Power grid environment risk assessment method and device for hot-line work robot

Abstract

The invention provides a power grid environment risk assessment method and device for a hot-line work robot. The method comprises the following steps: acquiring preprocessed multi-source data; the multi-source data comprises power grid environment three-dimensional point cloud object information, robot state information and robot pre-action path information; modeling the three-dimensional point cloud object information of the power grid environment to obtain a risk potential field model containing different object attributes; inputting the multi-source data and the risk potential field model into a risk assessment generation model to obtain a power grid environment risk assessment result; the risk assessment generation model is obtained based on the following steps: constructing a multi-level knowledge corpus of a power grid environment where the robot is located; training a large language model based on the multi-level knowledge corpus to obtain a risk assessment generation model; a trainable rank decomposition matrix is injected into a key layer bypass of the large language model. According to the method, the problems that the accuracy is insufficient and the operation safety and interpretability of the robot cannot be ensured during power grid environment risk assessment can be solved.

Description

Technical Field

[0001] This invention relates to the field of live-line work assessment technology, and specifically to a method and apparatus for assessing the power grid environment risk of a live-line work robot. Background Technology

[0002] With the rapid development of society and the economy, people's demands for a continuous and stable power supply in their production and daily lives are increasing. Currently, traditional power grid operations are mainly carried out manually. When workers are working on live lines, short circuits are easily caused, leading to accidents such as injuries and fatalities. Furthermore, this method of operation is often labor-intensive, inefficient, and dangerous, seriously affecting the continuous and stable operation of the power grid. Therefore, to ensure the safety of personnel working on live lines, improve the efficiency of live-line operations, and ensure the continuous and stable operation of the power grid, it is becoming increasingly important to replace manual labor in power grid-related tasks. Intelligent robots are changing traditional work patterns, improving production safety and efficiency, and playing a significant role in promoting high-quality social and economic development.

[0003] In live-line working environments, the complex and diverse electrical equipment (such as phase wires, current-carrying wires, and insulators) presents numerous safety hazards for traditional manual operations. Against this backdrop, the use of live-line working robots to replace manual labor has become an inevitable trend. Accurate risk quantification modeling and path generation for complex environments are of paramount importance, with environmental risk modeling and path planning becoming key technologies for ensuring safe operations. Furthermore, in power grid operations, power grid environmental risk assessment requires the integration of multi-source information, including trajectory data, environmental sensors, robot model, and numerous safety criteria. Accurately assessing risks and providing rational justification, identifying high-risk points, and providing visual warnings are crucial for ensuring the safety of robot operations and enhancing the operational confidence of power maintenance personnel.

[0004] Because power grid maintenance and repair work is generally located in open outdoor environments, the objects in these high-voltage environments are diverse in type and properties, exhibiting characteristics such as unstructured nature, high and low voltage, hard and soft surfaces, and dynamic and static states. Existing environmental modeling methods cannot fully consider the environmental differences caused by object properties, making traditional rule-based representation methods difficult to apply. Furthermore, existing large-scale models suffer from deficiencies in risk assessment of robot pre-action paths, such as a lack of knowledge about the power grid domain, which can easily lead to logical gaps. When handling specific power tasks, they face problems such as insufficient accuracy, a scarcity of high-quality domain datasets, and difficulties in knowledge updates.

[0005] In summary, the aforementioned problems with existing technical solutions result in insufficient accuracy in risk assessment of the power grid environment, and fail to guarantee the safety and interpretability of robot operations. Summary of the Invention

[0006] In view of this, it is necessary to provide a method and device for assessing the power grid environment risk of live-line working robots, so as to solve the technical problems that existing technical solutions have insufficient accuracy in assessing the power grid environment risk and cannot guarantee the safety and interpretability of robot operation.

[0007] To address the aforementioned problems, in a first aspect, the present invention provides a method for assessing the power grid environment risk of a live-line working robot, comprising: Acquire preprocessed multi-source data; the multi-source data includes: three-dimensional point cloud object information of the power grid environment in which the robot operates, robot state information, and robot pre-action path information; Three-dimensional risk potential field modeling is performed on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation to obtain a risk potential field model containing different object attributes. The multi-source data and the risk potential field model are input into the risk assessment generation model to obtain the power grid environmental risk assessment results. The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

[0008] In one possible implementation, the risk potential field model includes: a high-pressure equipment repulsive potential field, a flexible object dynamic potential field, a dynamic object prediction potential field, and a hardness-differentiated potential field. The high-pressure equipment repulsive potential field, the flexible object dynamic potential field, the dynamic object prediction potential field, and the hardness differentiation potential field are weighted and calculated using a combination of the analytic hierarchy process and the entropy weighting method to obtain the risk potential field model.

[0009] In one possible implementation, the expression for the repulsive potential field of the high-voltage equipment is: ; in, This represents the value of the repulsive potential field of the high-voltage equipment. This is the preset repulsion gain coefficient for high-voltage equipment. Current position of the robot To the obstacle distance, To ensure safety of the impact radius, This is the voltage level coefficient.

[0010] In one possible implementation, the expression for the dynamic potential field of the flexible object is: ; in, This represents the value of the dynamic potential field of the flexible object. This is a preset flexible gain coefficient used to control the response intensity of flexible obstacle avoidance. To consider the object's motion t Time Robot Location q To the obstacle location Euclidean distance, To ensure safety of the impact radius, To account for the safe distance extension for the swing amplitude of flexible objects.

