CN-122023968-A - Real scene and AIGC combination-based transformation defect image simulation method and system

CN122023968ACN 122023968 ACN122023968 ACN 122023968ACN-122023968-A

Abstract

The invention discloses a transformation defect image simulation method based on combination of a real scene and AIGC, which relates to the technical field of image simulation and comprises the steps of constructing a real transformer substation scene image sample library, containing transformer substation scene images under various equipment types and environmental conditions, constructing a transformer equipment real defect learning library, containing typical transformation defect images and corresponding defect types and morphological characteristics thereof, constructing a defect image generation model based on the real transformer substation scene image sample library and the transformer equipment real defect learning library, generating a preliminary defect simulation image in an input real scene image by utilizing a trained defect image generation model, carrying out authenticity optimization processing on the preliminary defect simulation image, and outputting the preliminary defect simulation image as a final transformation defect simulation image. The invention effectively solves the problems of scarcity of the transformation defect sample, low simulation fidelity and inadequate AIGC adaptation.

Inventors

XUE HAI
HU YANJIE
YU CONGCONG
WANG ZHEN
LIU ZIQUAN
LU YONGLING
LIU ZHENGYU
PAN JIANYA
YIN ZE
HU CHENGBO
LIU JIANJUN

Assignees

国网江苏省电力有限公司电力科学研究院
国网江苏省电力有限公司
江苏省电力试验研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20260409

Claims (10)

1. The transformation defect image simulation method based on combination of a real scene and AIGC is characterized by comprising the following steps of: Step1, constructing a real substation scene image sample library, and containing substation scene images under various equipment types and environmental conditions; step2, constructing a real defect learning library of the transformer equipment, wherein typical transformer defect images and corresponding defect types and morphological characteristics are contained; Step3, constructing a defect image generation model based on a real substation scene image sample library and a substation equipment real defect learning library; Step4, generating a model by utilizing the trained defect image, and generating a preliminary defect simulation image in the input real scene image; Step5, performing authenticity optimization processing on the preliminary defect simulation image, and outputting the preliminary defect simulation image as a final transformation defect simulation image.
2. The transformation defect image simulation method based on the combination of the real scene and AIGC as claimed in claim 1, wherein the construction of the real transformer substation scene image sample library comprises the following concrete steps: Respectively acquiring images of each power transformation device in a normal state under different environmental conditions; Labeling the acquired images, and screening out effective images with the application coefficients higher than a threshold value; And storing the effective images in a classified manner to form a real substation scene image sample library.
3. The transformation defect image simulation method based on the combination of the real scene and AIGC as claimed in claim 1, wherein the construction of the transformation equipment real defect learning library comprises the following steps: Collecting multisource original defect images to a real learning library of the power transformation equipment and preprocessing the multisource original defect images; and extracting a pixel-level outline Mask of the defect area from the original defect image as a morphological feature, and marking the defect type of the pixel-level outline Mask.
4. The transformation defect image simulation method based on the combination of the real scene and AIGC as claimed in claim 1, wherein the defect image generation model is constructed, and specifically comprises the following sub-steps: introducing a generating condition encoder based on a traditional AIGC model for processing defect generating conditions of different modes; the output of the generating condition encoder is sent to an adaptive fusion unit to generate a final condition control vector, and the final condition control vector is injected into a AIGC model backbone generating network; And combining the real substation scene image sample library and the real defect learning library of the power transformation equipment to perform segment training on the model.
5. The method for simulating a power transformation defect image based on a combination of a real scene and AIGC as defined in claim 4, wherein the generating condition encoder includes a scene structure branch, a defect morphology branch, and an environmental condition branch.
6. The method for simulating the power transformation defect image based on the combination of the real scene and AIGC as claimed in claim 4, wherein the adaptive fusion unit adopts an attention mechanism, and the fusion weight is generated by a small neural network.
7. The method for simulating the power transformation defect image based on the combination of the real scene and AIGC, which is characterized by combining a real substation scene image sample library and a power transformation equipment real defect learning library to train the model in a segmentation way, comprises the following concrete steps: a pre-training stage, namely fixing scene structure branches and environmental condition branches, and training defect morphology branches and a trunk generator by using defect image Mask and defect category labels in a real defect learning library; A full model fine tuning stage, namely thawing scene structure branches and environment condition branches by using double-library joint data and jointly training all branches; consistency strengthening training, in the training process, periodically calculating priori loss And shape loss Resampling and weighting the difficult samples with dominant loss items to strengthen the study of the model on the consistency of the physical laws; And when each batch of training is finished, calculating the average loss of all training samples in the batch, performing tuning iteration on the model parameters according to the average loss value, triggering early stop when the descending rate of the average loss reaches a threshold value, and storing the model version with the minimum loss value into the production environment.
8. The transformation defect image simulation method based on the combination of the real scene and AIGC as claimed in claim 1, wherein the training defect image generation model is used to generate a preliminary defect simulation image in the input real scene image, and specifically comprises the following sub-steps: Searching out a field Jing Tuxiang and an associated environmental condition label from a real substation scene image sample library according to the keywords; searching a corresponding defect image Mask from a real defect learning library of the transformer equipment according to the target defect type; inputting the retrieved scene image, the environment condition label, the target defect type and the defect image Mask into a trained defect image generation model to generate a preliminary defect simulation image.
9. The transformation defect image simulation method based on the combination of the real scene and AIGC as claimed in claim 1, wherein the preliminary defect simulation image is subjected to the authenticity optimization process, and specifically comprises the following sub-steps: decomposing the preliminary defect image and the corresponding original image into sub-bands with different scales by using a Laplacian pyramid; calculating an adaptive weight map according to the saliency of the preliminary defect image on different scales; carrying out self-adaptive fusion on sub-bands of each scale of the primary defect image and the original image based on the self-adaptive weight graph; and (3) reversely transforming the fused sub-bands with different scales to synthesize a final transformation defect simulation image.
10. The transformation defect image simulation system based on the combination of the real scene and AIGC is characterized by comprising a real transformer substation scene image sample library, a transformation equipment real defect learning library, a preliminary defect image generation module and a defect image optimization module; the real substation scene image sample library is used for storing substation scene images under various equipment types and environmental conditions; the real defect learning library of the power transformation equipment is used for storing images of typical power transformation defects and corresponding defect types and morphological characteristics of the images; the preliminary defect image generation module is used for generating a preliminary defect simulation image in the input real scene image by using the defect image generation model; And the defect image optimization module is used for carrying out authenticity optimization processing on the preliminary defect simulation image and outputting the preliminary defect simulation image as a final transformation defect simulation image.

