An insulator abnormal state identification method based on FA-YOLOv8

Abstract

This invention discloses an insulator anomaly state identification method based on FA-YOLOv8, belonging to the field of target recognition technology for power transmission components. The method includes data preparation and preprocessing, re-labeling the insulator dataset and adding labels to each image, constructing a dual-convolutional fusion module to replace some convolutional modules in the YOLOv8 model backbone network, performing the first-stage feature extraction, and building a multi-attention mechanism fusion framework for feature layer sampling; performing the second, third, and fourth stages of feature extraction, which are then sequentially input into the multi-attention mechanism fusion framework; performing feature transfer and task decoupling, and improving the loss function in the YOLOv8 model by designing a composite loss function. This invention, employing the aforementioned FA-YOLOv8-based insulator anomaly state identification method, solves the problems of high computational load, slow response speed, low identification efficiency, and poor adaptability of current target recognition methods when dealing with minor defects in transmission line insulators and strong background interference.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology for power transmission components, and in particular to a method for identifying abnormal states of insulators based on FA-YOLOv8. Background Technology

[0002] Object recognition, as an important application of artificial intelligence visual learning, mainly involves identifying and locating ideal targets in captured images. Currently, object recognition is widely used in various fields such as autonomous driving, robot vision, and human-computer interaction. With the rapid advancement of computers and the development of deep learning technology, more and more object recognition models based on deep learning algorithms have emerged, significantly improving recognition accuracy and efficiency.

[0003] Identifying abnormal conditions in insulators often relies on manual inspections, which consumes significant manpower, material resources, and financial resources. To reduce the frequency of manual tower climbing, avoid personnel hazards, save costs, and improve efficiency, deep learning-based target recognition models have been applied to insulator inspections. However, while existing target recognition models have achieved some research results, they still suffer from problems such as high computational cost, slow response speed, low identification efficiency, and poor adaptability when faced with minor defects and strong background interference. Therefore, how to solve these problems remains an important area of ​​urgent research. Summary of the Invention

[0004] The purpose of this invention is to provide an insulator abnormality identification method based on FA-YOLOv8, which solves the problems of large computational load, slow response speed, low identification efficiency and poor adaptability of the current target identification method when there are small defects in the insulators of transmission lines and strong background interference. It can fully extract shallow information, enhance the identification effect of small targets and improve the identification accuracy.

[0005] To achieve the above objectives, this invention provides a method for identifying abnormal states of insulators based on FA-YOLOv8, comprising the following steps:

[0006] S1. Data preparation and preprocessing: The insulator dataset was re-labeled according to the requirements of Pascal VOC2007, and label information was added to each image to ensure the accuracy and reliability of the training data. The 160×160×3 image was used as the network input.

[0007] S2. Construct a dual convolution fusion module by fusing GhostConv and DConv to replace part of the Conv2d_BN_SiLU convolution module in the backbone network of the YOLOv8 model.

[0008] S3. Input the labeled insulator image into the first stage of the FANet backbone network composed of ConvModule and dual convolutional fusion module to perform the first stage feature extraction.

[0009] S4. Combine the efficient channel attention mechanism, spatial attention mechanism and feature pyramid module to build a multi-attention mechanism fusion framework, and apply it to the feature layer sampling stage of the dual convolution fusion module.

[0010] S5. After completing the first stage of feature extraction in the FANet backbone network, the second, third and fourth stages of feature extraction are performed and then sequentially input into the multi-attention mechanism fusion framework.

[0011] S6. The output features of the multi-attention mechanism fusion framework are passed through the FPN and PAN networks and decoupled heads are used to process the classification and localization tasks in the recognition process respectively, which further enhances the expressive power of multi-scale features.

[0012] S7. Improve the loss function in the YOLOv8 model by replacing the CIOU loss in the original model with the normalized Wasserstein distance (NWD). Combine the normalized Wasserstein distance with the binary cross-entropy classification loss function and the distributed focus regression loss function to design a composite loss function.

[0013] Preferably, the main body of FA-YOLOv8 includes an input, a dual convolutional fusion module, a multi-attention mechanism fusion framework, a fast adaptive backbone network FANet, and a head recognition layer.

