CN-121305453-B - Equipment identification and quality grade joint evaluation method for substation inspection image

CN121305453BCN 121305453 BCN121305453 BCN 121305453BCN-121305453-B

Abstract

The invention discloses a substation inspection image-oriented equipment identification and quality grade joint evaluation method, which comprises the steps of inputting an acquired original substation image into an image identification and quality grade joint evaluation model, wherein the model comprises a detection sensing image enhancement module, a target detection module, an ROI extraction module and an image quality evaluation module which are sequentially connected, carrying out enhancement processing on the original substation image by using the detection sensing image enhancement module to obtain an enhanced image, carrying out target detection on the enhanced image by using the target detection module to obtain a power equipment category label and ROI frame coordinate information, cutting the original substation image by using the ROI extraction module according to the ROI frame coordinate information to obtain a corresponding equipment ROI image, and carrying out non-reference image quality evaluation on the equipment ROI image by using the image quality evaluation module to obtain a quality grade label.

Inventors

ZOU ZHIWEI
XUE BING
TIAN TENG
GENG JIAQI
WANG ENHUI
GAO QINGWEI
ZHAO DAWEI
LI YONGXI
SHI WEIHAO
XING LU
QIU RUJIA
ZHAO LONG
JIN YUNAN
CAO JUN
SUN WEI

Assignees

国网安徽省电力有限公司电力科学研究院

Dates

Publication Date: 20260508
Application Date: 20251211

Claims (15)

