Substation defect image simulation method and system based on real scene and AIGC combination

Abstract

The application discloses a power transformation defect image simulation method based on a real scene and AIGC, relates to the technical field of image simulation, and comprises the following steps: constructing a real power transformation station scene image sample library, which contains power transformation station scene images under various equipment types and environmental conditions; constructing a real power transformation equipment defect learning library, which contains typical power transformation defect images and corresponding defect types and morphological characteristics; constructing a defect image generation model based on the real power transformation station scene image sample library and the real power transformation equipment defect learning library; generating a preliminary defect simulation image in an input real scene image by using the trained defect image generation model; and performing authenticity optimization processing on the preliminary defect simulation image, so as to output a final power transformation defect simulation image. The application effectively solves the problems of a lack of power transformation defect samples, low simulation authenticity and insufficient AIGC adaptation.

Description

Technical Field

[0001] This invention relates to the field of image simulation technology, and in particular to a method and system for simulating substation defects based on a combination of real-world scenes and AIGC. Background Technology

[0002] As the core hub of the power system, the operating status of substations directly determines the safety and stability of the power grid. During long-term service, substation equipment is prone to typical defects such as insulator damage, terminal overheating, and bushing cracks due to environmental corrosion, mechanical wear, and electrical aging. Failure to detect and address these defects in a timely manner can lead to equipment failures or even large-scale power outages. Therefore, efficient and accurate substation defect detection technology is crucial for ensuring the reliable operation of the power grid, and high-quality defect image samples are the core foundation for training defect detection models (such as deep learning models) and improving detection accuracy.

[0003] Currently, the acquisition of substation defect images mainly relies on on-site manual collection and inspection equipment (such as drones and infrared thermal imagers). However, this method has significant limitations: on the one hand, the occurrence of substation defects is random and has a low probability. Most typical defects (such as transformer winding short circuits and partial discharges in GIS equipment) only occur sporadically under specific operating conditions, making it difficult to collect real defect images and resulting in a small number of images. This makes it difficult to cover defect scenarios under different equipment types and different operating environments (such as high temperature, rain and snow, and strong electromagnetic interference), hindering the construction of a real-scene sample library and a real-defect learning library. On the other hand, some high-risk defects (such as precursors to insulation breakdown in high-voltage equipment) may be accompanied by safety risks during the collection process, further limiting the supplementation of defect samples. This results in existing sample libraries having problems such as single-scene coverage, incomplete defect types, and insufficient feature representativeness.

[0004] To address the problem of insufficient samples, existing technologies attempt to simulate substation defect images through artificial synthesis. This includes manually adding defect markers to real-world images using tools like Photoshop, or generating defect images using traditional image generation algorithms (such as early versions of GANs). However, these methods have significant drawbacks: artificial synthesis relies on operator experience, resulting in poor integration of defect texture, lighting, and morphology with the real scene, leading to low realism and an inability to reflect the physical relationship between actual defects and equipment structure, ambient lighting, and shadows. Traditional generation algorithms lack deep learning of the real-world defect characteristics of substation equipment, often resulting in morphological distortions (e.g., insulator crack orientations not conforming to mechanical laws) and poor scene adaptability (e.g., infrared thermal imaging features of high-temperature defects not matching the actual temperature difference), making it difficult to meet the high-fidelity sample requirements of defect detection models.

[0005] In recent years, AIGC (Generative Artificial Intelligence) technology has made breakthrough progress in the field of image generation. It has the ability to learn features based on massive amounts of data and generate highly realistic images, providing a new technical direction for substation defect image simulation. However, the application of AIGC technology in substation defect scenarios is still in its infancy: existing AIGC image generation research focuses on general scenes (such as people and landscapes) and has not been customized for the special scenarios of substations (such as dense equipment layout, image noise under strong electromagnetic interference, and specific equipment appearance features), making it impossible to generate scene images that conform to the real substation environment; at the same time, there is a lack of learning libraries specifically for real substation defects, making it difficult for AIGC models to learn the physical characteristics, morphological patterns, and spatial relationships with equipment of different defects. This results in the inability to accurately generate typical substation defects, and even more so, the inability to achieve a natural integration of defects with real substation scenes. Ultimately, it is difficult to form effective substation defect simulation scenarios and provide high-quality samples for defect detection model training.