[0014] Preferably, the specific steps for constructing the dual convolutional fusion module using S2 are as follows:

[0015] S21. Using the feature map from the previous layer as the input to this layer, dilated convolutions DConv with dilation rates r1, r2, and r3 learn features from different receptive fields. This process is represented as follows:

[0016]

[0017] Among them, F input For the feature mapping of the previous layer, Let R1, R2, and R3 represent DConv with a spatial resolution of 3×3 and an expansion rate of r1, r2, and r3, respectively, and let R1, R2, and R3 represent the feature maps obtained from DConv with an expansion rate of r1, r2, and r3, respectively.

[0018] S22. To minimize feature redundancy, reduce the number of parameters, and lighten the model, GhostConv convolution and standard convolution are used to refine the input features; the specific process is as follows:

[0019]

[0020] Where GhostConv represents the phantom convolution GhostConv, Conv 1×1 F1 represents a standard convolution with a 1×1 kernel, F2 represents a feature map obtained by passing GhostConv only, and F3 represents a feature map obtained by passing GhostConv and standard convolution in turn.

[0021] S23. Concatenate the obtained feature maps R1, R2, R3, F1, and F2, and then perform dimensionality reduction using a standard convolution with a 1×1 kernel. The specific process is as follows:

[0022] F output =Conv 1×1 (Concat(R i ,F1,F2)) i=1,2,3 (3)

[0023] Among them, F output This represents the output feature map.

[0024] Preferably, the specific process of S3 feature extraction is as follows: ConvModule refines the features and performs downsampling, the dual convolutional fusion module uses channel redundancy to learn features and reduce computational load, and DConv learns local and non-local features from multiple receptive fields.

[0025] Preferably, the specific steps for building a multi-attention mechanism fusion framework using S4 are as follows:

[0026] S41. In the efficient channel attention mechanism, global average pooling is used to process the input feature map, and a standard convolution with kernel k=5 and dimension 1 is used to perform a convolution operation on the output feature map to extract image features. The sigmoid activation function is then used to calculate the weights of each channel.

[0027]

[0028] Among them, X c (i,j) represents the value at position (i,j) of the c-th channel, σ k Represents the Sigmoid operation, ω c Indicates the weights of different channels;

[0029] S42. Complete the residual connection, multiply the obtained weights element-wise with the corresponding original feature maps to obtain new feature maps;

[0030]

[0031] Among them, F ECA This represents the newly generated feature map;

[0032] S43. Based on the salient target recognition method and point cloud feature extraction method, calculate the spatial attention between feature layers of different scales in the feature pyramid module, and use SA to fuse two feature maps with different spatial resolutions:

[0033] M s (F x )=σ(Comv 1×1 ([AvgPool(x);MaxPool(x)])) (6)

[0034] M s (F y )=σ(Conv 1×1 ([AvgPool(y);MaxPool(y)])) (7)

[0035] F SA =Conv 1×1 (Concat(M s (F x ),M s (F y ))) (8)

[0036] Where σ represents the Sigmoid operation, AvgPool represents average pooling, MaxPool represents max pooling, and x and y are feature maps with two different spatial resolutions.

[0037] S44. The final feature map F output by the feature pyramid. FP Feature map F after channel attention processing CA The final output is obtained by connecting the components along the channel direction;

[0038] F output =Concat(F ECA ,F FP (9)

[0039] Concat represents the concatenation operation.

[0040] Preferably, the second, third, and fourth stages of the FANet backbone network in S5 are all composed of ConvModule, dual convolutional fusion module, and multi-attention mechanism fusion framework.

[0041] Preferably, the formula for calculating the normalized Wasserstein distance (NWD) in S7 is as follows:

[0042]

[0043] in, Represents the center coordinates (cx) of the recognition frame a a ,cy a ) and width and height It measures the distance between frames a and b, NWD(N a N b ) represents the standardized form of the exponential function, where exp is the exponential function and C is the dataset association constant.