[0011] In one possible implementation, the expression for the predicted potential field of a dynamic object is: ; in, This represents the value of the potential field of the dynamic object. The predicted time domain length represents the number of discrete time points predicted forward. For time decay weights, an exponentially decreasing weight is used, satisfying... , For prediction in the current t The time after A future moment The position of the object, This represents the standard potential field value.

[0012] In one possible implementation, the expression for the hardness-differentiated potential field is: ; in, Let be the value of the potential field of a rigid object. The preset hardness coefficient, This represents the standard potential field value.

[0013] In one possible implementation, the risk assessment generation model includes: Hybrid encoders combine sparse and dense coding methods to encode input data, resulting in sparse and dense coding results. The hybrid retrieval module is used to retrieve results based on sparse and dense encoding, and outputs the most relevant search results. The reordering module is used to sort the most relevant search results according to document relevance, resulting in a sorted set of documents; The prompt generation module is used to output the power grid environmental risk assessment results based on the sorted document set and according to the preset prompt framework.

[0014] In one possible implementation, the most relevant search results are sorted according to document relevance to obtain a sorted set of documents, including: A reciprocal ranking fusion strategy is adopted. After the most relevant search results are initially sorted by multiple documents, the bge-reranker-v2-m3 model is used to re-rank the single documents, and the documents most relevant to the query are ranked first, forming a sorted set of documents.

[0015] In one possible implementation, the pre-defined prompting framework is used for: In the sorted document collection, the documents that meet the requirements of the problem are identified as the most relevant documents; Extract key fragments from the most relevant documents and use these fragments to make inferences and explanations to obtain the power grid environmental risk assessment results.

[0016] Secondly, the present invention also provides a power grid environmental risk assessment device for a live-line working robot, comprising: The acquisition module is used to acquire preprocessed multi-source data; the multi-source data includes: three-dimensional point cloud object information of the power grid environment in which the robot operates, robot state information, and robot pre-action path information. The potential field modeling module is used to perform three-dimensional risk potential field modeling on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation, and obtain a risk potential field model containing different object attributes. The assessment module is used to input the multi-source data and the risk potential field model into the risk assessment generation model to obtain the power grid environmental risk assessment results. The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

[0017] The beneficial effects of adopting the above implementation method are as follows: The power grid environment risk assessment method and device for live-line working robots provided by the present invention perform three-dimensional risk potential field modeling on the three-dimensional point cloud object information of the power grid environment after preprocessing the robot operation, design differentiated potential field functions for objects with different attributes in the power grid environment, construct an environmental risk potential field map, realize the quantitative mapping from object attributes to risk level, and make the environmental modeling more in line with the actual power maintenance site, thereby improving the accuracy of risk assessment.

[0018] Furthermore, this invention addresses the shortcomings of current general-purpose large language models in terms of power industry expertise. It effectively integrates multi-source data to improve the robot's ability to handle knowledge-intensive tasks and enhance factual accuracy. By constructing a multi-level knowledge corpus, it improves the large model's understanding of three-dimensional power grid scenarios. This solves the problems of insufficient accuracy, lack of high-quality domain datasets, and difficulty in knowledge updating faced by large models when handling live-line operations. It achieves evidence-based assessment results, ensures the safety and interpretability of robot operations, and provides risk assessors with accurate, interpretable, and effective evidence with strong real-time performance.

[0019] Therefore, the present invention can solve the technical problems of insufficient accuracy and inability to guarantee the safety and interpretability of robot operations when conducting risk assessments of the power grid environment in existing technical solutions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of an embodiment of the power grid environment risk assessment method for live-line working robots provided by the present invention; Figure 2 A schematic diagram of a three-dimensional point cloud object in a power grid environment provided by this invention; Figure 3 A flowchart of another embodiment of the power grid environment risk assessment method for live-line working robots provided by the present invention; Figure 4 A schematic block diagram of an embodiment of the power grid environment risk assessment device for a live-line working robot provided by the present invention; Figure 5 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0022] 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 a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] In the description of the embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0024] In this embodiment of the invention, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, apparatus, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such process, method, product or device.

[0025] The naming or numbering of steps in the embodiments of the present invention does not mean that the steps in the method flow must be executed in the time / logical order indicated by the naming or numbering. The execution order of the named or numbered process steps can be changed according to the technical purpose to be achieved, as long as the same or similar technical effect can be achieved.

[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0027] In view of the shortcomings of the prior art, the purpose of this invention is to solve the following technical problems: Achieving refined environmental modeling: Constructing a power grid environment model that fully represents the differences in object attributes, accurately delineating risk areas, and quantifying the impact of different object attributes on risk levels; Simultaneously, integrating professional knowledge in the power field and safety regulations for live-line working robots: Through deep integration of the domain knowledge base and the large model, compensating for the deficiencies of general models in terms of power professional knowledge, effectively integrating multi-source data, and improving the processing capacity and factual accuracy of knowledge-intensive tasks; Constructing a risk assessment system: Establishing a risk assessment model that perceives the power grid environment in real time, realizing real-time risk assessment, prediction, and early warning, providing timely and accurate safety decision support for robot operations, and enhancing the operational confidence of power workers.

