Icing detection method for power transmission line

Abstract

The invention discloses a power transmission line icing detection method, which comprises the following steps of: firstly, giving an actual application scene reference example of DHAM-YOLO to explain the characteristics of an algorithm of the invention; then, a DHAM module is provided, multi-scale feature extraction is carried out on a feature map output by the Backbone layer, then depth separable convolution, fusion and weighting are carried out, and target channel information extraction is achieved through a sigmoid function; meanwhile, the MSRCM is added into the Bottleneck structure of the Neck layer, so that the accuracy of identifying the multi-scale target is improved. And finally, verifying the performance of DHAM-YOLO through comparison experiments of different attention mechanisms, ablation experiments, target detection and training loss curves. Experiments show that DHAM-YOLO recognition precision is high, performance detection speed is high, calculated amount is smaller, and the method is more suitable for actual icing detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method for detecting ice coating on a power transmission line based on a DHAM-YOLO algorithm model. Background Art

[0002] In the power network, transmission lines undertake important tasks of power transmission. Due to the diversity of the installation environment of transmission lines and the wide distribution of lines, some potential risks may arise, such as transmission lines covered with ice and snow, insulators icing, etc. Under different circumstances, icing poses a serious threat to the power grid system and may cause serious line sliding, phase tripping, etc. The occurrence of accidents also brings many inconveniences to production and daily work. Line icing disasters have occurred in many regions at home and abroad, causing serious damage to the power system of transmission lines. Therefore, real-time monitoring of transmission line icing plays a vital role in the power system's ability to maintain safe and stable power supply to users.

[0003] In order to improve the stability of the power grid system in cold environments, image detection is the mainstream method for non-contact ice detection. In the field of image detection, deep learning is widely used in intelligent inspection of power transmission lines because it has powerful feature learning and representation capabilities, can learn complex feature expressions from a large amount of data, and is more suitable for handling ice detection tasks in complex scenarios. Summary of the invention

[0004] The purpose of the present invention is to provide a method for detecting ice coating on transmission lines. Based on the DHAM-YOLO detection algorithm model, the present invention first provides a reference example of the actual application scenario of DHAM-YOLO to illustrate the characteristics of the algorithm in this chapter. Then, the DHAM module is proposed to extract multi-scale features from the feature map output by the Backbone layer, and then perform deep separable convolution, fusion and weighting, and then extract the target channel information through the sigmoid function. At the same time, the present invention adds MSRCM to the Bottleneck structure of the Neck layer to improve the accuracy of identifying multi-scale targets. Finally, the performance of DHAM-YOLO is verified by comparative experiments of different attention mechanisms, ablation experiments, target detection, and training loss curves. Experiments show that DHAM-YOLO has high recognition accuracy, fast detection performance, and less computational complexity, and is more suitable for actual ice coating detection.

[0005] The technical solution of the present invention is a method for detecting ice coating on a power transmission line, which is specifically implemented according to the following steps:

[0006] Step 1: Collect ice-covered images of power transmission lines, pre-process the collected ice-covered images of power transmission lines, and annotate the image data set to divide it into a training set, a verification set, and a test set;

[0007] Step 2: construct a deep feature extraction module DFEB, thereby introducing a multi-scale residual convolution module MSRCM;

[0008] Step 3: Propose a deep hybrid attention module DHAM to design and train the DHAM-YOLO detection algorithm model;

[0009] Step 4: Input the images in the test set of step 1 into the DHAM-YOLO detection algorithm model trained in step 3 for detection, and obtain the transmission line icing detection result map.

[0010] Furthermore, the step 1 is specifically implemented according to the following steps:

[0011] Step 1.1, obtain a large number of on-site images through drones or fixed cameras, including images of ice-covered transmission lines and insulators of various types, and ensure that images of ice and snow covering each type of transmission line and insulator are obtained through close-up and close-up images;

[0012] Step 1.2, data enhancement and expansion of the transmission line ice-covered image obtained in step 1.1, including mirroring, rotation, flipping, tone adjustment, contrast and brightness adjustment, so as to obtain an expanded transmission line ice-covered image;

[0013] Step 1.3: The transmission line ice-covered image obtained in step 1.1 and the expanded transmission line ice-covered image obtained in step 1.2 are taken together as a transmission line ice-covered image set, and divided into a training set, a validation set, and a test set according to a ratio of 8:1:1.