[0044] Preferably, the formula for the composite loss function in S7 is as follows:

[0045]

[0046] Among them, L BCE L is the binary cross-entropy classification loss, used to compare the classification probability distribution with the actual labels. DFL For distributed focus regression loss, L NWD To normalize the Wasserstein distance loss, L All This represents the overall loss.

[0047] Therefore, the present invention employs the above-mentioned FA-YOLOv8-based insulator abnormality state identification method, which has the following beneficial effects:

[0048] (1) The present invention constructs a dual convolutional fusion module that integrates GhostConv and Dconv, which can learn feature information on multiple receptive fields while reducing computational complexity;

[0049] (2) This invention combines efficient channel attention mechanism, spatial attention mechanism and feature pyramid module to build a simple and efficient multi-attention mechanism fusion framework, which extends the feature learning range to the surrounding environment of the insulator, enhances background features, improves abnormal state response and reduces feature degradation, and helps to improve the adaptability of YOLOv8 model to complex environmental changes.

[0050] (3) This invention combines the dual convolutional fusion module and the multi-attention mechanism fusion framework to propose the fast adaptive backbone network FANet. By replacing the original backbone network, it can reduce the computational burden of target recognition tasks, provide more powerful feature extraction and representation capabilities, and improve the accuracy and efficiency of target recognition.

[0051] (4) This invention uses the normalized Wasserstein distance to replace the CIOU loss and combines it with the binary cross-entropy classification loss and the distributed focus regression loss to propose a more adaptive composite loss function, which can improve the identification accuracy of insulator abnormal states and the robustness of the model.

[0052] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0053] Figure 1 This is a flowchart of an insulator abnormal state identification method based on FA-YOLOv8 according to the present invention;

[0054] Figure 2 This is a diagram showing the overall network structure of the FA-YOLOv8 model of this invention.

[0055] Figure 3 This is a network structure diagram of the dual convolutional fusion module of the present invention;

[0056] Figure 4 This is a network structure diagram of the multi-attention mechanism framework of the present invention;

[0057] Figure 5 The figures show a comparison of the iteration curves before and after the model improvement of this invention, where (a) is the loss function curve and (b) is the average accuracy curve.

[0058] Figure 6 This is a diagram showing the results of identifying abnormal insulator states under different conditions using the FA-YOLOv8 model of this invention.

[0059] Figure 7 This is a bar chart showing the comparison results between the FA-YOLOv8 model of this invention and various other models. Detailed Implementation

[0060] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0061] like Figure 1 As shown, this invention provides a method for identifying abnormal states of insulators based on FA-YOLOv8, comprising the following steps:

[0062] S1. Data preparation and preprocessing: The insulator dataset was re-labeled according to the requirements of Pascal VOC2007, and label information was added to each image to ensure the accuracy and reliability of the training data. The 160×160×3 image was used as the network input.

[0063] like Figure 2 As shown, the main body of FA-YOLOv8 includes input, a dual convolutional fusion module, a multi-attention mechanism fusion framework, a fast adaptive backbone network FANet, and a head recognition layer.

[0064] S2. Construct a dual convolution fusion module by fusing GhostConv and DConv, replacing part of the Conv2d_BN_SiLU convolution module in the backbone network of the YOLOv8 model.

[0065] The specific steps for constructing the dual convolutional fusion module are as follows:

[0066] S21. Using the feature map from the previous layer as the input to this layer, dilated convolutions DConv with dilation rates r1, r2, and r3 learn features from different receptive fields. This process is represented as follows:

[0067]

[0068] Among them, F input For the feature mapping of the previous layer, Let R1, R2, and R3 represent DConv with a spatial resolution of 3×3 and an expansion rate of r1, r2, and r3, respectively. Let R1, R2, and R3 represent the feature maps obtained from DConv with expansion rates of r1, r2, and r3, respectively.

[0069] To avoid excessive dilution of the original feature map information during the DConv operation, the original feature map is repeatedly used as input for the DConv operation. Meanwhile, to reduce computational complexity and achieve efficient recognition when processing higher spatial resolutions, an appropriate dilation coefficient needs to be selected. The combinations of dilation rates used for different feature maps are shown in Table 1.

