CN-122022771-A - Substation intelligent patrol data analysis method based on electric power large model

CN122022771ACN 122022771 ACN122022771 ACN 122022771ACN-122022771-A

Abstract

The application relates to the field of power data analysis, and particularly discloses an intelligent substation inspection data analysis method based on a power large model, which comprises the steps of capturing fine texture features of the surface of equipment by using a special visual encoder of power, and associating the visual features with static attributes of the equipment; and then, filtering noise and guiding visual attention points by using the searched authoritative knowledge to construct a multi-mode prompt with precisely aligned semantics, and finally driving a large model to generate a structured diagnosis report with physical fact basis and industry specification constraint, thereby effectively solving the problems of unexplainable analysis result and low accuracy.

Inventors

WANG JUNYI
GUO YA
Cheng Nuannuan
LI ZIJING
WANG XIAONA
LIU YOUNING
LI DONGHAI

Assignees

国网河南省电力公司信息通信分公司
河南九域腾龙信息工程有限公司
河南先瑞能源科技集团有限公司

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (7)

1. The intelligent substation inspection data analysis method based on the large electric power model is characterized by comprising the following steps of: Acquiring static attribute text, a target image, a unique device identifier and telemetry data; Fine-grained visual feature extraction based on power special ViT is carried out on the target image so as to obtain a visual feature vector; Performing visual cue-based associated knowledge retrieval on the visual feature vector and the static attribute text to obtain a retrieved knowledge context; performing vision-knowledge bidirectional attention alignment and prompt construction on the vision feature vector, the retrieved knowledge context and the telemetry data to obtain an aligned multi-modal prompt; And carrying out large-model-based structured reasoning on the aligned multi-mode prompt to obtain an original diagnosis report.
2. The method for analyzing intelligent patrol data of transformer substation based on large electric power model according to claim 1, wherein obtaining static attribute text, target image, unique device identification and telemetry data comprises: Acquiring an original image stream, a unique device identifier and telemetry data; Extracting key frames based on definition evaluation from the original image stream to obtain a target image; and taking the unique equipment identifier as an index key, and executing query operation in an equipment ledger sheet of the database to obtain the static attribute text.
3. The method for analyzing intelligent patrol data of transformer substation based on large electric power model according to claim 1, wherein the step of extracting fine-grained visual feature based on special ViT electric power for the target image to obtain visual feature vector comprises the steps of: Performing image block projection and position coding embedding on a target image to obtain an image block feature sequence containing position information; the image block feature sequence containing the position information is subjected to self-attention based deep feature coding to obtain a visual feature vector.
4. The method for analyzing intelligent patrol data of transformer substation based on large electric power model according to claim 1, wherein performing visual cue-based associated knowledge retrieval on the visual feature vector and the static attribute text to obtain the retrieved knowledge context comprises: inputting the static attribute text into a pre-trained text encoder to obtain a text attribute vector; Performing multi-mode fusion projection on the text attribute vector and the visual feature vector to obtain a search query vector; Vector space similarity calculation is carried out on the search query vector and all knowledge item key vectors stored in the database so as to obtain a Top-K knowledge item set; The Top-K knowledge item set is structured packaged to obtain a retrieved knowledge context.
5. The method for analyzing intelligent patrol data of transformer substation based on large electric power model according to claim 4, wherein performing multi-mode fusion projection on text attribute vector and visual feature vector to obtain search query vector comprises: Performing attribute-guided fine-grained visual attention aggregation on the text attribute vector and the visual feature vector to obtain an attribute-enhanced visual vector; performing abnormal feature decoupling based on orthogonal projection on the attribute-enhanced visual vector and the text attribute vector to obtain a visual residual vector; And performing context-residual adaptive gating fusion on the visual residual vector and the text attribute vector to obtain a search query vector.
6. The method for analyzing intelligent patrol data of a transformer substation based on a large electric power model according to claim 5, wherein the performing abnormal feature decoupling based on orthogonal projection on the attribute-enhanced visual vector and the text attribute vector to obtain a visual residual vector comprises performing abnormal feature decoupling based on orthogonal projection on the attribute-enhanced visual vector and the text attribute vector according to the following formula: Wherein, the The visual vector is enhanced for the attribute, In the form of a text attribute vector, Representing the modular length of the text attribute vector, Representing the calculated position The projection in the direction is such that, Is a visual residual vector.
7. The method for analyzing intelligent patrol data of a transformer substation based on a large electric power model according to claim 1, wherein performing vision-knowledge bidirectional attention alignment and prompt construction on the vision feature vector, the retrieved knowledge context and the telemetry data to obtain an aligned multi-modal prompt comprises: inputting telemetry data into a multi-layer perceptron to carry out vectorization coding so as to obtain telemetry embedded vectors; Performing bidirectional cross-attention interaction on the visual feature vector and the retrieved knowledge context to obtain bidirectional alignment features; and constructing the aligned multi-mode prompt based on the bidirectional alignment feature and the telemetering embedded vector.

