Heuristic search method for maintenance scene of inner bore of hydro-generator

Abstract

The invention provides a heuristic search method for a hydro-generator bore maintenance scene, and the method comprises the steps: collecting RGB-D image data of a hydro-generator bore through an inspection robot, and achieving the recognition of key parts through a convolutional neural network; calculating a three-dimensional space coordinate of the key component based on binocular vision and depth information, and converting the three-dimensional space coordinate into a world coordinate system with an origin of a robot coordinate system as a reference; according to the unified template, converting the spatial distribution, the structural relationship and the size parameters of the component into a structured semantic text, inputting the structured semantic text and a preset topological knowledge template into a large language model, and reasoning and outputting a potential spatial region and confidence of the target component; according to the confidence coefficient, dynamically selecting a next optimal view point (NBV) path plan adopting semantic enhancement or returning to a pure geometric NBV, adjusting task correlation weight, and generating a candidate view point set; and selecting an optimal view angle according to an NBV optimization result, driving the robot to autonomously generate and execute a patrol path, and completing positioning and data acquisition of the target component.

Description

Technical Field

[0001] The present invention relates to the field of operation and maintenance of large-scale power generation equipment, and specifically to a heuristic search strategy based on the combination of a large language model (LLM) and next-best view (NBV) path planning. The strategy is suitable for the maintenance operations of a hydraulic generator bore inspection robot and belongs to the field of intelligent inspection and navigation technology for power generation equipment. Background Art

[0002] As the capacity of a single hydro-turbine generator unit continues to increase, the number, types, and depth of components that need to be inspected have also increased, resulting in a more complex internal bore structure with a large number of obstructions and areas that are not directly visible. Currently, regular inspections of the internal bore still rely mainly on manual experience or a small number of fixed sensors, which have the following shortcomings: (1) Manual close-range operations are risky and inefficient; (2) Fixed sensors have a limited field of view, which easily creates monitoring blind spots; and (3) There is a lack of real-time three-dimensional positioning methods, making it difficult to accurately correspond to CAD coordinates or specific structural parts.

[0003] While existing mobile robots can replace manual inspections to a certain extent, their path planning often relies on traversal or geometric heuristic algorithms, such as the NBV (Next-Best View) planning algorithm. These algorithms are unable to distinguish target parts from unrelated areas, resulting in numerous invalid movements and large positioning errors. The NBV algorithm, currently the mainstream approach for autonomous exploration tasks, primarily selects paths and viewpoints based on geometric visibility gain, lacking understanding and guidance of the semantic characteristics of inspection targets. Therefore, its adaptability and effectiveness are significantly limited in industrial scenarios with complex structures and severe occlusion, particularly within turbine bores. Large language models (LLMs), on the other hand, have demonstrated excellent natural language reasoning capabilities. However, when directly applied to highly structured industrial scenarios such as turbine bores, the lack of structural knowledge, such as equipment topology, often leads to distorted reasoning results, making it difficult to guide actual inspection tasks.

[0004] Therefore, there is an urgent need for a heuristic search strategy that can effectively integrate the spatial topological relationship data of specific scenarios with the reasoning capabilities of large language models and deeply combine it with traditional exploration and planning algorithms, so as to achieve target positioning and path optimization in complex occlusion environments, and further improve the efficiency and intelligence level of inspection robots in the maintenance operations of the inner bore of hydro-turbine generators. Summary of the Invention

[0005] This paper addresses the complex, heavily occluded, and inefficient manual maintenance of hydro-turbine generators. It proposes a next-best-view (NBV) heuristic search method driven by a large language model (LLM) enhanced with a "fixed-topology knowledge template." Through semantic perception, knowledge template injection and spatial reasoning, semantically weighted NBV path planning, closed-loop execution, and incremental learning, this method guides robots or maintenance systems to efficiently locate key components and significantly shorten maintenance paths. This method combines the semantic reasoning advantages of a large language model with the spatial exploration capabilities of NBV, achieving efficient target search and path optimization in occluded environments, effectively enhancing the automation and intelligence of hydro-turbine generator maintenance operations.