[0070] Table 1. Combinations of dilation rates used for different feature maps.

[0071]

[0072]

[0073] S22. To minimize feature redundancy, reduce the number of parameters, and lighten the model, GhostConv convolution and standard convolution are used to refine the input features; the specific process is as follows:

[0074]

[0075] Where GhostConv represents the phantom convolution GhostConv, Conv 1×1 F1 represents a standard convolution with a 1×1 kernel, F2 represents a feature map obtained by passing GhostConv alone, and F3 represents a feature map obtained by passing GhostConv and standard convolution in turn.

[0076] S23. Concatenate the obtained feature maps R1, R2, R3, F1, and F2, and then perform dimensionality reduction using a standard convolution with a 1×1 kernel. The specific process is as follows:

[0077] F output =Conv 1×1 (Concat(R i ,F1,F2)) i=1,2,3 (3)

[0078] Among them, F output This represents the output feature map.

[0079] S3. Input the labeled insulator image into the first stage of the FANet backbone network, which consists of ConvModule and dual convolutional fusion module, to perform the first stage of feature extraction. During the feature extraction process, ConvModule refines the features and performs downsampling. The dual convolutional fusion module uses channel redundancy to learn features and reduce the computational load. At the same time, it uses DConv to learn local and non-local features from multiple receptive fields.

[0080] The output of the first-stage ConvModule in the FANet backbone network is input into the dual-convolutional fusion module to extract local and non-local feature information from multiple receptive fields, such as... Figure 3 As shown.

[0081] S4. Combine the efficient channel attention mechanism, spatial attention mechanism and feature pyramid module to build a multi-attention mechanism fusion framework, and apply it to the feature layer sampling stage of the dual convolution fusion module.

[0082] The input image feature maps extracted by the dual convolutional fusion module are fed into a multi-attention mechanism fusion framework to extract complex environmental information to assist in the identification of insulator abnormal states. The overall process is as follows: Figure 4 As shown.

[0083] The specific steps for building a multi-attention mechanism fusion framework are as follows:

[0084] S41. In the efficient channel attention mechanism, global average pooling is used to process the input feature map, and a standard convolution with kernel k=5 and dimension 1 is used to perform a convolution operation on the output feature map to extract image features. The sigmoid activation function is then used to calculate the weights of each channel.

[0085]

[0086] Among them, X c (i,j) represents the value at position (i,j) of the c-th channel, σ k Represents the Sigmoid operation, ω c Indicates the weights of different channels;

[0087] S42. Complete the residual connection, multiply the obtained weights element-wise with the corresponding original feature maps to obtain new feature maps;

[0088]

[0089] Among them, F ECA This represents the newly generated feature map;

[0090] S43. Based on the salient target recognition method and point cloud feature extraction method, calculate the spatial attention between feature layers of different scales in the feature pyramid module, and use SA to fuse two feature maps with different spatial resolutions:

[0091] M s (F x )=σ(Comv 1×1 ([AvgPool(x);MaxPool(x)])) (6)

[0092] M s (F y )=σ(Conv 1×1 ([AvgPool(y);MaxPool(y)])) (7)

[0093] F SA =Conv 1×1 (Concat(M s (F x ),M s (F y ))) (8)

[0094] Where σ represents the Sigmoid operation, AvgPool represents average pooling, MaxPool represents max pooling, and x and y are feature maps with two different spatial resolutions.

[0095] S44. The final feature map F output by the feature pyramid. FPFeature map F after channel attention processing CA The final output is obtained by connecting the components along the channel direction;

[0096] F output =Concat(F ECA ,F FP (9)

[0097] Concat represents the concatenation operation.

[0098] The multi-attention mechanism fusion framework integrates an efficient channel attention mechanism and a feature pyramid module with an embedded spatial attention mechanism. It also utilizes the features of the target and its surrounding environment. By embedding the spatial attention mechanism into the feature pyramid module, the inherent spatial attention feature pyramid retains more feature information, thus solving the feature loss problem that occurs as the network depth increases.