Description

Real scene and AIGC combination-based transformation defect image simulation method and system Technical Field The invention relates to the technical field of image simulation, in particular to a transformation defect image simulation method and system based on combination of a real scene and AIGC. Background The transformer substation is used as a core hub of the power system, and the running state of equipment directly determines the safety and stability of power supply of a power grid. In the long-term service process of the power transformation equipment, typical defects such as insulator damage, overheat of a wiring terminal, sleeve crack and the like are easily generated due to factors such as environmental corrosion, mechanical loss, electrical aging and the like, and if the power transformation equipment is not detected and processed in time, equipment faults and even large-area power failure accidents can be caused. Therefore, the efficient and accurate transformation defect detection technology is a key for guaranteeing the reliable operation of the power grid, and the high-quality defect image sample is a core foundation for training a defect detection model (such as a deep learning model) and improving the detection precision. Currently, the acquisition of the power transformation defect image mainly depends on the shooting of field manual acquisition and inspection equipment (such as an unmanned aerial vehicle and a thermal infrared imager). However, the method has the remarkable limitations that on one hand, the occurrence of the transformer defects has randomness and low probability, most typical defects (such as short circuit of a transformer winding and partial discharge of GIS equipment) are sporadic only under specific working conditions, so that the acquisition difficulty of real defect images is high, the number of the real defect images is rare, the real scene sample library and the real defect learning library are prevented from being constructed due to the difficulty in covering defect scenes under different equipment types and different operating environments (such as high temperature, rain and snow and strong electromagnetic interference), and on the other hand, part of high-risk defects (such as insulation breakdown precursors of high-voltage equipment) can be accompanied with safety risks in the acquisition process, the supplement of defect samples is further limited, and the problems of single scene, insufficient defect types and insufficient characteristic representativeness of the existing sample library are caused. To solve the sample starvation problem, the prior art attempts to simulate a transformation defect image by means of artificial synthesis, such as manually adding defect markers to the real scene image based on a Photoshop or the like, or generating the defect image using a conventional image generation algorithm (e.g., the early version of GANs). However, the method has the obvious defects that the artificial synthesis mode depends on experience of operators, the fusion degree of textures, illumination and morphology of defects and a real scene is poor, the sense of reality of a simulated image is low, physical association of the actual defects with equipment structures and environment light shadows cannot be reflected, the traditional generation algorithm lacks deep learning of real defect characteristics of the transformer equipment, the generated defects often have the problems of morphological distortion (such as that crack trend of an insulator does not accord with a mechanical rule), poor scene suitability (such as that the infrared thermal image characteristics of high-temperature defects are not matched with real environment temperature difference) and the like, and the requirements of a defect detection model on a high-reality sample are difficult to meet. In recent years, AIGC (generating artificial intelligence) technology has made breakthrough progress in the field of image generation, has the capability of generating high-reality images based on mass data learning characteristics, and provides a new technical direction for transformation defect image simulation. However, the application of the current AIGC technology in the transformer defect scene is still in a blank stage, the current AIGC image generation research focuses on general scenes (such as characters and landscapes), the special scenes (such as equipment intensive layout, image noise under strong electromagnetic interference and special equipment appearance characteristics) of the transformer substation are not subjected to customized optimization, scene images conforming to the real environment of the transformer substation cannot be generated, meanwhile, the AIGC model is lack of learning library support specially aiming at the transformer real defects, the physical characteristics, morphological rules and spatial association relation with equipment of different defects are difficult to