CN117197530A - Insulator defect identification method based on improved YOLOv8 model and cosine annealing learning rate decay method - Google Patents

Abstract

The application provides an insulator defect identification method based on an improved YOLOv8 model and a cosine annealing learning rate attenuation method, relates to the field of power system inspection, and provides an insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method. Acquiring an electric transmission line insulator image by an unmanned aerial vehicle, marking the acquired normal insulator and defect insulator images of different materials by using labelimg to obtain a marking frame of the insulator, and dividing the marked insulator images into training sets according to a certain proportion; putting the obtained training set into an improved YOLOv8 model to construct an insulator defect identification model; and combining the insulator defect identification model to identify the acquired insulator image.

Description

Insulator defect identification method based on improved YOLOv8 model and cosine annealing learning rate decay method

Technical Field

The application belongs to the technical field of image processing, and particularly relates to an insulator defect identification method based on an improved YOLOv8 model and a cosine annealing learning rate attenuation method.

Background

The insulator is an indispensable part of a power system as an important component element of the power transmission line in the power system, so that the demand of China for electric power is continuously increased, and the electric energy quality and the reliability are very important, so that the power transmission line is required to be inspected. When the insulator is affected by load or environmental factors, such as high temperature, ice coating and the like, various defects often occur in the insulator, so that the safe and reliable operation of the power system is affected. Because the power transmission line environment of the power system is very complex, the power transmission line environment is only inspected by the traditional manual inspection, the efficiency is low, the manpower and material resources are very consumed, and the requirements of the inspection of the modern power system cannot be met. Therefore, the method has important practical significance in defect identification of the insulator, and can improve the safety and stability of the power system.

Although the currently used artificial intelligence algorithm can rapidly locate the defect insulator in the image, different types of insulators, such as glass insulators, composite insulators, ceramic insulators and the like, and different types of defects, such as pollution flashover, breakage, self-explosion and the like, can be used. The variety is various, and erroneous judgment is easy to occur. If different insulators and different defect types are trained separately, the sample requirement is excessive and the workload is huge, so that the defect identification problem of various insulators needs to be solved.

Along with the continuous development of computer vision and deep learning, the image recognition based on the deep learning has far overtaken the traditional image recognition algorithm in terms of precision and real-time. Deep learning has strong adaptive and nonlinear modeling capabilities, and has been successful in image processing and pattern recognition. The YOLO (You Only Look Once) model is a target detection model based on deep learning, and has a higher running speed compared with the traditional machine learning algorithm, and is widely used in a real-time system. YOLOv8 is the next significant updated version of YOLOv5, open source No. 10, year 1, 2023, by Ultralytics corporation, YOLOv8 has the advantage of being faster, more accurate, and easier to use than other algorithms than other image recognition algorithms, and the accuracy versus accuracy table for other algorithms is shown in table 1.

The method can be superior to a real-time detection task of the inspection of the transmission line in the power system, and the algorithm of the method has very wide application and important practical significance in the task of identifying the defects of the insulator.

In chinese patent document CN109949318A, there is a description of a method, but this method does not involve segmentation of radio frequency data, nor does it mention an operation of fusing two tasks of segmentation and classification.

Disclosure of Invention

The application aims to provide an insulator defect identification method based on an improved YOLOv8 model and a cosine annealing learning rate attenuation method, which is used for accurately identifying a single or a plurality of normal insulators and defective insulators in an image, so that accurate detection of the insulators is realized. And further, the efficiency and the success rate of inspection of the power transmission line of the power system are improved.

In order to achieve the technical aim, the application provides an insulator defect identification method based on an improved YOLOv8 model and a cosine annealing learning rate attenuation method, which is realized through the following steps:

s1, constructing an original data set of an insulator through an insulator image acquired by an unmanned aerial vehicle, and marking an insulator region in the image to obtain a marked data set;

s2, carrying out data preprocessing on the marked data set, and dividing the marked data set to obtain a training set and a testing set.

S3, putting the divided data set into an improved YOLOv8 model for training, and constructing an insulator defect identification model; and placing the image to be identified into a model, and identifying the insulator in the picture.

In the preferred scheme, in the step S1, labelimg picture marking tools are used for marking different types of normal insulators and defective insulators acquired through the unmanned aerial vehicle.