[0099] S5. After completing the first stage of feature extraction in the FANet backbone network, the second, third and fourth stages of feature extraction are performed and then sequentially input into the multi-attention mechanism fusion framework.

[0100] The fast adaptive backbone network FANet can be divided into four stages. Except for the first stage, the second, third and fourth stages are composed of ConvModule, dual convolutional fusion module and multi-attention mechanism fusion framework.

[0101] S6. The output features of the multi-attention mechanism fusion framework are passed through the FPN and PAN networks and decoupled heads are used to process the classification and localization tasks in recognition respectively, which further enhances the expressive power of multi-scale features.

[0102] S7. The loss function in the YOLOv8 model is improved by replacing the CIOU loss in the original model with the normalized Wasserstein distance (NWD). In addition, non-maximum suppression (NMS) is performed to improve recognition accuracy and system robustness.

[0103] A composite loss function is designed by combining the normalized Wasserstein distance with the binary cross-entropy classification loss function and the distributed focus regression loss function.

[0104] The formula for calculating NWD is as follows:

[0105]

[0106] in, Represents the center coordinates (cx) of the recognition frame a a ,cy a ) and width and height It measures the distance between frames a and b, NWD(N a N b ) represents the standardized form of the exponential function, where exp is the exponential function and C is the dataset association constant.

[0107] To improve training accuracy, the formula for the composite loss function is as follows:

[0108]

[0109] Among them, L BCE L is the binary cross-entropy classification loss, used to compare the classification probability distribution with the actual labels. DFL For distributed focus regression loss, L NWD To normalize the Wasserstein distance loss, L All For the total loss, (L) DFL L NWD The bounding box regression loss can better measure the positional difference between the predicted box and the actual target box, which helps the model to more accurately predict the abnormal position of the insulator of the transmission line.

[0110] Example

[0111] The experimental environment was created using a deep learning framework built on Python 3.8, PyTorch 2.0.1, and CUDA 11.7. Computational acceleration was achieved using an NVIDIA RTX 3090 GPU. The SGD optimizer was used with a momentum parameter of 0.937, an initial learning rate of 0.01, 100 epochs, and a batch size of 64.

[0112] The model was trained, validated, and tested using a self-collected dataset, the details of which are shown in Table 2. The 2255 images of abnormal conditions of transmission line insulators were re-labeled according to the requirements of Pascal VOC2007, and accurate label information was added to each image to ensure the accuracy and reliability of the training data. The insulator dataset was divided into training, validation, and test sets in a 7:1.5:1.5 ratio to ensure reasonable model training and validation.

[0113] Table 2 Dataset Information

[0114]

[0115] During the identification process, the magnitude of the loss function value between the predicted result and the actual target is used to measure whether the target position has been successfully predicted. The loss function threshold is set to 0.6, that is, if the composite loss function is greater than 0.6, it is considered a correct prediction, otherwise it is considered a wrong prediction.

[0116] Precision (P), recall (R), mean average precision (mAP), F1-Score, and model floating point operations (FLOPs) were used as evaluation metrics.

[0117] Precision P is the proportion of truly correct samples in the set that are identified as positive. A higher value indicates a stronger ability of the model to identify outliers of a specific category, and a lower classification error rate. The calculation formula is as follows:

[0118]

[0119] Where TP is the number of positive samples identified as positive, and FP is the number of negative samples identified as positive.

[0120] Recall R is the ratio of the number of positive samples identified as positive by the model to all positive samples. A higher recall rate indicates a lower prediction error rate. The calculation formula is as follows:

[0121]

[0122] Where FN is the number of positive samples that are identified as negative.

[0123] Mean Accuracy (mAP) is the average accuracy across all classes, reflecting the overall recognition performance of the model. A higher mAP value indicates higher model accuracy. The calculation formula is as follows:

[0124]

[0125] The F1-Score is the weighted harmonic mean of precision and recall. The calculation formula is as follows:

[0126]

[0127] FLOPs is a computational complexity metric that measures the computational complexity of a model and is used to evaluate the model's computational requirements and performance.

