Edge-deployed gas insulated switch defect detection method and system

Abstract

The invention relates to an edge-deployed gas insulated switch defect detection method and system, and belongs to the technical field of power equipment inspection and intelligent fault diagnosis. The system comprises an image enhancement module, a model training module, a data storage module, a defect prediction module, a front-end interaction module and a feedback learning module. The method comprises the steps of performing quality evaluation on an input GIS image, and executing adaptive image enhancement according to an evaluation result; a detection network is constructed based on an RT-DETR architecture, MFM is introduced to enhance the multi-scale feature expression ability, and an LAMP algorithm is adopted to prune the model; carrying out structured storage on the enhanced image, model output and annotation information; defect detection is carried out on the edge device deployment model, and the defect category, position and confidence are output; carrying out feedback, re-labeling and incremental training on a detection failure sample to realize continuous optimization of the model; meanwhile, a human-computer interaction interface is provided to support image enhancement, model training and result visualization.

Description

Technical Field

[0001] This invention belongs to the field of power equipment inspection and intelligent fault diagnosis technology, and relates to a method and system for detecting defects in edge-deployed gas-insulated switches. Background Technology

[0002] Gas-insulated switchgear (GIS) is widely used in substations, power grid transmission and distribution systems, and industrial power distribution systems due to its compact structure, safe and reliable operation, and low maintenance workload. However, during long-term operation, GIS equipment may develop various defects such as cracks, discharge marks, metal particles, and insulator damage due to manufacturing defects, environmental aging, insulation degradation, or partial discharge. If these defects are not detected and addressed in a timely manner, they may lead to serious insulation failures or even power grid accidents. Therefore, efficient and accurate defect detection of GIS equipment is of great significance.

[0003] With the development of intelligent vision technology, some studies have attempted to introduce convolutional neural networks (CNNs) or traditional object detection algorithms for equipment defect identification. However, these methods still have significant limitations: on the one hand, GIS defect features often have characteristics such as small scale, complex background, and uneven lighting, making it easy for traditional models to miss or falsely detect defects; on the other hand, existing methods generally lack mechanisms for handling the problem of inconsistent image quality in actual working conditions, and also lack the ability for self-evolution of models based on sample feedback, thus limiting their engineering application. Therefore, a new GIS defect detection method is urgently needed to solve these problems. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and system for detecting defects in gas-insulated switches at the edge with high precision and adaptive optimization capabilities, so as to meet the long-term operation and maintenance and model iteration requirements of power equipment inspection scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Solution 1: A defect detection method for edge-deployed gas-insulated switches, specifically including the following steps: S1: Image Enhancement: Acquire image data of gas-insulated switchgear (GIS), assess its quality, and perform adaptive image enhancement based on the assessment results; S2: Model Training: A detection network is built based on the RT-DETR (Real-time Object Detection Model) architecture. MFM is introduced into the neck network to enhance the feature representation ability at multiple scales. At the same time, the LAMP algorithm is used to prune the model. Here, MFM represents the modulation fusion module, and LAMP represents the layer adaptive amplitude pruning algorithm. S3: Data storage: Stores the enhanced images, model prediction outputs, and annotation information in the database, supporting structured management and export; S4: Defect Prediction: Load the trained model weights, deploy them on vehicle or edge devices, perform inference on the acquired images, output the defect location, category and confidence level, feed back the failed prediction samples to the server for storage, and when the cumulative number reaches the preset threshold, re-label and incrementally train to generate new model weights for updating the deployment. S5: System Front-end Interaction: Provides a human-computer interaction interface for image enhancement, model training, and result display, enabling parameter configuration and visualization operations.

[0006] Furthermore, in step S1, image enhancement specifically includes the following steps: S11: Perform quality assessment on the acquired raw GIS images and classify them into different quality categories based on preset thresholds to provide a basis for subsequent targeted enhancement; S12: Sharpness Assessment: Sharpness is estimated using the Laplace operator variance method, and its mathematical expression is:

[0007] in, The result of the Laplacian transform of the image. This indicates variance calculation; if the sharpness score... If so, the image is determined to be blurry and degraded. Spatial sharpness threshold; S13: Noise Intensity Assessment: Image noise level is estimated statistically using wavelet coefficients, and the noise intensity is defined as:

[0008] in, This represents the high-frequency coefficients of the image in the wavelet transform domain. This indicates taking the median; if the noise score If so, the image is determined to be severely contaminated by noise. Noise intensity threshold; S14: Brightness Assessment: Image brightness level is measured using the global grayscale mean.