In a preferred scheme, in step S2, the labeled data set is subjected to data preprocessing, and the specific steps include:

s21, carrying out data cleaning on the obtained insulator data set to obtain a cleaned data set;

s22, performing rotation, scaling, translation and cutting on the obtained cleaned data set by using a data enhancement method, and performing data expansion on the data set to obtain an expanded data set.

In a preferred scheme, the divided data set is put into an improved YOLOv8 model for training, and in step S3, the improved YOLOv8 model comprises a backbond module, a characteristic fusion Neck module, a detection Head module and a cosine annealing learning rate attenuation method module;

the Backbone module consists of a plurality of convolution layers and a pooling layer and is used for extracting low-level and high-level characteristics from an input image;

the feature fusion Neck module consists of a plurality of convolution layers and an up-sampling layer and is used for fusing features from different layers of a backbone network;

the detection Head module consists of a plurality of convolution layers and a full connection layer and is used for predicting a target boundary box and class probability from the characteristics extracted from the Neck network;

the cosine annealing learning rate attenuation method module dynamically adjusts the learning rate according to training conditions during model training.

In a preferred scheme, data cleaning is denoising by using an FFDNet model; the data enhancement method uses a Mosaic data enhancement method.

In a preferred scheme, the backbox module consists of a Darknet-53 and comprises a convolution module CBS, a residual module C2f and a spatial pyramid pooling structure SPPF;

the convolution module CBS is used for firstly extracting the input feature map as a feature map with more obvious state features;

the residual error module C2f is used for lightening the model and obtaining richer gradient flow information;

the spatial pyramid pooling structure SPPF is used for performing multiple splicing on feature graphs obtained by three global maximum pooling and global average pooling to realize feature fusion.

In a preferred scheme, the feature fusion Neck module performs upsampling and downsampling operations by adopting SPPF and FPN-PAN modules: the feature map obtained after SPPF features of the spatial pyramid pooling structure are fused is subjected to up-sampling operation through an FPN module, feature information in the feature map is transferred to a next-stage feature map, and position information of a target to be detected in an image is fused to a previous-stage feature map through up-sampling operation through a PAN module; thus, feature maps with different levels and different sizes are obtained.

In the preferred scheme, the detection Head module uses a detection Head without an anchor, does not use a predefined anchor frame to match a target frame, and predicts the position and the size of the target frame directly on a feature map obtained by the feature fusion Neck module; adding a CBAM light attention mechanism into a network; the position and the size of the target frame are calculated by replacing the original CIoU+DFL loss function by using the Wise-IoUv3 loss function.

In a preferred scheme, the cosine annealing learning rate attenuation method module reduces the learning rate through a cosine function, and the formula is as follows:

wherein i represents a first iteration run; n is n _min And n _max Represents the maximum value and the minimum value of the set learning rate; t (T) _cur Representing how many epochs are currently executed; t (T) _i Representing the total epoch number in the ith run.

The application has the advantages and beneficial effects that: the application utilizes the most excellent image classification, object detection and instance segmentation task model YOLOv8 at present, has the advantages of high speed, high precision and the like, and meets the high requirement of real-time inspection of the power transmission line of the power system. In addition, the improved lightweight network of the YOLOv8 can be suitable for various devices, and has wide coverage and high practicability. The application combines the cosine annealing learning rate attenuation method with the improved YOLOv8 model, and improves the training speed and the training precision by changing the learning rate in the training process. In addition, the cosine annealing learning rate attenuation method can enable the improved YOLOv8 model to have a larger learning rate in the initial stage of training, so that the convergence rate of the model is increased, and the learning rate is gradually reduced in the later stage of training, so that the model is more stable.

Drawings

FIG. 1 is a flow chart of an embodiment of an insulator defect identification method based on an improved YOLOv8 model and a cosine annealing learning rate decay method;

FIG. 2 is a schematic diagram of a marker dataset tool (labelimg) of the present application;

FIG. 3 is a schematic diagram of an improved YOLOv8 network in accordance with the present application;

FIG. 4 is a graph showing the variation of learning rate in a modified YOLOv8 model for the cosine annealing learning rate decay method according to the present application;

FIG. 5 is a PR graph obtained by training a model in the present application;

FIG. 6 is a graph corresponding to each parameter obtained by training a model in the present application;

fig. 7 is a schematic diagram of identifying an insulator image acquired by an unmanned aerial vehicle using a trained model in the present application.