[0028] This invention provides a method and apparatus for assessing the power grid environment risk of a live-line working robot, which will be described below.

[0029] This invention provides a method for assessing the power grid environment risk of a live-line working robot. The method can be implemented by executing an application on a terminal or server. The terminal can be a mobile phone or computer, and the server can be an edge server or a cloud server. Figure 1 As shown, the method includes: S101. Obtain preprocessed multi-source data; the multi-source data includes: three-dimensional point cloud object information of the power grid environment for robot operation, robot status information, and robot pre-action path information.

[0030] Understandably, this involves collecting multi-source heterogeneous data from the robot's operating environment, including: on-site images acquired by vision sensors, 3D spatial information acquired by depth cameras, and pose / path / state data fed back by the robot's own sensors. The raw data is then preprocessed, including denoising, registration, and normalization, to construct a unified spatiotemporal data benchmark.

[0031] Among them, the three-dimensional spatial information acquired by the depth camera, namely the three-dimensional point cloud object information of the power grid environment, such as... Figure 2 As shown; the pose / path / state data fed back by the robot's body sensors, corresponding to the robot's state information and robot's pre-action path information.

[0032] S102. Perform three-dimensional risk potential field modeling on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation to obtain a risk potential field model containing different object attributes.

[0033] It is understandable to use the Artificial Potential Field (APF), which has a simple structure and low computational burden, for environmental risk potential field modeling. Existing environmental conditions include high pressure, different soft and hard properties of objects, and dynamic and static properties of objects. In order to distinguish the different characteristics of environmental factors, objects with unique characteristics are differentiated, and then a potential field function is designed for each influencing factor to construct a risk potential field model.

[0034] S103. Input the multi-source data and the risk potential field model into the risk assessment generation model to obtain the power grid environmental risk assessment results; The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

[0035] Understandably, large language models can use deep learning (such as CNNs or graph neural networks) to learn implicit risk fields from historical task data, or knowledge graph-based reasoning systems.

[0036] In some embodiments, the risk potential field model includes: a high-pressure equipment repulsive potential field, a flexible object dynamic potential field, a dynamic object prediction potential field, and a hardness-differentiated potential field. The high-pressure equipment repulsive potential field, the flexible object dynamic potential field, the dynamic object prediction potential field, and the hardness differentiation potential field are weighted and calculated using a combination of the analytic hierarchy process and the entropy weighting method to obtain the risk potential field model.

[0037] Understandably, a combined weighting strategy using the Analytic Hierarchy Process (AHP) and entropy weighting is employed to dynamically calculate the weights of each risk factor. Subjective weighting (AHP): Based primarily on power safety regulations and expert experience, a hierarchical structure of risk factors is constructed to determine the relative importance weight of each factor. .

[0038] Objective weighting (entropy weighting method): Based on real-time collected data, calculate the information entropy of each factor. ; in, For the first Information entropy of each risk factor This represents the total number of sampling points (sample size). For the first The factor in the first Normalized values ​​of each sampling point.

[0039] The objective weights are: , This represents the total number of risk factors.

[0040] Overall weighting: A weighted geometric average method is used to combine subjective and objective weights. ;in, For the first The overall weight of each element, Subjective weighting, For objective weighting, and As the balance coefficient, this invention takes , .

[0041] Construction of a comprehensive risk potential field: By weighted superposition of the potential field functions of various objects, a comprehensive risk potential field for the power grid environment is constructed: ; in, To comprehensively assess the risk potential value, This represents the total number of objects identified in the environment. For the first The overall weight of each object This involves defining the potential field functions for objects with different properties. This generates a three-dimensional risk potential field map, providing a foundation for subsequent risk assessment.

[0042] Integration and enhancement of domain knowledge and security charters: By introducing an improved retrieval augmentation generation (RAG) strategy that integrates hierarchical information augmentation, hybrid retrieval strategies, and a finely tuned risk assessment generation model, the risk of robot operation can be reasonably assessed by accurately and quickly combining safety regulations and industry knowledge.

[0043] In some embodiments, the expression for the repulsive potential field of the high-voltage equipment is: ; in, This represents the value of the repulsive potential field of the high-voltage equipment. This is the preset repulsion gain coefficient for high-voltage equipment. Current position of the robot To the obstacle distance, To ensure safety of the impact radius, This is the voltage level coefficient.

[0044] It is understandable that the voltage level coefficient is the voltage level coefficient of the power grid, which is obtained by normalizing the actual voltage value.

[0045] In some embodiments, the expression for the dynamic potential field of the flexible object is: ; in, This represents the value of the dynamic potential field of the flexible object. This is a preset flexible gain coefficient used to control the response intensity of flexible obstacle avoidance. To consider the object's motion t Time Robot Location q To the obstacle location Euclidean distance, To ensure safety of the impact radius, To account for the safe distance extension for the swing amplitude of flexible objects.

[0046] It is understandable that the swing amplitude here refers to the swing amplitude of the flexible object, which is obtained by estimating the object's trajectory through a tracking algorithm.

[0047] For flexible objects, i.e. flexible oscillating objects (such as cables and tree branches), considering the uncertainty of their motion trajectory, a dynamic extended potential field is adopted.