[0006] This method uses multi-source RGB-D visual perception and topological knowledge templates to drive LLM reasoning of regions of interest. It then optimizes the NBV scoring function using confidence weights to dynamically output the optimal search path. The system supports dynamic switching of confidence levels and automatically falls back to traditional geometric NBV when inference confidence is insufficient, ensuring the robustness and adaptability of the method. Closed-loop feedback continuously optimizes the knowledge base, enabling efficient and stable maintenance path planning and autonomous search.

[0007] In order to achieve the above technical features, the purpose of the present invention is to provide a heuristic search method for the maintenance scenario of the inner bore of a hydro-generator, comprising the following steps: Step 1: The inspection robot collects RGB-D image data of the turbine generator's inner chamber and uses a convolutional neural network-based object detection algorithm to identify key components, which serves as the perception input for the subsequent heuristic search strategy. Step 2: Calculate the three-dimensional spatial coordinates of the key components based on binocular vision and depth information, and convert them into a world coordinate system with the origin of the inspection system's own coordinate system as a reference, providing a spatial positioning basis for the heuristic search strategy; Step 3: Convert component spatial distribution, structural relationships, and dimensional parameters into structured semantic text based on a unified template; Step 4: Input the structured semantic text and the preset topological knowledge template into the large language model (LLM), and output the target component potential space area by LLM reasoning and confidence ; Step 5, when the confidence When the confidence level is lower than the preset threshold, the system falls back to the next best view (NBV) planning driven by pure geometric information; When the value is not lower than the threshold, the following sub-steps are executed: Step 5.1, and Serialize to binary file; Step 5.2: Using the binary file as the constraint input of the NBV path planning algorithm, dynamically adjusting the task relevance weights, and generating a set of candidate viewpoints guided by high-value areas; Step 6: Select the optimal viewing angle based on the NBV optimization results, and drive the robot to autonomously generate and execute the inspection path to complete the positioning and data collection of the target component.

[0008] Preferably, the key components include: a rotor, a stator, an excitation system, and anchor nuts.

[0009] Preferably, the target detection algorithm in step 1 is set within a range where the confidence threshold is greater than 0.6, and semantic recognition is performed by combining continuously collected RGB-D image frames with the real-time posture information of the robot.

[0010] Preferably, step 2 uses a camera extrinsic calibration matrix and a binocular matching algorithm to map pixel coordinates into three-dimensional coordinates and fuse multi-source perception information.

[0011] Preferably, the semantic text generated in step 3 includes at least: polar coordinates or Cartesian coordinates of the component center, relative radial distance and height difference from the reference structure, and function dependency description parameters.

[0012] Preferably, the preset topology knowledge template includes at least the following quantifiable spatial relationship information, including: the rotor and stator are coaxial and at the same height; the excitation system is located 0.5m above the rotor; the anchor nuts evenly surround the excitation system, with a radius of 5m and a height of 0.5m lower.

[0013] Preferably, the NBV path planning algorithm in step 5.2 adopts a dynamic task relevance weight factor , and adjust the sampling probability of candidate viewpoints in real time, so that the robot can give priority to exploring the target area when there is severe occlusion and limited space; It is the intersection of the field of view and the region of interest; is the task relevance amplification factor.

[0014] Preferably, the robot motion control process in step 6 includes obstacle avoidance control, path replanning and abnormal target relocation mechanisms, which are used to automatically adjust the inspection strategy when encountering obstacles or recognition errors during the inspection process.

[0015] Preferably, during the inspection process, the robot dynamically switches the imaging parameters and lighting intensity according to the NBV planning results and LLM reasoning results to adapt to the complex lighting conditions of the bore and ensure the observation quality of key components.

[0016] Preferably, the NBV planner calculates the candidate view angles Calculated composite score Satisfies the following formula: ; in, is information gain; is the path cost; It is the intersection of the field of view and the region of interest; is the task relevance amplification factor; The confidence of the region of interest output by the large language model; is the confidence weight coefficient.

