Electrical equipment defect detection method based on image recognition

Abstract

The invention discloses a power equipment defect detection method based on image recognition, and belongs to the technical field of power equipment defect detection, and the method comprises the steps: carrying out the defect simulation based on physical mechanism driving according to an equipment three-dimensional model and physical field simulation parameters, and obtaining a defect simulation data set; according to the defect simulation data set and the real inspection data, training a cross-modal deep learning network based on physical law constraint to obtain a defect identification model; performing time-space diagram neural network modeling according to the historical time sequence inspection data and the defect identification model to obtain a state evolution model; and inputting inspection data acquired in real time into the equipment health state evolution model, and performing online reasoning to obtain a defect detection result. The problems that an existing electrical equipment defect detection method excessively depends on scarce real defect samples, the generalization ability for complex working conditions is weak, and the defect evolution trend prediction ability is lacked are solved.

Description

Technical Field

[0001] This invention relates to the technical field of power equipment defect detection, and more particularly to a power equipment defect detection method based on image recognition. Background Technology

[0002] In recent years, image recognition-based defect detection technologies for power equipment have primarily relied on deep learning methods, especially convolutional neural networks and their variants. These technologies train deep learning models to automatically identify defect features in visible light, infrared, or ultraviolet images, such as insulator bursts, conductor overheating, and hardware corrosion. To improve model performance, existing technologies typically employ strategies such as data augmentation, transfer learning, or ensemble learning to expand the training dataset. Furthermore, some advanced solutions attempt to introduce multimodal fusion techniques, fusing visible light images with infrared thermal images at the feature level or decision level, aiming to obtain more reliable detection results than single-modal methods. These methods constitute the core technical route for current intelligent inspection of power equipment, alleviating the burden of manual inspection to some extent.

[0003] However, the aforementioned existing technologies still have significant drawbacks. First, their performance is highly dependent on a large number of labeled real-world defect samples. However, power equipment defects are characterized by low incidence and high risk, resulting in an extremely scarce pool of defect samples available for training. This makes it difficult for the model to effectively learn rare defect features, creating detection blind spots. Second, existing methods primarily focus on the classification and localization of static images, exhibiting weak generalization ability to complex on-site conditions such as lighting changes, weather conditions, and shooting angles, easily leading to false alarms and missed detections. Most importantly, existing technologies are essentially a form of "post-event identification," only able to determine whether the equipment is currently faulty, but unable to model the evolution of defects, lacking the ability to quantitatively assess the development trend of defect severity and predict remaining effective lifespan, thus failing to meet the strategic needs of power systems for predictive maintenance. These shortcomings severely restrict the application effectiveness and value of existing technologies in actual production. Summary of the Invention

[0004] The technical problem solved by this invention is that existing power equipment defect detection methods rely too heavily on scarce real defect samples, have weak generalization ability to complex operating conditions, and lack the ability to predict defect evolution trends.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: As a preferred embodiment of the image recognition-based power equipment defect detection method of the present invention, wherein: Based on the equipment's 3D model and physical field simulation parameters, a defect simulation dataset is obtained by performing defect simulation driven by physical mechanisms. Based on the defect simulation dataset and real inspection data, a cross-modal deep learning network was trained according to physical law constraints to obtain a defect recognition model; A spatiotemporal graph neural network model was constructed based on historical time-series inspection data and a defect identification model to obtain a state evolution model; The real-time collected inspection data is input into the equipment health status evolution model for online reasoning to obtain defect detection results.

[0006] Furthermore, it also includes, after inputting the real-time collected inspection data into the equipment health status evolution model for online reasoning, obtaining an uncertainty sample set by performing uncertainty estimation, and continuously optimizing the defect identification model and the status evolution model by incremental learning based on the uncertainty sample set and expert annotations.

[0007] Furthermore, based on the equipment's 3D model and physical field simulation parameters, a defect simulation dataset is obtained by performing defect simulation driven by physical mechanisms, including: A digital twin model of the target power equipment is obtained by performing a 3D modeling based on the 3D scanning data of the target power equipment. Based on the material properties, electrical load, and environmental parameters of the target power equipment and the digital twin model, a physical field coupling simulation is performed to simulate the physical field data during the defect evolution process. Based on the physical field data and the preset defect mapping rules, a visualization rendering is performed on the digital twin model of the target power equipment to generate an initial multimodal simulation image containing the defect morphology; the initial multimodal simulation image includes a visible light image, an infrared thermogram, and an ultraviolet discharge image; Based on the initial multimodal simulation image and the real inspection data, image fusion and enhancement processing are performed to obtain multimodal defect samples; The defect simulation dataset is constructed by labeling the multimodal defect samples with defect type, location, severity, and physical parameter labels.

