CN-122023242-A - Intelligent detection method, system, electronic equipment and medium for insulator defects

CN122023242ACN 122023242 ACN122023242 ACN 122023242ACN-122023242-A

Abstract

The invention belongs to the technical field of insulator detection, and provides an intelligent detection method, an intelligent detection system, electronic equipment and an intelligent detection medium for insulator defects, which comprise the steps of carrying out image preprocessing on an acquired original image, extracting characteristic images of the image by adopting a deep convolution neural network, guiding YOLOv a model to concentrate on key characteristic areas in the image by using an attention mechanism, automatically adjusting the size and the shape of a YOLOv model convolution kernel by using a dynamic convolution technology, enhancing the perception of a YOLOv model on a small target by using a small target characteristic extraction and multi-scale characteristic fusion method, and finally visualizing and storing detection results. According to the invention, through the introduction of an attention mechanism, a small target feature extraction method and dynamic convolution, the precision and stability of the YOLOv model in insulator defect detection are obviously improved, and the identification capability of micro defects and the robustness under a complex scene are obviously improved.

Inventors

LI QINYUN
ZHANG QIYUE
WU JIANING
Yang Zibi
CHEN ZIYUE
Song Lingmo
CHEN MINLE
LU YUXUAN
Xu Jiangjiao

Assignees

上海电力大学

Dates

Publication Date: 20260512
Application Date: 20251219

Claims (10)

1. An intelligent detection method for defects of an insulator is characterized by comprising the following steps: Image preprocessing, namely acquiring an insulator image and preprocessing an acquired original image; Deep feature extraction, namely extracting a feature image of the image by adopting a deep convolutional neural network; inputting the characteristic image into YOLOv model, and optimizing the characteristic extraction effect by using normalization processing mode; identifying and detecting defects in the feature images by adopting a YOLOv model, directing YOLOv model to concentrate on key feature areas in the images by using an attention mechanism, automatically adjusting the size and shape of a YOLOv model convolution kernel by using a dynamic convolution technology, and enhancing the perception of YOLOv model to small targets by using a small target feature extraction and multi-scale feature fusion method; and (3) visualizing and storing the detection result, namely displaying the detection result through a display interface, displaying the current running state of the insulator and possible defect positions, automatically recording image data, the detection result and related parameters, and storing the image data, the detection result and related parameters into a database.
2. The intelligent detection method of the insulator defect according to claim 1, wherein in the image preprocessing of the collected original image, denoising, contrast enhancement and local feature enhancement processing are performed on the original image, so that background noise in the image is removed, and contrast of a defect area is enhanced.
3. The intelligent detection method for insulator defects according to claim 1, wherein the attention mechanism is used for guiding YOLOv the model to focus on key feature areas in the image, and the processing procedure of the feature map after YOLOv the model joins the attention mechanism is represented by the following formula: Wherein, the The feature map after the attention mechanism adjustment is represented, F (x) is the original feature map, and A (x) is the attention weight.
4. The intelligent detection method of insulator defects according to claim 1, wherein the size and shape of the YOLOv model convolution kernel are automatically adjusted by using a dynamic convolution technology, after the dynamic convolution is introduced, the convolution kernel weight of the YOLOv model is adaptively adjusted according to the input characteristics, and the convolution operation of the dynamic convolution is expressed by the following formula: Where W i represents each dynamic convolution kernel, x i is the input feature, and N is the number of convolution kernels.
5. The intelligent detection method of insulator defects according to claim 1, wherein in the defect identification, a multi-scale feature fusion method uses a feature pyramid structure to avoid information loss caused by scale scaling, and a calculation formula is as follows: Wherein, the Is the first A feature map of the layer is provided, For the original feature map of this layer, C onv represents the convolution operation and U psample is the upsampling operation.
6. The intelligent detection method of insulator defects according to claim 1, wherein in the defect identification, the small target feature extraction method adopts a transducer-based local attention model to enhance YOLOv the perception of the small target by the model, and the transducer-based local attention model captures local features through a self-attention mechanism, and the calculation formula is as follows: Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector.
7. The intelligent detection method of insulator defects according to claim 1, wherein in the defect identification, stable detection capability is provided under different environments by combining a multi-channel dynamic feature fusion mode, and a formula of the multi-channel feature fusion is as follows: Wherein, the For the fused features, F i is the feature of the ith channel, Is the weight of the channel.
8. An intelligent detection system for defects of an insulator, the system comprising: The data acquisition module is used for acquiring an original image of the insulator; the data processing module is used for preprocessing the original image of the insulator; the deep learning analysis module is used for extracting high-level features of the preprocessed original image by adopting a deep convolutional neural network and providing clear input data for the defect recognition module; The image conversion module is used for converting the preprocessed image into an input format required by the YOLOv model and carrying out normalization processing; the defect identification module is used for identifying and detecting defects in the feature images by adopting a YOLOv model combined with an attention mechanism, small target feature extraction and dynamic convolution technology; And the result output module is used for displaying the detection result through a display interface, displaying the current running state of the insulator and possible defect positions, automatically recording image data, the detection result and related parameters, and storing the image data, the detection result and the related parameters into a database.
9. An electronic device comprising a processor and a memory communicatively coupled to the processor for storing instructions executable by the processor, wherein the processor is configured to perform the intelligent method for detecting an insulator defect of any one of the claims 1-7.
10. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the method for intelligent detection of an insulator defect according to any one of claims 1 to 7.