[0014] Furthermore, the step 2 is specifically implemented according to the following steps:

[0015] Step 2.1, firstly, a deep feature extraction module DFEB is proposed, which uses a combination of deep convolution and point convolution to extract multi-level features of objects of different scales;

[0016] The specific method is as follows: the DFEB module first outputs the deep convolution, and then applies a 1×1×C convolution kernel to mix and integrate the feature maps to generate the final output feature map; then, the DFEB module performs a normalization (BN) process; finally, the Relu linear activation function is applied to ensure that each layer of the network has consistent distribution characteristics as input; the calculation equation of the DFEB module is shown in formula (1):

[0017]

[0018] where x i Represents the input features of each channel; represents the combination of feature maps in the channel dimension; f Conv3×3Indicates that a depth convolution calculation with a convolution kernel size of 3×3 is performed; f Conv1×1 Indicates point convolution calculation with a convolution kernel size of 1×1; BN represents the normalization operation of the BN layer to avoid overfitting of the model; Relu represents the operation of executing the Relu activation function;

[0019] Step 2.2: Based on the DFEB module, MSRCM is introduced to improve the accuracy of multi-scale target recognition through residual connection to avoid the loss of small target detail information during the convolution process;

[0020] The specific method is as follows: Input features Figure X It is divided into four branches for feature extraction. The first branch X1 passes through two DFEB modules in turn, then performs a residual connection, and then passes through a Relu activation function. This is repeated 4 times to obtain the output X of the first branch X1. 1out , as shown in formula (3):

[0021] X 1-1 =X 2out =Relu{DB[DB(X1)]+X1} (2)

[0022] X 1out =X 1-4 [X 1-3 (X 1-2 (X 1-1 ))] (3)

[0023] Where: X 1-1 , X 1-2 , X 1-3 , X 1-4 They represent the results of performing the residual operation and then passing through the Relu activation function; X 1out Indicates the final output result of the branch;

[0024] The output of the second branch X2 is:

[0025] X 2out =Relu{DB[DB(X2)]+X2} (4)

[0026] The third branch uses a depthwise convolution with a kernel size of 3×3, followed by BN processing and Relu activation function to obtain the output X of the third branch X3. 3out :

[0027] X 3out =Relu[BN(Conv 3×3 (X3))] (5)

[0028] The fourth branch only processes the input features Figure X4 Perform a BN process to obtain the output X of the fourth branch X4 4out :

[0029] X 4out =BN(X4) (6)

[0030] The results of the above four branches are combined and the obtained results are output through the Relu activation function;

[0031] X*=Relu(X 1out +X 2out +X 3out +X 4out ) (7)

[0032] Where: X * A new feature map representing the output of the multi-scale residual convolution module;

[0033] Step 2.3, design MSRCM in the Bottleneck structure of the Neck layer to obtain an improved target detection network.

[0034] Furthermore, step 3 is specifically implemented according to the following steps:

[0035] 3.1. Propose a deep hybrid attention module DHAM. DHAM is divided into two parts: the spatial attention module SAM and the channel attention module CAM.

[0036] 3.2. SAM extracts input features from four branches at multiple scales.

[0037] The first branch inputs feature X 1in : For input features Figure X (X 1in , X 2in , X 3in ), (X∈R C×W×H ), first, the first branch input feature X is extracted from two perspectives using maximum pooling and average pooling 1in Features:

[0038] X max1 =f maxpool (X 1in ) (8)

[0039] X avg1 =f avgpool (X 1in ) (9)

[0040] Where: X max1 ∈R C×H / 2×W / 2 ;X avg1 ∈R C×H / 2×W / 2; Use maximum pooling to extract important information, and use average pooling to extract global information; Concat the results of maximum pooling and average pooling, and then output them through the Sigmoid function:

[0041] A c =σ(X max1 +X avg1 ) (10)

[0042] Where: σ represents the Sigmoid function; A c ∈R C×1×1 ; A c Each element in represents the importance level of its corresponding channel;

[0043] Second, the first branch is fed with feature X 1in It is divided into two paths: one path is fused with the output of the Sigmoid function; the other path first performs a 3×3 ordinary convolution, then extracts features through a depth-separable convolution and outputs it through the Relu function. The process is as follows:

