Insulator abnormal state identification method based on FA-YOLOv8

Abstract

The invention discloses an insulator abnormal state identification method based on FA-YOLOv8, and belongs to the technical field of power transmission component target identification, and the method comprises the steps: data preparation and preprocessing, re-labeling of an insulator data set, adding label information to each image, construction of a double-convolution fusion module to replace a part of convolution modules in a YOLOv8 model backbone network, and identification of the abnormal state of an insulator. Carrying out first-stage feature extraction, and constructing a multi-attention mechanism fusion framework to carry out feature layer sampling; performing feature extraction in the second stage, the third stage and the fourth stage, and inputting the features into a multi-attention mechanism fusion framework in sequence; feature transfer and task decoupling are carried out, a loss function in the YOLOv8 model is improved, and a composite loss function is designed; according to the insulator abnormal state identification method based on FA-YOLOv8, the problems that an existing target identification method is large in calculated amount, low in response speed, low in identification efficiency and poor in adaptability when power transmission line insulator tiny defects and strong background interference exist are solved.

Description

Technical Field

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

[0002] As an important application of artificial intelligence visual learning, target recognition mainly identifies and locates ideal targets in captured images. Currently, target recognition has been widely used in many fields such as autonomous driving, robot vision, and human-computer interaction. With the upgrading of computers and the rapid development of deep learning technology, more and more target recognition models based on deep learning algorithms have emerged, and the recognition accuracy and efficiency have also been greatly improved.

[0003] The identification of abnormal insulator conditions often relies on manual inspections, which consumes a lot of manpower, material resources and financial resources. In order to reduce the frequency of manual tower climbing, avoid personnel danger, save costs and improve efficiency, the target recognition model based on deep learning is applied in insulator inspection. However, although the existing target recognition models have achieved certain research results, when faced with tiny defects and strong background interference, the existing methods still have problems such as large computational complexity, slow response speed, low recognition efficiency and poor adaptability. Therefore, how to solve these problems is still an important direction that needs to be studied urgently. Summary of the invention

[0004] The purpose of the present invention is to provide an insulator abnormal state identification method based on FA-YOLOv8 to solve the problems of large computational complexity, slow response speed, low identification efficiency and poor adaptability of existing target recognition methods when there are tiny defects in transmission line insulators and strong background interference. The method can fully extract shallow information, enhance the recognition effect of small targets and improve the recognition accuracy.

[0005] To achieve the above object, the present invention provides an insulator abnormal state identification method based on FA-YOLOv8, comprising the following steps:

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

[0007] S2, fused ghost convolution GhostConv and dilated convolution DCnv to construct a double convolution fusion module, replacing part of the Conv2d_BN_SiLU convolution module in the YOLOv8 model backbone network;

[0008] S3, input the annotated insulator image into the first stage of the FANet backbone network consisting of ConvModule and double convolution 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 the first stage of feature extraction is completed in the FANet backbone network, the second, third and fourth stages of feature extraction are performed and are input into the multi-attention mechanism fusion framework in sequence;

[0011] S6. The output features of the multi-attention mechanism fusion framework are transmitted through the FPN and PAN networks and the decoupling head is used to process the classification and positioning tasks in recognition respectively, further enhancing the expression ability of multi-scale features;

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

[0013] Preferably, the main part of FA-YOLOv8 includes input, dual convolution fusion module, multiple attention mechanism fusion framework, fast adaptive backbone network FANet and head recognition layer.

[0014] Preferably, the specific steps of S2 constructing a dual convolution fusion module are as follows:

[0015] S21, take the feature map of the previous layer as the input of this layer, and use the expansion rate r 1 、r 2 and r 3 The dilated convolution DConv learns features from different receptive fields. This process is expressed as follows:

[0016]

[0017] Among them, F input is the feature map of the previous layer, They represent the spatial resolution of 3×3 and the expansion rate of r. 1 、r 2 and r 3 DConv, R 1 , R2 , R 3 They represent the expansion rate obtained by DConv and r 1 、r 2 and r 3 feature map;

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

[0019]

[0020] Among them, GhostConv represents ghost convolution GhostConv, Conv 1×1 represents a standard convolution with a 1×1 kernel, F 1 represents the feature map obtained only by GhostConv, F 2 Represents the feature map obtained by GhostConv and standard convolution in sequence;

[0021] S23, the obtained feature map R 1 , R 2 , R 3 、F 1 and F 2 The concatenation is performed and the dimensionality reduction is performed using a standard convolution with a 1×1 kernel. The specific process is as follows:

