GBRNet network-based ADSS optical fiber cable icing detection method

Abstract

The invention discloses an ADSS optical fiber cable icing detection method based on a GBRNet network, and the method comprises the steps: building a GENet feature extraction network and a BiPANet feature fusion network, obtaining an optical fiber cable image in which an icing region recognition frame is added according to an iced ADSS optical fiber cable image, obtaining the coordinates of an icing image recognition frame, and carrying out the recognition of the icing region recognition frame; and cutting the ADSS optical fiber cable images before and after icing, obtaining the pixel thickness of the optical cable in the ADSS optical fiber cable images before and after icing based on an edge extraction network, and finally obtaining the icing thickness of the ADSS optical fiber cable. According to the method, the problem of difficulty in ADSS optical cable icing detection under complex weather conditions can be solved, the accuracy and the recognition speed of the model during icing detection are improved through combination of the GENet feature extraction network and the BiPANet feature fusion network, and the accuracy of an icing thickness detection result is relatively high.

Description

Technical Field

[0001] The present invention relates to the technical field of transmission line detection, and particularly to an ADSS optical fiber cable icing detection method based on the GBRNet network. Background Art

[0002] ADSS optical fiber cables are widely used in overhead transmission lines due to their excellent performance and cost-effectiveness, supporting the normal operation of the national power grid communication network. However, in many areas where optical cables are deployed, annual icing has different impacts on power lines. This icing affects the tensile strength of the line and may cause excessive sagging of the optical fiber cable and an unsafe clearance from the ground. In severe cases, it may cause the iron core or cable to break, resulting in the interruption of information flow and road blockage. Currently, icing detection methods mainly focus on overhead transmission lines and have limited applications on ADSS optical cables. With the continuous development of smart grids, significant progress has been made in transmission line icing detection. The widespread application of deep learning algorithms such as convolutional neural networks marks a shift from traditional image processing algorithms to more advanced methods, while also overcoming the limitations of traditional algorithms in dealing with complex tasks. Traditional icing detection methods involve installing capacitive sensors on transmission lines, which can detect capacitance changes caused by the dielectric constant changes of air and ice to estimate the icing thickness. However, this method is vulnerable to environmental factors such as wind and other foreign objects, which can affect the accuracy of the sensors in actual detection, resulting in large errors in the inspection results of the icing thickness. Summary of the Invention

[0003] The present invention discloses an ADSS optical fiber cable icing detection method based on the GBRNet network to overcome the above technical problems.

[0004] To achieve the above object, the technical solution of the present invention is as follows:

[0005] An ADSS optical fiber cable icing detection method based on the GBRNet network includes the following steps:

[0006] S1: Obtain images of the area where the ADSS optical fiber cable is located, including images of the ADSS optical fiber cable before icing and after icing;

[0007] S2: Establish a GENet feature extraction network and a BiPANet feature fusion network to obtain an optical fiber cable image with an icing area recognition frame based on the image of the ADSS optical fiber cable after icing, and obtain the coordinates of the icing image recognition frame;

[0008] S3: According to the coordinates of the ice-covered image recognition frame, crop the ADSS optical fiber cable images before and after icing respectively, and input the cropped ADSS optical fiber cable images before and after icing into the edge extraction network to obtain the pixel thickness of the optical cable in the ADSS optical fiber cable image before icing and the pixel thickness of the optical cable in the ADSS optical fiber cable image after icing;

[0009] S4: Obtain the icing thickness of the ADSS optical fiber cable according to the pixel thickness of the optical cable in the ADSS optical fiber cable image before icing and the pixel thickness of the optical cable in the ADSS optical fiber cable image after icing.

[0010] Beneficial effects: An ADSS optical fiber cable icing detection method based on the GBRNet network of the present invention establishes a GENet feature extraction network and a BiPANet feature fusion network to obtain an optical fiber cable image with an icing area recognition frame according to the ADSS optical fiber cable image after icing, so as to obtain the coordinates of the ice-covered image recognition frame, crop the ADSS optical fiber cable images before and after icing, and based on the edge extraction network, obtain the pixel thickness of the optical cable in the ADSS optical fiber cable images before and after icing, and finally obtain the icing thickness of the ADSS optical fiber cable. The present invention can solve the problem of difficult icing detection of ADSS optical cables under complex weather conditions: by combining the GENet feature extraction network and the BiPANet feature fusion network, the accuracy and recognition speed of the model in detecting icing are improved, and the accuracy of the inspection result of the icing thickness is relatively high. The GBRNet network of the present invention can reduce the difficulty of transmission line detection, improve the maintainability of the power system, provide key support for the security and reliability of signal transmission, and ensure the reliability and stability of the power communication network. Description of the Drawings

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to these drawings without creative efforts.