[0048] In some embodiments, the expression for the predicted potential field of a dynamic object is: ; in, This represents the value of the potential field of the dynamic object. The predicted time domain length represents the number of discrete time points predicted forward. For time decay weights, an exponentially decreasing weight is used, satisfying... , For prediction in the current t The time after A future moment The position of the object, This represents the standard potential field value.

[0049] It is understandable that dynamic objects, i.e. moving objects, such as workers and vehicles, are predicted to have future trajectories based on Kalman filtering, and a time-varying potential field is constructed.

[0050] In some embodiments, the expression for the hardness differential potential field is: ; in, Let be the value of the potential field of a rigid object. The preset hardness coefficient, This represents the standard potential field value.

[0051] It is understandable that different collision risk coefficients are set for objects with different hardness, such as rigid objects. =1.5 (Severe consequences of collision), flexible object =0.8 (moderate contact is permissible).

[0052] In some embodiments, the risk assessment generation model includes: Hybrid encoders combine sparse and dense coding methods to encode input data, resulting in sparse and dense coding results. The hybrid retrieval module is used to retrieve results based on sparse and dense encoding, and outputs the most relevant search results. The reordering module is used to sort the most relevant search results according to document relevance, resulting in a sorted set of documents; The prompt generation module is used to output the power grid environmental risk assessment results based on the sorted document set and according to the preset prompt framework.

[0053] Understandably, a hybrid encoder combines sparse and dense coding methods to provide a rich semantic core for the text. It employs the bge-m3 semantic vector model, performing both sparse and dense coding simultaneously.

[0054] Hybrid search module: It combines sparse and dense search to provide the most relevant search results.

[0055] Re-ranking module: Ranks documents by relevance to optimize the accuracy and relevance of search results.

[0056] In some embodiments, the most relevant search results are sorted according to document relevance to obtain a sorted set of documents, including: A reciprocal ranking fusion strategy is adopted. After the most relevant search results are initially sorted by multiple documents, the bge-reranker-v2-m3 model is used to re-rank the single documents, and the documents most relevant to the query are ranked first, forming a sorted set of documents.

[0057] Understandably, document relevance ranking optimizes the accuracy and relevance of search results: 1) A Reverse Ranking fusion (RRF) strategy is used for initial multi-document ranking. The RRF formula is: ,in For document The fusion score, It is a document The ranking of a document in the i-th search engine. It is a constant. 1) Number of retrievers. 2) Use the bge-reranker-v2-m3 model for single-document re-ranking. 3) Rank the documents most relevant to the query first to form a document set.

[0058] bge-reranker-v2-m3 is a multilingual reranking model developed by BAAI. Based on the bge-m3 architecture, it is specifically designed for efficiently processing query and document relevance ranking.

[0059] In some embodiments, the preset prompt frame is used for: In the sorted document collection, the documents that meet the requirements of the problem are identified as the most relevant documents; Extract key fragments from the most relevant documents and use these fragments to make inferences and explanations to obtain the power grid environmental risk assessment results.

[0060] Understandably, the prompt generation module is based on a large model: a pre-defined prompt framework is established, and the steps are as follows: 1) Identify the most relevant documents: identify documents that meet the question's requirements. 2) Extract key document fragments: find the parts that can directly answer the question. 3) Provide detailed reasoning explanations: conduct in-depth analysis in conjunction with the documents. 4) Provide explicit answers: select the most appropriate answer based on the reasoning process.

[0061] In some embodiments, the technical problem to be solved by the present invention is to address the inability to accurately predict operational risks and insufficient environmental understanding capabilities of live-line working robots in operating environments such as distribution networks and substations. This is achieved through refined environmental modeling: constructing a power grid environment model that fully characterizes the differences in object attributes, accurately delineating risk areas, and quantifying the impact of different object attributes on risk levels; simultaneously, integrating power industry expertise and safe operating procedures for live-line working robots: through deep integration of a domain knowledge base and a large model, compensating for the deficiencies of general models in terms of power industry expertise, effectively integrating multi-source data, and improving the processing capacity and factual accuracy of knowledge-intensive tasks; finally, constructing a risk assessment system: establishing a risk assessment model that perceives the operating environment in real time, enabling real-time risk assessment, prediction, and early warning.

[0062] like Figure 3 As shown, this invention mainly provides a method for modeling and assessing the risks of live-line working robots in the power grid environment, which mainly includes four steps: multi-source data acquisition and preprocessing, risk force field modeling, fusion and enhancement of domain knowledge and safety regulations, and comprehensive risk assessment and early warning.

[0063] In some embodiments, the specific implementation of the present invention is as follows: 1) Multi-source data acquisition and preprocessing Multi-source heterogeneous data from the robot's operating environment are collected, including: on-site images acquired by vision sensors, 3D spatial information acquired by depth cameras, and pose / path / state data fed back by the robot's own sensors. The collected raw data undergoes preprocessing such as denoising, registration, and normalization to construct a unified spatiotemporal data benchmark.