[0017] The present invention has the following beneficial effects: 1. This invention uses a convolutional neural network (such as YOLOv8) fine-tuned with scene-specific data, combined with RGB-D depth information, to perform real-time detection and segmentation of core components such as the rotor, stator, excitation system, and anchor nuts. The pixel-level mask is then mapped to a unified three-dimensional coordinate system to generate a semantic point cloud containing attributes such as coordinates, categories, and dimensions, opening up a "perception-to-decision" data interface for subsequent reasoning and path planning.

[0018] 2. Unlike approaches that require graph database retrieval or extensive fine-tuning, this invention organizes common spatial relationships, functional dependencies, and dimensional specifications within a generator into no more than 30 fixed statements (templates), which are then injected into the large language model at system startup. This significantly shortens prompt length and enables the LLM to possess spatial reasoning capabilities tailored to industrial scenarios without any fine-tuning.

[0019] 3. This paper designs the NBV scoring formula of "confidence-IoU dual weight": ; in, is information gain; is the path cost; Measure the overlap between the candidate perspective field of view and the area of ​​interest; is the task relevance amplification factor; The confidence of the region of interest output by the large language model; is the confidence weight coefficient.

[0020] This formula allows NBV to consider both traditional geometric information gain and dynamically integrate semantic overlap and large model confidence, thereby automatically concentrating sampling on high-value areas.

[0021] 4. The present invention combines the interest region R with the confidence Serialized into a 32-bit floating point array and written into a binary file, the NBV planner can read it in milliseconds; When it falls below the set threshold, the system immediately disables the semantic weight and falls back to pure geometric NBV, ensuring stable planning and not being affected by occasional hallucinations of large models.

[0022] 5. After observing the target from the optimal perspective, the robot will automatically write the observed coordinates and LLM prediction deviations back to the template library, gradually correcting and enriching the topological knowledge, realizing online self-learning of semantic knowledge, and facilitating rapid migration between different models.

[0023] In summary, this invention requires no additional fine-tuning of large language models and can be deployed solely using lightweight prompts. It also provides anti-hallucination fallback and knowledge increment mechanisms, resulting in high reliability and scalability. It is suitable for autonomous inspection of the inner cavities of hydro-turbine generators and other large, shielded electromechanical equipment (such as steam turbines and wind turbine nacelles), significantly reducing manual risks and downtime, and has significant potential for engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be further described below with reference to the accompanying drawings and examples.

[0025] Figure 1 This is a schematic diagram of knowledge injection into a large language model. It describes the process of injecting structured knowledge such as the spatial relationships, functional dependencies, and specification descriptions of key components of industrial equipment into the LLM through a fixed topology template, and shows the mapping relationship between the knowledge template, relationship matrix, and reasoning process.

[0026] Figure 2 This is the overall flow chart of the method of the present invention, showing the complete process from environmental perception, knowledge template injection and spatial reasoning, NBV path planning based on semantic weights, to closed-loop execution and incremental knowledge learning.

[0027] Figure 3 These are screenshots of the simulation experiment and inspection results of the present invention in a typical hydro-generator bore scenario, demonstrating the path planning process and target positioning effect obtained by using the heuristic search strategy of the present invention in a complex occlusion environment. DETAILED DESCRIPTION