[0128] The mAP curves and loss function curves of the YOLOv8 model before and after the improvement are as follows: Figure 5 As shown, the results of insulator abnormality state identification under different conditions are as follows: Figure 6 As shown, the comparison results of the FA-YOLOv8 model with various other models are as follows: Figure 7As shown in Table 3, the results of the ablation experiment are shown in Table 3.

[0129] In the loss function curve, the loss function gradually decreases with each iteration and eventually levels off. The FA-YOLOv8 model with the added fast adaptive backbone network FANet shows significantly lower loss and a faster rate of descent, demonstrating stronger generalization ability. In the mean accuracy curve, the FA-YOLOv8 model has a higher mAP value, with the mAP value increasing by approximately 4.8% at the 100th iteration. Therefore, the FA-YOLOv8 model exhibits better robustness and superior performance in identifying insulator anomalies.

[0130] Table 3 Ablation Experiment Results

[0131]

[0132] Compared with representative models such as SSD, RT-DETR, and YOLOv8, the FA-YOLOv8 model achieved improvements in accuracy of 24.1%, 2%, and 2.4%, respectively; recall of 25%, 3.1%, and 4.9%, respectively; mean precision of 15.1%, 0.3%, and 1.7%, respectively; F1-score of 18.7%, 2%, and 4.1%, respectively; and reduced floating-point computation by 16.6G, 4.5G, and 7.1G, respectively.

[0133] In ablation experiments, after designing a dual-convolutional fusion module with dual-convolutional functionality, YOLOv8's precision, recall, mean precision (mAP), and F1-Score improved by 0.8%, 2.3%, 0.9%, and 1.5%, respectively, while reducing computational cost by 2.5G. After adding a multi-attention mechanism fusion framework module with ECA, feature pyramid, and SA, YOLOv8's precision, recall, mean precision (mAP), and F1-Score improved by 2.1%, 4.4%, 1.3%, and 3.2%, respectively, while reducing computational cost by 4.2G. Using a composite loss function, YOLOv8's precision, recall, mean precision (mAP), and F1-Score improved by 1.6%, 3.7%, 0.6%, and 2.6%, respectively, improving recognition accuracy and system robustness while reducing false negative and false positive rates.

[0134] Therefore, this invention employs the aforementioned FA-YOLOv8-based insulator anomaly state identification method. Utilizing a designed dual-convolutional fusion module, it achieves feature information learning across a wider receptive field while maintaining low computational complexity. Secondly, a multi-attention mechanism fusion framework enhances background features, improves anomaly state response, and reduces feature degradation, contributing to improved adaptability of the YOLOv8 model to complex environmental changes. Finally, by designing a more adaptive composite loss function, the identification accuracy of insulator anomalies and the robustness of the model are improved. Experimental results show that the FA-YOLOv8 model achieves accuracy, recall, mean precision, and F1-score of 98.7%, 98.8%, 99.2%, and 98.0%, respectively. Compared with SSD, RT-DETR, and YOLOv8 models, the proposed model exhibits stronger adaptability to various complex conditions and higher sensitivity to anomalies in insulators of different sizes and types, demonstrating significant advantages.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. 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 still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying abnormal states of insulators based on FA-YOLOv8, characterized in that, Includes the following steps: S1. Data preparation and preprocessing: The insulator dataset is re-labeled and label information is added to each image. The 160×160×3 image is used as the network input. S2. Construct a dual convolutional fusion module by fusing GhostConv and DConv to replace part of the Conv2d_BN_SiLU convolutional module in the backbone network of the YOLOv8 model. The specific steps for constructing the dual convolutional fusion module are as follows: S21. Use the feature map from the previous layer as the input to this layer, and use an expansion rate of... r 1. r 2 and r The dilated convolution DConv of type 3 learns features from different receptive fields, and this process is represented as follows: （1） in, For the feature mapping of the previous layer, , , These represent spatial resolutions of 3×3 and dilatation rates of 100%. r 1. r 2 and r 3 of DConv, , , The expansion rate obtained from DConv is respectively represented as r 1. r 2 and r 3. Feature mapping; S22. Use GhostConv convolution and standard convolution to refine the input features; the specific process is as follows: （2） in, GhostConv represents phantom convolution. This represents a standard convolution with a 1×1 kernel. This indicates the feature mapping obtained solely through GhostConv. This represents the feature map obtained by sequentially passing GhostConv and standard convolution; S23. Obtain the feature map , , , and The connections are then made, and dimensionality reduction is performed using a standard convolution with a 1×1 kernel. The specific process is as follows: （3） in, This represents the output feature map; S3. Input the labeled insulator image into the first stage of the FANet backbone network composed of ConvModule and dual convolutional fusion module to perform the first stage feature extraction. S4. Combine the efficient channel attention mechanism, spatial attention mechanism and feature pyramid module to build a multi-attention mechanism fusion framework, and apply it to the feature layer sampling stage of the dual convolution fusion module. S5. After completing the first stage of feature extraction, the second, third and fourth stages of feature extraction are performed and then sequentially input into the multi-attention mechanism fusion framework. S6. The output features of the multi-attention mechanism fusion framework are passed through the FPN and PAN networks and a decoupling head is used to process the classification and localization tasks in the recognition process respectively. S7. Improve the loss function in the YOLOv8 model by replacing the CIOU loss in the original model with normalized Wasserstein distance. Combine the normalized Wasserstein distance with the binary cross-entropy classification loss function and the distributed focus regression loss function to design a composite loss function.