[0012] Figure 1 It is a flowchart of the ADSS optical fiber cable icing detection method based on the GBRNet network of the present invention;

[0013] Figure 2 It is a schematic structural diagram of the GENet feature extraction network established in the embodiment of the present invention;

[0014] Figure 3Schematic diagram of the BiPANet feature fusion network structure established in the embodiment of the present invention;

[0015] Figure 4 Schematic diagram of the GE attention mechanism in the embodiment of the present invention;

[0016] Figure 5a Schematic diagram of the FPN structure in the embodiment of the present invention;

[0017] Figure 5b Schematic diagram of the PANet structure in the embodiment of the present invention;

[0018] Figure 5c Schematic diagram of the BiPANet structure in the embodiment of the present invention;

[0019] Figure 6 Schematic diagram of the RCF network structure in the embodiment of the present invention;

[0020] Figure 7 Schematic diagram of icing edge detection in the embodiment of the present invention;

[0021] Figure 8 Schematic diagram of the process of extracting the icing edge of the optical cable in the embodiment of the present invention;

[0022] Figure 9 Schematic diagram of the mAP0.5 training loss curve in the embodiment of the present invention;

[0023] Figure 10a Schematic diagram of the YOLOv8s confusion matrix in the embodiment of the present invention;

[0024] Figure 10b Schematic diagram of the GBRNet confusion matrix in the embodiment of the present invention;

[0025] Figure 11 Schematic diagram of GBRNet optical cable icing detection in the embodiment of the present invention. Detailed implementation manners

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] This embodiment introduces an ADSS optical fiber cable icing detection method based on the GBRNet network, as Figure 1 shown,

[0028] S1: Obtain images of the area where the ADSS optical fiber cable is located, including the image of the ADSS optical fiber cable before icing and the image of the ADSS optical fiber cable after icing;

[0029] Specifically, the image of the area where the ADSS optical fiber cable in this embodiment is located is obtained by a camera fixed on the pole tower, and the collected image is input into the GBRNet network for object detection.

[0030] S2: Establish a GENet feature extraction network (dual-channel feature extraction network) and a BiPANet feature fusion network to obtain an optical fiber cable image with an icing area recognition frame based on the image of the ADSS optical fiber cable after icing, so as to obtain the coordinates of the icing image recognition frame;

[0031] Specifically, the GENet feature extraction network is used to obtain the local features and global features of the optical fiber cable based on the image of the ADSS optical fiber cable after icing; the input of the GENet feature extraction network is the image of the ADSS optical fiber cable after icing; the outputs are the color-dominated branch output feature map, the depth-dominated branch output feature map, and the fused depth output feature map; the BiPANet feature fusion network is used to obtain the icing area recognition frame based on the local features and global features of the optical fiber cable extracted by the GENet feature extraction network, and at the same time obtain the coordinates of the icing area recognition frame. The inputs of the BiPANet feature fusion network are the color-dominated branch output feature map, the depth-dominated branch output feature map, and the fused depth output feature map; the output is the optical cable icing image recognition frame, that is, a red recognition area is added to the original input cable icing image, as shown in Figure 8 ;

[0032] Preferably, the GENet feature extraction network includes an input layer, a color-dominated branch, a depth-dominated branch, a feature interaction module, a GE attention mechanism module, and an output layer; as Figure 2 shown;

[0033] The input end of the color-dominated branch is connected to the input layer and is used to obtain the color-dominated branch output feature map based on the image of the ADSS optical fiber cable after icing;

[0034] The color-dominated branch is connected to the input end of the feature interaction module and is used to obtain the shallow features of the color branch based on the color semantic feature map output from the output end of the first five-convolution layer of the color-dominated branch; the output end of the feature interaction module is connected to the depth-dominated branch;

[0035] The input ends of the depth-dominant branch are respectively connected to the input layer and the output end of the feature interaction module, and are used to obtain a depth-dominant branch output feature map according to the iced ADSS optical fiber cable image and the color branch shallow features;

[0036] The input ends of the GE attention mechanism module are respectively connected to the output end of the color-dominant branch and the output end of the depth-dominant branch, and are used to obtain a fused depth output feature map according to the color-dominant branch output feature map and the depth-dominant branch output feature map;

[0037] The input end of the output layer is connected to the output end of the GE attention mechanism module.

[0038] Specifically, the feature interaction module in this embodiment adopts a conventional concat module.

[0039] Preferably, the color-dominant branch includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first upsampling layer, a first residual connection and a normalization layer, a second upsampling layer, a second residual connection and a normalization layer, a third upsampling layer, a third residual connection and a normalization layer, a fourth upsampling layer, a fourth residual connection and a normalization layer, and a fifth upsampling layer, a fifth residual connection and a normalization layer;

[0040] The input end of the first convolutional layer is connected to the input layer, and is used to obtain a color edge feature map according to the iced ADSS optical fiber cable image;

[0041] The input end of the second convolutional layer is connected to the output end of the first convolutional layer, and is used to obtain a color texture feature map according to the color edge feature map;

[0042] The input end of the third convolutional layer is connected to the output end of the second convolutional layer, and is used to obtain a color spot feature map according to the color texture feature map;

