstracted connecting edges conform to the ring main unit installation specifications, avoiding unreasonable "jump connections" in the topology; connectivity constraints ensure the closure of the topology paths, preventing isolated nodes during graph optimization; directional constraints specify the directional attributes of the paths, preventing the introduction of erroneous loop closures when loop closure detection errors occur. In actual implementation, these constraints are transformed into factors in factor graph optimization, so that pose calculation not only considers sensor observation data but also is subject to the consistency constraints of the topology, thereby enhancing the robustness of SLAM in repetitive scenarios.

[0033] S4. In the SLAM optimization process, the semantic topological constraints are introduced, and the pose calculation is optimized by combining the inertial measurement results to generate enhanced localization results. The specific implementation is as follows: The node positions, connections, and constraints in the topology graph are introduced into the SLAM backend optimization model, with keyframe poses calculated by vision and laser odometry used as optimization variables. First, the topological information obtained through semantic feature recognition is combined with traditional geometric information and fed into the SLAM backend optimization model. Specifically, the system introduces the position of each node in the semantic topology graph, the connections between adjacent nodes, and the constraints between nodes (such as fixed spacing and path direction) as constraint factors into the factor graph. These topological constraints do not directly replace the original geometric constraints but are added as additional optimization conditions, ensuring that the final pose calculation not only satisfies geometric consistency but also conforms to the scene's structural logic. For example, if two electrical cabinets are arranged in a fixed order in the actual environment, even if visual features are confused due to lighting or reflection, the optimization process can "pull back" their positions based on this topological relationship.

[0034] During pose calculation, keyframe poses generated by visual and laser odometry are used as optimization variables, while topological constraints and motion prior information provided by the inertial measurement unit (IMU) are superimposed to construct a joint optimization objective function. In the pose optimization process, the system integrates multiple data sources to construct a joint optimization objective function. This objective function not only includes keyframe pose variables provided by visual and laser odometry but also incorporates motion prior information from the IMU. In other words, the pose of each frame is no longer determined by a single sensor but is solved through comprehensive constraints from the laser, camera, and inertial sensors, while also satisfying semantic topological constraints. This joint optimization approach avoids positioning drift caused by the failure of a single sensor. For example, when insufficient lighting prevents reliable matching of visual features, the laser and IMU can still provide certain constraints, ensuring the stability of the overall optimization process.

[0035] The optimization model is solved iteratively to output pose calculation results that satisfy semantic topological constraints and motion continuity requirements. After constructing the joint optimization model, the system uses a nonlinear optimization algorithm (such as an iterative solution method based on factor graphs) to continuously adjust the pose variables of keyframes, ensuring that they simultaneously satisfy geometric constraints, motion priors, and semantic topological constraints. During each iteration, the system recalculates the error function and updates the variables until the convergence condition is met. The final pose calculation results not only retain the geometric accuracy of the original SLAM but also incorporate the structured information of the scene, making pose estimation more stable and loop closure detection less susceptible to interference in environments with highly repetitive features, such as ring main units. The generated enhanced localization results are directly used for map updates and navigation path planning, ensuring that the inspection robot achieves continuous and accurate autonomous navigation in complex power distribution environments.

[0036] S5. During operation, the positioning health is calculated in real time. Based on the loop closure consistency index, feature matching confidence, and inertial drift trend, the positioning risk is predicted. The specific implementation is as follows: Based on the loop closure detection results generated during pose calculation, a loop closure consistency index is extracted. In practical implementation, the inspection robot periodically checks for loops between the current environment and historical keyframes during operation, i.e., determining whether it has returned to previously traversed areas. To accomplish this, the system outputs loop closure detection comparison results during the SLAM graph optimization stage, such as the feature matching degree between the current frame and candidate historical frames, and geometric consistency error. These results are converted into a numerical loop closure consistency index; a higher value indicates a more reliable match and a higher confidence level in loop closure. A significant decrease in this value usually indicates a potential drift trend in the current pose calculation.

[0037] By combining the matching confidence scores of semantic feature points, the consistency between the current scene recognition and the map model is evaluated. In addition to geometric loop closure detection, the system also utilizes semantic features generated in step S2 (such as category labels for electrical cabinets, corridors, and cable trenches) for further evaluation. This involves matching the semantic points identified in the current frame with semantic keyframes stored in the map and calculating the confidence score of each match. If the confidence score distribution is stable and the value is high, it indicates that the current scene matches the established map model. Figure 1 Conversely, if most semantic points cannot find a reliable correspondence, it indicates that there is a deviation between the current environmental perception and the map model, which is usually related to robot pose drift.