[0006] In summary, the current field of substation defect detection faces a triple dilemma: scarcity of real defect samples, low realism of existing simulation methods, and insufficient adaptability of AIGC technology. There is an urgent need for a substation defect image simulation method that can establish a real-scene sample library and a real defect learning library, and achieve efficient fusion of defects and real-scenes through AIGC technology, so as to break through the existing technical bottlenecks and provide key support for the development of substation defect detection technology. Summary of the Invention

[0007] This invention provides a method for simulating substation defect images based on a combination of real-world scenarios and AIGC (AIGC), comprising:

[0008] Step 1: Construct a sample library of real substation scene images, containing substation scene images under various equipment types and environmental conditions;

[0009] Step 2: Construct a real defect learning library for substation equipment, containing typical substation defect images and their corresponding defect types and morphological features;

[0010] Step 3: Construct a defect image generation model based on a real substation scene image sample library and a real defect learning library for power equipment.

[0011] Step 4: Using the trained defect image generation model, generate preliminary defect simulation images from the input real scene images;

[0012] Step 5: Perform realism optimization processing on the preliminary defect simulation image, and output the final substation defect simulation image.

[0013] The substation defect image simulation method based on real-world scenarios and AIGC, as described above, involves constructing a sample library of real substation scene images, specifically through the following sub-steps:

[0014] Images of each power equipment under normal operating conditions were collected under different environmental conditions.

[0015] The collected images are labeled, and valid images with an applicability coefficient higher than the threshold are selected.

[0016] Effective images are categorized and stored to form a sample image library of real substation scenes.

[0017] The substation defect image simulation method based on real-world scenarios and AIGC, as described above, involves constructing a real-world defect learning library for substation equipment, specifically through the following sub-steps:

[0018] Collect original defect images from multiple sources into a real learning library of substation equipment and perform preprocessing;

[0019] The pixel-level contour mask of the defect region is extracted from the original defect image as a morphological feature, and its defect type is labeled.

[0020] The substation defect image simulation method based on real-world scenarios and AIGC, as described above, involves constructing a defect image generation model, which specifically comprises the following sub-steps:

[0021] A generative condition encoder is introduced on the basis of the traditional AIGC model to handle the defect generation conditions of different modalities;

[0022] The output of the generating conditional encoder is fed into the adaptive fusion unit to generate the final conditional control vector, which is then injected into the AIGC model backbone generation network.

[0023] The model is trained in segments using a combined database of real substation scene images and a database of real defects in power equipment.

[0024] The substation defect image simulation method described above, which combines real-world scenarios with AIGC, involves segmented training of the model using a combined database of real substation scene image samples and a database of real substation equipment defects. This training is specifically divided into the following sub-steps:

[0025] Pre-training phase: Fix scene structure branches and environmental condition branches, use defect image masks and defect category labels from the real defect learning library to train defect morphology branches and backbone generators;

[0026] Full model fine-tuning stage: Using joint data from two libraries, unfreeze the scene structure branch and environment condition branch, and jointly train all branches;

[0027] Consistent reinforcement training: During the training process, the prior loss is calculated periodically. and shape loss Furthermore, the model resamples and weights the difficult samples where the loss terms dominate, thereby strengthening the model's learning of the consistency of physical laws.

[0028] At the end of each training batch, the average loss of all training samples in that batch is calculated. The model parameters are then fine-tuned iteratively based on the average loss value. When the rate of decrease of the average loss reaches a threshold, early stopping is triggered, and the model version with the smallest loss value is stored and deployed to the production environment.

[0029] The substation defect image simulation method based on real-world scenarios and AIGC, as described above, involves realism optimization processing of the initial defect simulation image, specifically divided into the following sub-steps:

[0030] The initial defect image and the corresponding original image were decomposed into sub-bands of different scales using the Laplacian pyramid.

[0031] An adaptive weight map is calculated based on the saliency of the initial defect image at different scales.