[0028] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0029] Example 1: A heuristic search method for the maintenance scenario of the inner bore of a hydro-generator includes the following steps: In step 1, the inspection robot collects RGB-D image data of the turbine generator's inner bore and uses a convolutional neural network-based object detection algorithm to identify key components such as the rotor, stator, excitation system, and anchor nuts. This data serves as the perception input for the subsequent heuristic search strategy. Step 2: Calculate the three-dimensional spatial coordinates of the key components based on binocular vision and depth information, and convert them into a world coordinate system with the origin of the inspection system's own coordinate system as a reference, providing a spatial positioning basis for the heuristic search strategy; Step 3: Convert component spatial distribution, structural relationships, and dimensional parameters into structured semantic text based on a unified template; Step 4: Input the structured semantic text and the preset topological knowledge template into the large language model (LLM), and output the target component potential space area by LLM reasoning and confidence ; Step 5, when the confidence When the confidence level is lower than the preset threshold, the system falls back to the next best view (NBV) planning driven by pure geometric information; When the value is not lower than the threshold, the following sub-steps are executed: Step 5.1, and Serialize to binary file; Step 5.2: Using the binary file as the constraint input of the NBV path planning algorithm, dynamically adjusting the task relevance weights, and generating a set of candidate viewpoints guided by high-value areas; Step 6: Select the optimal viewing angle based on the NBV optimization results, and drive the robot to autonomously generate and execute the inspection path to complete the positioning and data collection of the target component.

[0030] Furthermore, the target detection algorithm in step 1 is set within a range where the confidence threshold is greater than 0.6, and semantic recognition is performed by combining continuously collected RGB-D image frames with the real-time posture information of the robot.

[0031] Furthermore, step 2 uses the camera extrinsic calibration matrix and binocular matching algorithm to map pixel coordinates into three-dimensional coordinates and fuse multi-source perception information.

[0032] Furthermore, the semantic text generated in step 3 includes at least: polar coordinates or Cartesian coordinates of the component center, relative radial distance and height difference from the reference structure, and function dependency description parameters.

[0033] Furthermore, the preset topology knowledge template includes at least the following quantifiable spatial relationship information, including: the rotor and stator are coaxial and at the same height; the excitation system is located 0.5m above the rotor; the anchor nuts evenly surround the excitation system, with a radius of 5m and a height of 0.5m lower.

[0034] Furthermore, the NBV path planning algorithm in step 5.2 adopts a dynamic task relevance weight factor , and adjusts the sampling probability of candidate viewpoints in real time, so that the robot can give priority to exploring the target area when the occlusion is severe and the space is limited.

[0035] Furthermore, the robot motion control process in step 6 includes obstacle avoidance control, path replanning and abnormal target relocation mechanisms, which are used to automatically adjust the inspection strategy when encountering obstacles or recognition errors during the inspection process.

[0036] Furthermore, during the inspection process, the robot dynamically switches imaging parameters and lighting intensity based on the NBV planning results and LLM reasoning results to adapt to the complex lighting conditions in the bore and ensure the observation quality of key components.

[0037] Furthermore, the NBV planner calculates candidate viewpoints Calculated composite score Satisfies the following formula: ; in, is information gain; is the path cost; It is the intersection of the field of view and the region of interest; is the task relevance amplification factor; The confidence of the region of interest output by the large language model; is the confidence weight coefficient.

[0038] Example 2: In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is described in further detail below.

[0039] 1. Hardware and software platform: This invention utilizes a quadruped robot platform with autonomous locomotion and obstacle-crossing capabilities as the mainstay of the inspection system. A commercially available quadruped robot, such as the Jueying X30, is preferred, as it possesses the ability to maintain stable movement and maintain posture in narrow, complex industrial environments, such as the inner chamber of a hydroelectric generator, where there are stairs and other obstacles.

[0040] The quadruped robot is equipped with a robotic arm with six or more degrees of freedom. A high-resolution RGB-D camera can be mounted at the end of the arm, enabling flexible, multi-angle observation of key components. The arm supports autonomous planning, precise manipulation, and real-time posture adjustment, meeting the needs of localized, close-range observation and component identification in complex spatial environments.

[0041] The robot itself integrates an industrial-grade edge computing unit with high computing power and low power consumption, enabling on-site environmental perception, data preprocessing, path planning, and motion control. The computing unit interacts with cloud servers via high-speed wireless communications (such as 5G and WiFi 6), uploading scene descriptions, 3D point clouds, and images from the inspection process to the cloud in real time.

[0042] The cloud server is deployed with a large-scale pre-trained language model (LLM), which is responsible for receiving inputs such as environmental semantic information, three-dimensional coordinates of key components, and knowledge templates from the robot, performing spatial reasoning and area of ​​interest identification, and returning the reasoning results and confidence weights to the robot locally for optimizing NBV path planning and task decision-making.