[0043] The input end of the fourth convolutional layer is connected to the output end of the third convolutional layer, and is used to obtain a color stripe feature map according to the color spot feature map;

[0044] The input end of the fifth convolutional layer is connected to the output end of the fourth convolutional layer, and is used to obtain a color semantic feature map according to the color stripe feature map;

[0045] The output end of the fifth convolutional layer is connected to the input end of the feature interaction module, and is used to obtain color branch shallow features according to the color semantic feature map;

[0046] The input end of the first upsampling layer is connected to the output end of the fifteenth convolutional layer, and is used to obtain a feature map with the same resolution as the color stripe feature map according to the color semantic feature map;

[0047] The input ends of the first residual connection and normalization layer are respectively connected to the output end of the fourteenth convolutional layer and the output end of the first upsampling layer, and are used to obtain a residual connection output feature map containing color stripe features according to the color stripe feature map and the feature map with the same resolution as the color stripe feature map;

[0048] The input end of the second upsampling layer is connected to the output end of the first residual connection and normalization layer, and is used to obtain a feature map with the same resolution as the color spot feature map according to the residual connection output feature map containing color stripe features;

[0049] The input ends of the second residual connection and normalization layer are respectively connected to the output end of the thirteenth convolutional layer and the output end of the second upsampling layer, and are used to obtain a residual connection output feature map containing color spot features according to the color spot feature map and the feature map with the same resolution as the color spot feature map;

[0050] The input end of the third upsampling layer is connected to the output end of the second residual connection and normalization layer, and is used to obtain a feature map with the same resolution as the color texture feature map according to the residual connection output feature map containing color spot features;

[0051] The input ends of the third residual connection and normalization layer are respectively connected to the output end of the twelfth convolutional layer and the output end of the third upsampling layer, and are used to obtain a residual connection output feature map containing color texture features according to the color texture feature map and the feature map with the same resolution as the color texture feature map;

[0052] The input end of the fourth upsampling layer is connected to the output end of the third residual connection and normalization layer, and is used to obtain a feature map with the same resolution as the color edge feature map according to the residual connection output feature map containing color texture features;

[0053] The input ends of the fourth residual connection and normalization layer are respectively connected to the output end of the eleventh convolutional layer and the output end of the fourth upsampling layer, and are used to obtain a residual connection output feature map containing color edge features according to the color edge feature map and the feature map with the same resolution as the color edge feature map;

[0054] The input end of the first five upsampling layer is connected to the output end of the first four residual connection and normalization layer, and is used to obtain a feature map with the same resolution as the color input feature map according to the residual connection output feature map containing color edge features;

[0055] The input ends of the first five residual connection and normalization layer are respectively connected to the input layer and the output end of the first five upsampling layer, and are used to obtain a color-dominated branch output feature map according to the iced ADSS optical fiber cable image and the feature map with the same resolution as the color input feature map;

[0056] The output end of the first five residual connection and normalization layer is connected to the GE attention mechanism module.

[0057] Preferably, the depth-dominated branch includes a second one convolutional layer, a second two convolutional layer, a second three convolutional layer, a second four convolutional layer, a second five convolutional layer, a second one upsampling layer, a second one residual connection and normalization layer, a second two upsampling layer, a second two residual connection and normalization layer, a second three upsampling layer, a second three residual and upsampling layer, a second four upsampling layer, a second four residual connection and normalization layer, a second five upsampling layer, and a second five residual connection and normalization layer;

[0058] The input end of the second one convolutional layer is connected to the input layer, and is used to obtain a depth edge feature map according to the iced ADSS optical fiber cable image;

[0059] The input end of the second two convolutional layer is connected to the output end of the second one convolutional layer, and is used to obtain a depth texture feature map according to the depth edge feature map;

[0060] The input end of the second three convolutional layer is connected to the output end of the second two convolutional layer, and is used to obtain a depth spot feature map according to the depth texture feature map;

[0061] The input end of the second four convolutional layer is connected to the output end of the second three convolutional layer, and is used to obtain a depth stripe feature map according to the depth spot feature map;

[0062] The input end of the second five convolutional layer is connected to the output end of the second four convolutional layer, and is used to obtain a depth semantic feature map according to the depth stripe feature map;

[0063] The input end of the second one upsampling layer is respectively connected to the output end of the second five convolutional layer and the output end of the feature interaction module, and is used to obtain a feature map with the same resolution as the depth stripe feature map according to the depth semantic feature map and the shallow color branch features;

[0064] The input ends of the second residual connection and normalization layer are respectively connected to the output end of the second four-convolution layer and the output end of the second one-upsampling layer, and are used to obtain a residual connection output feature map containing depth stripe features according to the depth stripe feature map and a feature map having the same resolution as the depth stripe feature map;

[0065] The input end of the second two-upsampling layer is connected to the output end of the second one residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth spot feature map according to the residual connection output feature map containing depth stripe features;