[0008] Furthermore, the physical field coupling simulation includes thermal-electric coupling simulation, fluid-solid coupling simulation, and electrostatic field simulation; wherein, based on the material properties, electrical load, and environmental parameters of the digital twin model and the target power equipment, physical field coupling simulation is performed to simulate the physical field data during the defect evolution process, including: Based on the digital twin model, according to the geometry and material properties of the target power equipment, a thermo-electric coupling simulation is performed using applied electrical load and ambient temperature parameters to calculate the temperature field distribution data of the target power equipment under normal and defective states; the defective states include the state of increased resistance caused by connection point contact and the state of blocked internal carrier pathways; Based on the digital twin model, fluid-solid coupling simulation is performed according to the geometry, surface characteristics, ambient wind speed and particulate matter concentration parameters of the target power equipment to simulate the adsorption, accumulation process and distribution pattern of pollutant particles on the surface of the target power equipment, and output the surface pollutant density distribution data of the target power equipment. Based on the digital twin model, electrostatic field simulation is performed according to the geometry, material properties, applied voltage, and surface contamination density distribution data of the target power equipment. The electric field intensity distribution data on the surface of the target power equipment is calculated, and the potential occurrence area and intensity of corona discharge or arc discharge are determined based on the field emission principle. The temperature field distribution data, the surface contamination density distribution data, and the electric field intensity distribution data are integrated to form multiphysics field data in the defect evolution process.

[0009] Furthermore, based on the defect simulation dataset and real inspection data, a cross-modal deep learning network is trained according to physical constraints to obtain a defect identification model, including: The multimodal defect samples in the defect simulation dataset are mixed with real inspection data at a preset ratio and pixel-level registration is performed to construct a multimodal training dataset; each sample unit in the multimodal training dataset contains a registered visible light image, an infrared thermogram, and an ultraviolet discharge image. A cross-modal deep learning network with three independent branches is constructed. Each branch processes an image of a different modality and extracts features to obtain visible light features, infrared thermal features, and ultraviolet discharge features. The extracted visible light features, infrared thermal features, and ultraviolet discharge features are input into a physical law constraint module for feature fusion; the physical law constraint module is used to generate a physical attention weight map based on the preset prior knowledge of the device's physical structure and the visible light features, infrared thermal features, and ultraviolet discharge features. Using the multimodal training dataset, with defect classification and localization as the supervised tasks, the cross-modal deep learning network and the physical law constraint module are jointly trained until the model converges, thus obtaining the defect recognition model.

[0010] Furthermore, based on historical time-series inspection data and the defect identification model, a spatiotemporal graph neural network model is performed to obtain a state evolution model, including: A spatiotemporal graph is constructed based on the spatial topology of the target power equipment; wherein, each key component of the target power equipment is defined as a graph node, and the physical connections or functional associations between components are defined as graph edges; Historical time-series inspection data of each graph node is extracted and processed using the defect identification model to generate defect status feature vectors for each node at historical time points, forming a node feature time-series sequence. The spatiotemporal graph, the time series sequence of defect state feature vectors of each node, and the corresponding environmental parameter time series data are input into the spatiotemporal graph neural network model. With the goal of predicting the defect state features at future time points, the spatiotemporal graph neural network model is trained until it converges, thus obtaining the state evolution model that can predict the defect evolution trend. The spatiotemporal graph neural network model captures the spatial correlation between nodes through a graph attention mechanism and uses a temporal convolutional network or a recurrent neural network to learn the evolution of the defect state characteristics of each node over time. The output of the state evolution model includes a predicted value of the defect severity at a specific future time point and an estimated value of the remaining life of the equipment.

[0011] Furthermore, the real-time collected inspection data is input into the equipment health status evolution model for online reasoning to obtain defect detection results, including: The real-time collected inspection data, including visible light, infrared and ultraviolet modes, is preprocessed and registered to form multimodal input data at the current time point; The multimodal input data at the current time point is input into the defect identification model to extract the defect state feature vectors of each key component at the current time point; The feature vector sequence of the defect status of each key component at the current time point and the historical time point, the environmental parameter sequence, and the spatiotemporal diagram are all input into the state evolution model. Online reasoning is performed using the state evolution model to output defect detection results; The defect detection results include: classification and location information of the current defect type of each key component, quantitative assessment of the severity of the defect, prediction of the evolution trend of the defect in the future within a preset time period, and estimated value of the remaining effective life of the affected key components.

[0012] Furthermore, the step of performing online reasoning through the state evolution model and outputting defect detection results includes: After receiving the spatiotemporal graph, defect state feature vector sequence, and environmental parameter sequence from the state evolution model, the spatial dependency weights between key component nodes at the current time point are first calculated through the graph attention network layer inside the state evolution model. Based on the spatial dependency weights, the defect state feature vector sequence of each key component is modeled temporally using a temporal convolutional network or a recurrent neural network to analyze the evolution pattern of the defect state of each key component over time and predict the defect state feature vector at a future preset time point. The predicted future defect state feature vector is mapped to a specific quantitative value of defect severity, and the estimated remaining effective life of the affected key components is calculated based on the preset mapping relationship between defect severity and remaining effective life. Based on the spatial dependency weight analysis results, the current defect status, the predicted future defect evolution trend, and the estimated remaining effective lifetime, defect detection results are generated.

[0013] Furthermore, after inputting the real-time collected inspection data into the equipment health status evolution model for online inference, an uncertainty sample set is obtained by performing uncertainty estimation. Among them, results with a model inference confidence level higher than a preset model inference confidence level threshold are directly output as deterministic conclusions, while results with a confidence level lower than a preset threshold are marked as uncertain samples in the corresponding real-time collected inspection data, thus obtaining an uncertainty sample set.