[0044] X 1-1 =A c *X 1in

[0045] X 1-2 =f conv3 (X 1in )

[0046] X 1-3 =Relu[f conv1×1 (f conv3×3 (X 1-2 ))] (11)

[0047] Where: X 1-1 Represents X 1in With A c The result of fusion; Conv3×3 Represents a 3×3 ordinary convolution operation;

[0048] Finally, X 1-1 With X 1-3 Perform fusion output:

[0049] F DM =X 1-out =X 1-1 +X 1-3 (12)

[0050] Where: X 1-out and F DM Represents the first branch input feature X after fusion 1inThe output feature map of the above feature map extraction process is named as Deep-extraction Module (DM);

[0051] The second branch of the SAM structure inputs feature X 2in And the third branch input feature X 3in After 2x upsampling and 4x upsampling respectively, the input is input into the DM structure for feature extraction, and then after maximum pooling and average pooling respectively, the second branch input feature X is obtained. 2in And the third branch input feature X 3in Output characteristics Figure X 2-out and X 3-out :

[0052] X 2-out =f maxpool [F DM (UP2(X 2in ))]

[0053] X 3-out =f maxpool [F DM (UP4(X 2in ))] (13)

[0054] Where: UP2 represents the input feature Figure X 2in The width and height are upsampled by 2 times; UP4 means the input features Figure X 3in The width and height of F are upsampled by 4 times; DM It means that the upsampled feature maps are input into the DM structure to extract features; f maxpool and f avgpool Respectively represent the extraction of F from the perspective of maximum pooling and average pooling DM Features;

[0055] The last part of SAM is to combine the output features of the three branches Figure X 1-out , X 2-out , X 3-out After fusion, the weighted sum is performed with the input X to obtain the weight of the target:

[0056] X A =X×(X 1-out +X 2-out +X 3-out ) (14)

[0057] Where: X A Represents the output result of SAM;

[0058] This method can enhance the spatial information of the target object while keeping a relatively small number of parameters by performing multi-scale extraction and fusion of images;

[0059] 3.3. The CAM part of the DHAM module extracts input features X from two different angles through maximum pooling and average pooling A After the extraction is completed, a deep separable convolution operation is performed, and then the obtained results are fused. The fused features are combined with the input X A Weighted, get the weight of the target; finally, through the sigmoid function to get the weight between (0,1), extract the target channel information, that is:

[0060] X'=σ{X A ×[DW(f maxpool (X A ))+DW(f avgpool (X A ))]} (15)

[0061] Where: σ represents the sigmoid function; X'∈R C×1×1 ; Each element in X' represents the importance level of its corresponding channel, and DW represents the depth convolution (f cov3x3 ) and point convolution (f cov1x1 ) Extraction process;

[0062] 3.4. Embedding the DHAM module into the Neck layer can adaptively assign weights to specific target areas to improve the model's detection capabilities for small targets and complex backgrounds, and avoid missed detections and false detections.

[0063] Furthermore, step 4 is specifically implemented according to the following steps:

[0064] The performance of DHAM-YOLO is verified through comparative experiments of different attention mechanisms, ablation experiments, target detection, and training loss curves;

[0065] 4.1. Under the same experimental conditions, different attention mechanisms are compared based on various indicators and further verified using heat maps;

[0066] 4.2. Add DHAM and MSRCM modules to the initial module in turn, conduct ablation experiments, and compare and analyze the ice detection performance of each module;

[0067] 4.3. Parameters, FPS, FLOPs and mAP of the mainstream algorithms Faster RCNN, SSD, Centernet and YOLO are compared with the algorithm in this paper. At the same time, the convergence performance and overfitting problem of the model of the present invention are verified by comparing the training loss curves and verification loss curves of different models.

[0068] 4.4. Select images of iced transmission lines and iced insulators in different complex environments for testing.

[0069] The beneficial effects of the present invention are:

[0070] 1. The present invention adds a Multi-Scale Residual Convolution Module (MSRCM) and a Deep Hybrid Attention Module (DHAM) based on YOLO. MSRCM not only maintains its advantages of fast convergence and efficient training in high-dimensional networks, but also makes full use of the advantages of reduced computation brought by deep separable convolution, thereby improving the recognition accuracy of multi-scale targets and avoiding the loss of detail information of small targets during the convolution process. DHAM improves the accuracy and feature integrity of detection by enhancing the model's attention to ice-covered areas and reducing background interference. In addition, DHAM can also effectively integrate features of different scales, enabling the model to simultaneously focus on information at multiple scales, thereby improving the global perception and speed of detection.