[0009] in, M and N These represent the height and width of the image, respectively; if the brightness score... If the brightness is insufficient, it is determined to be insufficient; If so, it is judged as overexposure; This is the lower limit threshold for brightness. This is the upper limit threshold for brightness; S15: Based on the image quality assessment results, classify the input image according to the following rules:

[0010] in, For the quality classification results, This indicates that it is normal. Indicates ambiguity. Indicates noise. Indicates low light intensity. It has been exposed; S16: Adaptive image enhancement execution: Normal: Output directly without any enhancement processing; Blurred: The image is deblurred using Wiener filtering or blind deconvolution algorithms to restore image details; Noisy: Noise suppression is achieved using either the block matching 3D denoising algorithm (BM3D) or the nonlocal mean denoising algorithm, while preserving image structural information; LowLight: Uses adaptive histogram equalization (CLAHE) or Retinex series algorithms for brightness compensation and contrast enhancement; OverExposed: Brightness recovery is achieved using adaptive gamma correction or highlight suppression algorithms; S17: Store the enhanced images in a database for use in model training or prediction.

[0011] Furthermore, in step S2, model training specifically includes the following steps: S21: Using RT-DETR as the overall detection framework, its end-to-end query encoder-decoder structure is used as the skeleton for feature extraction and target decoding, avoiding error accumulation and delay caused by additional candidate generation and non-maximum suppression (NMS), and improving the consistency and real-time performance of training-inference. S22: In the neck network of RT-DETR, MFM is introduced to replace the original Concat module. Through the weight learning mechanism, multi-scale feature maps from different paths are adaptively fused to enhance the response of key features and suppress redundant information. S23: The LAMP algorithm is used to prune the final improved RT-DETR model to reduce the number of model parameters, and the pruned model is fine-tuned to restore detection accuracy.

[0012] Furthermore, step S22 specifically includes the following steps: S221: MFM receives multi-scale feature maps from different paths as input; S222: Adaptive weights are generated by a weight generation submodule; the weight generation submodule includes a global average pooling layer, a multilayer perceptron network, and a Softmax normalization layer; The global average pooling layer performs global average pooling on each input feature map in its spatial dimension, compressing each feature map into a channel feature vector to obtain its channel-level global semantic statistical information. The Multilayer Perceptron (MLP) network inputs the feature vectors of each channel obtained by global average pooling into a shared two-layer MLP for nonlinear transformation; the first layer of the MLP performs dimensionality reduction, the second layer restores the original number of channels, and finally outputs the unnormalized initial weight vector; The Softmax normalization layer concatenates the weight vectors corresponding to the same batch from all input paths along the path dimension and applies the Softmax function for normalization to generate a modulation coefficient matrix corresponding to each input feature map, thereby characterizing the importance of features at different scales. S223: Perform channel-by-channel and element-by-element multiplication on the modulation coefficient matrix obtained in step S222 and the corresponding original input feature map to obtain the modulated feature map. Then, concatenate all the modulated feature maps on the channel and then perform cross-channel information integration and dimensionality reduction through a standard 1×1 convolutional layer to finally output an enhanced fused feature.

[0013] Option 2: A system for edge-deployed gas-insulated switch defect detection, comprising: The image enhancement module is used to acquire gas-insulated switch (GIS) image data and perform quality assessment on it. For quality problems such as blurring, noise, low light or overexposure identified in the assessment, targeted image enhancement processing is performed. The model training module constructs a detection network based on the RT-DETR (Real-time Object Detection Model) architecture, introduces MFM into the neck network to enhance feature representation capabilities, and uses the LAMP algorithm to prune the model. The data storage module is used to store image enhancement results, model prediction outputs, and annotation information in the database, supporting structured management and export. The defect prediction module is used to load the trained model weights, deploy them on vehicle-mounted or edge devices, perform inference on the acquired images, output the defect location, category and confidence level, and feed back the failed prediction samples to the server for storage. When the cumulative number reaches a preset threshold, it performs re-labeling and incremental training to generate new model weights for updating the deployment. The front-end interaction module provides a human-computer interaction interface for image enhancement, model training and result display, enabling parameter configuration and visualization operations. The feedback learning module is configured to collect failed prediction samples and trigger relabeling and incremental training.