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

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

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

[0025] Preferably, the specific steps of S4 to build a multi-attention mechanism fusion framework are as follows:

[0026] S41. In the efficient channel attention mechanism, the global average pooling operation is used to process the input feature map, and the output feature map is convolved using a standard convolution with a convolution kernel of k=5 and a dimension of 1 to extract image features, and the Sigmoid activation function is used to calculate the weight of each channel:

[0027]

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

[0029] S42, completing the residual connection, multiplying the obtained weights by the corresponding original feature maps, and obtaining a new feature map;

[0030]

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

[0032] S43. According to the salient target recognition method and the point cloud feature extraction method, the spatial attention between feature layers of different scales is calculated in the feature pyramid module, and SA is used to fuse two feature maps of 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] Among them, σ represents the Sigmoid operation, AvgPool represents average pooling, MaxPool represents maximum pooling, and x and y are feature maps with two different spatial resolutions respectively;

[0037] S44, the feature map F finally output by the feature pyramid FP And the feature map F after channel attention processing CA Connect in series along the channel direction to obtain the final output;

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

[0039] Among them, Concat represents a concatenation operation.

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

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

[0042]

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

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

[0045]

[0046] Among them, L BCE is the binary cross entropy classification loss, which is used to compare the classification probability distribution with the actual label, L DFL is the distributed focal regression loss, L NWD is the normalized Wasserstein distance loss, L All for overall loss.

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

[0048] (1) The present invention constructs a dual convolution fusion module that integrates GhostConv and Dconv, which can reduce the computational complexity while learning feature information on multi-layer receptive fields;

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

[0050] (3) The present invention combines the dual convolution fusion module and the multiple attention mechanism fusion framework to propose a fast adaptive backbone network FANet, which can reduce the computational burden for target recognition tasks, provide more powerful feature extraction and representation capabilities, and improve the accuracy and efficiency of target recognition by replacing the original backbone network.

[0051] (4) The present invention uses the normalized Wasserstein distance to replace the CIOU loss, and combines it with the binary cross entropy classification loss and the distributed focal regression loss to propose a more adaptable composite loss function, which can improve the recognition accuracy of the abnormal state of the insulator and the robustness of the model.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0054] Figure 2 It is the overall network structure diagram of the FA-YOLOv8 model of the present invention;

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

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

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

[0058] Figure 6 This is a diagram showing the abnormal state identification results of insulators under different conditions using the FA-YOLOv8 model of the present invention;

[0059] Figure 7 It is a bar chart of the comparison results between the FA-YOLOv8 model of the present invention and various models. DETAILED DESCRIPTION

[0060] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] like Figure 1As shown, the present invention provides an insulator abnormal state identification method based on FA-YOLOv8, comprising the following steps:

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

[0063] like Figure 2 As shown in the figure, the main parts of FA-YOLOv8 include input, dual convolution fusion module, multiple attention mechanism fusion framework, fast adaptive backbone network FANet and head recognition layer.

[0064] S2, fuses ghost convolution GhostConv and dilated convolution DConv to construct a dual convolution fusion module, replacing part of the Conv2d_BN_SiLU convolution modules in the YOLOv8 model backbone network.

[0065] The specific steps to construct the dual convolution fusion module are as follows:

[0066] S21, take the feature map of the previous layer as the input of this layer, and use the expansion rate r 1 、r 2 and r 3 The dilated convolution DConv learns features from different receptive fields. This process is expressed as follows:

[0067]

[0068] Among them, F input is the feature map of the previous layer, They represent the spatial resolution of 3×3 and the expansion rate of r. 1 、r 2 and r 3 DConv, R 1 , R 2 , R 3 They represent the expansion rate obtained by DConv and r 1 、r 2 and r 3 The feature map of .

[0069] In order to avoid excessive dilution of the original feature map information during the DConv operation, the original feature map is repeatedly called as the input of the DConv operation. At the same time, in order to reduce the computational complexity and achieve efficient recognition when processing higher spatial resolution, it is necessary to select a suitable expansion coefficient. The expansion rate combinations used by different feature maps are shown in Table 1.

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

[0071]

[0072]

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

[0074]

[0075] Among them, GhostConv represents ghost convolution GhostConv, Conv 1×1 represents a standard convolution with a 1×1 kernel, F 1 represents the feature map obtained only by GhostConv, F 2 Represents the feature map obtained by GhostConv and standard convolution in sequence.