[0066] The input ends of the second two residual connection and normalization layer are respectively connected to the output end of the second three-convolution layer and the output end of the second two-upsampling layer, and are used to obtain a residual connection output feature map containing depth spot features according to the depth spot feature map and a feature map having the same resolution as the depth spot feature map;

[0067] The input end of the second three-upsampling layer is connected to the output end of the second two residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth texture feature map according to the residual connection output feature map containing depth spot features;

[0068] The input ends of the second three residual connection and normalization layer are respectively connected to the output end of the second two-convolution layer and the output end of the second three-upsampling layer, and are used to obtain a residual connection output feature map containing depth texture features according to the depth texture feature map and a feature map having the same resolution as the depth texture feature map;

[0069] The input end of the second four-upsampling layer is connected to the output end of the second three residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth edge feature map according to the residual connection output feature map containing depth texture features;

[0070] The input ends of the second four residual connection and normalization layer are respectively connected to the output end of the second one-convolution layer and the output end of the second four-upsampling layer, and are used to obtain a residual connection output feature map containing depth edge features according to the depth edge feature map and a feature map having the same resolution as the depth edge feature map;

[0071] The input end of the second five-upsampling layer is connected to the output end of the second four residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth input feature map according to the residual connection output feature map containing depth edge features;

[0072] The input ends of the second five residual connection and normalization layer are respectively connected to the output ends of the input layer and the second five upsampling layer, and are used to obtain a depth-dominated branch output feature map according to the iced ADSS optical fiber cable image and a feature map with the same resolution as the depth input feature map;

[0073] The output end of the second five residual connection and normalization layer is connected to the GE attention mechanism module.

[0074] Preferably, as Figure 3 shown; the BiPANet feature fusion network structure includes: a first upsampling layer, a first convolutional layer, a second upsampling layer, a second convolutional layer, a first feature interaction layer, a second feature interaction layer, a third feature interaction layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first residual connection and normalization layer, a second residual connection and normalization layer, and an output layer;

[0075] The input end of the first upsampling layer is connected to the GENet feature extraction network. Specifically, it is connected to the GE attention mechanism module of the GENet feature extraction network, and is used to obtain a fused depth output feature map with the same resolution as the depth-dominated branch output feature map according to the fused depth output feature map;

[0076] The input ends of the first feature interaction layer are respectively connected to the output end of the first upsampling layer and the GENet feature extraction network. Specifically, it is connected to the depth-dominated branch of the GENet feature extraction network, and is used to obtain a shallow fusion feature map according to the depth-dominated branch output feature map and a fused depth output feature map with the same resolution as the depth-dominated branch output feature map;

[0077] The input end of the first convolutional layer is connected to the output end of the first feature interaction layer, and is used to obtain shallow feature information according to the shallow fusion feature map;

[0078] The input end of the second upsampling layer is connected to the output end of the first convolutional layer, and is used to obtain a shallow feature map with the same resolution as the color-dominated branch output feature map according to the shallow feature information;

[0079] The input ends of the second feature interaction layer are respectively connected to the output end of the second upsampling layer and the GENet feature extraction network. Specifically, it is connected to the color-dominated branch of the GENet feature extraction network, and is used to obtain a deep fusion feature map according to the color-dominated branch output feature map and a shallow feature map with the same resolution as the color-dominated branch output feature map;

[0080] The input end of the second convolutional layer is connected to the output end of the second feature interaction layer, and is used to obtain deep feature information according to the deep fusion feature map;

[0081] The input ends of the third feature interaction layer are respectively connected to the output end of the second convolutional layer and the GENet feature extraction network, and are used to obtain deep color feature information according to the feature map output by the color dominant branch and the deep feature information;

[0082] The input end of the third convolutional layer is connected to the output end of the third feature interaction layer, and is used to obtain the first-size object detection head according to the deep color feature information;

[0083] The input ends of the first residual connection and normalization layer are respectively connected to the output end of the third convolutional layer and the GENet feature extraction network, and are used to obtain the feature map with the original depth dominant branch output feature information according to the feature map output by the depth dominant branch and the first-size object detection head;

[0084] The input end of the fourth convolutional layer is connected to the output end of the first residual connection and normalization layer, and is used to obtain the second-size object detection head according to the feature map with the original depth dominant branch output feature information;

[0085] The input ends of the second residual connection and normalization layer are respectively connected to the output end of the fourth convolutional layer and the GENet feature extraction network, and are used to obtain the feature map with the original fusion depth output information according to the fused depth output feature map and the second-size object detection head;

[0086] The input end of the fifth convolutional layer is connected to the output end of the second residual connection and normalization layer, and is used to obtain the third-size object detection head according to the feature map with the original fusion depth output information;

[0087] The output layer is used to obtain and output the optical fiber cable image with the icing area recognition frame added according to the first-size object detection head, the second-size object detection head, and the third-size object detection head.