[0064] 2) Risk Force Field Modeling Environmental risk potential field modeling is performed using an artificial potential field (APF), which is simple in structure and has low computational burden. Existing environmental conditions include high pressure, different soft and hard properties of objects, and dynamic and static properties of objects. To distinguish the different characteristics of environmental factors, objects with unique characteristics are differentiated, and a potential field function is designed for each influencing factor to construct a risk potential field model.

[0065] Standard repulsive potential field: ; in, The standard repulsive potential field value, Repulsion gain coefficient, Robot position q to obstacle position Euclidean distance, The radius of influence of the potential field (safe distance threshold).

[0066] 1. Design corresponding potential field functions for objects with different properties: ① Repulsive potential field of high-voltage equipment: For high-voltage equipment, an enhanced repulsive potential field function is adopted: ; in, The magnitude of the repulsive potential field of the high-voltage equipment. This refers to the repulsion gain coefficient of high-voltage equipment. Current position of the robot To the obstacle distance, To ensure safety of the impact radius, This is the voltage level coefficient (obtained by normalizing the actual voltage value).

[0067] ② Dynamic potential field of a flexible object: For flexible, oscillating objects (such as cables or tree branches), considering the uncertainty of their trajectory, a dynamically extended potential field is used: ; in, The magnitude of the dynamic potential field of the flexible object. This is the flexible gain coefficient, which controls the response intensity of flexible obstacle avoidance. To consider the robot's position q after the object moves to the obstacle's position Euclidean distance, To ensure safety of the impact radius, To account for the safe distance extension of the swing amplitude of the flexible object, the object's motion trajectory is estimated using a tracking algorithm.

[0068] ③ Predicting the potential field of dynamic objects: For moving objects (such as workers or vehicles), their future trajectories are predicted based on Kalman filtering, and a time-varying potential field is constructed: ; in, This represents the potential field value of a dynamic object. The prediction time domain length (prediction steps) represents the number of discrete time points predicted forward. For time decay weights, an exponentially decreasing weight is used, satisfying... . For the predicted first The position of an object at a future moment. This represents the standard potential field value.

[0069] ④ Hardness differential potential field: Different collision risk coefficients are set for objects of different hardness: ; in, For a rigid object, Hardness coefficient: for rigid objects (Severe consequences of collision), flexible objects (Approachable with moderate contact) This represents the standard potential field value.

[0070] 2. Adaptive allocation of risk factor weights: A combined weighting strategy using the Analytic Hierarchy Process (AHP) and entropy weighting is employed to dynamically calculate the weights of each risk factor. Subjective weighting (AHP): Based primarily on power safety regulations and expert experience, a hierarchical structure of risk factors is constructed to determine the relative importance weight of each factor. .

[0071] Objective weighting (entropy weighting method): Based on real-time collected data, calculate the information entropy of each factor. ; in, For the first Information entropy of each risk factor This represents the total number of sampling points (sample size). For the first The factor in the first Normalized values ​​of each sampling point.

[0072] The objective weights are: , This represents the total number of risk factors.

[0073] 3. Overall Weighting: A weighted geometric average method is used to combine subjective and objective weights. ; in, For the first The overall weight of each element, Subjective weighting, For objective weighting, and As a balance coefficient, the present invention can take , .

[0074] Construction of a comprehensive risk potential field: By weighted superposition of the potential field functions of various objects, a comprehensive risk potential field for the power grid environment is constructed: ; in, To comprehensively assess the risk potential value, This represents the total number of objects identified in the environment. For the first The overall weight of each object This involves defining the potential field functions for objects with different properties. This generates a three-dimensional risk potential field map, providing a foundation for subsequent risk assessment.

[0075] 3) Integration and enhancement of domain knowledge and security charters By introducing an improved retrieval augmentation generation (RAG) strategy that integrates hierarchical information augmentation, hybrid retrieval strategies, and a finely tuned risk assessment generation model, the risk of robot operation can be reasonably assessed by accurately and quickly combining safety regulations and industry knowledge.

[0076] 1. Data Preparation Stage: Design a dataset construction method based on a three-level hierarchical architecture. To address the lack of a high-quality dataset of safety regulations specifically for power robots, this study constructs structured and hierarchical documents based on important regulations such as power robot operation regulations, power safety operation procedures for power practitioners, and emergency response guidelines. This results in a multi-level knowledge corpus for intelligent reasoning. The safety regulations are semantically decomposed into three layers: an L1 framework index layer (storing meta-information such as chapter titles, scope of application, and key terms, using a structured index); an L2 clause and rule layer (storing specific safety clauses, quantitative standards, and constraints, extracting numerical rule templates such as "IF voltage level = 10kV AND distance < 0.7m THEN risk = high"); and an L3 case knowledge layer (storing historical operation cases, violation events, and expert experience, constructing scene feature vectors and case reasoning indexes). Each layer employs a differentiated indexing strategy to support a progressive retrieval path from coarse-grained framework positioning to fine-grained clause matching and case verification. This ensures that each text segment has independence and flexibility, laying the foundation for subsequent stages.