[0076] S23, the obtained feature map R 1 , R 2 , R 3 、F 1 and F 2 The concatenation is performed and the dimensionality reduction is performed using a standard convolution with a 1×1 kernel. The specific process is as follows:

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

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

[0079] S3. The labeled insulator image is input into the first stage of the FANet backbone network consisting of ConvModule and double convolution fusion module for first stage feature extraction. During the feature extraction process, ConvModule refines the features and performs downsampling. The double convolution fusion module uses channel redundancy for feature learning to reduce the computational load. At the same time, DConv is used to learn local and non-local features from multiple receptive fields.

[0080] The output results of the first stage ConvModule in the FANet backbone network are input into the double convolution fusion module to extract local and non-local feature information in multiple receptive fields, such as Figure 3 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 map extracted by the dual convolution fusion module is input into the 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 shown.

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

[0084] S41. In the efficient channel attention mechanism, the global average pooling operation is used to process the input feature map, and the output feature map is convolved using a standard convolution with a convolution kernel of k=5 and a dimension of 1 to extract image features, and the Sigmoid activation function is used to calculate the weight of each channel:

[0085]

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

[0087] S42, completing the residual connection, multiplying the obtained weights by the corresponding original feature maps, and obtaining a new feature map;

[0088]

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

[0090] S43. According to the salient target recognition method and the point cloud feature extraction method, the spatial attention between feature layers of different scales is calculated in the feature pyramid module, and SA is used to fuse two feature maps of 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] Among them, σ represents the Sigmoid operation, AvgPool represents average pooling, MaxPool represents maximum pooling, and x and y are feature maps with two different spatial resolutions respectively;

[0095] S44, the feature map F finally output by the feature pyramid FP And the feature map F after channel attention processing CA Connect in series along the channel direction to obtain the final output;

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

[0097] Among them, Concat represents a concatenation operation.

[0098] The multi-attention mechanism fusion framework integrates an efficient channel attention mechanism and a feature pyramid module embedded with a spatial attention mechanism, while utilizing 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 and solves the problem of feature loss that occurs as the network depth increases.

[0099] S5. After completing the first stage feature extraction in the FANet backbone network, the second, third and fourth stages of feature extraction are carried out and are input into the multiple attention mechanism fusion framework in turn.

[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 convolution fusion module and multiple attention mechanism fusion framework.

[0101] S6. The output features of the multi-attention mechanism fusion framework are transmitted through the FPN and PAN networks and use the decoupling head to process the classification and positioning tasks in recognition respectively, further enhancing the expression ability of multi-scale features.

[0102] S7. Improve the loss function in the YOLOv8 model by using the normalized Wasserstein distance (NWD) instead of the CIOU loss in the original model. In addition, perform the non-maximum suppression (NMS) operation to improve the recognition accuracy and system robustness.

[0103] The normalized Wasserstein distance is combined with the binary cross entropy classification loss function and the distributed focal regression loss function to design a composite loss function.

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

[0105]

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

[0107] In order to improve the accuracy of training, the formula of the composite loss function is as follows:

[0108]

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

[0110] Example

[0111] Create an experimental environment, a deep learning framework built based on Python 3.8, PyTorch 2.0.1, and CUDA 11.7, use the NVIDIA RTX 3090 GPU for computing acceleration, use the SGD optimizer, the momentum parameter is 0.937, the initial learning rate is 0.01, the epoch is 100, and the batch size is 64.

[0112] The self-collected data set is used for model training, verification and testing. The data set information is shown in Table 2. The 2255 transmission line insulator abnormal state images are re-labeled according to the requirements of Pascal VOC2007, and label information is accurately added to each image to ensure the accuracy and reliability of the training data. The insulator data set is divided into training set, verification set and test set according to the ratio of 7:1.5:1.5 to ensure reasonable model training and verification.

[0113] Table 2 Dataset information

[0114]

[0115] During the recognition process, the loss function value between the predicted result and the actual target is used to measure whether the target position is 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 recorded as a correct prediction, otherwise it is recorded as an incorrect prediction.

[0116] The evaluation indicators are precision (Precision, P), recall (Recall, R), mean Average Precision (mAP), comprehensive index (F1-Score) and model calculation amount (model Floating point operations, FLOPs).

[0117] Precision P is the proportion of truly correct samples in the sample set identified as positive. The larger the value, the stronger the model's ability to identify abnormal states of a specific category and the smaller the 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] The recall rate R is the ratio of the number of positive samples identified as positive by the model to all positive samples. The larger the value, the smaller the model prediction error rate. The calculation formula is as follows:

[0121]

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

[0123] The mean average precision (mAP) is the average precision of all categories, reflecting the overall recognition performance of the model. The larger the value, the higher the model accuracy. The calculation formula is as follows:

[0124]

[0125] F 1 -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 computational requirements and performance of the model.