[0043] The sensors used in the system include at least: a) High-resolution RGB-D camera (for 3D environment perception and semantic segmentation); b) IMU inertial measurement unit (to realize robot body posture perception and trajectory calculation); c) LiDAR (assisted mapping, positioning, and dynamic obstacle avoidance).

[0044] Based on the ROS 1 system architecture, the software platform integrates deep learning object detection models, semantic segmentation networks, edge point cloud processing, and NBV path planning algorithms. It also provides a secure communication interface with cloud-based inference services, enabling local-cloud collaboration. The overall system supports remote monitoring, task dispatching, and dynamic parameter tuning, facilitating large-scale industrial deployment and application.

[0045] 2. System overall structure and functional modules: The system consists of the following modules: a) Environmental Perception Module: This module collects and fuses multi-source sensor data. It uses a convolutional neural network (such as YOLOv8) fine-tuned for industrial scenarios to detect and segment key components such as the rotor, stator, excitation system, and anchor nuts in real time. It then outputs pixel-level masks, maps them to a unified 3D coordinate system, and generates a semantic point cloud with attribute labels.

[0046] b) Knowledge Template Injection and Spatial Reasoning Module: This module uses fixed natural language templates to input pre-built information such as component spatial relationships, functional dependencies, and specification descriptions into the Large Language Model (LLM). This knowledge template is injected through the Prompt project to efficiently express key spatial topological relationships. The LLM combines the real-time semantic point cloud with the injected knowledge to output high-confidence target regions of interest and their spatial distribution.

[0047] c) Semantic Weighted NBV Path Planning Module: Based on the ROI and confidence scores output by the LLM, the module calculates the NBV scoring function (taking into account information gain, path cost, IoU overlap between the ROI and candidate viewpoints, and LLM confidence weights). It then outputs the optimal movement path and observation pose, enabling prioritized exploration of the target area. When the confidence score falls below a set threshold, the module automatically falls back to pure geometric NBV to ensure algorithm robustness.

[0048] d) Closed-Loop Execution and Incremental Knowledge Learning Module: The robot navigates to candidate viewpoints based on the optimal path, acquires new observation information, and verifies the accuracy of LLM reasoning in real time. If errors or new spatial relationships are discovered, the corrected information is written back to the knowledge template library, enabling online correction and incremental accumulation of knowledge, supporting adaptive optimization and cross-machine migration.

[0049] During system operation, these functional modules collaborate with each other to achieve efficient and autonomous inspection of the turbine generator's bore in complex obstructed environments. To ensure the robot's adaptability to specific scenarios and its ability to transfer knowledge, this paper further proposes a method for constructing and injecting fixed topology knowledge templates. The detailed process is described in the following section.

[0050] 3. Fixed topology knowledge template injection process (Attachment Figure 1 ): To achieve efficient spatial reasoning of large language models in complex industrial scenarios, this paper proposes the following five-step knowledge template injection process: a) Based on the actual structure of the turbine generator's inner chamber, select typical components such as the excitation system, anchor nuts, stator windings, and upper frame baffles. Sorting out key attributes such as their spatial distribution, quantity, height difference, and functional dependencies, and constructing an entity relationship network. For example, using the excitation system as the reference point, determine the number, distribution radius, and relative height of the anchor nuts.

[0051] b) Use standardized, machine-parseable natural language templates to clearly describe the relationships between the above components. For example, "The height of the excitation system is 1 meter, and 10 anchor nuts are evenly distributed on a circle with a radius of 5 meters and a height 0.5 meters lower than the excitation system." The template covers multi-dimensional features such as spatial coordinates, geometric distribution, and connection relationships.