[0014] Furthermore, incremental learning is performed to continuously optimize the defect identification model and state evolution model based on the uncertainty sample set and expert annotations, including: The uncertain sample set is pushed to the expert annotation platform to obtain the experts' annotations on the defect type, location and severity of the samples in the uncertain sample set, thus forming an annotated sample set; The labeled sample set is mixed with the multimodal training dataset used to train the defect recognition model according to a preset ratio to obtain the mixed dataset; Using the current parameters of the defect identification model as pre-training weights, the defect identification model is incrementally trained using the mixed dataset to update the network parameters of the defect identification model, thus obtaining the updated defect identification model. Using the updated defect identification model, the defect state feature vectors are re-extracted from the historical time-series inspection data to generate an updated node feature time-series sequence. Using the updated node feature time sequence as input and the current parameters of the state evolution model as pre-training weights, the state evolution model is incrementally trained to update the parameters of the spatiotemporal graph neural network of the state evolution model, thus obtaining the updated state evolution model. The updated defect identification model and the updated state evolution model were deployed to the online inference system to replace the original model, thus completing the optimization and iteration of the defect identification model and the state evolution model. The beneficial effects of this invention are: 1. This invention addresses the problems of existing power equipment defect detection methods, such as over-reliance on scarce real defect samples, weak generalization ability to complex operating conditions, and lack of defect evolution trend prediction ability, by training a cross-modal deep learning network using a defect simulation dataset driven by physical mechanisms and real inspection data, combined with the modeling ability of spatiotemporal graph neural networks for equipment state evolution. It achieves accurate real-time detection of power equipment defects and prediction of state evolution trends, improving the accuracy and timeliness of defect detection. Simultaneously, the training strategy based on physical constraints enhances the model's generalization ability under complex operating conditions. 2. This invention generates high-fidelity defect evolution data through multi-physics coupling simulation driven by a digital twin model. Combined with a training strategy constrained by the physical laws of real inspection data, it achieves physical attention weight fusion of visible light, infrared and ultraviolet multimodal features, which significantly improves the accuracy of defect type identification and spatial positioning accuracy, and solves the problem of easy missed diagnosis and misdiagnosis in traditional single-modal detection.

[0015] 3. This invention constructs a state evolution model based on the spatiotemporal graph modeling of equipment topology and the ability to capture the temporal evolution patterns of temporal convolutional or recurrent neural networks. This model can accurately predict the dynamic changes in the severity of defects in key components and their remaining effective lifespan, achieving a leapfrog upgrade from passive detection to proactive predictive maintenance. 4. This invention forms an uncertain sample set through a model inference confidence threshold screening mechanism, and constructs a closed-loop iterative system by combining an incremental learning strategy with expert annotation. This enables the defect identification model and the state evolution model to continuously evolve under complex working conditions, solving the problem of generalization ability decay caused by the solidification of traditional models, and ensuring the stability and reliability of the system in long-term operation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the basic process of a power equipment defect detection method based on image recognition, provided as an embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example 1

[0018] like Figure 1 As shown in the figure, this embodiment introduces a power equipment defect detection method based on image recognition, including: Step 1: Based on the equipment's 3D model and physical field simulation parameters, perform defect simulation based on physical mechanisms to obtain a defect simulation dataset.

[0019] This invention, based on physical mechanism-driven defect simulation, couples a three-dimensional model with multi-physics parameters to generate a high-fidelity defect dataset containing the evolution of multiple physics fields. It realizes multi-dimensional data characterization from defect morphology to physical parameters, solves the problem of the disconnect between traditional simulation data and actual defect evolution laws, and provides training samples with strong physical consistency and covering the entire defect evolution cycle for cross-modal deep learning.

[0020] Step 2: Based on the defect simulation dataset and real inspection data, train a cross-modal deep learning network to obtain a defect recognition model based on physical law constraints.

[0021] This invention employs a cross-modal deep learning network training strategy constrained by physical laws. Simulated data and real inspection data are jointly trained after pixel-level registration. By fusing multimodal features using a physical attention weight map, precise coupling of defect type identification and spatial localization is achieved. Simultaneously, the physical law constraint module effectively suppresses overfitting under complex working conditions, improving the generalization ability of the defect identification model in unknown scenarios. Step 3: Based on historical time-series inspection data and the defect identification model, perform spatiotemporal graph neural network modeling to obtain the state evolution model.

[0022] This invention is based on spatiotemporal graph neural network modeling of equipment spatial topology. It captures the spatial dependencies between nodes through graph attention mechanism and combines temporal convolutional network to learn the temporal evolution law of defect state features, and constructs a state evolution model. This realizes the dimensionality improvement from single-point defect detection to prediction of the evolution trend of the health status of the entire equipment, and provides a quantitative basis for predictive maintenance of remaining lifetime.

[0023] Step 4: Input the real-time collected inspection data into the equipment health status evolution model, perform online reasoning, and obtain defect detection results.

[0024] This invention integrates real-time multimodal inspection data with historical time-series features through the online reasoning mechanism of the state evolution model. It utilizes a spatiotemporal graph neural network to achieve integrated output of dynamic assessment of defect severity, prediction of future evolution trends, and estimation of remaining life, forming a full-chain real-time decision support, which significantly improves the timeliness and scientific nature of power equipment defect detection. Example 2

[0025] This embodiment describes the implementation steps of an image recognition-based power equipment defect detection method, including: Step 1: Based on the equipment's 3D model and physical field simulation parameters, perform defect simulation based on physical mechanisms to obtain a defect simulation dataset.