Description

Intelligent detection method, system, electronic equipment and medium for insulator defects Technical Field The invention belongs to the technical field of insulator detection, and particularly relates to an intelligent insulator defect detection method, an intelligent insulator defect detection system, electronic equipment and a medium. Background Insulators are one of the important electrical devices in an electric power system, which support an electric power line and ensure normal current flow. As the scale of the power system is continuously increased, the insulator is subjected to greater and greater pressure in the high-voltage line, and various types of defects (such as cracks, dirt and the like) can appear on the surface of the insulator along with long-term use, which can cause the power system to malfunction. Therefore, the method can timely and accurately detect the defects of the insulators, and maintain and repair the defects, and becomes a key task for guaranteeing the safe operation of the power system. The traditional insulator detection methods depend on manual inspection or algorithms based on traditional image processing, the methods generally depend on manual experience, detection accuracy and efficiency are low, and efficient real-time monitoring cannot be performed under complex environmental conditions. However, with the advent of deep learning, automated detection techniques based on computer vision have become the mainstream. Particularly, the application of the Convolutional Neural Network (CNN) in the aspect of image recognition obviously improves the detection precision and efficiency. Currently, insulator defect detection mainly depends on a traditional image processing method and a target detection method based on deep learning, wherein the most representative method is a traditional YOLO target detection method. YOLO (You Only Look Once) is a highly efficient target detection algorithm, which has been widely used in the field of industrial detection, including insulator defect identification. These models enable quick identification of the defect location of the insulator in the image by learning features from a large amount of data. However, the existing YOLO model still has the problems of insufficient precision, difficult extraction of small target features and easy interference from complex environments. The conventional YOLO mainly performs insulator defect detection by: (1) And (3) processing an input image, namely carrying out normalization processing on the input insulator image so as to adapt to the input requirement of the neural network. (2) Feature extraction, namely extracting deep features of the image by using a deep convolutional neural network (such as Darknet). (3) Target detection and framing by full connection layer prediction bounding box (Bounding Box) and class score and non-maximum suppression (NMS) is employed to remove overlapping boxes. (4) And (3) defect classification and output, namely determining defect types according to classification results, and outputting detection results, wherein the detection results comprise defect positions, confidence scores and the like. YOLOv8 is used as the latest version of the YOLO series, and is optimized in terms of detection speed, precision and model weight, so that the method has strong real-time performance and adaptability in industrial detection scenes. However, the existing YOLOv model still has the following drawbacks: 1. the precision is not enough: the existing YOLOv model has low precision in identifying defects with complex details, and particularly has the problem that missed detection or false detection easily occurs in the detection of small defects such as cracks, dirt and the like with irregular shapes. 2. Small target feature extraction is difficult: Insulator defects tend to be smaller and have lower contrast, especially in complex backgrounds or under different lighting conditions, where the target is weaker. Although the existing YOLOv model has been optimized for small objects, difficulties remain in extracting features of these small objects. The root cause is that the network structure of YOLOv is mainly used for extracting the characteristics through global information, and the local detail grabbing capability of the small target characteristics is relatively weak. Particularly, in the case of a low resolution image or a large noise interference, the edge information of the small object may be ignored, which results in inaccurate feature extraction and thus affects the accuracy of detection. 3. Is easy to be interfered by complex environments: due to factors such as light change and complex background, the existing YOLOv model is easily affected by environmental noise, so that the identification effect is unstable. In summary, the existing YOLOv model has the problems of insufficient precision, difficult extraction of small target features and easy interference of complex environments i