Detailed Description

As shown in fig. 1 to 7, an insulator defect identification method based on an improved YOLOv8 model and a cosine annealing learning rate attenuation method comprises the following steps:

step 1: constructing an original insulator data set through an insulator image acquired by the unmanned aerial vehicle, and marking an insulator region in the image to obtain a marked data set;

step 2: and carrying out data preprocessing on the marked data set, and dividing the marked data set to obtain a training set and a testing set.

Step 3: putting the divided data set into an improved YOLOv8 model for training, and constructing an insulator defect identification model; and placing the image to be identified into a model, and identifying the insulator in the picture.

Step 4: performance evaluation is performed on the well-trained insulator defect recognition model to obtain a rating index including accuracy (Pre), recall (recall, rec), average accuracy (average precision, AP) of each class of targets, average accuracy mean (mean average precision, mAP) and transmission frames per second (frames per second, FPS).

Specifically, step 1 includes:

and step 1-1, shooting and sampling insulators in the power transmission line by using the unmanned aerial vehicle under different complex backgrounds to obtain original atlas of various normal insulators and defective insulators.

And step 1-2, marking different types of normal insulators and defective insulators (glass insulators, composite insulators, ceramic insulators and defects) acquired through the unmanned aerial vehicle by using a labelimg picture marking tool to obtain the PascalVOC format data tag. And converting the data tag to obtain txt format data which can be used by the model, wherein the txt format data comprises index of the category of the target insulator, and labeling information such as x coordinate, y coordinate, width and height of the center point of the boundary box of the target insulator, and the like, so as to obtain a labeled data set.

Specifically, step 2 includes:

step 2-1, performing data cleaning on the obtained insulator data set to obtain a cleaned data set;

the FFDNet model based on the deep neural network is used for denoising noise influence existing in the insulator image;

step 2-2, performing data expansion on the obtained operation multi-data set subjected to rotation, scaling, translation and cutting on the cleaned data set by using a data enhancement method;

the data enhancement method uses a mosaics data enhancement method, 4 pictures are randomly used for combination and splicing each time, and the operation of overturning and changing the color gamut is performed on each spliced picture, so that each spliced picture can be provided with various insulators and defects thereof to increase the model generalization capability.

Step 2-3, according to 8, the data set after data cleaning and data enhancement: the 2 scale is divided into training and testing sets.

Specifically, step 3 includes:

and 3-1, the Backbone network of the model is divided into a backhaul module, a characteristic fusion Neck module and a detection Head module.

Step 3-2, a backfone module is a part of the model for extracting image features and consists of a plurality of convolution layers and pooling layers, and is used for extracting low-level and high-level features from an input image;

the operation steps of the backup module are as follows: firstly, the preprocessed input image 640 x 3 passes through a first convolution module, the convolution kernel size is 3*3, the step length is 2, the padding is 1, and a 320 x 64 feature image is output; sequentially passing through a first CBS module, wherein the convolution kernel size is 3*3, the step length is 2, the padding is 1, and the first C2f module outputs a 160×160×128 feature map; then through a second CBS module, the convolution kernel size is 3 multiplied by 3, the step length is 1, the padding is 1, and a second C2f module outputs a characteristic diagram of 80×80×256; the characteristic diagram of 40 x 512 is output through a third CBS module and a third C2f module; through a fourth CBS module and a fourth C2f module, outputting a 20 x 512 feature map; finally, 80 x 256, 40 x 512 and 20 x 512 feature maps are input to the neg layer through the added CBAM attention mechanism and SPPF layer.

The convolution module CBS is used for firstly extracting the input feature map as a feature map with more obvious state features; the residual error module C2f is used for lightening the model and obtaining richer gradient flow information; the spatial pyramid pooling structure SPPF is used for performing multiple splicing on feature graphs obtained by three global maximum pooling and global average pooling to realize feature fusion.