[0032] Adaptive fusion of sub-bands at each scale between the preliminary defect image and the original image is performed based on an adaptive weight map.

[0033] The fused sub-bands of each scale are inversely transformed to synthesize the final simulated image of substation defects.

[0034] The present invention also provides a substation defect image simulation system based on real scene and AIGC, including: a real substation scene image sample library, a real defect learning library of substation equipment, a preliminary defect image generation module, and a defect image optimization module;

[0035] A sample library of real substation scene images, used to store substation scene images under various equipment types and environmental conditions;

[0036] A real-world defect learning library for power equipment, used to store images of typical power equipment defects and their corresponding defect types and morphological features;

[0037] The preliminary defect image generation module is used to generate preliminary defect simulation images from input real scene images using a defect image generation model;

[0038] The defect image optimization module is used to perform realism optimization processing on the preliminary defect simulation image, and the output is the final substation defect simulation image.

[0039] The beneficial effects achieved by this invention are as follows: It effectively solves the problems of scarce substation defect samples, low simulation realism, and insufficient AIGC adaptation. By constructing a real-scene sample library and a defect learning library, and combining a multi-branch conditional control AIGC model, it ensures that the location and shape of defects conform to physical laws; further, through Laplace pyramid optimization, it enhances the realism of defect-scene fusion, provides high-quality samples for the defect detection model, assists in the accurate detection of power grid defects, and ensures the safety and stability of the power grid. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0041] Figure 1 This is a flowchart of a substation defect image simulation method based on real-world scenarios and AIGC, provided in Embodiment 1 of this application. Detailed Implementation

[0042] 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 some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1:

[0044] like Figure 1 As shown, Embodiment 1 of this application provides a method for simulating substation defect images based on a combination of real-world scenarios and AIGC, including:

[0045] Step S110: Construct a sample library of real substation scene images, containing substation scene images under various equipment types and environmental conditions;

[0046] Constructing a sample image library of real substation scenes covering various equipment types and environmental conditions includes the following sub-steps:

[0047] Step S111: Collect images of each power equipment under normal operating conditions under different environmental conditions;

[0048] The types of substation equipment to be acquired include, but are not limited to, transformers, insulators, circuit breakers, disconnect switches, instrument transformers, surge arresters, bushings, and terminals. Image acquisition equipment includes, but is not limited to, visible light cameras, infrared thermal imagers, drone inspection systems, and handheld shooting devices. For the same type of equipment, images must be acquired from multiple angles (front view, side view, top view, and bottom view) to cover its typical structural and appearance features. The image acquisition environment must meet the following conditions:

[0049] Different lighting conditions: sunny day, cloudy day, night (auxiliary lighting), dusk;

[0050] Different weather conditions: rain, snow, fog, strong wind;

[0051] Different seasons and temperature conditions: high temperatures in summer, low temperatures in winter, and periods of significant temperature differences;

[0052] Different electromagnetic environment areas: strong electromagnetic interference area, weak interference area.

[0053] Step S112: Label the acquired images and select valid images with an applicability coefficient higher than the threshold;

[0054] The annotation information includes: equipment type, shooting time, environmental conditions, and shooting angle; the applicability coefficient quantifies the applicability of the current image to the substation defect simulation scenario from three dimensions: image quality, image content integrity, and domain specificity. Its calculation formula is expressed as:

[0055] Where S is the applicability coefficient calculation result, , , These are image quality items. Image content complete item Special items in the field The weight, , , , It is a function that calculates the variance of the Laplacian operator for image I and performs Min-Max normalization. Where i is the number of pixels with grayscale value i, and N is the total number of pixels. The area occupied by the device in the current image. This represents the total area of ​​the current image. It is the average visibility of the key structures of the equipment (the average of the confidence scores of each key structure output by the target detection model). ), This is a meteorological condition weight. If the image was captured under special meteorological conditions (such as rain, snow, or fog), it is assigned a higher fixed value because these conditions are difficult to perfectly simulate with CGI and are extremely valuable for enhancing the generalization ability of the AIGC model. Under normal weather conditions, this value is 0. This is an electromagnetic interference noise weight. For images acquired in areas with strong electromagnetic interference, the noise power in the high-frequency region is calculated and normalized. The more pronounced the noise characteristics, the better. The higher the value, , , , These are adjustable weighting coefficients, and + =1, + =1.