[0014] The system data flow collaborates as follows: the defect prediction module sends the prediction failure samples generated during the inference process to the feedback learning module; the feedback learning module stores the collected failure samples in the data storage module, and when the cumulative number of failure samples reaches a preset threshold, it sends an incremental training instruction to the model training module; the model training module responds to the incremental training instruction, retrieves the failure samples and their associated data from the data storage module, and combines them with the results after manual review and annotation by the front-end interaction module to perform incremental training to generate new model weights; the model training module updates the generated new model weights to the defect prediction module, completing the model deployment update.

[0015] Furthermore, in the data storage module, the database includes an image information table and a prediction result table; the image information table is used to record information such as training images, acquisition time, and device number; the prediction result table is used to record defect categories, confidence levels, and model version numbers.

[0016] Furthermore, the defect prediction module specifically includes: loading the trained and optimized model weight file and deploying it on an in-vehicle edge computing device or industrial inspection terminal; performing defect detection on GIS images acquired in real time or input in batches, and outputting the category, location coordinates, and confidence level of each defect target; filtering out detection results with low confidence based on a preset confidence threshold and pushing them as samples to be reviewed to the manual review stage; periodically feeding back failed samples to the server, and triggering re-annotation and incremental training when the cumulative number reaches a preset number to update the model weights; saving the inference results and generating a visualization output file containing defect location annotations.

[0017] Furthermore, the front-end interaction module specifically includes: providing a graphical user interface that supports image uploading, image enhancement algorithm selection, model training control, and result browsing; supporting image zooming, target selection, category filtering, and result highlighting in the detection result display interface to facilitate defect area location analysis; supporting manual review and annotation of failed inference samples, and submitting the annotation results to the server for model updates.

[0018] The beneficial effects of this invention are as follows: 1) This invention dynamically selects enhancement strategies based on image quality, performs targeted enhancement processing on low-brightness, overexposed, blurred, and noise-contaminated images, improves the ability to express texture details in defective areas, significantly enhances recognition robustness in complex environments, and overcomes the problem of poor adaptability of traditional fixed enhancement methods; 2) This invention is based on the RT-DETR real-time detection framework. MFM is introduced into its neck network to enhance the expressive power of multi-scale features through an adaptive weight learning mechanism. The LAMP structured pruning algorithm is used to compress the model, which significantly reduces the model complexity and computational cost while ensuring high detection accuracy. 3) This invention constructs an application system that integrates image enhancement, model training, database management and result visualization. It can be deployed on vehicle terminals or edge devices and supports automatic feedback and incremental training mechanisms for failed detection samples, enabling continuous model optimization. It has high precision and adaptive optimization capabilities, meeting the long-term application needs of power equipment inspection scenarios.

[0019] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a diagram showing the overall structure of the system of the present invention; Figure 2 This is a diagram of the image enhancement strategy structure. Figure 3 To improve the combined flowchart of RT-DETR model and LAMP pruning. Detailed Implementation

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0022] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0023] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0024] Example 1: Please see Figures 1-3 This embodiment provides a defect detection method for edge-deployed GIS, including the following steps: S1: Image enhancement steps: Acquire GIS image data, perform quality assessment on the input image, and perform targeted image enhancement processing for quality problems such as blurriness, noise, low light, or overexposure identified in the assessment.

[0025] The image enhancement steps are as follows: S11: Perform quality assessment on the acquired raw GIS images and classify them into different quality categories based on preset thresholds to provide a basis for subsequent targeted enhancement.

[0026] S12: Sharpness Assessment: Sharpness is estimated using the Laplace operator variance method, and its mathematical expression is:

[0027] in, The result of the Laplacian transform of the image. This indicates variance calculation; if the sharpness score... If so, the image is determined to be blurry and degraded. This is the spatial sharpness threshold.

[0028] S13: Noise Intensity Assessment: Image noise level is estimated statistically using wavelet coefficients, and the noise intensity is defined as:

[0029] in, This represents the high-frequency coefficients of the image in the wavelet transform domain. This indicates taking the median; if the noise score If so, the image is determined to be severely contaminated by noise.

[0030] S14: Brightness Assessment: Image brightness level is measured using the global grayscale mean.