[0088] Among them, the relationship between the first-size object detection head, the second-size object detection head, and the third-size object detection head is: the size of the second-size object detection head = the size of the first-size object detection head * 2, and the size of the third-size object detection head = the size of the second-size object detection head * 2;

[0089] S3: According to the coordinates of the icing image recognition frame, respectively crop the ADSS optical fiber cable image before icing and the ADSS optical fiber cable image after icing, and respectively input the cropped ADSS optical fiber cable image before icing and the cropped ADSS optical fiber cable image after icing into the edge extraction network to obtain the pixel thickness of the optical cable in the ADSS optical fiber cable image before icing and the pixel thickness of the optical cable in the ADSS optical fiber cable image after icing;

[0090] Among them, the edge extraction network is an existing technology in the field, and the edge extraction network will not be described in detail here.

[0091] S4: According to the pixel thickness of the optical cable in the ADSS optical fiber cable image before icing and the pixel thickness of the optical cable in the ADSS optical fiber cable image after icing, and using the standard diameter of the optical cable, the icing thickness is obtained.

[0092] Specifically, traditional attention mechanisms, such as the Convolutional Block Attention Module (CBAM), introduce large-scale convolutional kernels to extract spatial features, but often ignore long-range dependencies. Although some other attention modules solve this problem, they often have a higher number of parameters and complexity compared to the more common GE module.

[0093] To enhance the model's ability to represent the iced optical cable in different scenarios, a GENet feature extraction network (dual-channel feature extraction network) is built in this embodiment, as Figure 2 . The GENet feature extraction network adopts a multi-scale feature pyramid structure. These pyramids cover feature maps at different levels and use upsampling and downsampling operations to flexibly adjust the scale of the feature pyramid. Among them, upsampling can improve the resolution and is conducive to accurately capturing the details of small targets; downsampling can reduce the resolution and helps to grasp the overall features of large targets. At the end of the network, a GE attention module is additionally added to strengthen the features, and finally the GENet network is constructed. The GENet model can maximize spatial attention to better explore the feature correlation between different convolutional layers, thereby improving the detection accuracy and efficiency of the network.

[0094] Specifically, the GE module is a lightweight module similar to the Squeeze-and-Excitation (SE) attention mechanism. In the GE module, the aggregation operator is used to extract features from local spatial positions, and the excitation operator scales these features and restores them to their original dimensions. As Figure 4 shown.

[0095] In Figure 4 , the excitation processes of these two operations are given, where ξ G and ξ E represent the aggregation and excitation operations respectively. First, is responsible for aggregating responses, performing extensive spatial feature responses in a large spatial neighborhood to capture a wider context. Subsequently, the aggregation result of ξ G and the original input tensor are passed to ξ E。The purpose of this step is to adjust the feature responses of the original input tensor according to the captured context information to obtain better feature representations. Through this process, the performance of the CNN network in capturing context information is improved. Among them, the selection operator i(u,e) = eu + δ: δ ∈ [-[(2e - 1) / 2], [2e - 1 / 2]] 2 is to evaluate the impact on the size of the spatial region where aggregation occurs:

[0096]

[0097] where: i(u,e) represents the selection operator; u represents the output position; e represents the weight coefficient; δ represents the bias term; : represents the value range of the variable; represents information aggregation; x represents the input vector; ⊙ represents the Hadamard product; represents the indicator tensor;

[0098] where, u ∈ 1,…,H′×1,…,W′, c ∈ 1,…,C, f:R H′×W′×C →[0,1] H×W×C , l represents the indicator tensor. It indicates that at each output position u of channel c, the aggregation operator has the receptive field of the input within a single channel. H′ represents the height of the feature map at output position u; W′ represents the width of the feature map at output position u; c represents the number of channels; C represents the maximum value of the output channels; f represents the activation operation function; R represents the three-dimensional feature map; H represents the height of the output feature map after mapping; W represents the width of the output feature map after mapping;

[0099]

[0100] In the formula: represents the output result of the activation operation; represents the information aggregation vector; represents the activation operation of the information aggregation vector;

[0101] where f:R H′×W′×C →[0,1] H×W×C is the mapping responsible for scaling and distributing the signals from the aggregation, → represents the mapping process;

[0102] Specifically, in object detection, multi-scale feature representation is crucial for improving the algorithm performance. Traditional detectors usually directly make predictions based on the pyramid features extracted from the backbone network. However, this method may lead to insufficient utilization of multi-scale features and semantic difference problems. To solve these problems, the Feature Pyramid Network (FPN) is introduced, which combines multi-scale features in a top-down manner, such as Figure 5aAs shown. This method improves the performance and robustness of obtaining semantic information. However, FPN still faces semantic differences between different levels. PANet introduces an additional bottom-up path on the basis of FPN, as Figure 5b shown, which helps with more comprehensive information fusion, but also makes the structure relatively complex because it requires an additional information transmission path. According to PANet, the feature fusion network BiPANet built in this embodiment, as Figure 5c shown, this network introduces learnable weights to evaluate the importance of different input features. BiPANet repeatedly utilizes top-down and bottom-up multi-scale feature fusion to more effectively process multi-scale features. In addition, BiPANet adds a residual structure for merging context information and multiplies each residual structure by a corresponding weight, thereby fusing more features without significantly increasing the computational cost. BiPANet improves the original algorithm by introducing more effective bidirectional cross-scale connections and fusing weighted feature maps, enhancing its ability to process multi-scale features and introduce context information. The feature fusion calculation expression of BiPANet at the P6 layer is shown as follows.