[0038] Simultaneously, the system analyzes the deviation trend between the current trajectory and the inertial-predicted trajectory using acceleration and angular velocity information output by the inertial measurement unit (IMU). To avoid relying solely on visual or laser feature matching, the system utilizes inertial information provided by the IMU for auxiliary judgment. IMU data, through integration, yields an inertial-predicted trajectory. Although this trajectory itself has accumulated errors, it is relatively stable over short periods. The system compares the SLAM-derived trajectory with the IMU-predicted trajectory in real time, including translational and rotational errors. If the difference shows a continuously increasing trend, it indicates that the SLAM module may be malfunctioning, requiring an early risk signal.

[0039] The above multidimensional indicators are weighted and fused to generate a comprehensive health index. The health index adopts a weighted fusion strategy, and the formula is as follows: H = w1 C_loop + w2 C_sem + w3 C_imu Wherein, C_loop is the consistency index for loop closure detection, C_sem is the confidence score for semantic feature matching, and C_imu is the inertial trajectory deviation; w1, w2, and w3 are dynamically adjusted weights, which are adaptively allocated based on scene lighting conditions, number of features, and sensor status.

[0040] The health threshold is updated in real time based on the historical mean and standard deviation of the indicators to achieve adaptive judgment. The health indicator is based on a weighted fusion of loop closure consistency, semantic feature matching confidence, and inertial trajectory deviation. Specifically, it includes: calculating the matching similarity of loop closure detection as the first indicator, extracting the semantic feature matching success rate of the current frame as the second indicator, and analyzing the deviation between the inertial trajectory and the SLAM trajectory as the third indicator. Through a dynamic weight allocation strategy, the weights of each indicator are adjusted according to illumination, environmental texture, and the number of features to generate a comprehensive health score, and an adaptive threshold is set for risk assessment. The trend of this indicator is used to determine whether there is a potential positioning offset risk. When the risk exceeds the preset threshold, a drift warning signal is output for subsequent relocalization module calls. Finally, the system will perform weighted fusion of data from three sources—loop closure consistency indicator, semantic feature matching confidence, and IMU deviation trend—to form a comprehensive health score H. The weighting strategy can be fixed weights or dynamically adjusted according to scene characteristics. For example, when the illumination changes significantly, the weight of visual features in the health score is reduced. When the health score is within a preset range, it indicates stable positioning. If the score continues to decline and falls below a threshold, the system will output a drift risk signal. This signal will trigger subsequent repositioning processes, such as QR code recognition and semantic keyframe correction, to prevent the robot from completely losing its positioning.

[0041] S6. When the health level is below the threshold, the system triggers relocation, corrects the pose calculation results using the cabinet's QR code identifier and keyframe semantic features, and replans the inspection path based on the corrected pose. The specific implementation is as follows: When the health index reaches the trigger condition, the repositioning process is initiated. This involves identifying the QR code on the cabinet to obtain the absolute position information corresponding to the QR code, and then aligning this position information with the current pose. The system first determines whether a repositioning operation is necessary based on the positioning health index. If the continuously monitored health value falls below a preset threshold during system operation, indicating a potential significant drift risk in the current positioning result, the repositioning process is immediately triggered. This health index is calculated by fusing multiple features such as loop closure detection consistency, semantic matching confidence, and inertial trajectory deviation, reflecting the stability of pose calculation in real time. This mechanism allows for proactive correction before positioning errors expand, rather than waiting for complete failure before recovery, thus ensuring the continuity of the robot's inspection tasks.

[0042] During the relocalization process, the system first identifies the QR code markers installed along the inspection path on the cabinets, extracts the absolute position information corresponding to the QR codes, and compares this position information with the current SLAM pose. After the relocalization process is initiated, the system first scans the pre-set QR code markers on the cabinets along the inspection path using a front-end vision sensor. Each QR code marker corresponds to a unique code, which is associated with the system's built-in map index table. After successful identification, the system decodes the QR code information and maps it to the global coordinate system to obtain the absolute position coordinates corresponding to the QR code. Subsequently, the system aligns this absolute position information with the pose obtained from the current SLAM solution, and corrects the initial solution of the current pose through rigid body transformation or pose fusion algorithms to ensure that subsequent optimization processes start from a reliable reference position.