[0031] in, M and N These represent the height and width of the image, respectively; if the brightness score... If the brightness is insufficient, it is determined to be insufficient; If so, it is judged as overexposure.

[0032] Among them, threshold T s , T n , T l , T h These thresholds are spatial clarity threshold, noise intensity threshold, lower brightness threshold, and upper brightness threshold. These thresholds are preset by statistically analyzing the quality indicators of a historical normal GIS image sample library and taking a specific quantile (such as the 5th or 95th quantile) of its distribution as the initial value. They can also be manually fine-tuned through the system front-end interactive module to adapt to different deployment environments.

[0033] S15: Based on the image quality assessment results, classify the input image according to the following rules:

[0034] S16: Adaptive image enhancement execution: Normal: Output directly without any enhancement processing.

[0035] Blurred: The image is deblurred using Wiener filtering or blind deconvolution algorithms to restore image details.

[0036] Noisy: Noise suppression is achieved using either the block matching 3D denoising algorithm (BM3D) or the nonlocal mean denoising algorithm, while preserving image structural information.

[0037] LowLight: Uses Adaptive Histogram Equalization (CLAHE) or Retinex series algorithms for brightness compensation and contrast enhancement.

[0038] OverExposed: Brightness is restored using adaptive gamma correction or highlight suppression algorithms.

[0039] S17: Store the enhanced images in a database for use in model training or prediction.

[0040] S2: Model training steps: Construct a detection network based on the RT-DETR architecture, introduce a modulation fusion module (MFM) in the neck network to enhance the multi-scale feature representation capability, and prune the model using layer adaptive amplitude pruning (LAMP).

[0041] The specific steps for model training are as follows: S21: Based on the end-to-end detection framework of RT-DETR, its core includes: Hybrid encoder: Receives multi-scale feature maps (such as C3, C4, and C5) extracted by the backbone network and processes them through an internally designed cross-scale feature fusion mechanism.

[0042] The Transformer decoder interacts with the encoder output features using object queries, directly generating target prediction boxes through a cross-attention mechanism. This framework avoids post-processing steps such as candidate box generation and non-maximum suppression (NMS) in traditional detectors, reducing error accumulation and improving the consistency and real-time performance of training and inference.

[0043] S22: In the feature fusion part of the RT-DETR neck network, the MFM module is used to replace the original simple concat operation. This module is implemented through the following structure and steps: S221: The MFM module receives feature maps of at least two levels with different spatial resolutions from the preceding nodes of the encoder as input.

[0044] S222: Adaptive Weight Generation. This step is implemented by a weight generation submodule, the structure of which is as follows: Global Average Pooling (GAP) layer: For each input feature map, global average pooling is performed in its spatial dimension, compressing each feature map into a channel feature vector to obtain its channel-level global semantic statistics.

[0045] Multilayer Perceptron (MLP) network: The feature vectors of each channel obtained by global average pooling are input into a shared two-layer MLP for nonlinear transformation. The first layer of the MLP performs dimensionality reduction, the second layer restores the original number of channels, and finally outputs an unnormalized initial weight vector.

[0046] Softmax normalization layer: The weight vectors corresponding to the same batch from all input paths are concatenated along the path dimension and normalized using the Softmax function to generate a modulation coefficient matrix corresponding to each input feature map, thereby representing the importance of features at different scales.

[0047] S223: Perform channel-wise and element-wise multiplication operations on the modulation coefficient matrix obtained in step S222 and the corresponding original input feature map to obtain the modulated feature map. Then, concatenate all the above modulated feature maps in the channel dimension, and then perform cross-channel information integration and dimensionality reduction through a standard 1×1 convolutional layer to finally output an enhanced fusion feature.

[0048] S23: The improved RT-DETR model trained above is pruned using LAMP, with the following steps: S231: Load the model weights after training convergence, and score the importance of each connection in all layers to be pruned (e.g., weights of fully connected layers and convolutional layers) in the network based on the LAMP criterion. The LAMP score is calculated as follows:

[0049] in, W This represents the weight tensor of that layer. W [ u ] indicates index u The weight values ​​of the mapping, W [ v ] indicates that index v is greater than or equal to u The weight value. The score satisfies:

[0050] S232: After calculating the LAMP score for each connection in all layers to be pruned, sort all scores globally. Based on a pre-defined global model sparsity (i.e., the overall pruning ratio), determine a global score threshold. Determine the number of connections below this threshold as unimportant connections and set their weights to zero, thus achieving structured pruning.