And 3-3, a feature fusion Neck module which is a part for connecting the back bone module and the detection Head module in the model. The system consists of a plurality of convolution layers and an up-sampling layer, and is used for fusing the characteristics from different layers of a backbone network;

wherein, firstly, the 20 x 512 characteristic diagram output by the SPPF in the backup is processed by one time of upsample and the 40 x 512 characteristic diagram output at the same time to obtain the 40 x 512 characteristic diagram, performing Concat operation on the feature map of 80 x 256 and the feature map of 80 x 256 through one C2f convolution, and outputting the feature map of 80 x 256 obtained through one C2f convolution to a Head; the obtained 80 x 256 feature map is the same as the feature map 40 x 512 obtained by performing a Concat operation on the cbS operation and the above feature map 40 x 512, and the feature map 40 x 512 obtained by performing a C2f convolution is input into a Head; and performing Concat operation on the above characteristic diagram of 40 x 512 by performing C2f convolution operation and the characteristic diagram of 20 x 512 output by the SPPF to obtain the characteristic diagram of 20 x 512, and finally performing C2f convolution to obtain the characteristic diagram of 20 x 512 to be input into a Head.

And 3-4, a detection Head module which is a part for predicting the target bounding box and the class probability in the model. It consists of multiple convolution layers and full connection layers for predicting target bounding boxes and class probabilities from features extracted from the Neck network.

The feature map of 80×80×256, 40×40×512 and 20×20×512 obtained from the neg is subjected to two-dimensional convolution after 2 CBS operations, BCE Loss classification Loss calculation and Distribution Focal Loss and replaced Wise-IoUv3 regression Loss calculation are respectively performed to obtain a pre-selected frame, and finally, a non-maximum suppression NMS operation is adopted to screen a plurality of prediction frames to determine a prediction frame of the final insulator defect and output the prediction frame.

Step 3-5, wise-IoUv3 can focus on the anchor frame with general quality. At the same time, there is a faster speed because the aspect ratio is not calculated.

Defining an outlier to describe the quality of the anchor box, which is defined as:

a small outlier means that the anchor box is of high quality, which is assigned a small gradient gain to bring the bounding box back into focus on the anchor box of normal quality. Assigning smaller gradient gains to anchor boxes with larger outliers will effectively prevent low quality examples from producing larger detrimental gradients. Constructing a non-monotonic Focusing coefficient by using beta, and constructing WIoU v3:

the Wise-IoU v3 algorithm utilizes beta to construct a non-monotonic focusing coefficient r, so that a plurality of negative influences are effectively shielded in the training process, and the model precision is further improved.

Step 3-6, a cosine annealing learning rate attenuation method module reduces the learning rate through a cosine function, can dynamically change the learning rate during model training, and enables the learning rate to adapt to different training states, and the formula is as follows:

wherein i represents a first iteration run; n is n _min And n _max Represents the maximum value and the minimum value of the set learning rate; t (T) _cur Representing how many epochs are currently executed; t (T) _i Representing the total epoch number in the ith run.

Specifically, step 4 includes:

step 4-1, putting the training set finished in the step 2 into an improved YOLOv8 model for training to obtain an insulator defect identification model; secondly, placing the sample image to be detected into a trained insulator identification model to obtain an insulator identification result in the image to be identified;

step 4-2, firstly, saving a weight file after training is completed, then putting the test set image into a model for recognition, and then comparing the test set image with the true position marked in the original image to obtain a corresponding evaluation index; secondly, the original YOLOv8 model automatically puts each evaluation index into a specified folder; finally, the evaluation index and the overall evaluation generated in each iteration and described in the step 4 can be obtained in the CSV file of the folder, and the formula is as follows:

accuracy (Pre) refers to the ratio of the number of documents retrieved to the total number of documents retrieved, i.e., the proportion of documents retrieved that are actually related to the retrieved documents. The formula is:

the recall (recall, rec) refers to the ratio of the number of relevant documents retrieved to the number of relevant documents missed, i.e. the ratio of documents retrieved that are actually relevant to the documents that are actually relevant. The formula is:

where TP represents the positive example and is predicted as the number of samples of the positive example. FP represents the number of samples of the negative example but predicted to be positive. FN represents the number of samples of positive examples but predicted as negative examples. TN represents the number of samples that are predicted as negative examples.

And the F1 value equation taking into account the accuracy and recall is:

the average precision mean (mean average precision, mAP) refers to the average of all class AP values. The formula is:

wherein n represents the number of categories, AP _i Mean precision value representing the ith class

The number of transmission frames per second (frames per second, FPS) refers to how many frames of images per second can be processed in video processing. The higher the FPS, the smoother the video, the formula:

where pre-process represents image preprocessing time, reference represents inference speed, refer to the time from the image input model to the model output result after preprocessing, and NMS represents post-processing time.