[0043] Subsequently, pose correction is performed based on semantic feature matching of historical keyframes, combined with cabinet sequence features and corridor topology. After QR code position correction, the system further incorporates historical keyframes and semantic features for optimization. Specifically, the system calls the historical keyframe database to retrieve keyframes highly similar to the current scene features and establishes a correspondence through feature descriptor matching. Simultaneously, the system incorporates electrical cabinet sequence features and corridor topology as constraints into the graph optimization model to improve the global consistency of the matching results. This process is equivalent to further enhancing the robustness of optimization by utilizing scene semantics and structural information on top of the QR code constraints, avoiding secondary positioning errors caused by missing or misidentified QR codes.

[0044] After pose correction, the inspection path is recalculated, and the navigation target point is updated based on the corrected pose. Immediately after pose correction, the system recalculates the inspection path. The path planning module, based on the latest pose information and current task status, regenerates the optimal path from the current position to the target inspection point and verifies key nodes in the path to ensure consistency with the semantic topology graph. Subsequently, the system sends the new navigation path to the motion control module, updates the navigation target point, and clears the historical path cache caused by drift. This update process ensures that the inspection robot can restore the correct operating trajectory without terminating the task and continue executing subsequent inspection tasks.

[0045] See Figure 1After extracting image and point cloud features separately during the encoding stage, a feature alignment module is used for cross-modal fusion to form a unified feature representation, reducing registration errors caused by differences in viewpoints. The inspection robot first acquires color images and depth information from a vision sensor, and simultaneously acquires 3D point cloud data from a LiDAR scanner. To ensure the accuracy of subsequent processing, the system first extracts features from the image and point cloud separately during the encoding stage. For the image, a convolutional neural network model is used to extract two-dimensional features such as color, texture, and edges; for the point cloud, a point cloud deep learning network (such as a model based on the PointNet structure) is used to extract spatial geometric features. After the two types of feature extraction are completed, they are not simply concatenated, but fused across modally using a specially designed feature alignment module. This module maps the image features and point cloud features in a feature space, enabling them to match in a unified coordinate system, thereby forming a unified feature representation. This reduces registration errors caused by differences in viewpoints and sampling methods, providing a more reliable data foundation for subsequent semantic segmentation and topological analysis.

[0046] During semantic annotation, a dual-branch structure is employed. One branch performs pixel-level classification for 2D images, while the other branch performs point-level classification for 3D point clouds. Pixel labels are then mapped to point labels based on coordinate mapping. After cross-modal feature fusion, the system proceeds to the semantic annotation stage. This embodiment uses a dual-branch neural network structure for semantic classification. One branch processes 2D image data, classifying each pixel and outputting pixel-level semantic labels, such as labeling the area as belonging to an electrical cabinet, wall, passageway, or cable trench. The other branch processes the 3D point cloud, identifying the category of each point and obtaining point-level semantic labels, such as whether the point belongs to a cabinet structure or obstacle. After completing these two independent classification tasks, the system uses a coordinate mapping strategy to map image pixel labels to point cloud labels, i.e., associating 2D pixels and 3D point coordinates based on depth information, achieving consistent integration of semantic information. In this way, the robot can not only "know" what category a region belongs to but also accurately grasp its specific location in 3D space.

[0047] The generated semantic feature set contains object category identifiers, 3D center point coordinates, bounding boxes, and topological attribute labels in the form of a data structure, providing input for subsequent topological mapping. After the above processing, the system organizes all identified object information into a structured semantic feature set. In this embodiment, the set adopts a standardized data structure, and each object entry includes a category identifier (e.g., electrical cabinet, wall), 3D center point coordinates, bounding box dimensions, and topological attribute labels. This information not only describes the object's category and location but also provides a foundation for subsequent semantic topological mapping, such as identifying which nodes belong to the same corridor and whether the spatial relationships between nodes conform to connection rules. The final generated semantic feature set will serve as input to the topology modeling module and provide key constraints in the subsequent SLAM optimization process, further improving localization robustness and map accuracy.