[0026] Step 1.1: Perform 3D modeling based on the 3D scanning data of the target power equipment to obtain a digital twin model of the target power equipment.

[0027] This embodiment constructs a digital twin model based on 3D scanning data to achieve accurate mapping of the geometric shape and topology of the target power equipment, ensuring the physical consistency of the simulation of the defect evolution process.

[0028] Step 1.2: Based on the material properties, electrical load, and environmental parameters of the target power equipment and the digital twin model, perform physical field coupling simulation to obtain physical field data during the defect evolution process.

[0029] This embodiment generates physical field data for the entire cycle of defect evolution through multi-physics coupling simulation, realizing a dynamic characterization of the entire chain from the initiation of micro-defects to the development of macro-faults, and providing interpretable evidence at the physical mechanism level for cross-modal deep learning networks.

[0030] Step 1.2.1: Based on the digital twin model, according to the geometry and material properties of the target power equipment, a thermo-electric coupling simulation is performed by applying electrical load and ambient temperature parameters to calculate the temperature field distribution data of the target power equipment under normal and defective states; the defective states include the state of increased resistance caused by connection point contact and the state of blocked internal carrier pathways.

[0031] This embodiment uses thermo-electric coupling simulation to quantify the three-dimensional temperature field distribution under normal / defective conditions through electrical load and ambient temperature parameters. It accurately captures the temperature rise characteristics of connection point contact defects and carrier path obstruction scenarios, providing a physical parameter verification benchmark for infrared thermographic defect identification.

[0032] Step 1.2.2: Based on the digital twin model, perform fluid-solid coupling simulation according to the geometry, surface characteristics, ambient wind speed and particulate matter concentration parameters of the target power equipment to simulate the adsorption, accumulation process and distribution pattern of pollutant particles on the surface of the target power equipment, and output the surface pollutant density distribution data of the target power equipment.

[0033] This embodiment uses fluid-solid coupling simulation to simulate the adsorption and accumulation process of dirt particles through environmental parameters, and outputs surface dirt density distribution data, providing a physical mechanism explanation of flow field-electric field coupling for locating corona / arc discharge regions in ultraviolet discharge images.

[0034] Step 1.2.3: Based on the digital twin model, perform electrostatic field simulation according to the geometry, material properties, applied voltage, and surface contamination density distribution data of the target power equipment, calculate the electric field intensity distribution data on the surface of the target power equipment, and determine the potential occurrence area and intensity of corona discharge or arc discharge based on the field emission principle.

[0035] This embodiment calculates the electric field intensity distribution based on surface contamination density and applied voltage parameters using electrostatic field simulation, and determines the potential corona discharge region by combining the field emission principle, thus providing physical constraints on the electric field-discharge correlation for ultraviolet modal defect detection.

[0036] Step 1.2.4: Integrate the temperature field distribution data, the surface contamination density distribution data, and the electric field intensity distribution data to form multiphysics field data in the defect evolution process.

[0037] This embodiment uses multi-physics data fusion to form a defect evolution data cube, realizing a complete mechanism characterization from a single physical field to multi-field coupling, enhancing the dimensional richness and physical realism of the simulation data.

[0038] Step 1.3: Based on the physical field data and the preset defect mapping rules, perform visualization rendering on the digital twin model of the target power equipment to generate an initial multimodal simulation image containing the defect morphology; the initial multimodal simulation image includes a visible light image, an infrared thermogram, and an ultraviolet discharge image.

[0039] This embodiment generates a multimodal simulation image containing visible light, infrared, and ultraviolet features based on the visualization rendering of physical field data and defect mapping rules. It realizes cross-modal mapping from physical parameter space to visual feature space, providing physically consistent multimodal training samples for cross-modal learning.

[0040] Step 1.4: Perform image fusion and enhancement processing based on the initial multimodal simulation image and the real inspection data to obtain multimodal defect samples.

[0041] This embodiment generates multimodal defect samples through image fusion and enhancement processing, which enhances the generalization adaptability of the samples under complex working conditions and improves the robustness of the model to real inspection environments.

[0042] Step 1.5: Label the multimodal defect samples with defect type, location, severity and physical parameter labels to construct the defect simulation dataset.

[0043] This embodiment performs four-dimensional annotation on multimodal defect samples to construct a defect simulation dataset containing physical parameter constraints, realizing bidirectional interpretable annotation from visual features to physical parameters, and providing training supervision signals with physical law constraints for cross-modal learning.

[0044] Step 2: Based on the defect simulation dataset and real inspection data, train a cross-modal deep learning network to obtain a defect recognition model based on physical law constraints.

[0045] Step 2.1: Mix the multimodal defect samples in the defect simulation dataset with the real inspection data according to a preset ratio, and perform pixel-level registration to construct a multimodal training dataset; each sample unit in the multimodal training dataset contains a registered visible light image, an infrared thermogram, and an ultraviolet discharge image.

[0046] This embodiment constructs a multimodal training dataset by hybrid registration of defect simulation data and real inspection data, realizing spatiotemporal alignment and feature complementarity between simulation data and real scene data, enhancing the model's generalization and adaptability to multimodal defect features under complex working conditions, and providing training samples driven by both physical parameters and visual features for cross-modal deep learning.

[0047] Step 2.2: Construct a cross-modal deep learning network with three independent branches. Each branch processes an image of a different modality and extracts features to obtain visible light features, infrared thermal features, and ultraviolet discharge features.