[0056] If the applicability coefficient S of the current image is higher than the preset threshold, it is considered a valid image; otherwise, it is filtered out in this step and will not participate in subsequent implementation steps.

[0057] Step S113: Classify and store the valid images to form a sample image library of real substation scenes;

[0058] The system categorizes and stores valid images according to their annotation information (device type, environmental conditions, and shooting angle) in multiple dimensions; it also builds a retrieval interface to support quick retrieval of image samples by keywords such as device type, environmental conditions, and shooting angle; and it regularly expands and updates the sample library to include images collected under new device types or new environmental conditions.

[0059] Step S120: Construct a real defect learning library for substation equipment, containing typical substation defect images and their corresponding defect types and morphological features;

[0060] Construct a real-world defect learning library containing typical substation defects. This library provides high-quality supervision information for subsequent AIGC models to learn the physical morphology and spatial patterns of defects. The specific steps include the following:

[0061] Step S121: Collect multi-source original defect images into the real learning library of substation equipment and perform preprocessing;

[0062] We collect real defect image data of power equipment from various sources, including historical maintenance reports and archives, online monitoring systems, and industry shared databases. We then perform preprocessing on the collected image data, such as format standardization and quality enhancement.

[0063] Step S122: Extract the pixel-level contour mask of the defect region from the original defect image as a morphological feature, and label its defect type;

[0064] A team of image annotators was formed to use image segmentation tools (such as LabelMe and CVAT) to accurately delineate the pixel-level outline mask of the defect area (hereinafter referred to as the defect image mask), add defect type labels to the image, and record the device type in the image for retrieval.

[0065] Step S130: Construct a defect image generation model based on a real substation scene image sample library and a real defect learning library for power equipment;

[0066] The core of the defect image generation model lies in its multi-branch conditional control architecture and physical law consistency constraint mechanism, which enables it to deeply understand the structural characteristics and ambient lighting conditions of the substation scene, and accurately learn the physical morphological laws of defects, thereby achieving high-fidelity and rational integration of defects and the scene; specifically, it includes the following sub-steps:

[0067] Step S131: Introduce a generative condition encoder based on the traditional AIGC model to handle the defect generation conditions of different modalities;

[0068] The generating condition encoder includes the following three network branches:

[0069] Scene structure branch: The input consists of a scene image extracted from a real substation scene image sample library and a target defect type label (i.e., the defect type label to be generated). After passing through the defect probability prediction network, the output is a prior probability map P of the defect location. location The defect probability prediction network is trained using a semantic segmentation network as the base model. It takes the original defect images from the real defect learning library of substation equipment as input and the corresponding defect masks as training targets, and is used to guide the location and geometric context of defect generation.

[0070] Defect Morphology Branch: The input consists of a defect image Mask extracted from a real defect learning database of substation equipment and its corresponding defect type label. An encoder network fuses the geometric morphology information of the Mask with the defect type label, and encodes the output morphological feature tensor F of the defect. morp This is used to precisely control the shape, type, and fine structure of the generated defects; the encoder network encodes the mask image into a feature vector F through a lightweight CNN encoder. mask The defect type label is converted into a word vector F through an embedding layer. label splicing F mask and F label Then, the final tensor F is output through fusion and dimensionality transformation using a multilayer perceptron. morp .

[0071] The environmental condition branch includes an embedding layer and a projection layer. The input is the environmental condition label corresponding to the scene image. The embedding layer maps the input label into a high-dimensional embedding vector, and the projection layer projects the embedding vector onto a vector space that matches the feature dimensions of other branches. The output is the environmental condition feature vector F. envi It is used to control the apparent properties of generated defects (such as the wetness of defects in rainy weather, and the thermal radiation effect of defects in high temperature).