[0103]

[0104] In the formula: represents the output feature of the bottom-up path of the sixth level. represents the intermediate feature of the top-down path of the sixth level; Conv(·) represents the convolution operation; ω′1 represents the weight coefficient of the input feature map of the second convolutional layer ; represents the input feature map of the second convolutional layer; ω′2 represents the weight coefficient of the feature fusion feature map; ω′3 represents the weight coefficient of the output feature map of the second feature interaction layer; represents the operation of resizing the feature map; represents the output feature map of the second feature interaction layer; ε represents a positive number approaching 0, used to avoid division by zero; represents the input feature map of the third feature interaction layer; ω1 represents the weight coefficient of the input feature map of the second convolutional layer ; ω2 represents the weight coefficient of the input feature map of the third feature interaction layer ;

[0105] Specifically, ω i is a learnable weight, and ε = 0.0001 is used to avoid numerical instability.

[0106] Specifically, as Figure 6As shown, at each stage, the RCF detection network introduces a 1×1 convolutional layer with a channel depth of 21 after each convolutional layer, and applies another 1×1 convolution to accumulate features. Then deconvolution is performed to upsample the feature map. At the end of each stage, a loss layer is added to calculate the loss. Finally, all the upsampling layers are interconnected across layers and fused with 1×1 convolutional kernels, and the output image is obtained after passing through the loss layer.

[0107] In this embodiment, after edge detection of the ice-covered optical cable, the ice thickness needs to be calculated. Therefore, first, the ice-covered area of the optical cable is extracted, and then the ice thickness on the optical cable is measured, as shown in the figure. The method for calculating the ice thickness is as Figure 7 shown. Half of the difference between D and d in the figure is the calculated ice thickness. D and d represent the pixel thicknesses of the ice and the optical cable in the image respectively. The formula is as follows:

[0108]

[0109] In the formula: y represents the average thickness of the ice; D represents the pixel measurement value of the outer diameter after icing, that is, the pixel thickness of the optical cable in the image of the ADSS optical fiber cable after icing; d represents the pixel measurement value of the outer diameter without icing, that is, the pixel thickness of the optical cable in the image of the ADSS optical fiber cable before icing; h represents the standard diameter of the optical cable.

[0110] A specific embodiment of the present invention is as follows:

[0111] The optical cable dataset comes from cameras fixed on transmission towers. The images are taken at hourly intervals and positions every day. These images show significant feature differences before and after the ADSS optical cable is iced.

[0112] The dataset used in the experiment was provided by the State Grid Corporation of China. It consists of images taken in real time by cameras installed under various transmission lines for monitoring the status of ADSS optical cables. The dataset was sorted and screened, and a total of 1274 images containing ice-covered optical cables and normal optical cables were obtained. The dataset was manually annotated using LabelImg software with labels "ice" and "cable" to generate corresponding XML files. The dataset was divided into a training set and a validation set in a ratio of 9:1.

[0113] The experiment was carried out on a server equipped with the Ubuntu16.04 operating system, and three NVIDIA GeForce RTX 3090 GPUs were used for training. The training parameters included 300 epochs, using the Stochastic Gradient Descent (SGD) optimizer, an initial learning rate of 0.01, a batch size of 48, and the input image size was set to 640×640 pixels.

[0114] Among them, the algorithm performance evaluation metrics selected in this embodiment include P (test accuracy), R (recall rate), mAP (mean average precision), and FPS (frames per second). Accuracy is the proportion of true samples among all positive samples predicted by the model. Recall rate is the proportion of true samples predicted by the model among all true samples. Generally, there is a trade-off between accuracy and recall rate, where one increases while the other decreases. To balance the effects of accuracy and recall rate and evaluate the model more comprehensively, the mean average precision is introduced as a comprehensive evaluation metric. mAP is actually the area under the P-R curve, and a larger value indicates better model performance. The size of FPS directly reflects the speed of model inference, and a higher FPS means the model can process input data faster. The relevant formulas are as follows:

[0115]

[0116] mAP = ∫0 1 P(R)dR

[0117] In the formula: TP represents the optical cable instances correctly identified as ice-covered by the model, FP represents the optical cable instances misidentified as ice-covered by the model, and FN represents the partial optical cable instances not identified as ice-covered by the model. P(R) represents the function of the relationship between P and R in the P-R curve;

[0118] Specifically, to more deeply evaluate the performance of the GENet feature extraction network in this embodiment, accuracy, recall rate, and mAP0.5 are selected as the evaluation metrics for model recognition accuracy. Considering the demand for lightweight detection in actual tower deployments, the size and detection time of the model are selected as performance evaluation metrics. The specific comparison is shown in Table 1. By comparing the detection results of different models, it can be seen that the FPS values of the SSD and YOLOv5s models are relatively high, but their accuracy and mean average precision are relatively low. The YOLOv5m model has the highest FPS value, but its model size is larger than that of the YOLOv8s model, which is not conducive to future lightweight deployments. Compared with the YOLOv8m and YOLOv8s models, the GENet feature extraction network has improved in both detection accuracy and recall rate, with the mean average precision increasing by 2.2% and 0.6% respectively. When inferring a single image, the FPS of the network has increased by 16.7%. This indicates that the improved GENet feature extraction network performs best in detecting ice-covered optical cables.

