CN-121982742-A - Power grid wiring diagram verification method based on multi-mode large model

CN121982742ACN 121982742 ACN121982742 ACN 121982742ACN-121982742-A

Abstract

The application discloses a power grid wiring diagram verification method based on a multi-mode large model, which comprises the steps of identifying a power grid wiring diagram image to be verified through a computer vision model to obtain preliminary primitive information, adding non-shielding vision marks to corresponding primitives in the power grid wiring diagram image based on the preliminary primitive information to generate a vision enhancement image, searching relevant target naming rules from a power grid naming standard knowledge base through an improved Monte Carlo tree search algorithm based on the preliminary primitive information, and identifying error information in the preliminary primitive information through the multi-mode large model based on the vision enhancement image and the target naming rules to complete power grid wiring diagram verification. The application realizes high-precision pattern-to-pattern conversion and logic verification of the wiring diagram.

Inventors

WANG ZHENCHUAN
YAN BINYU
LEI WENQIANG
HUANG CHEN

Assignees

四川大学

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (10)

1. The utility model provides a power grid wiring diagram verification method based on a multi-mode large model, which is characterized by comprising the following steps: identifying a power grid wiring diagram image to be verified through a computer vision model, and obtaining preliminary primitive information, wherein the preliminary primitive information comprises primitive categories and primitive positions; based on the preliminary primitive information, adding non-shielding visual marks to corresponding primitives in the power grid wiring diagram image, and generating a visual enhancement image; Based on the preliminary primitive information, retrieving related target naming rules from a power grid naming specification knowledge base through an improved Monte Carlo tree search algorithm, wherein the improved Monte Carlo tree search algorithm introduces a large language model in the search process to evaluate semantic relevance of child nodes of a current node, and differentially setting initial values and access times of the child nodes based on evaluation results; and based on the visual enhancement image and the target naming rule, identifying error information in the preliminary primitive information through a multi-mode large model, and completing verification of the power grid wiring diagram.
2. The method for verifying a power grid wiring diagram based on a multi-mode large model according to claim 1, wherein the adding of non-occlusion visual markers to corresponding primitives in the power grid wiring diagram image comprises: converting the power grid wiring diagram image into a gray scale image For each primitive, calculating the background whiteness score of each candidate region in a plurality of candidate regions outside the boundary frame of the primitive, wherein the calculation formula of the background whiteness score is as follows: Wherein, the The background white space is scored for the purpose of, For the gray value of the image at that pixel, For the width of the visual indicia, Height for visual indicia; and selecting the candidate area with the highest score to place the visual mark.
3. The method for verifying a power grid wiring diagram based on a multi-mode large model according to claim 1, wherein the improved monte carlo tree search algorithm utilizes a large language model to rank semantic relevance of child nodes of a current node in an initialization stage of a search tree, and performs non-zero initialization on top k child nodes after ranking according to ranks, so that initial values and initial access times of child nodes ranked in an ith position satisfy: Wherein, the As a total number of child nodes, For the initial value of the product, In order to have access to the number of times, Rank the relevance to The child node index of the bit.
4. A multi-modal large model based grid wiring diagram verification method as in claim 3, wherein the improved monte carlo tree search algorithm selects child nodes to be expanded according to an upper confidence formula at a node selection stage, the formula being: Wherein, the As the current average value of the node, In order to explore the constants, For the total number of accesses by the parent node, For the number of accesses of the current child node, For the parent node of the node to be a parent node, Is its child node.
5. The method for verifying a power grid wiring diagram based on a multi-modal large model according to claim 1, wherein the multi-modal large model is obtained by training a intra-group phase strategy optimization algorithm, and a reward function used in training is a weighted sum of multidimensional verifiable topological rewards, and the reward function is defined as: Wherein, the As a function of the reward, In order to be in the bonus format, For the interval to identify the reward, Rewards are given to the hierarchical structure of the game, Rewards are given for the type of component, The number of components is rewarded for the number of the components, To the point of Is the weight coefficient of each dimension.
6. The method for verifying a power grid wiring diagram based on a multi-modal large model according to claim 1, wherein the identifying error information by the multi-modal large model specifically comprises performing a double-flow thinking chain reasoning process: the transverse comparison flow is used for comparing the number and types of the components at the same level in different electrical intervals, and judging that the components are not reported or redundant if the components are inconsistent; And (3) longitudinal checking flow, namely combining the target naming rule and the type of a position component in the middle of the adjacent interval, carrying out logic checking on the primitive category in the preliminary primitive information, and judging that the primitive category is misreported if the primitive category conflicts.
7. The multi-modal large model-based grid wiring diagram verification method as set forth in claim 1, further comprising generating a structured verification report including at least error types, associated primitive identifications, and error locations.
8. A multi-modal large model based grid wiring diagram verification device in accordance with any one of claims 1-7, wherein said device comprises: the information acquisition module is used for identifying a power grid wiring diagram image to be verified through a computer vision model to obtain preliminary primitive information, wherein the preliminary primitive information comprises primitive types and primitive positions; the visual enhancement module is used for adding non-shielding visual marks to corresponding primitives in the power grid wiring diagram image based on the preliminary primitive information to generate a visual enhancement image; The specification retrieval module is used for retrieving related target naming rules from a power grid naming specification knowledge base through an improved Monte Carlo tree search algorithm based on the preliminary primitive information, wherein the improved Monte Carlo tree search algorithm introduces a large language model to evaluate semantic relevance of a child node of a current node in the search process, and differentially sets the initial value and access times of the child node based on an evaluation result, and the specification retrieval module is used for performing the semantic relevance evaluation on the child node based on the evaluation result And the model verification module is used for recognizing error information in the preliminary primitive information through a multi-mode large model based on the visual enhancement image and the target naming rule, and completing power grid wiring diagram verification.
9. An electronic device comprising a processor, a memory and a computer program stored on the memory and executable on the processor, the processor implementing the multi-modal large model-based grid wiring diagram verification method of any one of claims 1 to 7 when the program is executed by the processor.
10. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, implements a multi-modal large model-based grid wiring diagram verification method as claimed in any one of claims 1 to 7.