[0051] S233: Pruning disrupts the original weight distribution of the model, leading to a decrease in accuracy. To restore model performance, the pruned sparse model is fine-tuned. During fine-tuning, the original training dataset is used, and knowledge distillation is introduced. The original high-performance model before pruning is used as the teacher model to guide the training of the pruned student model. A low learning rate and an optimization objective that includes distillation loss are used during fine-tuning to ensure that the model maintains fast inference while its detection accuracy approaches that of the original model.

[0052] S3: Data storage steps: Store image enhancement results, model prediction outputs, and annotation information in the database, supporting structured management and export.

[0053] The specific steps for data storage are as follows: S31: During the training phase, the image enhancement results, model prediction outputs, and annotation information are stored in a relational database in a structured form.

[0054] S32: To achieve the classification, storage and hierarchical management of training samples and model output data, the relational database sets up two core data tables: an image information table and a prediction result table.

[0055] S33: The image information table is used to store basic attribute information of training images. Its fields include image unique identifier ID, image storage path, acquisition time, acquisition device number, image resolution, data batch number, data augmentation type, and original image source. By setting up the image information table, traceable management of training data sources can be achieved, and data queries can be performed based on conditions such as time and acquisition device, thereby improving data scheduling efficiency.

[0056] S34: The prediction result table is used to record the inference output information generated by the model during the training or validation phase. Its fields include image ID (associated with the image information table), target defect category label, bounding box coordinates (xmin, ymin, xmax, ymax), confidence score, inference time, loss function value, and model version number. By recording the model version number, horizontal comparability of output results from different model versions can be achieved, providing data support for model optimization.

[0057] S4: Defect prediction steps: Load the trained model weights, deploy them on vehicle-mounted or edge devices, perform inference on the acquired images, output the defect location, category and confidence level, feed back the failed prediction samples to the server for storage, and when the cumulative number reaches a preset threshold, re-label and incrementally train to generate new model weights for updating deployment.

[0058] The specific steps for defect prediction are as follows: S41: Load the trained and optimized model weight file and deploy it on an in-vehicle edge computing device or industrial inspection terminal.

[0059] S42: Perform defect detection on GIS images acquired in real time or input in batches, and output the category, location coordinates and confidence level of each defect target.

[0060] S43: A reliability assessment mechanism is implemented on the detection results output by the model. The detection results are filtered according to a preset confidence threshold to identify false positives and false negatives. For detection results with low reliability, the system marks the corresponding image and detection data as failure samples and stores them. At the same time, it automatically records information such as inference time, environmental parameters, and model version for subsequent model performance backtracking.

[0061] S44: Failed samples can be periodically uploaded to the server for storage and used as part of the model's continuous learning samples. When the number of failed samples on the server reaches a preset threshold, the system automatically triggers a sample screening and manual re-labeling process. The corrected set of failed samples is then incorporated into the incremental training dataset. The original detection model is incrementally trained and updated through fine-tuning, thereby dynamically optimizing the model weights and improving the model's ability to identify defects under complex working conditions.

[0062] S45: Save the reasoning results and generate a visual output file containing defect location annotations.

[0063] S5: System front-end interaction steps: Provides a human-computer interaction interface for image enhancement, model training and result display, and realizes parameter configuration and visualization operations.

[0064] The specific steps for system front-end interaction are as follows: S51 provides a graphical user interface that supports image uploading, image enhancement algorithm selection, model training control, and result browsing.

[0065] S52. The detection result display interface supports image zooming, target selection, category filtering, and result highlighting to facilitate defect area location and analysis.

[0066] S53. Supports manual review and annotation of failed inference samples, and submits the annotation results to the server for model updates.

[0067] Example 2: This embodiment provides an edge-deployed GIS defect detection system for implementing the method described in Embodiment 1. The system includes: Image enhancement module, configured to perform the image enhancement steps described in S1; The model training module is configured to perform the model training steps described in S2. A data storage module, configured to perform the data storage steps described in S3; A defect prediction module, configured to perform the defect prediction steps described in S4; The front-end interaction module is configured to execute the system front-end interaction steps described in S5. The feedback learning module is configured to collect failed prediction samples and trigger relabeling and incremental training.