[0048] This embodiment constructs a three-branch cross-modal deep learning network to achieve independent feature extraction for visible light, infrared, and ultraviolet modes. By using modality-specific branches, the physical and semantic integrity of the original features of each modality is preserved, providing multi-dimensional and high-fidelity feature inputs for subsequent feature fusion and improving the comprehensiveness and fine-grainedness of defect feature expression.

[0049] Step 2.3: Input the extracted visible light features, infrared thermal features, and ultraviolet discharge features into a physical law constraint module for feature fusion; the physical law constraint module is used to generate a physical attention weight map based on the preset prior knowledge of the device's physical structure and the visible light features, infrared thermal features, and ultraviolet discharge features.

[0050] This embodiment generates an attention weight map based on prior knowledge of the physical structure of the equipment using a physical law constraint module, thereby achieving physical consistency fusion of multimodal features. The weight map guides the model to focus on key feature regions related to defects, enhancing the interpretability of the feature fusion process and its anti-interference ability in complex backgrounds.

[0051] Step 2.4: Using the multimodal training dataset, with defect classification and localization as the supervised task, jointly train the cross-modal deep learning network and the physical law constraint module until the model converges, and obtain the defect recognition model.

[0052] This embodiment trains a cross-modal network and a physical constraint module through a joint supervised task of defect classification and localization, thereby achieving end-to-end optimization of the defect recognition model. The collaborative training of the physical constraint module and the feature extraction branch improves the model's comprehensive recognition capabilities for defect type discrimination, spatial localization, and physical parameter correlation, ensuring the model's detection accuracy and robustness in practical applications.

[0053] Step 3: Based on historical time-series inspection data and the defect identification model, perform spatiotemporal graph neural network modeling to obtain the state evolution model.

[0054] Step 3.1: Construct a spatiotemporal graph based on the spatial topology of the target power equipment; wherein, each key component of the target power equipment is defined as a graph node, and the physical connections or functional associations between components are defined as graph edges.

[0055] This embodiment constructs a spatiotemporal graph based on the spatial topology of the equipment. Through the node-edge structure, it accurately depicts the physical connections and functional associations between key components, providing a spatial dependency modeling foundation for the spatiotemporal graph neural network that conforms to the actual operating logic of the equipment, and enhancing the collaborative analysis capability of the state evolution model for the overall health status of the equipment.

[0056] Step 3.2: Extract the historical time-series inspection data of each graph node, and process it using the defect identification model to generate the defect status feature vector of each node at each historical time point, forming a node feature time-series sequence.

[0057] This embodiment utilizes a defect identification model to process historical time-series inspection data, generating a time-series sequence of node-level defect state feature vectors. This achieves semantic dimensionality enhancement from raw multimodal data to a high-dimensional feature space, providing a time-series input with physical semantic consistency for the spatiotemporal graph neural network and improving the state evolution model's ability to capture the fine-grained laws of defect evolution.

[0058] Step 3.3: Input the spatiotemporal graph, the time series sequence of defect state feature vectors of each node, and the corresponding environmental parameter time series data into the spatiotemporal graph neural network model. With the goal of predicting the defect state features at future time points, train the spatiotemporal graph neural network model until it converges, and obtain the state evolution model that can predict the defect evolution trend.

[0059] This embodiment integrates spatiotemporal graph structure, node feature time series, and environmental parameter time series data through a spatiotemporal graph neural network. It captures the spatial correlation between nodes using a graph attention mechanism and combines a temporal convolutional network to learn the temporal evolution pattern of defect state features. It constructs a state evolution model driven by three dimensions of "space-time-environment" to achieve integrated output of dynamic prediction of defect severity, evolution trend deduction, and remaining lifetime estimation, thereby improving the scientificity and foresight of predictive maintenance decisions.

[0060] In this embodiment, the spatiotemporal graph neural network model captures the spatial correlation between nodes through a graph attention mechanism, and uses a temporal convolutional network or a recurrent neural network to learn the evolution of the defect state characteristics of each node over time.

[0061] In this embodiment, the output of the state evolution model includes a predicted value of the defect severity at a specific future time point and an estimated value of the remaining life of the equipment.

[0062] Step 4: Input the real-time collected inspection data into the equipment health status evolution model, perform online reasoning, and obtain defect detection results.

[0063] Step 4.1: Preprocess and register the real-time collected inspection data containing visible light, infrared and ultraviolet modes to form multimodal input data at the current time point.

[0064] This embodiment ensures accurate alignment of visible light, infrared, and ultraviolet images in spatial and temporal dimensions through preprocessing and pixel-level registration of multimodal inspection data. This provides the defect recognition model with structurally consistent and semantically matched multimodal input, improves the ability to capture the correlation between defect morphology and physical parameters in the feature extraction stage, and enhances the robustness of the model under complex working conditions.

[0065] Step 4.2: Input the multimodal input data at the current time point into the defect identification model to extract the defect state feature vectors of each key component at the current time point.

[0066] This embodiment utilizes the three-branch feature extraction architecture of the defect identification model to transform the registered multimodal data into a high-dimensional defect state feature vector, thereby achieving semantic dimensionality enhancement from the original image to physical semantic features, preserving the physical integrity of each modality feature, providing interpretable temporal feature input for the state evolution model, and improving the fine-grained analysis capability of defect state evolution.