[0119] Table 1 Comparative experiments of different models

[0120]

[0121]

[0122] The results of the comparative experiments in Table 1 show that the existing detection models for cable icing pay less attention to iced optical cables, and their recognition of icing and the surrounding environment is relatively limited. In contrast, the GBRNet network jointly formed by the GENet, BiPANet, and RCF networks in this factual example has a significant improvement in the ability to focus on iced optical cables and also pays more attention to the surrounding environment. Therefore, the improved algorithm can be better applied to the detection of iced optical cables.

[0123] The process of icing detection is as Figure 8 shown. In Figure 8 , first, the GENet network and the BiPANet network are used to identify the iced optical cable part, and then the icing area is intercepted and input into the RCF edge detection network, as shown. Finally, the edge extraction image of the iced optical cable is obtained, simplifying the thickness calculation.

[0124] In the measurement of the ice thickness on the optical cable, some important influencing factors are observed, especially those related to the areas in the image that are far from the camera focus and have a small pixel thickness. These factors may lead to a relatively large error rate in the measurement. Measuring a smaller pixel thickness usually results in a more significant error. This can be attributed to environmental factors such as changes in lighting conditions and reflectivity, which have a great impact on the measurement. In actual measurement, it is noted that the error rate varies significantly. In the areas far from the camera focus, the error rate is usually higher, ranging from 3.3% to 3.5%, while in other areas, the error rate is relatively lower, approximately between 0.9% and 2.3%. This difference emphasizes the important influence of the measurement location on the error rate. The experimental results are shown in Table 2.

[0125] Table 2 Ice thickness calculation

[0126]

[0127] Considering the observed measurement errors for different optical cables, in practice, the average total error rate is approximately 2.5%. This analysis emphasizes the effectiveness of the optical cable thickness measurement, especially in the areas far from the camera region and with a small pixel thickness. To further improve the accuracy of ice and snow thickness estimation, other measures are taken to reduce the error rate and ensure more reliable early warning and preventive measures to reduce the potential adverse effects of severe ice and snow coverage.

[0128] Specifically, Figure 9Shows the comparison of the mAP training curves of the YOLOv8s model and the GBRNet model. Since the complexity of the feature extraction network in the GBRNet model is higher than that of the CSPDarknet of the YOLOv8s model, the detection accuracy and convergence speed are relatively low before 100 epochs. However, after 125 epochs, the mAP0.5 of the GBRNet model gradually approaches that of the YOLOv8s model and exceeds the YOLOv8s model during the convergence stage.

[0129] Specifically, Figure 10a and Figure 10b respectively show the confusion matrices of the YOLOv8s model and the GBRNet model. For the ice label, the prediction accuracy of the YOLOv8s model is 72%, while another 11% is misclassified as the background. In contrast, the GBRNet model accurately predicts 73%, and only 10% is mispredicted as the background. For the cable label, the prediction accuracy of the YOLOv8s model is 61%, while another 16% is misclassified as the background. In contrast, the GBRNet model accurately predicts 67%, and only 13% is mispredicted as the background. In general, compared with the YOLOv8s model, the GBRNet model shows better detection performance in the optical cable icing detection task.

[0130] Specifically, the results of using the GBRNet model to identify the images and icing states of ADSS optical cables are as Figure 11 shown. It can be seen that in the large-scale complex scene of Figure 11 , the model can accurately identify the optical cable and icing, and the confidence levels of identifying icing and the optical cable are higher than 0.8 and 0.6 respectively.