[0068] The defect prediction module, feedback learning module, data storage module, model training module, and front-end interaction module collaborate according to the following data flow: The defect prediction module sends the failed prediction samples generated during the inference process to the feedback learning module; the feedback learning module stores the collected failed samples in the data storage module, and when the cumulative number of failed samples reaches a preset threshold, it sends an incremental training instruction to the model training module; the model training module responds to the incremental training instruction, retrieves the failed samples and their associated data from the data storage module, and combines them with the results after manual review and annotation by the front-end interaction module to perform incremental training to generate new model weights; the model training module updates the generated new model weights to the defect prediction module, completing the model deployment update.

[0069] Finally, 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 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 edge-deployed gas-insulated switches, characterized in that, The method specifically includes the following steps: S1: Image Enhancement: Acquire image data of gas-insulated switchgear, assess its quality, and perform adaptive image enhancement based on the assessment results; S2: Model Training: A detection network is built based on the RT-DETR architecture. MFM is introduced into the neck network to enhance the multi-scale feature representation capability. At the same time, the LAMP algorithm is used to prune the model. Here, MFM represents modulation fusion module, LAMP represents layer adaptive amplitude pruning algorithm, and RT-DETR represents real-time target detection model. S3: Defect Prediction: Load the trained model weights, deploy them on vehicle or edge devices, perform inference on the acquired images, output the defect location, category and confidence level, feed back the failed prediction samples to the server for storage, and when the cumulative number reaches a preset threshold, re-label and incrementally train to generate new model weights for updating deployment.

2. The method for detecting defects in edge-deployed gas-insulated switches according to claim 1, characterized in that, In step S1, image enhancement specifically includes the following steps: S11: Perform quality assessment on the acquired raw GIS images and classify them into different quality categories based on preset thresholds; S12: Sharpness estimation is performed using the Laplace operator variance method, and its mathematical expression is: in, The result of the Laplacian transform of the image. This indicates variance calculation; if the sharpness score... If so, the image is determined to be blurry and degraded. Spatial sharpness threshold; S13: Estimate the image noise level using wavelet coefficient statistics, defining the noise intensity as: in, This represents the high-frequency coefficients of the image in the wavelet transform domain. This indicates taking the median; if the noise score If so, the image is determined to be severely contaminated by noise. Noise intensity threshold; S14: Use the global grayscale mean to measure the image brightness level. in, M and N These represent the height and width of the image, respectively; if brightness is scored... If the brightness is insufficient, it is determined to be insufficient; if If so, it is determined to be overexposed; This is the lower limit threshold for brightness. This is the upper limit threshold for brightness. S15: Based on the image quality assessment results, classify the input image according to the following rules: in, For the quality classification results, This indicates that it is normal. Indicates ambiguity. Indicates noise. Indicates low light intensity. It has been exposed; S16: Adaptive image enhancement execution: Normal: Direct output without enhancement processing; Blur: Wiener filtering or blind deconvolution algorithms are used for deblurring to restore image details; Noise: Block matching 3D denoising algorithm or nonlocal mean denoising algorithm are used to suppress noise while preserving image structural information; Low illumination: Adaptive histogram equalization or Retinex series algorithms are used for brightness compensation and contrast enhancement; Overexposure: Brightness recovery is achieved using adaptive gamma correction or highlight suppression algorithms; S17: Store the enhanced images in a database for use in model training or prediction.

3. The method for detecting defects in edge-deployed gas-insulated switches according to claim 1, characterized in that, In step S2, model training specifically includes the following steps: S21: Using RT-DETR as the overall detection framework, its end-to-end query encoder-decoder structure is used as the skeleton for feature extraction and target decoding, avoiding error accumulation and delay caused by additional candidate generation and non-maximum suppression. S22: In the feature fusion part of the neck network of RT-DETR, MFM is used to replace the original Concat module, and the multi-scale feature maps from different paths are adaptively fused through the weight learning mechanism. S23: The LAMP algorithm is used to prune the final improved RT-DETR model to reduce the number of model parameters, and the pruned model is fine-tuned to restore detection accuracy.