[0048] See Figure 1 After the QR code is decoded, the unique identifier obtained is queried against a preset map index table and bound to a predefined global coordinate system to obtain the absolute position coordinates. In specific implementation, QR code labels are pre-attached to the surface of cabinets along the inspection path. Each QR code corresponds to a unique code, which corresponds one-to-one with the map index table inside the system. During the robot's operation, when the health index triggers the relocalization mechanism, the camera is controlled to photograph the QR code of the target cabinet, and the decoding algorithm is called to parse the unique identifier of the QR code. This identifier is not directly used for pose calculation, but is used as a key index value to quickly query the map index table stored in the system to obtain the preset coordinate information of the corresponding cabinet in the global coordinate system. After coordinate transformation, this absolute position coordinate is unified into the robot's current SLAM coordinate system and serves as the initial reference point for relocalization calculation, ensuring that the pose calculation can quickly converge to a globally consistent position.

[0049] During pose correction, the location information provided by the QR code is first added to the graph optimization model as a strong constraint factor, and then the keyframe matching result based on semantic features is added to the optimization as a weak constraint factor to ensure computational stability. Pose correction is not simply replacing the current pose, but rather readjusting the current keyframe nodes through the graph optimization model. When building the factor graph, the system adds a new constraint factor to represent the absolute position provided by the QR code. This factor has a high weight and is used as a "strong constraint" to directly affect the global position calculation of the current keyframe. At the same time, the system searches for historical keyframes with high semantic similarity in the local keyframe library. These keyframes contain feature information such as electrical cabinets and channels. The relative poses of these keyframes and the current frame are calculated through semantic feature matching and introduced into the optimization model as "weak constraint" factors. This combination of strong and weak constraints allows the optimization process to quickly converge to the global position provided by the QR code while maintaining the continuity and consistency of the trajectory using local semantic information, avoiding jumps caused by single-point positioning.

[0050] A priority matching strategy is adopted for cabinet sequence features and topological relationships. When the QR code is unavailable, and after a preset number of consecutive QR code recognition failures, the relocation strategy switches to a cabinet sequence-based estimation mode, predicting the current position by combining the inspection direction and historical travel path. If the health index still has not recovered and QR code recognition continues to fail, iterative estimation based on the topology structure is activated, updating the pose through node connectivity constraints until the QR code becomes available again or manual intervention is needed. The current position is then estimated by combining the cabinet arrangement order with the known path topology structure, and the factor connectivity relationships in the SLAM graph are iteratively updated. In practical applications, QR code recognition may be temporarily unavailable due to factors such as dust obstruction and camera angle. In this case, the system automatically switches to a backup relocation strategy. This strategy is based on the sequence features and preset topological relationships of the electrical cabinets in the scene. The inspection path is usually linear or branching, and the arrangement order between cabinets has strong regularity. The system uses the cabinet number confirmed at the previous location, combined with the inspection direction and historical mileage data, to predict the cabinet position where the robot should currently be. Simultaneously, the system gradually adjusts the calculated value of the current pose based on the constraint relationships between nodes in the topology graph to ensure that path connectivity is not disrupted. During this process, the factor connectivity relationships of the SLAM graph are dynamically updated, and the newly calculated pose is added as a temporary node to the optimization until QR code recognition is restored or other strong constraint mechanisms are triggered, completing position correction and maintaining trajectory continuity.

[0051] Example 2: Figure 2 The present invention provides an autonomous navigation system for a ring main unit inspection robot based on SLAM, comprising: The multi-source data acquisition module, including lidar, vision sensor and inertial measurement unit, is used to acquire 3D point cloud, color image, depth information and inertial data, and complete time synchronization through a unified timestamp mechanism to output multi-source sensing data stream; The scene semantic parsing module is used to receive synchronous data, perform image feature and point cloud feature extraction, perform pixel-level semantic annotation on the image, perform point-level semantic annotation on the point cloud, and generate a semantic feature set containing object category and three-dimensional coordinates. The semantic topology modeling module is used to construct a topology graph based on semantic features, abstracting cabinets, corridors, etc. as nodes, paths as edges, and establishing distance constraints, connectivity constraints, and directional constraints between nodes. The topology constraint optimization module is used to introduce the topology graph as a constraint into the SLAM backend optimization model, and combine laser, visual odometry calculation and inertial prior to output the optimized pose calculation result. The positioning health assessment module is used to generate health indicators based on loop closure detection consistency, semantic feature matching confidence, and inertial trajectory deviation, and to trigger a relocation signal when the indicator is below a threshold. The relocation execution module is used to correct the pose, update the SLAM graph factor connections, and replan the inspection path when triggered, by recognizing the cabinet QR code and combining semantic keyframes and topological features.