[0052] c) To reduce the reasoning burden of large models and avoid "hallucination" interference, all components are divided into several independent groups based on space and function. Each group only includes a small number of closely related and interrelated components. For example, the excitation system, anchor nuts, and stator windings are in one group, while the stator windings and upper frame baffle are in another group. The internal relationships within each group are described in detail, and the connections between groups are reduced to avoid complex cross-group logic and chain dependencies. Subsequently, using the Prompt project, templates are injected into each group one by one to build a structured knowledge base. This enables the LLM to robustly perform spatial reasoning within the group and reduce the risk of misjudgment.

[0053] d) During the actual inspection process, the system can present real-time intra-group spatial reasoning tasks to the LLM (e.g., "Based on the excitation system at (0,0,1) m, estimate the coordinates of the anchor nuts?"). The LLM uses intra-group templates and simple geometric relationships to perform reasoning and output structured spatial coordinates. These results are automatically compared with on-site point cloud perception data, assisting NBV path planning and robot action decision-making.

[0054] e) During robot inspections, if the actual observation results do not match the inference output, the system can immediately correct the template entries in the relevant group, feed the new data back into the knowledge base, continuously optimize the spatial knowledge structure, and support migration across different devices / models.

[0055] 4. Task Execution Process (Attachment Figure 2 ): The task execution process of the present invention includes five major steps: perception input, intelligent reasoning, NBVP path decision, execution action and closed-loop correction. Each module works together to ensure efficient autonomous inspection in complex occlusion scenarios. The specific process is as follows: a) After receiving navigation and task instructions, the inspection robot first enters the perception input phase, collecting environmental data through multi-source sensors (including high-resolution RGB-D cameras and LiDAR). This sensor data is processed by an object detection model and a camera projection model to achieve real-time detection and spatial positioning of key target components within the scene.

[0056] b) The environmental perception module structures the categories and three-dimensional attributes of detected components and inputs them into a knowledge template library. The system automatically determines whether relevant objects are detected. If so, it triggers a structured thinking dimension prompt. The Large Language Model (LLM) performs spatial reasoning and outputs the estimated location and confidence level of the target work area. After successful LLM reasoning, high-value areas of interest are spatially annotated to provide prior constraints for subsequent path optimization.

[0057] c) The action execution module analyzes the current scene coverage based on the existing knowledge map and ROI. It generates a set of candidate viewpoints using methods such as the RRT tree. It then calculates and selects the optimal next observation viewpoint by combining information gain, path cost, the overlap (IoU) between the ROI and the candidate viewpoints, and the confidence weights output by the LLM. The robot then moves its arm or sensor to the target area along the planned path.

[0058] d) After the robot reaches the target area, it collects new observation data in real time to determine whether the work target has been fully observed. If the area of ​​interest is not fully covered or the LLM confidence level decreases, the knowledge map is reanalyzed and the viewpoint generation, path optimization, and action execution steps are repeated until all work targets are efficiently observed and marked.

[0059] e) After the inspection task is completed, the system compares the actual observation results with the LLM inference results. If errors are found in the location of the area of ​​interest or target identification, the relevant knowledge template is automatically corrected and the corrected data is written back to the knowledge base, enabling dynamic self-learning and continuous optimization of semantic knowledge. This corrected knowledge can be transferred and reused in subsequent tasks or other devices, improving the long-term adaptability and intelligence of the system.

[0060] (Note: The flowchart primarily illustrates the main processes of perception, reasoning, and decision-making for autonomous inspection tasks. The knowledge correction and incremental learning mechanisms, as system extensions, can be dynamically triggered based on observation errors during implementation, enhancing the system's adaptability and long-term intelligence.) V. Application Examples and Effect Verification (Attachment Figure 3 ): In a simulation environment (based on the Gazebo platform), the proposed method was applied to the internal inspection of a typical hydro-turbine generator. Results showed that compared to the traditional NBV algorithm, the proposed method reduced the total path length by approximately 49.8% for the same task, significantly reducing the number of invalid moves and achieving efficient and rapid positioning of obstructed components.