[0071] 2. The present invention improves the model's feature extraction capability for ice-covered targets of different scales by introducing a multi-scale residual convolution module (MSRCM), ensuring accurate recognition of small targets and complex backgrounds. At the same time, the deep hybrid attention module (DHAM) is embedded to enhance the model's attention to ice-covered areas, reduce background interference, and ensure the integrity and global perception of detection. Experiments show that the algorithm of the present invention achieves an mAP of 97.58%, which is significantly better than other mainstream algorithms, and has a faster detection speed, making it suitable for practical applications. In addition, the model performs stably in the training loss curve and the verification loss curve, converges quickly, and has no overfitting phenomenon, showing good robustness and generalization ability. The algorithm exhibits high detection accuracy and real-time processing capability in complex environments, verifying its great application potential in ice detection of transmission lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is the DHAM-YOLO network structure diagram;

[0073] Figure 2 It is a structural diagram of the efficient extraction module;

[0074] Figure 3 It is a structural diagram of the multi-scale residual convolution module;

[0075] Figure 4 It is an improved Bottleneck network structure diagram;

[0076] Figure 5 This is the DHAM network structure diagram;

[0077] Figure 6 This is the EHAM distribution map in the Neck layer;

[0078] Figure 7 This is a sample graph of the dataset;

[0079] Figure 8 This is a comparison of the training loss curves of different models;

[0080] Fig. 9 This is a comparison of the validation loss curves of different models;

[0081] Fig.10 This is the ice detection result diagram of different algorithms; DETAILED DESCRIPTION

[0082] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] The present invention is based on the DHAM-YOLO detection algorithm model of the power transmission line ice detection method, such as Figure 1 As shown, the specific implementation steps are as follows:

[0084] Step 1: Collect ice-covered images of power transmission lines, pre-process the collected ice-covered images of power transmission lines, and annotate the image data set to divide it into a training set, a verification set, and a test set;

[0085] Step 2: construct a deep feature extraction module DFEB, thereby introducing a multi-scale residual convolution module MSRCM;

[0086] Step 3: Propose a deep hybrid attention module DHAM to design and train the DHAM-YOLO detection algorithm model;

[0087] Step 4: Input the images in the test set of step 1 into the DHAM-YOLO detection algorithm model trained in step 3 for detection, and obtain the transmission line icing detection result map.

[0088] Furthermore, step 1 is specifically implemented according to the following steps:

[0089] Step 1.1, obtain a large number of on-site images through drones or fixed cameras, including images of ice-covered transmission lines and insulators of various types, and ensure that images of ice and snow covering each type of transmission line and insulator are obtained through close-up and close-up images;

[0090] Step 1.2, data enhancement and expansion of the transmission line ice-covered image obtained in step 1.1, including mirroring, rotation, flipping, tone adjustment, contrast and brightness adjustment, so as to obtain an expanded transmission line ice-covered image;

[0091] Step 1.3: The transmission line ice-covered image obtained in step 1.1 and the expanded transmission line ice-covered image obtained in step 1.2 are taken together as a transmission line ice-covered image set, and divided into a training set, a validation set, and a test set according to a ratio of 8:1:1.

[0092] Furthermore, if Figure 2 As shown, step 2 is implemented in the following steps:

[0093] Step 2.1, firstly, a deep feature extraction module DFEB is proposed, which uses a combination of deep convolution and point convolution to extract multi-level features of objects of different scales;

[0094] The specific method is as follows: the DFEB module first outputs the deep convolution, and then applies a 1×1×C convolution kernel to mix and integrate the feature maps to generate the final output feature map; then, the DFEB module performs a normalization (BN) process; finally, the Relu linear activation function is applied to ensure that each layer of the network has consistent distribution characteristics as input; the calculation equation of the DFEB module is shown in formula (1):

[0095]

[0096] where x i Represents the input features of each channel; represents the combination of feature maps in the channel dimension; f Conv3×3 Indicates that a depth convolution calculation with a convolution kernel size of 3×3 is performed; f Conv1×1 Indicates point convolution calculation with a convolution kernel size of 1×1; BN represents the normalization operation of the BN layer to avoid overfitting of the model; Relu represents the operation of executing the Relu activation function;