[0072] Step S132: The output of the generating conditional encoder is fed into the adaptive fusion unit to generate the final conditional control vector, and then injected into the AIGC model backbone generation network.

[0073] Generate the P output of the conditional encoder location F morp F envi The final conditional control vector C is generated through fusion via an adaptive fusion unit. fused , will C fused The cross-attention layer is inserted into each layer (including downsampling block, intermediate block, and upsampling block) of the original AIGC model as a key and value, and the query Q is the feature map of that layer; through the attention mechanism, each operation of the AIGC model can "pay attention" to the scene structure, defect morphology and environmental information contained in the condition vector, thereby achieving pixel-level precise control; the adaptive fusion unit adopts the attention mechanism, and its fusion weights are generated by a small neural network.

[0074] The improved AIGC model is stored as a defect image generation model, and its loss function L is given for an input sample. total for:

[0075] ,in It is an input image of a real scene. This is the injected conditional control vector. A generator for generating defect image models. The discriminator for the defect image generation model. , These are adjustable weighting coefficients. , To generate a prior probability map of defect locations output by the scene structure branch in the conditional encoder. The mask image output after the defect simulation image generated for the model is processed by the defect probability prediction network.

[0076] x and y are the horizontal and vertical coordinate indices of the pixels in the Mask image, respectively, and W and H are the width and height of the Mask image, respectively. It is the pixel weight of the pixel coordinate (x, y) (higher weight is given to pixels at the edge of the defect, and vice versa for those inside). This represents the pixel value at coordinates (x, y) in the input defect image Mask. The pixel value at coordinates (x, y) in the defect simulation image Mask.

[0077] Step S133: Combine the real substation scene image sample library and the real defect learning library of power equipment to train the above model in segments;

[0078] Pre-training phase: Fix scene structure branches and environmental condition branches, use defect image masks and defect category labels from the real defect learning library to train defect morphology branches and backbone generators, so that the model first learns to generate various types of high-fidelity isolated defect images based on the masks;

[0079] Full model fine-tuning phase: using joint data (scene images sampled from a database of real substation scene images). The image "and its labeled environmental conditions" are compared with the defect image "Mask" sampled from a real defect learning library of power equipment. (And defect type labels are randomly combined to form training samples), unfreeze the scene structure branch and environmental condition branch and train all branches together. In this stage, the model learns to integrate the defects of a specified form into the appropriate position of the specified scene according to the environmental conditions.

[0080] Consistency reinforcement training: During the training process, periodic calculations are performed. and Furthermore, the model resamples or weights the difficult samples where the loss terms dominate, thereby enhancing the model's learning of the consistency of physical laws.

[0081] At the end of each training batch, the average loss of all training samples in that batch is calculated. The model parameters are then fine-tuned iteratively based on the average loss value. When the rate of decrease of the average loss reaches a threshold, early stopping is triggered, and the model version with the smallest loss value is stored and deployed to the production environment.

[0082] Of course, a model evaluation phase can also be introduced to further ensure the model's generalization ability; no restrictions are placed here.

[0083] Step S140: Using the trained defect image generation model, generate preliminary defect simulation images from the input real scene images;

[0084] First, scene images and their associated environmental condition labels are retrieved from a real substation scene image sample library by keywords. Then, the corresponding defect image mask is retrieved from a real substation equipment defect learning library by target defect type. Subsequently, the retrieved scene image, environmental condition label, target defect type, and defect image mask are input into the trained defect image generation model to generate a preliminary defect simulation image.

[0085] Step S150: Perform realism optimization processing on the preliminary defect simulation image, and output the final substation defect simulation image;

[0086] Realism optimization processing refers to intelligently selecting which information to retain from the preliminary defect simulation image (hereinafter referred to as the preliminary defect image) and the original real scene image (hereinafter referred to as the original image) at different scales. Specifically, at low-frequency scales (lighting, color), the realism of the original scene is fully trusted; at high-frequency scales (details, texture), the defect details provided by the preliminary defect image need to be adaptively fused. This is divided into the following sub-steps:

[0087] Step S151: Use the Laplacian pyramid to decompose the preliminary defect image and the corresponding original image into sub-bands of different scales;

[0088] The high-frequency and low-frequency detail subbands at various scales are extracted from the two images. The low-frequency detail subbands are directly retained from the original image, while the high-frequency detail subbands are selectively retained through the following fusion step. The high-frequency detail subbands at the k-th scale extracted from the original image and the preliminary defect image are denoted as follows: , .