[0061] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A heuristic search method for the maintenance scenario of the inner bore of a hydro-generator, characterized in that: The steps include: Step 1: The inspection robot collects RGB-D image data of the turbine generator's inner chamber and uses a convolutional neural network-based object detection algorithm to identify key components, which serves as the perception input for the subsequent heuristic search strategy. Step 2: Calculate the three-dimensional spatial coordinates of the key components based on binocular vision and depth information, and convert them into a world coordinate system with the origin of the inspection system's own coordinate system as a reference, providing a spatial positioning basis for the heuristic search strategy; Step 3: Convert component spatial distribution, structural relationships, and dimensional parameters into structured semantic text based on a unified template; Step 4: Input the structured semantic text and the preset topological knowledge template into the large language model (LLM), and output the target component potential space area by LLM reasoning and confidence ; Step 5, when the confidence When the confidence level is lower than the preset threshold, the system falls back to the next best view (NBV) planning driven by pure geometric information; When the value is not lower than the threshold, the following sub-steps are executed: Step 5.1, and Serialize to binary file; Step 5.2: Using the binary file as the constraint input of the NBV path planning algorithm, dynamically adjusting the task relevance weights, and generating a set of candidate viewpoints guided by high-value areas; Step 6: Select the optimal viewing angle based on the NBV optimization results, and drive the robot to autonomously generate and execute the inspection path to complete the positioning and data collection of the target component.

2. The heuristic search method for the hydro-generator inner bore maintenance scenario according to claim 1 is characterized by: The key components include: a rotor, a stator, an excitation system, and a foundation nut.

3. The heuristic search method for the maintenance scenario of the inner bore of a hydro-generator according to claim 1 is characterized by: The target detection algorithm described in step 1 is set within the range of a confidence threshold greater than 0.6, and semantic recognition is performed by continuously acquiring RGB-D image frames combined with the real-time posture information of the robot.

4. The heuristic search method for the hydro-generator inner bore maintenance scenario according to claim 1 is characterized by: Step 2 uses the camera extrinsic calibration matrix and binocular matching algorithm to map pixel coordinates into three-dimensional coordinates and fuse multi-source perception information.

5. The heuristic search method for the hydro-generator inner bore maintenance scenario according to claim 1 is characterized by: The semantic text generated in step 3 includes at least: polar coordinates or Cartesian coordinates of the component center, relative radial distance and height difference from the reference structure, and function dependency description parameters.

6. The heuristic search method for the hydro-generator inner bore maintenance scenario according to claim 1 is characterized by: The preset topology knowledge template includes at least the following quantifiable spatial relationship information, including: the rotor and the stator are coaxial and at the same height; the excitation system is located 0.5m above the rotor; the anchor nuts evenly surround the excitation system, with a radius of 5m and a height of 0.5m lower.

7. The heuristic search method for the hydro-generator inner bore maintenance scenario according to claim 1 is characterized by: The NBV path planning algorithm in step 5.2 uses a dynamic task relevance weight factor , and adjust the sampling probability of candidate viewpoints in real time, so that the robot can give priority to exploring the target area when there is severe occlusion and limited space; in: It is the intersection of the field of view and the region of interest; is the task relevance amplification factor.

8. The heuristic search method for the hydro-generator inner bore maintenance scenario according to claim 1 is characterized by: The robot motion control process in step 6 includes obstacle avoidance control, path replanning, and abnormal target relocation mechanisms, which are used to automatically adjust the inspection strategy when encountering obstacles or recognition errors during the inspection process.

9. The heuristic search method for the maintenance scenario of the inner bore of a hydro-generator according to claim 1, characterized in that: During the inspection process, the robot dynamically switches imaging parameters and lighting intensity based on the NBV planning results and LLM reasoning results to adapt to the complex lighting conditions in the bore and ensure the observation quality of key components.

10. The heuristic search method for the maintenance scenario of the inner bore of a hydro-generator according to claim 1, characterized in that: NBV planner for candidate perspectives Calculated composite score Satisfies the following formula: ； in, is information gain; is the path cost; It is the intersection of the field of view and the region of interest; is the task relevance amplification factor; The confidence of the region of interest output by the large language model; is the confidence weight coefficient.

Images