[0077] 2. Model Fine-tuning Stage: Addressing the specialized knowledge requirements of specific power grid operation scenarios, a combination of efficient parameter tuning techniques and progressive data augmentation strategies is employed to enable the general-purpose large language model to possess professional reasoning capabilities in the power sector. Technically, Low-Rank Adaptive (LoRA) technology is used for supervised instruction tuning. This method embeds a trainable rank decomposition matrix in the bypass of key model layers, requiring only about 1% of the model weights to complete domain transfer. Compared to full parameter tuning, this significantly reduces computational resource consumption and accelerates training speed. In terms of training strategy, drawing on the core concept of Retrieval Enhancement Fine-tuning (RAF), a three-stage curricular training scheme is designed, progressing from easy to difficult: the initial stage uses high-purity procedural text to build a basic knowledge system; the intermediate stage proportionally mixes in semantically similar noisy documents to simulate a real retrieval environment; and the final stage significantly increases the proportion of interfering information and introduces adversarial examples. Through progressively increasing task difficulty, the model's accurate filtering and deep reasoning capabilities in complex information environments are strengthened. Details are as follows: ① Lightweight Fine-tuning Technology Application: The Low-Rank Matrix Injection (LoRA) method is used to perform supervised domain adaptation. This scheme implants rank constraint matrix branches (typically rank values ​​r=8-16) into the bypass of key layers of the model. Knowledge transfer can be completed simply by updating the weights of these branches, reducing memory usage by 85% and accelerating the training process by 3-5 times compared to full parameter tuning. ② Multi-Question Generation Module Automatically Outputs Fine-tuning Datasets: A prompt-driven question-answer pair generation pipeline is established, comprising three key steps: 1) Implementing a few-sample demonstration design, compiling 3-5 standard format question-answer examples for different query types (fact retrieval, numerical judgment, etc.) as model learning references; 2) Performing context-enhanced generation, inputting the target paragraph of the procedure document (typically 500-800 words) as background knowledge into the pre-trained language model, driving it to produce structured question-answer instances based on the example format; 3) Conducting quality control screening, with domain experts verifying the accuracy and format compliance of the automatically generated samples, eliminating samples with logical errors, ambiguous expressions, or redundant information.

[0078] The LoRA fine-tuning module retains the original pre-trained weights W, which include two low-rank matrices A and B. Knowledge transfer can be completed simply by updating these branch weights, effectively reducing the number of parameters. The formula is as follows.

[0079] ; in For the updated pre-trained weights, and It is a trainable low-rank matrix. and These are the output and input dimensions, respectively. For the original pre-trained weights, the rank parameter The value ranges from 8 to 64, much smaller than the dimension of the original weight matrix. The values ​​are frozen during training. Update only and The number of parameters is from Reduce to The parameter compression ratio reaches over 90%.

[0080] 3. Retrieval and Reasoning Stage: A two-stage architecture of "offline index building + online real-time query" enables efficient semantic retrieval and intelligent reasoning generation of power grid regulations. The multi-level structured regulation documents (including L1 framework layer, L2 clause layer, and L3 case layer) output from the data preparation module are fed into the hybrid encoding module. Next, the system enters the online real-time query stage. After the user inputs the data before evaluation, the system executes a four-step reasoning process: query understanding and encoding, hybrid retrieval recall, re-ranking and filtering, and enhanced generation and reasoning.

[0081] ① Hybrid Encoder: Combining sparse and dense coding methods to provide a rich semantic core for the text. It employs the bge-m3 semantic vector model, performing both sparse and dense coding simultaneously.

[0082] ② Hybrid Search Module: This module comprehensively utilizes sparse and dense retrieval methods to provide the most relevant search results. 1) It batch imports the vectorized set of procedure documents into memory. 2) Given a user query, the encoder performs both sparse and dense retrieval. 3) It calculates similarity using cosine similarity. Hybrid search formula: ,in: Cosine distance It is a query vector. It is a document vector. This represents the mixed similarity score. These are the weighting coefficients. For sparse retrieval scores, Score for dense retrieval.

[0083] ③ Re-ranking module: Document relevance ranking to optimize the accuracy and relevance of search results: 1) Initial multi-document ranking is performed using the Reverse Rank Fusion (RRF) strategy. RRF formula: ,in For document The fusion score, It is a document The ranking of a document in the i-th search engine. It is a constant. 1) Number of retrievers. 2) Use the bge-reranker-v2-m3 model for single-document re-ranking. 3) Rank the documents most relevant to the query first to form a document set.

[0084] ④ Hint Generation Based on a Large Model: A hint framework is pre-defined, with the following steps: 1) Identify the most relevant documents: Identify documents that meet the question's requirements. 2) Extract key document fragments: Find the parts that can directly answer the question. 3) Provide detailed reasoning explanations: Conduct in-depth analysis in conjunction with the documents. 4) Provide explicit answers: Select the most appropriate answer based on the reasoning process.

[0085] 4) Comprehensive risk assessment and early warning Based on the multi-source data and environmental potential field map obtained from the above steps, the data is output to a large model in the fusion domain knowledge base to conduct a risk assessment on the robot's preset path. Finally, the risk display of the path points, the judgment basis (safety regulations, industry standards), and operation suggestions are generated.

[0086] The key point of this invention is: 1. To address the problem of insufficient safety assessment of live-line working robots and the difficulty for operators to have strong confidence in their operation, this invention constructs a complete risk modeling and assessment process. By integrating multi-source data (3D environmental information, robot pose / path / state, live-line working safety regulations in the power grid field and power industry standards), it outputs risk display, judgment basis and operation suggestions for robot preset path points, which can improve real-time risk assessment, prediction and early warning capabilities and enhance operator confidence.