4. The method for detecting defects in edge-deployed gas-insulated switches according to claim 3, characterized in that, Step S22 specifically includes the following steps: S221: MFM receives multi-scale feature maps from different paths as input; S222: Adaptive weights are generated by a weight generation submodule; the weight generation submodule includes a global average pooling layer, a multilayer perceptron network, and a Softmax normalization layer; The global average pooling layer performs global average pooling on each input feature map in its spatial dimension, compressing each feature map into a channel feature vector to obtain its channel-level global semantic statistical information. The multilayer perceptron network takes the feature vectors of each channel obtained by global average pooling and inputs them into a shared two-layer multilayer perceptron for nonlinear transformation; the first layer of the multilayer perceptron performs dimensionality reduction, the second layer restores the original number of channels, and finally outputs the unnormalized initial weight vector; The Softmax normalization layer concatenates the weight vectors corresponding to the same batch from all input paths along the path dimension and applies the Softmax function for normalization to generate a modulation coefficient matrix corresponding to each input feature map. S223: Perform channel-wise and element-wise multiplication operations on the modulation coefficient matrix obtained in step S222 and the corresponding original input feature map to obtain the modulated feature map. Then, concatenate all the modulated feature maps on the channel and then perform cross-channel information integration and dimensionality reduction through a standard 1×1 convolutional layer to finally output an enhanced fused feature.

5. The method for detecting defects in edge-deployed gas-insulated switches according to claim 1, characterized in that, The method also includes a data storage step, which stores the enhanced image, model prediction output, and annotation information in a database, supporting structured management and export.

6. The method for detecting defects in edge-deployed gas-insulated switches according to claim 1, characterized in that, The method also includes a system front-end interaction step, providing a human-computer interaction interface for image enhancement, model training and result display, and enabling parameter configuration and visualization operations.

7. A system applicable to the defect detection method for edge-deployed gas-insulated switches as described in any one of claims 1 to 6, characterized in that, The system includes: The image enhancement module is used to acquire image data of gas-insulated switches and evaluate their quality. For those with blurriness, noise, low light, or overexposure identified in the evaluation, targeted image enhancement processing is performed. The model training module constructs a detection network based on the RT-DETR architecture, introduces MFM into the neck network to enhance feature representation capabilities, and uses the LAMP algorithm to prune the model. The data storage module is used to store image enhancement results, model prediction outputs, and annotation information in the database, supporting structured management and export. The defect prediction module is used to load the trained model weights, deploy them on vehicle-mounted or edge devices, perform inference on the acquired images, output the defect location, category and confidence level, and feed back the failed prediction samples to the server for storage. When the cumulative number reaches a preset threshold, it performs re-labeling and incremental training to generate new model weights for updating the deployment. The front-end interaction module provides a human-computer interaction interface for image enhancement, model training and result display, enabling parameter configuration and visualization operations. The feedback learning module is configured to collect failed prediction samples and trigger re-labeling and incremental training. The defect prediction module sends the failed prediction samples generated during the inference process to the feedback learning module. The feedback learning module stores the collected failed samples in the data storage module, and when the cumulative number of failed samples reaches a preset threshold, it sends an incremental training instruction to the model training module. The model training module responds to the incremental training instruction by calling the failed samples and their associated data from the data storage module, and combines them with the results after manual review and annotation by the front-end interaction module to perform incremental training to generate new model weights. The model training module updates the defect prediction module with the generated new model weights, completing the model deployment update.

8. The system according to claim 7, characterized in that, In the data storage module, the database includes an image information table and a prediction result table; the image information table is used to record training images, acquisition time, and device number; the prediction result table is used to record defect category, confidence level, and model version number.

9. The system according to claim 7, characterized in that, The defect prediction module specifically includes: loading the trained and optimized model weight file and deploying it on an in-vehicle edge computing device or industrial inspection terminal; performing defect detection on GIS images acquired in real time or input in batches, and outputting the category, location coordinates, and confidence level of each defect target; filtering out detection results with low confidence based on a preset confidence threshold and pushing them as samples to be reviewed to the manual review stage; periodically feeding back failed samples to the server, and triggering re-annotation and incremental training when the cumulative number reaches a preset number to update the model weights; saving the inference results and generating a visualization output file containing defect location annotations.

10. The system according to claim 7, characterized in that, The front-end interaction module specifically includes: providing a graphical user interface that supports image uploading, image enhancement algorithm selection, model training control, and result browsing; supporting image zooming, target selection, category filtering, and result highlighting in the detection result display interface to facilitate defect area location analysis; supporting manual review and annotation of failed inference samples, and submitting the annotation results to the server for model updates.

Images