[0128] The mAP curve and loss function curve results of the YOLOv8 model before and after improvement are as follows Figure 5 As shown in Figure 2, the abnormal state identification results of insulators under different conditions are as follows: Figure 6 As shown in Figure 2, the comparison results of the FA-YOLOv8 model with various models are as follows: Figure 7 The ablation experiment results are shown in Table 3.

[0129] In the loss function curve, the loss function gradually decreases as the iteration proceeds and eventually stabilizes. The FA-YOLOv8 model with the addition of the fast adaptive backbone network FANet has significantly lower loss, a faster decline rate, and exhibits stronger generalization ability. In the average accuracy curve, the mAP value of the FA-YOLOv8 model is higher, and the mAP value of the 100th iteration increased by about 4.8%. Therefore, the FA-YOLOv8 model has better robustness for identifying abnormal insulator states and better performance.

[0130] Table 3 Ablation experiment results

[0131]

[0132] Compared with the representative SSD, RT-DETR and YOLOv8 models, the accuracy of the FA-YOLOv8 model is improved by 24.1%, 2% and 2.4% respectively; the recall rate is improved by 25%, 3.1% and 4.9% respectively; the average precision is improved by 15.1%, 0.3% and 1.7% respectively; the F1-Score is improved by 18.7%, 2% and 4.1% respectively; and the floating point operations are reduced by 16.6G, 4.5G and 7.1G respectively.

[0133] In the ablation experiment, after designing a dual convolution fusion module with dual convolution function, the accuracy, recall rate, average precision mAP and F1-Score of YOLOv8 increased by 0.8%, 2.3%, 0.9% and 1.5% respectively, and the amount of calculation was reduced by 2.5G. After adding the multi-attention mechanism fusion framework module with ECA, feature pyramid module and SA, the accuracy, recall rate, average precision mAP and F1-Score of YOLOv8 increased by 2.1%, 4.4%, 1.3% and 3.2% respectively, and the amount of calculation was reduced by 4.2G. After using the composite loss function, the accuracy, recall rate, average precision mAP and F1-Score of YOLOv8 increased by 1.6%, 3.7%, 0.6% and 2.6% respectively, which reduced the missed detection rate and false detection rate while improving the recognition accuracy and system robustness.

[0134] Therefore, the present invention adopts the above-mentioned insulator abnormal state identification method based on FA-YOLOv8, and uses the designed double convolution fusion module to realize the learning of feature information on a wider receptive field while maintaining low computational complexity. Secondly, the multi-attention mechanism fusion framework is used to enhance the background features, improve the abnormal state response, and reduce feature degradation, which helps to improve the adaptability of the YOLOv8 model to complex environmental changes. Finally, by designing a more adaptable composite loss function, the recognition accuracy of the insulator abnormal state and the robustness of the model are improved. The experimental results show that the model accuracy, recall rate, average precision and F1-Score of FA-YOLOv8 reach 98.7%, 98.8%, 99.2% and 98.0% respectively. Compared with SSD, RT-DETR and YOLOv8 models, the model proposed in the present invention has stronger adaptability to various complex conditions, higher sensitivity to abnormal states of insulators of different sizes and types, and has obvious advantages.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for identifying abnormal state of insulator based on FA-YOLOv8, characterized in that: The following steps are involved: S1. Data preparation and preprocessing: re-label the insulator dataset and add label information to each image. Use 160×160×3 images as network input. S2, fused ghost convolution GhostConv and dilated convolution DCnv to construct a double convolution fusion module, replacing part of the Conv2d_BN_SiLU convolution module in the YOLOv8 model backbone network; S3, input the annotated insulator image into the first stage of the FANet backbone network consisting of ConvModule and double convolution 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 the first stage of feature extraction is completed, the second, third and fourth stages of feature extraction are performed and are input into the multi-attention mechanism fusion framework in sequence; S6, the output features of the multi-attention mechanism fusion framework are passed through the FPN and PAN networks and the decoupling head is used to handle the classification and positioning tasks in recognition respectively; S7. Improve the loss function in the YOLOv8 model by using the normalized Wasserstein distance to replace the CIOU loss in the original model. Combine the normalized Wasserstein distance with the binary cross entropy classification loss function and the distributed focal regression loss function to design a composite loss function.