[0097] Step 2.2: Based on the DFEB module, MSRCM is introduced to improve the accuracy of multi-scale target recognition through residual connection to avoid the loss of small target detail information during the convolution process. Figure 3 As shown;

[0098] The specific method is as follows: Input features Figure X It is divided into four branches for feature extraction. The first branch X1 passes through two DFEB modules in turn, then performs a residual connection, and then passes through a Relu activation function. This is repeated 4 times to obtain the output X of the first branch X1.1out , as shown in formula (3):

[0099] X 1-1 =X 2out =Relu{DB[DB(X1)]+X1} (2)

[0100] X 1out =X 1-4 [X 1-3 (X 1-2 (X 1-1 ))] (3)

[0101] Where: X 1-1 , X 1-2 , X 1-3 , X 1-4 They represent the results of performing the residual operation and then passing through the Relu activation function; X 1out Indicates the final output result of the branch;

[0102] The output of the second branch X2 is:

[0103] X 2out =Relu{DB[DB(X2)]+X2} (4)

[0104] The third branch uses a depthwise convolution with a kernel size of 3×3, followed by BN processing and Relu activation function to obtain the output X of the third branch X3. 3out :

[0105] X 3out =Relu[BN(Conv 3×3 (X3))] (5)

[0106] The fourth branch only processes the input features Figure X 4 Perform a BN process to obtain the output X of the fourth branch X4 4out :

[0107] X 4out =BN(X4) (6)

[0108] The results of the above four branches are combined and the obtained results are output through the Relu activation function;

[0109] X*=Relu(X 1out +X 2out +X 3out +X 4out ) (7)

[0110] Where: X * A new feature map representing the output of the multi-scale residual convolution module;

[0111] Step 2.3, design MSRCM in the Bottleneck structure of the Neck layer to obtain an improved target detection network, such as Figure 4 shown.

[0112] Furthermore, if Figure 5 As shown, step 3 is implemented in the following steps:

[0113] 3.1. Propose a deep hybrid attention module DHAM. DHAM is divided into two parts: the spatial attention module SAM and the channel attention module CAM.

[0114] 3.2. SAM extracts input features from four branches at multiple scales.

[0115] The first branch inputs feature X 1in : For input features Figure X (X 1in , X 2in , X 3in ), (X∈R C×W×H );

[0116] First, we use maximum pooling and average pooling to extract input features X from two perspectives. 1in Features:

[0117] X max1 =f max pool (X 1in ) (8)

[0118] X avg1 =f avgpool (X 1in )(9)

[0119] Where: X max1 ∈R C×H / 2×W / 2 ;X avg1 ∈R C×H / 2×W / 2 ; Use maximum pooling to extract important information, and use average pooling to extract global information; Concat the results of maximum pooling and average pooling, and then output them through the Sigmoid function:

[0120] A c =σ(X max1 +X avg1 ) (10)

[0121] Where: σ represents the Sigmoid function; A c ∈R C×1×1 ; A c Each element in represents the importance level of its corresponding channel;

[0122] Second, the first branch is fed with feature X1in It is divided into two paths. One path is fused with the output of the Sigmoid function. The other path first performs a 3×3 ordinary convolution, then extracts features through a depth-separable convolution and outputs it through the Relu function. The process is as follows:

[0123] X 1-1 =A c *X 1in

[0124] X 1-2 =f conv3 (X 1in )

[0125] X 1-3 =Relu[f conv1×1 (f conv3×3 (X 1-2 ))] (11)

[0126] Where: X 1-1 Represents X 1in With A c The result of fusion; Conv3×3 Represents a 3×3 ordinary convolution operation;

[0127] Finally, X 1-1 With X 1-3 Perform fusion output:

[0128] F DM =X 1-out =X 1-1 +X 1-3 (12)

[0129] Where: X 1-out and F DM Represents the first branch input feature X after fusion 1in The output feature map of the above feature map extraction process is named as Deep-extraction Module (DM);

[0130] The second branch of the SAM structure inputs feature X 2in And the third branch input feature X 3in After 2x upsampling and 4x upsampling respectively, the input is input into the DM structure for feature extraction, and then after maximum pooling and average pooling respectively, the second branch input feature X is obtained. 2in And the third branch input feature X 3in Output characteristics Figure X 2-out and X 3-out :