The verification effect is determined using four parameters of F1, recall, mAP and FPS as shown in Table 1:

TABLE 1 Effect of different algorithms

The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method can identify the conditions of normal insulators and defective insulators (breakage, pollution flashover and self-explosion), can meet the requirement of real-time detection of a power transmission line of a power system, and solves the problems of low efficiency and low speed during manual inspection.

The above embodiments are only for illustrating the technical solution of the present application, and are not limiting; although the application has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims (9)

1. An insulator defect identification method based on an improved YOLOv8 model and a cosine annealing learning rate attenuation method is characterized by comprising the following steps:

s1, constructing an original data set of an insulator through an insulator image acquired by an unmanned aerial vehicle, and marking an insulator region in the image to obtain a marked data set;

s2, carrying out data preprocessing on the marked data set, and dividing the marked data set to obtain a training set and a testing set.

S3, putting the divided data set into an improved YOLOv8 model for training, and constructing an insulator defect identification model; and placing the image to be identified into a model, and identifying the insulator in the picture.

2. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 1, which is characterized in that: in the step S1, labelimg picture marking tools are used for marking different types of normal insulators and defect insulators acquired through the unmanned aerial vehicle.

3. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 1, which is characterized in that: in the step S2, the marked data set is subjected to data preprocessing, and the specific steps include:

s21, carrying out data cleaning on the obtained insulator data set to obtain a cleaned data set;

s22, performing rotation, scaling, translation and cutting on the obtained cleaned data set by using a data enhancement method, and performing data expansion on the data set to obtain an expanded data set.

4. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 1, which is characterized in that: putting the divided data set into an improved YOLOv8 model for training, wherein the improved YOLOv8 model in step S3 comprises a backbox module, a characteristic fusion Neck module, a detection Head module and a cosine annealing learning rate attenuation method module;

the Backbone module consists of a plurality of convolution layers and a pooling layer and is used for extracting low-level and high-level characteristics from an input image;

the feature fusion Neck module consists of a plurality of convolution layers and an up-sampling layer and is used for fusing features from different layers of a backbone network;

the detection Head module consists of a plurality of convolution layers and a full connection layer and is used for predicting a target boundary box and class probability from the characteristics extracted from the Neck network;

the cosine annealing learning rate attenuation method module dynamically adjusts the learning rate according to training conditions during model training.

5. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 3, wherein the insulator defect identification method is characterized by comprising the following steps: the data cleaning is to use an FFDNet model for denoising; the data enhancement method uses a Mosaic data enhancement method.

6. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 4, which is characterized in that: the backbox module consists of a Darknet-53 and comprises a convolution module CBS, a residual module C2f and a space pyramid pooling structure SPPF;

the convolution module CBS is used for firstly extracting an input feature map as a feature map with more obvious state features;

the residual error module C2f is used for lightening the model and obtaining richer gradient flow information;

the SPPF is used for splicing the feature graphs obtained by three global maximum pooling and global average pooling for multiple times to realize feature fusion.

7. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 4, which is characterized in that: the feature fusion Neck module performs up-sampling and down-sampling operations by adopting an SPPF and FPN-PAN module: the feature map obtained after SPPF features of the spatial pyramid pooling structure are fused is subjected to up-sampling operation through an FPN module, feature information in the feature map is transferred to a next-stage feature map, and position information of a target to be detected in an image is fused to a previous-stage feature map through up-sampling operation through a PAN module; thus, feature maps with different levels and different sizes are obtained.

8. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 4, which is characterized in that: the detection Head module is used for detecting the position and the size of the target frame directly on the feature map obtained by the feature fusion Neck module without using a predefined anchor frame to match the target frame; the CBAM light attention mechanism is added into the network, so that the model can more notice the place with the most abundant information in each insulation sub-image. The position and the size of the target frame are calculated by replacing the original CIoU+DFL loss function with the Wise-IoUv3 loss function with higher precision and higher speed.

9. The insulator defect identification method based on the improved YOLOv8 model and the cosine annealing learning rate attenuation method according to claim 4, which is characterized in that: the cosine annealing learning rate attenuation method module reduces the learning rate through a cosine function, and the formula is as follows:

wherein i represents a first iteration run; n is n _min And n _max Represents the maximum value and the minimum value of the set learning rate; t (T) _cur Representing how many epochs are currently executed; t (T) _i Representing the total epoch number in the ith run.