[0067] Step 4.3: Input the feature vector sequence of the defect status of each key component at the current time point and the historical time point, the environmental parameter sequence, and the spatiotemporal diagram into the state evolution model.

[0068] This embodiment integrates the defect state feature vector sequence, environmental parameter sequence, and spatiotemporal graph structure of the current and historical time points to construct input data driven by the three dimensions of "space-time-environment". By fusing equipment topology dependence, temporal evolution law and environmental influencing factors through spatiotemporal graph neural network, it provides a comprehensive basis for equipment health state evolution analysis for state evolution model and enhances the ability to predict defect evolution trends under complex working conditions.

[0069] Step 4.4: Perform online reasoning using the state evolution model and output the defect detection results.

[0070] Step 4.4.1: After receiving the spatiotemporal graph, defect state feature vector sequence, and environmental parameter sequence from the state evolution model, the spatial dependency weights between key component nodes at the current time point are first calculated through the graph attention network layer inside the state evolution model.

[0071] This embodiment utilizes a graph attention network layer to dynamically calculate the spatial dependency weights between key component nodes. By guiding the weight allocation through prior knowledge of physical structure, it focuses on the spatial correlation of defect-related components, enhances the collaborative analysis capability of the state evolution model for the overall health status of the equipment, and improves the physical consistency of spatial dependency modeling.

[0072] Step 4.4.2: Based on the spatial dependency weights, perform temporal modeling on the defect state feature vector sequence of each key component using a temporal convolutional network or a recurrent neural network, analyze the evolution pattern of the defect state of each key component over time, and predict the defect state feature vector at a future preset time point.

[0073] This embodiment uses spatially dependent weights and temporal convolutional networks or recurrent neural networks to perform temporal modeling of the defect state feature vector sequence, capturing the changing pattern of the defect state of each component over time. Combined with temporal data of environmental parameters, it improves the prediction accuracy and dynamic adaptability of defect evolution trends.

[0074] Step 4.4.3: Map the predicted future defect state feature vector to a specific quantitative value of defect severity, and calculate the estimated value of the remaining effective life of the affected key components based on the preset mapping relationship between defect severity and remaining effective life.

[0075] This embodiment maps the predicted future defect state feature vector to a quantitative value of the specific defect severity and an estimated value of the remaining effective life. Through a preset physical parameter mapping relationship, the feature space is transformed into engineering indicators, providing an operable decision basis for predictive maintenance and enhancing the engineering application value of the model output.

[0076] Step 4.4.4: Based on the spatial dependency weight analysis results, the current defect status, the prediction of future defect evolution trends, and the estimated remaining effective lifetime, generate the defect detection results.

[0077] This embodiment integrates spatial dependency weight analysis results, current defect status, future evolution trend prediction, and remaining lifetime estimation to generate comprehensive detection results including defect type, location, severity, and remaining lifetime. It achieves integrated output of defect detection and health status assessment, providing real-time decision support for power equipment operation and maintenance across the entire chain and multiple dimensions.

[0078] In this embodiment, the defect detection results include: classification and location information of the current defect type of each key component, quantitative assessment of the severity of the defect, prediction of the evolution trend of the defect in the future within a preset time period, and estimated value of the remaining effective life of the affected key components.

[0079] Step 5: After inputting the real-time collected inspection data into the equipment health status evolution model for online reasoning, an uncertainty sample set is obtained by performing uncertainty estimation.

[0080] In this embodiment, results with a model inference confidence level higher than a preset model inference confidence level threshold are directly output as deterministic conclusions. For results with a confidence level lower than a preset threshold, the corresponding real-time collected inspection data are marked as uncertain samples, thus obtaining an uncertain sample set.

[0081] This embodiment uses a model inference confidence threshold screening mechanism to quantitatively evaluate the uncertainty of online inference results, marking low-confidence samples as uncertainty sample sets, realizing the reliability classification output of defect detection results, and providing targeted optimized sample sources for subsequent incremental learning, thereby enhancing the system's adaptive decision-making ability under complex working conditions.

[0082] Step 6: Continuously optimize the defect identification model and state evolution model through incremental learning based on the uncertainty sample set and expert annotations.

[0083] Step 6.1: Push the uncertain sample set to the expert annotation platform to obtain the experts' annotations on the defect type, location and severity of the samples in the uncertain sample set, and form an annotated sample set.

[0084] This embodiment pushes uncertain samples to an expert annotation platform to obtain professional annotations and form a high-quality annotation sample set, ensuring the physical and semantic accuracy and engineering interpretability of incremental learning data, and providing supervision signals that conform to actual working conditions for model optimization.

[0085] Step 6.2: Mix the labeled sample set with the multimodal training dataset used to train the defect recognition model according to a preset ratio to obtain the mixed dataset.

[0086] This embodiment constructs an augmented dataset containing new scene knowledge by hybrid expansion of labeled sample sets and original training data. While retaining the original knowledge of the model, it incorporates new sample features to achieve a smooth transition and knowledge accumulation in incremental training.

[0087] Step 6.3: Using the current parameters of the defect identification model as pre-training weights, incrementally train the defect identification model using the mixed dataset, update the network parameters of the defect identification model, and obtain the updated defect identification model.

[0088] This embodiment uses pre-trained weights for incremental training, leverages the advantages of transfer learning to accelerate model convergence, and updates the physical attention mechanism parameters of the defect recognition network to improve the generalization ability to recognize new defect patterns and complex working conditions.