[0131] X 2-out =fmaxpool [F DM (UP2(X 2in ))]

[0132] X 3-out =f maxpool [F DM (UP4(X 2in ))] (13)

[0133] Where: UP2 represents the input feature Figure X 2in The width and height are upsampled by 2 times; UP4 means the input features Figure X 3in The width and height of F are upsampled by 4 times; DM It means that the upsampled feature maps are input into the DM structure to extract features; f maxpool and f avgpool Respectively represent the extraction of F from the perspective of maximum pooling and average pooling DM Features;

[0134] The last part of SAM is to combine the output features of the three branches Figure X 1-out , X 2-out , X 3-out After fusion, the weighted sum is performed with the input X to obtain the weight of the target:

[0135] X A =X×(X 1-out +X 2-out +X 3-out ) (14)

[0136] Where: X A Represents the output result of SAM;

[0137] This method can enhance the spatial information of the target object while keeping a relatively small number of parameters by performing multi-scale extraction and fusion of images;

[0138] 3.3. The CAM part of the DHAM module extracts input features X from two different angles through maximum pooling and average pooling A After the extraction is completed, a deep separable convolution operation is performed, and then the obtained results are fused. The fused features are combined with the input X A Weighted, get the weight of the target; finally, through the sigmoid function to get the weight between (0,1), extract the target channel information, that is:

[0139] X'=σ{X A ×[DW(f maxpool (X A))+DW(f avgpool (X A ))]} (15)

[0140] Where: σ represents the sigmoid function; X'∈R C×1×1 ; Each element in X' represents the importance level of its corresponding channel, and DW represents the depth convolution (f cov3x3 ) and point convolution (f cov1x1 ) Extraction process;

[0141] 3.4. Embedding the DHAM module into the Neck layer can adaptively assign weights to specific target areas, improve the model's detection capabilities for small targets and complex backgrounds, and avoid missed detections and false detections, such as Figure 6 shown.

[0142] Furthermore, step 4 is specifically implemented according to the following steps:

[0143] The performance of DHAM-YOLO is verified through comparative experiments of different attention mechanisms, ablation experiments, target detection, and training loss curves;

[0144] 4.1. The deep learning framework of the experiment of the present invention is Pytorch, the training platform is Windows 10, the CPU is Inteli9-11900k, the GPU is NVIDIA GTX 3090 (24GB), the CUDA version is 11.3, the Pytorch version is 1.8, the batch size is 24, the Epoch is 400, and the initial learning rate is 0.001;

[0145] 4.2. Under the same experimental conditions, different attention mechanisms are compared based on various indicators and further verified using heat maps;

[0146] Table 1 Comparison results of different attention mechanisms

[0147]

[0148] 4.3. Add DHAM and MSRCM modules to the initial module in turn, conduct ablation experiments, and compare and analyze the ice detection performance of each module;

[0149] Table 2 Statistical results of ablation experiments

[0150]

[0151]

[0152] 4.4. Compare the parameters, FPS, FLOPs and mAP of the mainstream algorithms such as Faster RCNN, SSD, Centernet and YOLO with the algorithm in this paper. At the same time, compare the training loss curves and validation loss curves of different models, such as Figure 8 , Fig. 9 As shown, the convergence performance and over-fitting problem of the model of the present invention are verified;

[0153] Table 3 Experimental results of different algorithms

[0154]

[0155] 4.5. Select images of iced transmission lines and iced insulators in different complex environments for testing.

[0156] In order to effectively compare the advantages of the DHAM-YOLO algorithm with various other algorithms, we used consistent training and testing methods to conduct comparative experiments on other algorithms and the model proposed in this invention. For specific results, see Fig.10 .

Claims

1. A method for detecting ice coating on a power transmission line, characterized in that: Follow the steps below to implement it: Step 1: Collect ice-covered images of power transmission lines, pre-process the collected ice-covered images of power transmission lines, and annotate the image data set to divide it into a training set, a verification set, and a test set; Step 2: construct a deep feature extraction module DFEB, thereby introducing a multi-scale residual convolution module MSRCM; Step 3: Propose a deep hybrid attention module DHAM to design and train the DHAM-YOLO detection algorithm model; Step 4: Input the images in the test set of step 1 into the DHAM-YOLO detection algorithm model trained in step 3 for detection, and obtain the transmission line icing detection result map.