[0052] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0053] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0054] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0055] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0056] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0057] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0058] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0059] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A SLAM-based autonomous navigation method for a ring main unit inspection robot, characterized in that, Comprising: S1, collecting point cloud data, image data and inertial data of the ring network cabinet inspection environment by laser radar, visual sensor and inertial measurement unit, and performing time synchronization processing; S2, performing deep learning processing on the collected data, identifying the outlines and spatial positions of electrical cabinets, corridor paths, wall surfaces and cable trenches, and generating a semantic feature set containing category labels and three-dimensional coordinates; S3, constructing a semantic topological map according to the semantic features, abstracting the cabinet nodes and corridor paths into a topological structure, and establishing the distance constraints and path connection relationships between nodes; S4, introducing the semantic topological constraints in the SLAM optimization process, combining the inertial measurement results to constrain and optimize the pose solution, and generating an enhanced positioning result; S5, calculating the positioning health degree in real time during operation, comprehensively considering the loop consistency index, feature matching confidence and inertial drift trend to predict whether there is a risk of displacement; S6, when the health degree is lower than the threshold, the system triggers repositioning, corrects the pose solution result using cabinet two-dimensional code identification and key frame semantic features, and replans the inspection path according to the corrected pose.

2. The SLAM-based autonomous navigation method for ring main unit inspection robots according to claim 1, characterized in that, S1 includes: Performing all-around or multi-plane scanning through the laser radar configured on the inspection robot to obtain three-dimensional point cloud reflecting the spatial geometric form; Collecting scene data containing color images and depth information through the visual sensor installed on the inspection robot; Collecting acceleration and angular velocity information in real time through the inertial measurement unit to form inertial data describing the motion state of the robot; Synchronizing the laser, visual and inertial data through a unified timestamp mechanism and hardware clock to output time-consistent multi-source perception data stream.

3. The SLAM-based navigation method for a ring main unit inspection robot according to claim 1, wherein S2 Including: Inputting the image data and point cloud data processed by time synchronization into a deep learning model, performing convolution feature extraction and target classification on the image data, and performing spatial feature extraction and category identification on the point cloud data; Performing pixel-level semantic labeling and point-level semantic labeling on the extracted results to identify the outlines and spatial positions of electrical cabinets, corridor paths, wall surfaces, cable trenches and other static structures; Combining the three-dimensional spatial coordinates of the target objects calculated based on depth information or laser point cloud to generate a semantic feature set containing category labels and spatial position information.

4. The SLAM-based navigation method for a ring main unit inspection robot according to claim 1, wherein S3 Including: According to the semantic features, extracting the three-dimensional center coordinates and category attributes of the identified electrical cabinets, corridor paths, wall surfaces and cable trenches as basic information of the topological nodes, judging whether the connectivity condition is met according to the spatial distance, relative orientation and path width between nodes and other scene constraints, and if the condition is met, abstracting the passable path between the nodes into an edge in the topological graph; When constructing the topological relationship, determining the connection relationship between adjacent nodes based on the cabinet arrangement rules and scene geometric constraints, and labeling the node category attributes and spatial orientation information; Constructing a topological structure containing distance constraints, connectivity constraints and directionality constraints to form a topological graph that can be used to constrain the pose solution.

5. The SLAM-based navigation method for a ring main unit inspection robot according to claim 1, wherein S4 Including: Introducing the node positions, connection relationships and constraint conditions in the topological graph into the SLAM backend optimization model, taking the key frame poses calculated by the visual and laser odometry as optimization variables; In the process of pose solving, the key frame poses generated by visual and laser odometry are taken as optimization variables, and the topological constraints and motion prior information provided by the inertial measurement unit are superimposed to construct a joint optimization objective function; By iteratively solving the optimization model, the pose solving result that meets the semantic topological constraints and motion continuity requirements is output.