[0087] 2. This invention designs differentiated potential field functions for objects with different properties (high voltage / low voltage, soft / hard, dynamic / static) in the power grid environment, and at the same time adopts an adaptive weight allocation mechanism to construct a comprehensive risk potential field and build an environmental risk potential field map.

[0088] 3. This invention addresses the shortcomings of current general-purpose large models in terms of power industry expertise. It effectively integrates multi-source data to improve the robot's ability to handle knowledge-intensive tasks and the accuracy of factual information. By constructing a hierarchical information-enhanced RAG framework, it improves the understanding ability of large models in three-dimensional power grid scenarios. This solves the problems of insufficient accuracy, lack of high-quality domain datasets, and difficulty in knowledge updating faced by large models when handling live-line operations by robots. It achieves evidence-based evaluation results and ensures the safety and interpretability of robot operations.

[0089] 4. A complete set of dataset construction methods for extracting power standards and industry knowledge has been developed, which can provide risk assessors with accurate, interpretable and effective evidence with strong real-time performance.

[0090] This invention has the following three advantages: 1. Existing technology assessment results lack traceability (e.g., power operation emergency alarm systems only output risk factors without providing charter references). This invention not only outputs risk levels but also provides judgment criteria (specific safety charter clauses) and operational suggestions, enhancing the trust and decision support capabilities of power maintenance personnel.

[0091] 2. Existing technologies do not fully consider the differences in object attributes (such as the risk assessment method for live-line work based on multi-source data, which only constructs a knowledge graph and a grid risk propagation model). This invention designs a special potential field function for different attributes such as high voltage, soft and hard, dynamic and static, to realize the quantitative mapping of object attributes to risk levels, and the environmental modeling is more in line with the actual situation of power maintenance site.

[0092] 3. Existing general large-scale models lack knowledge in the power field (e.g., power operation safety violation identification methods based on embodied intelligence rely on general deep neural networks and are not optimized for power regulations). However, this invention deeply embeds the power robot safety operation regulations and safety standards into the model reasoning process through hierarchical dataset construction, LoRA fine-tuning, and hybrid retrieval RAG strategy, thereby filling knowledge gaps and improving the accuracy and interpretability of the assessment facts.

[0093] like Figure 4 As shown, the present invention also provides a power grid environmental risk assessment device 400 for live-line working robots, comprising: The acquisition module 401 is used to acquire preprocessed multi-source data; the multi-source data includes: three-dimensional point cloud object information of the power grid environment in which the robot operates, robot state information, and robot pre-action path information. The potential field modeling module 402 is used to perform three-dimensional risk potential field modeling on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation, and obtain a risk potential field model containing different object attributes. The assessment module 403 is used to input the multi-source data and the risk potential field model into the risk assessment generation model to obtain the power grid environmental risk assessment results. The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

[0094] The power grid environment risk assessment device for live-line working robots provided in the above embodiments can realize the technical solutions described in the embodiments of the power grid environment risk assessment method for live-line working robots. The specific implementation principles of each module or unit can be found in the corresponding content in the embodiments of the power grid environment risk assessment method for live-line working robots, and will not be repeated here.

[0095] like Figure 5 As shown, the present invention also provides an electronic device 500. The electronic device 500 includes a processor 501, a memory 502, and a display 503. Figure 5 Only some components of the electronic device 500 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0096] In some embodiments, memory 502 may be an internal storage unit of electronic device 500, such as a hard disk or memory of electronic device 500. In other embodiments, memory 502 may also be an external storage device of electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 500.

[0097] Furthermore, the memory 502 may include both internal storage units of the electronic device 500 and external storage devices. The memory 502 is used to store application software and various types of data installed on the electronic device 500.

[0098] In some embodiments, processor 501 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 502 or process data, such as the power grid environment risk assessment method for live-line working robots in this invention.

[0099] In some embodiments, display 503 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 503 is used to display information from electronic device 500 and to display a visual user interface. Components 501-503 of electronic device 500 communicate with each other via a system bus.

[0100] In some embodiments of the present invention, when the processor 501 executes the power grid environment risk assessment program for the live-line working robot stored in the memory 502, the following steps can be implemented: Acquire preprocessed multi-source data; the multi-source data includes: three-dimensional point cloud object information of the power grid environment in which the robot operates, robot state information, and robot pre-action path information; Three-dimensional risk potential field modeling is performed on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation to obtain a risk potential field model containing different object attributes. The multi-source data and the risk potential field model are input into the risk assessment generation model to obtain the power grid environmental risk assessment results. The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

[0101] It should be understood that when the processor 501 executes the power grid environment risk assessment program for the live-line working robot in the memory 502, in addition to the functions mentioned above, it can also perform other functions, as can be found in the description of the corresponding method embodiments above.