[0089] Step 6.4: Using the updated defect identification model, extract the defect state feature vectors from the historical time-series inspection data again to generate an updated node feature time-series sequence.

[0090] This embodiment re-extracts historical time-series features based on the updated defect identification model, generating a node feature time-series sequence containing new identification capabilities. This ensures that the input data of the state evolution model remains consistent with the feature space of the updated model, thereby improving the evolution prediction accuracy of the spatiotemporal graph neural network.

[0091] Step 6.5: Using the updated node feature time sequence as input and the current parameters of the state evolution model as pre-training weights, perform incremental training on the state evolution model to update the parameters of the spatiotemporal graph neural network of the state evolution model, and obtain the updated state evolution model.

[0092] This embodiment incrementally trains the state evolution model by updating the node feature time sequence, optimizes the graph attention weights and time evolution module parameters of the spatiotemporal graph neural network, and enhances the model's ability to dynamically capture the evolution trend of equipment health status and the accuracy of remaining life estimation.

[0093] Step 6.6: Deploy the updated defect identification model and the updated state evolution model to the online inference system, replace the original model, and complete the optimization and iteration of the defect identification model and the state evolution model.

[0094] This embodiment deploys the optimized model to an online inference system, enabling seamless iterative upgrades of the model, ensuring continuous improvement in defect detection and health assessment capabilities, and guaranteeing the long-term scientific rigor and reliability of power equipment operation and maintenance decisions.

[0095] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting defects in power equipment based on image recognition, characterized in that, include: Based on the equipment's 3D model and physical field simulation parameters, a defect simulation dataset is obtained by performing defect simulation driven by physical mechanisms. Based on the defect simulation dataset and real inspection data, a cross-modal deep learning network was trained according to physical law constraints to obtain a defect recognition model; A spatiotemporal graph neural network model was constructed based on historical time-series inspection data and a defect identification model to obtain a state evolution model; The real-time collected inspection data is input into the equipment health status evolution model for online reasoning to obtain defect detection results.

2. The power equipment defect detection method based on image recognition as described in claim 1, characterized in that, It also includes, after inputting the real-time collected inspection data into the equipment health status evolution model for online reasoning, obtaining an uncertainty sample set by performing uncertainty estimation, and continuously optimizing the defect identification model and the status evolution model by incremental learning based on the uncertainty sample set and expert annotations.

3. The power equipment defect detection method based on image recognition as described in claim 2, characterized in that, Based on the equipment's 3D model and physical field simulation parameters, a defect simulation dataset is obtained through physical mechanism-driven defect simulation, including: A digital twin model of the target power equipment is obtained by performing a 3D modeling based on the 3D scanning data of the target power equipment. Based on the material properties, electrical load, and environmental parameters of the target power equipment and the digital twin model, a physical field coupling simulation is performed to simulate the physical field data during the defect evolution process. Based on the physical field data and the preset defect mapping rules, a visualization rendering is performed on the digital twin model of the target power equipment to generate an initial multimodal simulation image containing the defect morphology; the initial multimodal simulation image includes a visible light image, an infrared thermogram, and an ultraviolet discharge image; Based on the initial multimodal simulation image and the real inspection data, image fusion and enhancement processing are performed to obtain multimodal defect samples; The defect simulation dataset is constructed by labeling the multimodal defect samples with defect type, location, severity, and physical parameter labels.

4. The power equipment defect detection method based on image recognition as described in claim 3, characterized in that, The physical field coupling simulation includes thermal-electric coupling simulation, fluid-solid coupling simulation, and electrostatic field simulation; wherein, based on the material properties, electrical load, and environmental parameters of the digital twin model and the target power equipment, physical field coupling simulation is performed to simulate the physical field data during the defect evolution process, including: Based on the digital twin model, according to the geometry and material properties of the target power equipment, a thermo-electric coupling simulation is performed using applied electrical load and ambient temperature parameters to calculate the temperature field distribution data of the target power equipment under normal and defective states; the defective states include the state of increased resistance caused by connection point contact and the state of blocked internal carrier pathways; Based on the digital twin model, fluid-solid coupling simulation is performed according to the geometry, surface characteristics, ambient wind speed and particulate matter concentration parameters of the target power equipment to simulate the adsorption, accumulation process and distribution pattern of pollutant particles on the surface of the target power equipment, and output the surface pollutant density distribution data of the target power equipment. Based on the digital twin model, electrostatic field simulation is performed according to the geometry, material properties, applied voltage, and surface contamination density distribution data of the target power equipment. The electric field intensity distribution data on the surface of the target power equipment is calculated, and the potential occurrence area and intensity of corona discharge or arc discharge are determined based on the field emission principle. The temperature field distribution data, the surface contamination density distribution data, and the electric field intensity distribution data are integrated to form multiphysics field data in the defect evolution process.