6. The SLAM-based navigation method for a ring main unit inspection robot according to claim 1, wherein S5 It includes: Based on the loop closure detection results generated in the pose solving process, the loop consistency index is extracted; Combined with the matching confidence of semantic feature points, the consistency between the current scene recognition and the map model is evaluated; At the same time, the acceleration and angular velocity information output by the inertial measurement unit is used to analyze the deviation trend between the current trajectory and the inertial calculated trajectory; The above multi-dimensional indicators are fused through weighted fusion to generate a comprehensive health degree index, and the trend of the index is used to determine whether there is a potential positioning deviation risk, and when the risk exceeds the preset threshold, a drift warning signal is output for subsequent repositioning module calling.

7. The SLAM-based navigation method for a ring main unit inspection robot according to claim 1, wherein S6 It includes: When the health degree index reaches the trigger condition, the repositioning process is started, the absolute position information corresponding to the two-dimensional code is obtained by recognizing the two-dimensional code on the cabinet, and the position information is aligned with the current pose for calculation; During the repositioning process, first, the two-dimensional code on the cabinet along the inspection path is recognized, the absolute position information corresponding to the two-dimensional code is extracted, and the position information is compared with the current SLAM pose; Subsequently, based on the historical key frame semantic feature matching, the pose correction is performed based on the cabinet sequence features and the corridor topological relationship; After completing the pose correction, the inspection path is recalculated, and the navigation target point is updated according to the corrected pose.

8. The SLAM-based navigation method for a ring main unit inspection robot according to claim 3, characterized in that, It includes: After the image and point cloud features are extracted in the encoding stage, a cross-modal fusion is performed using a feature alignment module to form a unified feature representation, which is used to reduce the registration error caused by the difference in viewing angle; When performing semantic labeling, a double-branch structure is used, one branch implements pixel-level classification for two-dimensional images, and the other branch implements point-level classification for three-dimensional point clouds, and then the pixel labels and point labels are mapped based on coordinates; The generated semantic feature set includes object class identification, three-dimensional center point coordinates, bounding box, and topological attribute label in the form of data structure, providing input for subsequent topological mapping.

9. The SLAM-based navigation method for a ring main unit inspection robot according to claim 7, characterized in that, It includes: After decoding the two-dimensional code, the unique identification code obtained by decoding is queried with the preset map index table, and is bound to the predefined global coordinate system to obtain the absolute position coordinates; During the pose correction process, first, the position information provided by the two-dimensional code is added as a strong constraint factor to the graph optimization model, and then the key frame matching results based on semantic features are added as weak constraint factors to the optimization to ensure the stability of the calculation; The priority matching strategy is used for the cabinet sequence features and topological relationship, when the two-dimensional code is not available, the current position is calculated based on the cabinet arrangement sequence and the known path topological structure, and the factor connection relationship in the SLAM graph is iteratively updated.

10. A SLAM-based autonomous navigation system for a ring main unit inspection robot, for implementing the SLAM-based autonomous navigation method of any one of claims 1-9, characterized in that, It includes: A multi-source data acquisition module including a laser radar, a visual sensor and an inertial measurement unit is used to acquire three-dimensional point clouds, color images, depth information and inertial data, and time synchronization is completed through a unified timestamp mechanism to output multi-source perception data stream; The scene semantic analysis module is configured to receive the synchronization data, perform image feature and point cloud feature extraction, perform pixel-level semantic labeling on the image, perform point-level semantic labeling on the point cloud, and generate a semantic feature set containing object categories and three-dimensional coordinates. The semantic topology modeling module is configured to construct a topology graph according to the semantic features, abstract the cabinet body and the corridor as nodes, abstract the path as an edge, and establish distance constraints, connectivity constraints and direction constraints between the nodes. The topology constraint optimization module is configured to introduce the topology graph as a constraint into a SLAM backend optimization model, combine laser and visual odometry calculation and inertial priori, and output an optimized pose calculation result. The positioning health degree evaluation module is configured to generate a health degree index based on loop detection consistency, semantic feature matching confidence and inertial trajectory deviation, and trigger a repositioning signal when the index is lower than a threshold. The repositioning execution module is configured to, under the triggering condition, correct the pose by identifying the cabinet two-dimensional code and combining the semantic key frame and the topology feature, update the SLAM graph factor connection, and re-plan the inspection path.

Images