[0089] Step S152: Calculate the adaptive weight map based on the saliency of the preliminary defect image at different scales;

[0090] The adaptive weight map determines the proportion of information from the original image and the preliminary defect image in the final output sub-band at position (x, y). The formula for calculating the adaptive weight map at the k-th scale is expressed as:

[0091] in It represents the weight value at position (x, y) in the adaptive weight graph at the k-th scale. () is the Sigmoid function. It is the scale attenuation coefficient. It is the value at position (x, y) in the high-frequency detail subband at the k-th scale of the preliminary defect image. It is a minimum value that avoids the denominator being zero. It is the total number of elements in the high-frequency detail subband at the k-th scale of the preliminary defect image.

[0092] Step S153: Adaptively fuse the sub-bands of each scale of the preliminary defect image and the original image based on the adaptive weight map;

[0093] For the high-frequency detail subband at the k-th scale, the fusion formula is:

[0094] ,in Let be the value at position (x, y) in the high-frequency detail subband at the k-th scale after fusion. It is the value at position (x, y) in the high-frequency detail subband at the k-th scale of the original image.

[0095] Step S154: Perform inverse transformation on the fused sub-bands of each scale to synthesize the final simulated image of substation defects;

[0096] The fused high-frequency and low-frequency detail subbands at each scale are inversely transformed to synthesize the final substation defect simulation image I. final .

[0097] Example 2:

[0098] Embodiment 2 of this application provides a substation defect image simulation system based on the combination of real scene and AIGC, including: a sample library of real substation scene images, a learning library of real defects of substation equipment, a preliminary defect image generation module, and a defect image optimization module;

[0099] (1) A sample library of real substation scene images, used to store substation scene images under various equipment types and environmental conditions;

[0100] (2) A real defect learning library for substation equipment, used to store images of typical substation defects and their corresponding defect types and morphological characteristics;

[0101] (3) Preliminary defect image generation module, used to generate preliminary defect simulation images from the input real scene images using the defect image generation model;

[0102] (4) Defect image optimization module; used to perform realism optimization processing on the preliminary defect simulation image and output the final substation defect simulation image.

[0103] Corresponding to the above embodiments, the present invention provides a computer storage medium, including: at least one memory and at least one processor;

[0104] The memory is used to store one or more program instructions;

[0105] The processor is used to run one or more program instructions to execute a substation defect image simulation method based on real-world scenarios and AIGC.

[0106] Corresponding to the above embodiments, this embodiment of the invention provides a computer-readable storage medium containing one or more program instructions, which are executed by a processor to provide a substation defect image simulation method based on a combination of real-world scenarios and AIGC.

[0107] The embodiments disclosed in this invention provide a computer-readable storage medium storing computer program instructions. When the computer program instructions are executed on a computer, the computer performs the above-described method for simulating substation defect images based on a combination of real-world scenarios and AIGC.