Description

Power grid wiring diagram verification method based on multi-mode large model Technical Field The application relates to the technical field of artificial intelligence and natural language processing, in particular to a power grid wiring diagram verification method based on a multi-mode large model. Background With the deep advancement of the construction of a novel electric power system and the panoramic perception of digital twinning, the analog-digital integrated construction of a power grid graph becomes a key link for improving the dispatching operation and asset management efficiency. In the process of constructing a "one-piece" grid, the topology logic and device parameters of the core are mainly carried in a huge amount of storage wiring diagrams, and these documents usually exist in the form of unstructured grid images (such as JPG, PNG) or PDF scan pieces. Specifically, the primary wiring diagram of the plant not only comprises high-density electrical primitives (such as a breaker, a transformer and a transformer), but also defines a strict power grid topological structure and dispatch naming logic through complex connection lines and text labels. However, implementing automatic high-precision recognition of wiring diagrams faces serious robustness and generalization challenges, and the related art still has a number of shortcomings, specifically as follows: Prior art one The current automatic identification of the power grid station wiring diagram mainly depends on the combination of the traditional computer vision technology and the deep learning algorithm, and the core flow is that firstly, the electrical equipment graphic element is identified through the target detection model, the position information is obtained, then the text information such as the equipment number is extracted through the text detection and identification algorithm, and finally, the graphic association matching is completed based on the geometric position relation or the global optimization strategy. The technology has obvious defects that firstly, the generalization capability of the cross-station is weak, the model is influenced by the difference of drawing styles of different stations, the adaptability to 'domain difference' is poor, the false detection and omission of the graphic element easily occur when a new station drawing is processed, secondly, the technology lacks of electric topology logic understanding, only the geometrical distance is matched with a text and the graphic element, the mismatch of equipment numbers and the graphic element easily occurs under a compact drawing layout scene, and the generation accuracy of the graphic model is influenced. Two prior art In order to solve the problems, the related technology tries to introduce a universal multi-Mode Large Language Model (MLLM) and combine a retrieval enhancement generation (RAG) technology to assist in power grid pattern analysis and verification, and the process comprises the steps of inputting a wiring diagram image and basic information extracted by a traditional computer vision model, acquiring related specification fragments from a knowledge base through a vector retrieval technology, and inputting the specification fragments and the image into the multi-mode large model for verification. The technology has the defects that firstly, the general RAG dependency vector similarity retrieval is difficult to accurately position and adapt to the structural naming rule of the current wiring scene, secondly, the general multi-mode large model has insufficient symbol sensitivity to the professional extraction graphic element of the power grid, identification errors or illusions are easy to occur, thirdly, the difficulty of directly analyzing the complex topological connection relationship through vision is high, line intersection and crossing are difficult to distinguish, and topology reasoning errors are caused. Disclosure of Invention In order to solve the problems, the application provides a power grid wiring diagram verification method based on a multi-mode large model, which aims to solve the problems that a traditional computer vision model is poor in generalization and lacks of electric topology logic understanding in cross-station wiring diagram identification, and a general multi-mode large model has vision illusion and cannot be matched with structural dispatching naming standards accurately when a graphic element symbol is processed. The first aspect of the embodiment of the invention provides a power grid wiring diagram verification method based on a multi-mode large model, which comprises the following steps: identifying a power grid wiring diagram image to be verified through a computer vision model, and obtaining preliminary primitive information, wherein the preliminary primitive information comprises primitive categories and primitive positions; based on the preliminary primitive information, adding non-shielding visual marks to corresp