5. The power equipment defect detection method based on image recognition as described in claim 4, characterized in that, Based on the defect simulation dataset and real inspection data, a cross-modal deep learning network is trained according to physical constraints to obtain a defect recognition model, including: The multimodal defect samples in the defect simulation dataset are mixed with real inspection data at a preset ratio and pixel-level registration is performed to construct a multimodal training dataset; each sample unit in the multimodal training dataset contains a registered visible light image, an infrared thermogram, and an ultraviolet discharge image. A cross-modal deep learning network with three independent branches is constructed. Each branch processes an image of a different modality and extracts features to obtain visible light features, infrared thermal features, and ultraviolet discharge features. The extracted visible light features, infrared thermal features, and ultraviolet discharge features are input into a physical law constraint module for feature fusion; the physical law constraint module is used to generate a physical attention weight map based on the preset prior knowledge of the device's physical structure and the visible light features, infrared thermal features, and ultraviolet discharge features. Using the multimodal training dataset, with defect classification and localization as the supervised tasks, the cross-modal deep learning network and the physical law constraint module are jointly trained until the model converges, thus obtaining the defect recognition model.

6. The power equipment defect detection method based on image recognition as described in claim 5, characterized in that, Based on historical time-series inspection data and a defect identification model, a spatiotemporal graph neural network model is used to obtain a state evolution model, including: A spatiotemporal graph is constructed based on the spatial topology of the target power equipment; wherein, each key component of the target power equipment is defined as a graph node, and the physical connections or functional associations between components are defined as graph edges; Historical time-series inspection data of each graph node is extracted and processed using the defect identification model to generate defect status feature vectors for each node at historical time points, forming a node feature time-series sequence. The spatiotemporal graph, the time series sequence of defect state feature vectors of each node, and the corresponding environmental parameter time series data are input into the spatiotemporal graph neural network model. With the goal of predicting the defect state features at future time points, the spatiotemporal graph neural network model is trained until it converges, thus obtaining the state evolution model that can predict the defect evolution trend. The spatiotemporal graph neural network model captures the spatial correlation between nodes through a graph attention mechanism and uses a temporal convolutional network or a recurrent neural network to learn the evolution of the defect state characteristics of each node over time. The output of the state evolution model includes a predicted value of the defect severity at a specific future time point and an estimated value of the remaining life of the equipment.

7. The power equipment defect detection method based on image recognition as described in claim 6, characterized in that, The real-time collected inspection data is input into the equipment health status evolution model for online inference to obtain defect detection results, including: The real-time collected inspection data, including visible light, infrared and ultraviolet modes, is preprocessed and registered to form multimodal input data at the current time point; The multimodal input data at the current time point is input into the defect identification model to extract the defect state feature vectors of each key component at the current time point; The feature vector sequence of the defect status of each key component at the current time point and the historical time point, the environmental parameter sequence, and the spatiotemporal diagram are all input into the state evolution model. Online reasoning is performed using the state evolution model to output defect detection results; The defect detection results include: classification and location information of the current defect type of each key component, quantitative assessment of the severity of the defect, prediction of the evolution trend of the defect in the future within a preset time period, and estimated value of the remaining effective life of the affected key components.

8. The power equipment defect detection method based on image recognition as described in claim 7, characterized in that, The online reasoning through the state evolution model, outputting defect detection results, includes: After receiving the spatiotemporal graph, defect state feature vector sequence, and environmental parameter sequence from the state evolution model, the spatial dependency weights between key component nodes at the current time point are first calculated through the graph attention network layer inside the state evolution model. Based on the spatial dependency weights, the defect state feature vector sequence of each key component is modeled temporally using a temporal convolutional network or a recurrent neural network to analyze the evolution pattern of the defect state of each key component over time and predict the defect state feature vector at a future preset time point. The predicted future defect state feature vector is mapped to a specific quantitative value of defect severity, and the estimated remaining effective life of the affected key components is calculated based on the preset mapping relationship between defect severity and remaining effective life. Based on the spatial dependency weight analysis results, the current defect status, the predicted future defect evolution trend, and the estimated remaining effective lifetime, defect detection results are generated.

9. The power equipment defect detection method based on image recognition as described in claim 8, characterized in that, After inputting the real-time collected inspection data into the equipment health status evolution model for online inference, an uncertainty sample set is obtained by performing uncertainty estimation. Among them, results with a model inference confidence level higher than the preset model inference confidence level threshold are directly output as deterministic conclusions, while results with a confidence level lower than the preset threshold are marked as uncertain samples in the corresponding real-time collected inspection data, thus obtaining the uncertainty sample set.

10. The power equipment defect detection method based on image recognition as described in claim 9, characterized in that, Incremental learning is used to continuously optimize the defect identification model and state evolution model based on the uncertain sample set and expert annotations, including: The uncertain sample set is pushed to the expert annotation platform to obtain the experts' annotations on the defect type, location and severity of the samples in the uncertain sample set, thus forming an annotated sample set; The labeled sample set is mixed with the multimodal training dataset used to train the defect recognition model according to a preset ratio to obtain the mixed dataset; Using the current parameters of the defect identification model as pre-training weights, the defect identification model is incrementally trained using the mixed dataset to update the network parameters of the defect identification model, thus obtaining the updated defect identification model. Using the updated defect identification model, the defect state feature vectors are re-extracted from the historical time-series inspection data to generate an updated node feature time-series sequence. Using the updated node feature time sequence as input and the current parameters of the state evolution model as pre-training weights, the state evolution model is incrementally trained to update the parameters of the spatiotemporal graph neural network of the state evolution model, thus obtaining the updated state evolution model. The updated defect identification model and the updated state evolution model are deployed to the online inference system to replace the original model, thus completing the optimization and iteration of the defect identification model and the state evolution model.

Images