[0108] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0109] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0110] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0111] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0112] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0113] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0114] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0115] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for simulating substation defect images based on a combination of real-world scenarios and AIGC (AIGC), characterized in that, include: Step 1: Construct a sample library of real substation scene images, containing substation scene images under various equipment types and environmental conditions; Step 2: Construct a real defect learning library for substation equipment, which contains typical substation defect images and their corresponding defect types and morphological features; Step 3: Construct a defect image generation model based on a real substation scene image sample library and a real defect learning library of power equipment. Step 4: Using the trained defect image generation model, generate preliminary defect simulation images from the input real scene images; Step 5: Perform realism optimization processing on the preliminary defect simulation image, and output the final substation defect simulation image; The construction of a defect image generation model consists of the following sub-steps: A generative condition encoder is introduced on the basis of the traditional AIGC model to handle the defect generation conditions of different modalities; The output of the generating conditional encoder is fed into the adaptive fusion unit to generate the final conditional control vector, which is then injected into the AIGC model backbone generation network. The above model is trained in segments by combining a real substation scene image sample library and a real defect learning library of power equipment. The above model is trained in segments using a combined database of real substation scene images and a database of real defects in power equipment. This process involves the following sub-steps: Pre-training phase: Fix scene structure branches and environmental condition branches, use defect image masks and defect category labels from the real defect learning library to train defect morphology branches and backbone generators; Full model fine-tuning stage: Using joint data from two libraries, unfreeze the scene structure branch and the environment condition branch, and jointly train all branches; Consistent reinforcement training: During the training process, the prior loss is calculated periodically. and shape loss Furthermore, the model resamples and weights the difficult samples where the loss terms dominate, thereby strengthening the model's learning of the consistency of physical laws. At the end of each training batch, the average loss of all training samples in that batch is calculated, and the model parameters are fine-tuned iteratively based on the average loss value. When the rate of decrease of the average loss reaches the threshold, early stopping is triggered, and the model version with the smallest loss value is stored and put into the production environment. The initial defect simulation image undergoes realism optimization processing, which is specifically divided into the following sub-steps: The initial defect image and the corresponding original image were decomposed into sub-bands of different scales using the Laplacian pyramid. An adaptive weight map is calculated based on the saliency of the initial defect image at different scales. Adaptive fusion of sub-bands at each scale between the preliminary defect image and the original image is performed based on an adaptive weight map. The fused sub-bands of each scale are inversely transformed to synthesize the final simulated image of substation defects.

2. The substation defect image simulation method based on real-world scenarios and AIGC as described in claim 1, characterized in that, Building a sample image library of real substation scenes involves the following sub-steps: Images of each power equipment under normal operating conditions were collected under different environmental conditions. The acquired images are labeled, and valid images with an applicability coefficient higher than the threshold are selected. Effective images are categorized and stored to form a sample image library of real substation scenes.

3. The substation defect image simulation method based on real-world scenarios and AIGC as described in claim 1, characterized in that, The construction of a real-world defect learning library for power equipment involves the following sub-steps: Collect original defect images from multiple sources into a real learning library of power equipment and perform preprocessing; The pixel-level contour mask of the defect region is extracted from the original defect image as a morphological feature, and its defect type is labeled.

4. The substation defect image simulation method based on real-world scenarios and AIGC as described in claim 1, characterized in that, The generation condition encoder includes a scene structure branch, a defect morphology branch, and an environmental condition branch.

5. The substation defect image simulation method based on real-world scenarios and AIGC as described in claim 1, characterized in that, The adaptive fusion unit employs an attention mechanism, and its fusion weights are generated by a small neural network.

6. The substation defect image simulation method based on real-world scenarios and AIGC as described in claim 1, characterized in that, Using a trained defect image generation model, preliminary defect simulation images are generated from input real-world scene images. This process involves the following sub-steps: Retrieve scene images and their associated environmental condition tags from a database of real substation scene images using keywords; Retrieve the corresponding defect image Mask from the real defect learning library of power equipment according to the target defect type; The retrieved scene image, environmental condition label, target defect type, and defect image mask are input into the trained defect image generation model to generate a preliminary defect simulation image.

7. A substation defect image simulation system based on real-world scenarios and AIGC (AIGC), characterized in that, The substation defect image simulation system is used to implement the substation defect image simulation method according to any one of claims 1-6 above; The substation defect image simulation system includes: a sample library of real substation scene images, a learning library of real defects in substation equipment, a preliminary defect image generation module, and a defect image optimization module. A sample library of real substation scene images, used to store substation scene images under various equipment types and environmental conditions; A real-world defect learning library for power equipment, used to store images of typical power equipment defects and their corresponding defect types and morphological features; The preliminary defect image generation module is used to generate preliminary defect simulation images from input real scene images using a defect image generation model; The defect image optimization module is used to perform realism optimization processing on the preliminary defect simulation image, and the output is the final substation defect simulation image.

Images