[0102] Furthermore, the embodiments of the present invention do not specifically limit the type of electronic device 500 mentioned. Electronic device 500 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, electronic device 500 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0103] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the power grid environment risk assessment method for live-line working robots provided by the methods described above, the method comprising: Acquire preprocessed multi-source data; the multi-source data includes: three-dimensional point cloud object information of the power grid environment in which the robot operates, robot state information, and robot pre-action path information; Three-dimensional risk potential field modeling is performed on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation to obtain a risk potential field model containing different object attributes. The multi-source data and the risk potential field model are input into the risk assessment generation model to obtain the power grid environmental risk assessment results. The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

[0104] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0105] The above provides a detailed description of the power grid environment risk assessment method and device for live-line working robots provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for assessing the power grid environmental risk of a live-line working robot, characterized in that, include: Obtain preprocessed multi-source data; The multi-source data includes: three-dimensional point cloud object information of the power grid environment in which the robot operates, robot status information, and robot pre-action path information; Three-dimensional risk potential field modeling is performed on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation to obtain a risk potential field model containing different object attributes. The multi-source data and the risk potential field model are input into the risk assessment generation model to obtain the power grid environmental risk assessment results. The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

2. The method for assessing the power grid environmental risk of a live-line working robot according to claim 1, characterized in that, The risk potential field model includes: high-pressure equipment repulsive potential field, flexible object dynamic potential field, dynamic object prediction potential field, and hardness differential potential field. The high-pressure equipment repulsive potential field, the flexible object dynamic potential field, the dynamic object prediction potential field, and the hardness differentiation potential field are weighted and calculated using a combination of the analytic hierarchy process and the entropy weighting method to obtain the risk potential field model.

3. The method for assessing the power grid environmental risk of a live-line working robot according to claim 2, characterized in that, The expression for the repulsive potential field of the high-voltage equipment is: ； in, This represents the value of the repulsive potential field of the high-voltage equipment. This is the preset repulsion gain coefficient for high-voltage equipment. Current position of the robot To the obstacle distance, To ensure safety of the impact radius, This is the voltage level coefficient.

4. The method for assessing the power grid environmental risk of a live-line working robot according to claim 2, characterized in that, The expression for the dynamic potential field of the flexible object is: ； in, This represents the value of the dynamic potential field of the flexible object. This is a preset flexible gain coefficient used to control the response intensity of flexible obstacle avoidance. To consider the object's motion t Time Robot Location q To the obstacle location Euclidean distance, To ensure safety of the impact radius, To account for the safe distance extension for the swing amplitude of flexible objects.

5. The method for assessing the power grid environmental risk of a live-line working robot according to claim 2, characterized in that, The expression for predicting the potential field of a dynamic object is: ； in, This represents the value of the potential field of the dynamic object. The predicted time domain length represents the number of discrete time points predicted forward. For time decay weights, an exponentially decreasing weight is used, satisfying... , For prediction in the current t The time after A future moment The position of the object, This represents the standard potential field value.

6. The method for assessing the power grid environmental risk of a live-line working robot according to claim 2, characterized in that, The expression for the hardness-differentiated potential field is: ； in, Let be the value of the potential field of the rigid object. The preset hardness coefficient, This represents the standard potential field value.

7. The method for assessing the power grid environmental risk of a live-line working robot according to any one of claims 1-6, characterized in that, The risk assessment generation model includes: Hybrid encoders combine sparse and dense coding methods to encode input data, resulting in sparse and dense coding results. The hybrid retrieval module is used to retrieve results based on sparse and dense encoding, and outputs the most relevant search results. The reordering module is used to sort the most relevant search results according to document relevance, resulting in a sorted set of documents; The prompt generation module is used to output the power grid environmental risk assessment results based on the sorted document set and according to the preset prompt framework.

8. The method for assessing the power grid environmental risk of a live-line working robot according to claim 7, characterized in that, The most relevant search results are sorted according to document relevance, resulting in a sorted set of documents, including: A reciprocal ranking fusion strategy is adopted. After the most relevant search results are initially sorted by multiple documents, the bge-reranker-v2-m3 model is used to re-rank the single documents, and the documents most relevant to the query are ranked first, forming a sorted set of documents.

9. The method for assessing the power grid environmental risk of a live-line working robot according to claim 7, characterized in that, The default prompt frame is used for: In the sorted document collection, the documents that meet the requirements of the problem are identified as the most relevant documents; Extract key fragments from the most relevant documents and use these fragments to make inferences and explanations to obtain the power grid environmental risk assessment results.

10. A power grid environmental risk assessment device for a live-line working robot, characterized in that, include: The acquisition module is used to acquire preprocessed multi-source data; The multi-source data includes: three-dimensional point cloud object information of the power grid environment in which the robot operates, robot status information, and robot pre-action path information; The potential field modeling module is used to perform three-dimensional risk potential field modeling on the three-dimensional point cloud object information of the power grid environment of the preprocessed robot operation, and obtain a risk potential field model containing different object attributes. The assessment module is used to input the multi-source data and the risk potential field model into the risk assessment generation model to obtain the power grid environmental risk assessment results; The risk assessment generation model is obtained based on the following steps: A multi-level knowledge corpus of the power grid environment in which the robot operates is constructed based on the safety regulations for live-line working and industry standards; the multi-level knowledge corpus includes: a framework index layer, a clause and rule layer, and a case knowledge layer; Based on the multi-level knowledge corpus and using a progressive data augmentation strategy to train a large language model, a risk assessment generation model is obtained; the key layers of the large language model are injected with trainable rank decomposition matrices.

Images