2. The insulator abnormality state identification method based on FA-YOLOv8 according to claim 1, characterized in that: The main components of FA-YOLOv8 include the input, a dual convolutional fusion module, a multi-attention mechanism fusion framework, a fast adaptive backbone network FANet, and a head recognition layer.

3. The insulator abnormality state identification method based on FA-YOLOv8 according to claim 1, characterized in that: The specific process of S3 feature extraction is as follows: ConvModule refines features and performs downsampling; the dual convolutional fusion module uses channel redundancy to learn features and reduce computational load; and DConv learns local and non-local features from multiple receptive fields.

4. The insulator abnormality state identification method based on FA-YOLOv8 according to claim 1, characterized in that, The specific steps for building a multi-attention mechanism fusion framework in S4 are as follows: S41. In the efficient channel attention mechanism, global average pooling is used to process the input feature map, and a standard convolution with kernel k=5 and dimension 1 is used to perform a convolution operation on the output feature map to extract image features. The sigmoid activation function is then used to calculate the weights of each channel. （4） in, Indicates the first One channel The value of the position, This represents the Sigmoid operation. Indicates the weights of different channels; S42. Complete the residual connection, multiply the obtained weights element-wise with the corresponding original feature maps to obtain new feature maps; （5） in, This represents the newly generated feature map; S43. Calculate spatial attention between feature layers of different scales in the feature pyramid module, and use SA to fuse two feature maps with different spatial resolutions: （6） （7） （8） in, This represents the Sigmoid operation. Indicates average pooling. This indicates max pooling. , These are two feature maps with different spatial resolutions; S44. The final feature map output by the feature pyramid. Feature map with channel attention processing The final output is obtained by connecting the components along the channel direction; （9） in, This indicates a splicing operation.

5. The insulator abnormality state identification method based on FA-YOLOv8 according to claim 1, characterized in that: In S5, the second, third, and fourth stages of the FANet backbone network are all composed of ConvModule, dual convolutional fusion module, and multi-attention mechanism fusion framework.

6. The insulator abnormality state identification method based on FA-YOLOv8 according to claim 1, characterized in that, The formula for calculating the normalized Wasserstein distance in S7 is as follows: （10） （11） in,( ) represents the recognition framework center coordinates ( ) and width and height ( ), It is a measure of the recognition frame and The distance between them Represents the standardized form of the exponential function. It is an exponential function. It is a dataset association constant.

7. The insulator abnormality state identification method based on FA-YOLOv8 according to claim 6, characterized in that, The formula for the composite loss function in S7 is as follows: （12） in, For binary cross-entropy classification loss, For distributed focus regression loss, To normalize the Wasserstein distance loss, This represents the overall loss.

Images