2. A method for detecting ice coating on a power transmission line according to claim 1, characterized in that: The step 1 is specifically implemented according to the following steps: Step 1.1, obtain a large number of on-site images through drones or fixed cameras, including images of ice-covered transmission lines and insulators of various types, and ensure that images of ice and snow covering each type of transmission line and insulator are obtained through close-up and close-up images; Step 1.2, data enhancement and expansion of the transmission line ice-covered image obtained in step 1.1, including mirroring, rotation, flipping, tone adjustment, contrast and brightness adjustment, so as to obtain an expanded transmission line ice-covered image; Step 1.3: The transmission line ice-covered image obtained in step 1.1 and the expanded transmission line ice-covered image obtained in step 1.2 are taken together as a transmission line ice-covered image set, and divided into a training set, a validation set, and a test set according to a ratio of 8:1:

1.

3. A method for detecting ice coating on a power transmission line according to claim 2, characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1, firstly, a deep feature extraction module DFEB is proposed, which uses a combination of deep convolution and point convolution to extract multi-level features of objects of different scales; The specific method is as follows: the DFEB module first passes the output through deep convolution, and then applies a 1×1×C convolution kernel to mix and integrate the feature maps to generate the final output feature map; Then, the DFEB module performs a normalization (BN) process; Finally, the Relu linear activation function is applied to ensure that each layer of the network has consistent distribution characteristics as input; the calculation equation of the DFEB module is shown in formula (1): where x i Represents the input features of each channel; represents the combination of feature maps in the channel dimension; f Conv3×3 Indicates that a depth convolution calculation with a convolution kernel size of 3×3 is performed; f Conv1×1 Indicates point convolution calculation with a convolution kernel size of 1×1; BN represents the normalization operation of the BN layer to avoid overfitting of the model; Relu represents the operation of executing the Relu activation function; Step 2.2: Based on the DFEB module, MSRCM is introduced to improve the accuracy of multi-scale target recognition through residual connection to avoid the loss of small target detail information during the convolution process; The specific method is as follows: Divide the input feature map X into four branches for feature extraction. The first branch X1 passes through two DFEB modules in sequence, then performs a residual connection, and then passes through a Relu activation function, and repeats 4 times to obtain the output X of the first branch X1. 1out , as shown in formula (3): X 1-1 =X 2out =Relu{DB[DB(X1)]+X1} (2) X 1out =X 1-4 [X 1-3 (X 1-2 (X 1-1 ))] (3) Where: X 1-1 , X 1-2 , X 1-3 , X 1-4 They represent the results of performing the residual operation and then passing through the Relu activation function; X 1out Indicates the final output result of the branch; The output of the second branch X2 is: X 2out =Relu{DB[DB(X2)]+X2} (4) The third branch uses a depthwise convolution with a kernel size of 3×3, followed by BN processing and Relu activation function to obtain the output X of the third branch X3. 3out : X 3out =Relu[BN(Conv 3×3 (X3))] (5) The fourth branch only performs BN processing on the input feature map X4 once, and obtains the output X of the fourth branch X4. 4out : X 4out =BN(X4) (6) The results of the above four branches are combined and the obtained results are output through the Relu activation function; X*=Relu(X 1out +X 2out +X 3out +X 4out ) (7) Where: X * A new feature map representing the output of the multi-scale residual convolution module; Step 2.3, design MSRCM in the Bottleneck structure of the Neck layer to obtain an improved target detection network.

4. A method for detecting ice coating on a power transmission line according to claim 3, characterized in that: The step 3 is specifically implemented according to the following steps: 3.

1. Propose a deep hybrid attention module DHAM. DHAM is divided into two parts: the spatial attention module SAM and the channel attention module CAM. 3.