1. The equipment identification and quality grade joint evaluation method for the substation inspection image is characterized by comprising the following steps of: Inputting an acquired original substation image into an image recognition and quality grade joint evaluation model, wherein the model comprises a detection sensing image enhancement module, a target detection module, an ROI extraction module and an image quality evaluation module which are sequentially connected, and the detection sensing image enhancement module comprises a semantic segmentation network, a backbone network and an enhancement network; The method comprises the steps of carrying out enhancement processing on an original substation image by using a detection perception image enhancement module to obtain an enhanced image, adopting a semantic segmentation network to process the original substation image to generate a corresponding semantic mask for representing that pixel points belong to an equipment area or a background area, adopting a backbone network to carry out feature extraction on the original substation image to obtain a feature map of the original substation image, generating semantic sensitive features based on the semantic mask and the feature map, inputting the semantic sensitive features into an enhancement network based on a U-Net architecture to carry out end-to-end enhancement propagation to obtain the enhanced image, wherein the generation of the semantic sensitive features based on the semantic mask and the feature map comprises carrying out global average pooling based on the feature map and the semantic mask after downsampling, calculating the duty vector of the equipment area on each channel of the feature map, mapping the duty vector into channel attention weight, and multiplying the channel attention weight with the feature map to generate the semantic sensitive features; performing target detection on the enhanced image by using a target detection module to obtain power equipment category labels and ROI frame coordinate information; Cutting out the corresponding equipment ROI image from the original substation image by utilizing an ROI extraction module according to the ROI frame coordinate information; And performing non-reference image quality evaluation on the equipment ROI image by using an image quality evaluation module to obtain a quality grade label.
2. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 1, wherein the original substation image comprises a visible light image and an infrared image; Accordingly, the enhanced image includes a visible light enhanced image and an infrared enhanced image.
3. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 1, wherein the target detection module comprises a backbone network, a multi-mode feature dynamic fusion network and a detection head; The method for detecting the target of the enhanced image by using the target detection module to obtain the power equipment category label and the coordinate information of the ROI frame comprises the following steps: the method comprises the steps of respectively extracting features of a visible light enhanced image and an infrared enhanced image by utilizing a backbone network to obtain visible light features and infrared features; fusing the visible light characteristics and the infrared characteristics by utilizing a multi-mode characteristic dynamic fusion network to obtain a fusion characteristic diagram; And processing the fusion feature map by using a detection head to obtain the position and width and height of the target center point, and outputting the type label of the power equipment and the coordinate information of the ROI frame.
4. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 3, wherein the backbone network is a four-scale feature extraction network obtained based on CSPREPRESNET structure expansion in a PP-YOLOE network model; the multi-mode feature dynamic fusion network is a four-scale feature fusion network obtained based on PAN network expansion in a PP-YOLOE network model.
5. The substation inspection image-oriented device identification and quality level joint evaluation method according to claim 1, wherein after the ROI extracting module is configured to clip the corresponding device ROI image from the original substation image according to the ROI frame coordinate information, the method further comprises: judging whether the confidence coefficient of the equipment ROI image is larger than or equal to a confidence coefficient threshold value; if yes, reserving the equipment ROI image, and finishing the equipment ROI image by utilizing a semantic mask guiding mechanism to obtain a finished ROI image; if not, a saliency mask generation method based on the intermediate feature image is called, and whether the semantic saliency region is contained in the ROI image region of the equipment is judged; if yes, reserving the equipment ROI image, and finishing the equipment ROI image by utilizing a semantic mask guiding mechanism to obtain a finished ROI image; If not, rejecting the ROI image of the equipment.
6. The substation inspection image-oriented device identification and quality level joint evaluation method according to claim 5, wherein the refining the device ROI image by using the semantic mask guidance mechanism to obtain a refined ROI image comprises: extracting a salient region of the power equipment by using a backbone characteristic network of YOLO; processing the saliency area by adopting a saliency prediction head to generate a saliency mask map; and performing logical AND operation on the mask map and the boundary box area of the equipment ROI image to obtain the refined ROI image.
7. The substation inspection image-oriented device identification and quality level joint evaluation method according to claim 1, wherein the performing non-reference image quality evaluation on the device ROI image by using the image quality evaluation module to obtain a quality level label comprises: Performing non-reference image quality evaluation on the equipment ROI image by using a quality evaluation module to obtain a quality score; and determining the quality score as a corresponding quality grade label according to quality thresholds of different power equipment types.
8. The substation inspection image-oriented device identification and quality level joint evaluation method according to claim 7, wherein the quality evaluation module comprises a semantic weighting layer, a feature extraction layer, a distortion branch and a quality prediction branch, and the quality evaluation module is used for performing non-reference image quality evaluation on the device ROI image to obtain quality scores, and the method comprises the following steps: Using a semantic weighting layer, giving 1.5 times weight to pixels belonging to a device region in a device ROI image based on a saliency mask map, and giving 0.3 times weight to pixels belonging to a background region; Extracting the characteristics of the equipment ROI image by utilizing the characteristic extraction layer to obtain equipment characteristics; Processing the equipment characteristics by using the distortion branches to obtain distortion characteristic vectors; And taking the distortion feature vector as auxiliary semantics of the quality prediction branch, and processing the equipment feature by utilizing the quality prediction branch to obtain the quality score.
9. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 1, wherein the detection perception image enhancement module and the target detection module adopt end-to-end joint training, and the detection perception image enhancement module adopts a joint loss function during training The method comprises the following steps: in the formula, To enhance pixel level reconstruction loss of the front and back images, Represent the first The detection of the loss of class-equipment, For example-level feature consistency loss, 、、 In order to lose the weight coefficient(s), Is the category difficulty weight.
10. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 9, wherein the calculation formula of the category difficulty weight is as follows: in the formula, For the momentum coefficient of the EMA, For the category difficulty weight of the last iteration, Lower values indicate greater difficulty in category detection The larger the size of the container, Representing an exponential function.
11. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 9, wherein the calculation formula of the instance level feature consistency loss is as follows: in the formula, For the same class of instance distances, For the distance of the heterogeneous instance, Is the spacing parameter in the instance-level feature consistency penalty.
12. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 1, wherein the detection perception image enhancement module and the target detection module adopt end-to-end joint training, and a loss function with fault perception weight adopted during training of the target detection module The method comprises the following steps: in the formula, In order to classify the loss of the device, The total loss is predicted for the bounding box, For the set of anomaly regions noted in the infrared map, For the failure-associated region weight, Is the first Failure association loss for each abnormal region.
13. The substation inspection image-oriented equipment identification and quality level joint evaluation method according to claim 12, wherein the calculation formula of the boundary box prediction total loss is as follows: in the formula, As the weight coefficient of the light-emitting diode, To measure the loss of the degree of overlap between the predicted and real frames, Is a loss of coordinate regression accuracy for constraining the bounding box.
14. The utility model provides a device identification and quality level joint evaluation system towards transformer substation's inspection image which characterized in that includes: The image input unit is used for inputting the acquired original substation image into a pre-deployed image recognition and quality grade joint evaluation model, the model comprises a detection sensing image enhancement module, a target detection module, an ROI extraction module and an image quality evaluation module which are sequentially connected, and the detection sensing image enhancement module comprises a semantic segmentation network, a backbone network and an enhancement network; The image recognition and quality grade joint evaluation model is used for carrying out enhancement processing on an original transformer substation image by utilizing a detection perception image enhancement module to obtain an enhanced image, carrying out target detection on the enhanced image by utilizing a target detection module to obtain a power equipment type label and ROI frame coordinate information, and cutting out the corresponding equipment ROI image from the original transformer substation image by utilizing an ROI extraction module according to the ROI frame coordinate information; The method comprises the steps of carrying out enhancement processing on an original substation image by using a detection perception image enhancement module to obtain an enhanced image, adopting a semantic segmentation network to process the original substation image to generate a corresponding semantic mask for representing that pixel points belong to a device area or a background area, adopting a backbone network to carry out feature extraction on the original substation image to obtain a feature map of the original substation image, generating semantic sensitive features based on the semantic mask and the feature map, inputting the semantic sensitive features into an enhancement network based on a U-Net architecture to carry out end-to-end enhancement propagation to obtain the enhanced image, wherein the generation of the semantic sensitive features based on the semantic mask and the feature map comprises carrying out global average pooling based on the feature map and the semantic mask after downsampling, calculating the duty vector of the device area on each channel of the feature map, mapping the duty vector into channel attention weight, and multiplying the channel attention weight with the feature map to generate the semantic sensitive features.
15. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method according to any of claims 1-13.