2. The method for identifying abnormal state of insulator based on FA-YOLOv8 according to claim 1, characterized in that: The main parts of FA-YOLOv8 include input, dual convolution fusion module, multiple attention mechanism fusion framework, fast adaptive backbone network FANet and head recognition layer.

3. The method for identifying abnormal state of insulator based on FA-YOLOv8 according to claim 1, characterized in that: The specific steps of S2 to construct a dual convolution fusion module are as follows: S21, take the feature map of the previous layer as the input of this layer, and use the dilated convolution DConv with dilation rates r1, r2 and r3 to learn features from different receptive fields. This process is expressed as follows: Among them, F input is the feature map of the previous layer, They represent DConv with spatial resolution of 3×3 and dilation rates of r1, r2, and r3 respectively. R1, R2, and R3 represent feature maps obtained by DConv with dilation rates of r1, r2, and r3 respectively. S22. Use ghost convolution GhostConv and standard convolution to refine the input features. The specific process is as follows: Among them, GhostConv represents ghost convolution GhostConv, Conv 1×1 represents a standard convolution with a 1×1 kernel, F1 represents the feature map obtained only by GhostConv, and F2 represents the feature map obtained by GhostConv and standard convolution in sequence; S23, concatenate the obtained feature maps R1, R2, R3, F1 and F2, and use standard convolution with a 1×1 kernel for dimensionality reduction. The specific process is as follows: F output =Conv 1×1 (Concat(R i ,F1,F2))i=1,2,3 (3) Among them, F output Represents the output feature map.

4. The method for identifying abnormal state of insulator based on FA-YOLOv8 according to claim 1, characterized in that: The specific process of S3 feature extraction is that ConvModule refines features and performs downsampling, the dual convolution fusion module uses channel redundancy for feature learning to reduce the computational load, and DConv is used to learn local and non-local features from multiple receptive fields.

5. The method for identifying abnormal state of insulator based on FA-YOLOv8 according to claim 1, characterized in that: The specific steps of S4 to build a multi-attention mechanism fusion framework are as follows: S41. In the efficient channel attention mechanism, the global average pooling operation is used to process the input feature map, and the output feature map is convolved using a standard convolution with a convolution kernel of k=5 and a dimension of 1 to extract image features, and the Sigmoid activation function is used to calculate the weight of each channel: Among them, X c (i, j) represents the value of the cth channel (i, j) position, σ k Represents Sigmoid operation, ω c Represents the weights of different channels; S42, completing the residual connection, multiplying the obtained weights by the corresponding original feature maps, and obtaining a new feature map; Among them, F ECA Represents the newly generated feature map; S43. Calculate the spatial attention between feature layers of different scales in the feature pyramid module, and use SA to fuse two feature maps of different spatial resolutions: M s (F x )=σ(Comv 1×1 (|AvgPool(x);MaxPool(x)])) (6) M s (F y )=σ(Conv 1×1 ([AvgPool(y);MaxPool(y)])) (7) F SA =Conv 1×1 (Concat(M s (F x ),M s (F y ))) (8) Among them, σ represents the Sigmoid operation, AvgPool represents average pooling, MaxPool represents maximum pooling, and x and y are feature maps with two different spatial resolutions respectively; S44, the feature map F finally output by the feature pyramid FP And the feature map F after channel attention processing CA Connect in series along the channel direction to obtain the final output; F output =Concat(F ECA ,F FP ) (9) Among them, Concat represents a concatenation operation.

6. The method for identifying abnormal state of insulator based on FA-YOLOv8 according to claim 1, characterized in that: The second, third and fourth stages in the FANet backbone network in S5 are composed of ConvModule, dual convolution fusion module and multiple attention mechanism fusion framework.

7. The method for identifying abnormal state of insulator based on FA-YOLOv8 according to claim 1, characterized in that: The calculation formula of normalized Wasserstein distance in S7 is as follows: in, Represents the center coordinates of the recognition frame a (cx a ,cy a ) and width and height It is a measure of the distance between the recognition frames a and b, NWD(N a ,N b ) represents the standardized form of the exponential function, exp is the exponential function, and C is the constant associated with the data set.

8. The method for identifying abnormal state of insulator based on FA-YOLOv8 according to claim 7, characterized in that: The formula of the composite loss function in S7 is as follows: Among them, L BCE is the binary cross entropy classification loss, L DFL is the distributed focal regression loss, L NWD is the normalized Wasserstein distance loss, L All for overall loss.

Images