2. SAM extracts input features from four branches at multiple scales. The first branch inputs feature X 1in :For the input feature map X(X 1in , X 2in , X 3in ), (X∈R C×W×H ), first, the first branch input feature X is extracted from two perspectives using maximum pooling and average pooling 1in Features: X max1 =f maxpool (X 1in ) (8) X avg1 =f avgpool (X 1in ) (9) Where: X max1 ∈R C×H / 2×W / 2 ;X avg1 ∈R C×H / 2×W / 2 ; Use maximum pooling to extract important information, and use average pooling to extract global information. Concat the results of maximum pooling and average pooling, and then output them through the Sigmoid function: A c =σ(X max1 +X avg1 ) (10) Where: σ represents the Sigmoid function; A c ∈R C×1×1 ; A c Each element in represents the importance level of its corresponding channel; Second, the first branch is fed with feature X 1in It is divided into two paths: one path is fused with the output of the Sigmoid function; the other path first performs a 3×3 ordinary convolution, then extracts features through a depth-separable convolution and outputs it through the Relu function. The process is as follows: X 1-1 =A c *X 1in X 1-2 =f conv3 (X 1in ) X 1-3 =Relu[f conv1×1 (f conv3×3 (X 1-2 ))] (11) Where: X 1-1 Represents X 1in With A c The result of fusion; Conv3×3 Represents a 3×3 ordinary convolution operation; Finally, X 1-1 With X 1-3 Perform fusion output: F DM =X 1-out =X 1-1 +X 1-3 (12) Where: X 1-out and F DM Represents the first branch input feature X after fusion 1in The output feature map of the above feature map extraction process is named as Deep-extraction Module (DM); The second branch of the SAM structure inputs feature X 2in And the third branch input feature X 3in After 2x upsampling and 4x upsampling respectively, the input is input into the DM structure for feature extraction, and then after maximum pooling and average pooling respectively, the second branch input feature X is obtained. 2in And the third branch input feature X 3in The output feature map X 2-out and X 3-out : X 2-out =f maxpool [F DM (UP2(X 2in ))] X 3-out =f maxpool [F DM (UP4(X 2in ))] (13) Where: UP2 represents the input feature map X 2in The width and height of the input feature map X are upsampled by 2 times; UP4 means 3in The width and height of F are upsampled by 4 times; DM It means that the upsampled feature maps are input into the DM structure to extract features; f maxpool and f avgpool Respectively represent the extraction of F from the perspective of maximum pooling and average pooling DM Features; The last part of SAM is to transform the output feature maps X of the three branches 1-out , X 2-out , X 3-out After fusion, the weighted sum is performed with the input X to obtain the weight of the target: X A =X×(X 1-out +X 2-out +X 3-out ) (14) Where: X A Represents the output result of SAM; This method can enhance the spatial information of the target object while keeping a relatively small number of parameters by performing multi-scale extraction and fusion of images; 3.

3. The CAM part of the DHAM module extracts input features X from two different angles through maximum pooling and average pooling A characteristic information; After the extraction is completed, a depth-separable convolution operation is performed, and then the obtained results are fused. The fused features are combined with the input X A Weighted, get the weight of the target; finally, through the sigmoid function to get the weight between (0,1), extract the target channel information, that is: X'=σ{X A ×[DW(f maxpool (X A ))+DW(f avgpool (X A ))]} (15) Where: σ represents the sigmoid function; X'∈R C×1×1 ; Each element in X' represents the importance level of its corresponding channel, and DW represents the depth convolution (f cov3x3 ) and point convolution (f cov1x1 ) Extraction process; 3.

4. Embedding the DHAM module into the Neck layer can adaptively assign weights to specific target areas to improve the model's detection capabilities for small targets and complex backgrounds, and avoid missed detections and false detections.

5. A method for detecting ice coating on a power transmission line according to claim 4, characterized in that: The step 4 is specifically implemented according to the following steps: The performance of DHAM-YOLO is verified through comparative experiments of different attention mechanisms, ablation experiments, target detection, and training loss curves; 4.

1. Under the same experimental conditions, different attention mechanisms are compared based on various indicators and further verified using heat maps; 4.

2. Add DHAM and MSRCM modules to the initial module in turn, conduct ablation experiments, and compare and analyze the ice detection performance of each module; 4.

3. Parameters, FPS, FLOPs and mAP of the mainstream algorithms Faster RCNN, SSD, Centernet and YOLO are compared with the algorithm in this paper. At the same time, the convergence performance and overfitting problem of the model of the present invention are verified by comparing the training loss curves and verification loss curves of different models. 4.

4. Select images of iced transmission lines and iced insulators in different complex environments for testing.

Images