Description

Substation intelligent patrol data analysis method based on electric power large model Technical Field The application relates to the field of power data analysis, in particular to an intelligent substation inspection data analysis method based on a large power model. Background With the comprehensive promotion of smart power grid construction, the safety and stability of the running state of the transformer substation serving as a hub of a power system are important. In order to overcome the defects of low efficiency, many blind areas, large influence of bad weather and the like of the traditional manual inspection, an intelligent inspection system constructed by utilizing an inspection robot, an unmanned aerial vehicle and a fixed camera is widely applied, and massive multi-mode inspection data are generated. The intelligent substation inspection data analysis scheme based on the electric power large model is constructed, and transition from image perception to fault diagnosis report generation is realized through strong semantic understanding and reasoning capability of the large model, so that the method is a necessary trend of solving the bottleneck of mass data analysis, improving the intelligent level of operation and inspection and ensuring the safe operation of power grid equipment. However, the existing intelligent patrol data analysis scheme of the transformer substation still has significant technical limitations in practical application. The traditional mainstream scheme relies on general computer vision algorithms (such as YOLO series) to locate equipment and defects, but the output is limited to coordinate frames and class labels, and a structural diagnosis report meeting the strict specifications of the power industry cannot be generated. The emerging common multi-modal large model in recent years has graphics context generation capability, but often faces severe challenges of visual-semantic alignment bias in the face of highly specialized power scenarios. The method is characterized in that the general model lacks depth understanding of fine textures and physical states of the power equipment and is not strictly constrained by industry standard knowledge, so that the generated inspection report often has illusions, namely the model can misjudge the equipment states according to image background or the generated defect description is inconsistent with real safety regulations. That is, the subtle differences between the device body features and the defective abnormal features cannot be distinguished, resulting in the generated diagnostic report often referencing erroneous protocol terms or giving recommendations that are not compliant with safety regulations, and it is difficult to meet the stringent requirements of industrial field applications. The problem that visual features cannot be accurately mapped to professional semantic space makes it difficult for the prior art to meet the industrial-level requirements of substation inspection on high accuracy and high interpretability of diagnosis results. Therefore, an optimized substation intelligent patrol data analysis scheme based on a large power model is desired. Disclosure of Invention The present application has been made to solve the above-mentioned technical problems. The embodiment of the application provides an intelligent substation inspection data analysis method based on a large electric power model. According to one aspect of the application, there is provided a substation intelligent patrol data analysis method based on a large electric power model, comprising: Acquiring static attribute text, a target image, a unique device identifier and telemetry data; Fine-grained visual feature extraction based on power special ViT is carried out on the target image so as to obtain a visual feature vector; Performing visual cue-based associated knowledge retrieval on the visual feature vector and the static attribute text to obtain a retrieved knowledge context; performing vision-knowledge bidirectional attention alignment and prompt construction on the vision feature vector, the retrieved knowledge context and the telemetry data to obtain an aligned multi-modal prompt; And carrying out large-model-based structured reasoning on the aligned multi-mode prompt to obtain an original diagnosis report. The intelligent substation inspection data analysis method based on the electric power large model comprises the steps of capturing fine texture features of the surface of equipment by utilizing an electric power special visual encoder, associating the visual features with static attributes of the equipment, introducing a feature decoupling mechanism, carrying out mathematical separation on inherent attributes representing equipment bodies and abnormal residual errors representing faults in a feature space, accurately searching matched industry standards and historical cases in a knowledge base according to the mathematical separation, then carryi