Description

Equipment identification and quality grade joint evaluation method for substation inspection image Technical Field The invention relates to the technical field of power equipment inspection and image processing, in particular to a substation inspection image-oriented equipment identification and quality grade joint evaluation method. Background Along with the intelligent development of the power industry, the automation degree and the intelligent degree of substation equipment are continuously improved, the stable operation of the substation equipment is crucial to the reliability of a power system, and equipment identification and state monitoring become operation and maintenance key links. Deep learning has remarkable results in the field of image recognition, but the traditional model is still limited in application to substation equipment recognition, and actual operation and maintenance face multiple challenges. As the core of the power grid, the thermal faults (such as overheat of joints and discharge of insulators) of the equipment account for over half of the faults of the electrical equipment, so that power failure accidents are easy to cause, and the requirements on real-time and accurate detection and state evaluation of the equipment are urgent. The infrared thermal imaging becomes the main stream of thermal fault detection due to the 'four-no' advantage, however, the problem of the traditional inspection mode is prominent, the manual inspection is faced with a huge infrared image library, the efficiency is low, the omission ratio is high, the infrared image is easy to be interfered by electromagnetic noise and atmospheric attenuation, the temperature detection is inaccurate, the accuracy is reduced due to the fact that the image enhancement is not associated with detection requirements in the prior art, the quality evaluation depends on traditional indexes, the image characteristics of equipment are not adapted, the misjudgment ratio is high, the multi-scale target of a transformer substation is difficult to detect, large-scale equipment (such as a lightning arrester) is difficult to integrally photograph, small-scale equipment (such as an insulator) is difficult to effectively identify, and the robustness of the traditional deep learning model is insufficient. The current intelligent study of electric power inspection is progressed, but a short plate still exists, namely the adaptability of a detection model is poor, an edge end of a large model is difficult to deploy, the light model is used for distinguishing small targets and similar equipment, the precision is to be improved, an image enhancement and detection task is disjointed, only visual effects are optimized, equipment edge blurring and false detection are easy to cause, the quality evaluation is poor in scene adaptability, the traditional NR-IQA algorithm is not optimized for equipment heat distribution and fault texture, the classification is lack of differentiation standard, and the false judgment rate is high. In the related art, in order to solve the problem of inaccurate defect identification caused by poor quality of inspection images of a transformer substation, a patent application document with publication number CN116228774A proposes a thought of 'precursor quality evaluation and post-identification', wherein an image with quality higher than a threshold value is screened out and then is input into a defect identification model for detection by designing a lightweight image quality evaluation network, so that the system essentially is a data filtering process, mainly focuses on quality screening before identification, the quality evaluation and an identification module are not related to each other at a characteristic level, the quality evaluation is only a pre-screening device, has no inverse optimization function on a detection task and does not relate to a cooperative optimization relation between detection and quality evaluation, and the 'precursor quality evaluation and post-identification' method directly discards low-quality images, cannot realize real-time enhancement or self-adaptive optimization, and therefore, the system cannot intelligently compensate 'bad images' and cannot reflect the quality level of equipment images, so that defect detection results are unstable. In literature (research on defect recognition algorithm of typical equipment of Transformer substation based on improved PP-YOLOE, sun Dedong, the university of Liaoning engineering technology, shuoshi, the research on an algorithm model for performing defect recognition by using a Swin-transform semantic segmentation auxiliary improved PP-YOLOE model is focused on how to improve defect recognition accuracy by using a deep learning model in a complex scene of the Transformer substation, therefore, the scheme performs local optimization around the precision and robustness of the PP-YOLOE model, such as enhancing semantic perception by introducing a t