[0131] In this embodiment, aiming at the problem of low icing detection accuracy of all-dielectric self-supporting (ADSS) optical cables on overhead transmission lines, it solves the problem that the ADSS optical fiber cables are iced under harsh environmental conditions, which hinders timely inspection and interferes with their normal operation. First, with the help of the dual-branch backbone network of the GENet feature extraction network, the image is disassembled into components of multiple resolutions. This operation not only enhances the image details but also optimizes the low-frequency information, creating good conditions for deeply extracting the detail and semantic information of the image. The GENet feature extraction network uses a dual-backbone network architecture to extract deep features. Its encoding-decoding structure enables interaction between features of different scales, thereby obtaining richer deep features, enhancing its adaptability to various environments, and improving the recognition ability of icing states in different scenarios. In addition, the established BiPANet feature fusion network endows the model with stronger global feature and local information capture capabilities. The BiPANet feature fusion network can utilize top-down and bottom-up multi-scale feature fusion to more effectively process multi-scale features and improve the detection accuracy. When the GENet feature extraction network and the BiPANet feature fusion network are fused with each other, the feature information of different scales is efficiently integrated, the object localization and detection accuracy of the model are significantly improved, and its detection ability is also greatly enhanced. Finally, using the RCF edge detection network, the edge features of the iced optical cable are obtained in the RGB image, the edge features of the iced optical cable are obtained and the thickness of the ice is calculated, and its accuracy can meet the early warning monitoring of the optical cable icing. Finally, the ice thickness is calculated according to the icing image. The experimental results show that compared with the YOLOv8s network, the icing detection accuracy of this improved network is increased by 5.9%, the recall rate is increased by 5%, and mAP0.5 is increased by 2.2%. The detection speed of this network reaches 147FPS, and the detection error rate of the ice thickness is 2.5%, providing valuable reference for the monitoring of ADSS optical cables, ensuring the normal operation of the power signal network, and the icing warning of ADSS optical cables. This embodiment has comprehensive performances of fewer parameters, faster inference speed, and higher detection accuracy. This is beneficial to subsequent ice thickness detection operations and can be easily deployed on detection devices with limited resources.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting ice coating on ADSS optical fiber cables based on GBRNet network, characterized in that: The steps include: S1: Acquire images of the area where the ADSS optical fiber cable is located, including images of the ADSS optical fiber cable before and after ice covering; S2: Establish a GENet feature extraction network and a BiPANet feature fusion network to obtain an optical fiber cable image with an ice-covered area identification frame added according to the ADSS optical fiber cable image after ice-covered, so as to obtain the coordinates of the ice-covered image identification frame; S3: According to the coordinates of the ice-covered image recognition frame, the ADSS optical fiber cable image before ice-covered and the ADSS optical fiber cable image after ice-covered are cropped respectively, and the cropped ADSS optical fiber cable image before ice-covered and the cropped ADSS optical fiber cable image after ice-covered are input into the edge extraction network to obtain the pixel thickness of the optical cable in the ADSS optical fiber cable image before ice-covered and the pixel thickness of the optical cable in the ADSS optical fiber cable image after ice-covered; S4: Obtain ice thickness of the ADSS optical fiber cable according to the pixel thickness of the optical cable in the ADSS optical fiber cable image before ice coating and the pixel thickness of the optical cable in the ADSS optical fiber cable image after ice coating.

2. According to a GBRNet network-based ADSS optical fiber cable ice detection method according to claim 1, it is characterized in that: The GENet feature extraction network includes an input layer, a color-dominant branch, a depth-dominant branch, a feature interaction module, a GE attention mechanism module, and an output layer; The input end of the color-dominant branch is connected to the input layer, and is used to obtain a color-dominant branch output feature map according to the ADSS optical fiber cable image after ice coating; The color dominant branch is connected to the input end of the feature interaction module, and is used to obtain the shallow features of the color branch according to the color semantic feature map; The input end of the depth-dominant branch is connected to the output end of the input layer and the feature interaction module respectively, and is used to obtain a depth-dominant branch output feature map according to the ADSS optical fiber cable image after ice coating and the shallow features of the color branch; The input end of the GE attention mechanism module is connected to the output end of the color-dominant branch and the output end of the depth-dominant branch, respectively, for obtaining a fused depth output feature map according to the color-dominant branch output feature map and the depth-dominant branch output feature map; The input end of the output layer is connected to the output end of the GE attention mechanism module.

3. The method for detecting ice coating on ADSS optical fiber cables based on GBRNet network according to claim 2 is characterized in that: The color dominant branch includes a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, a fifth convolution layer, a first upsampling layer, a first residual connection and a normalization layer, a second upsampling layer, a second residual connection and a normalization layer, a third upsampling layer, a third residual connection and a normalization layer, a fourth upsampling layer, a fourth residual connection and a normalization layer, a fifth upsampling layer, a first fifth residual connection and a normalization layer; The input end of the first convolutional layer is connected to the input layer, and is used to obtain a color edge feature map according to the ADSS optical fiber cable image after ice coating; The input end of the first second convolutional layer is connected to the output end of the first first convolutional layer, and is used to obtain a color texture feature map according to the color edge feature map; The input end of the first three convolutional layers is connected to the output end of the first two convolutional layers, and is used to obtain a color spot feature map according to the color texture feature map; The input end of the first four convolutional layers is connected to the output end of the first three convolutional layers, and is used to obtain a color stripe feature map according to the color spot feature map; The input end of the first five convolutional layers is connected to the output end of the first four convolutional layers, and is used to obtain a color semantic feature map according to the color stripe feature map; The output end of the first five convolutional layers is connected to the input end of the feature interaction module, and is used to obtain the shallow features of the color branch according to the color semantic feature map; The input end of the first upsampling layer is connected to the output end of the first five convolutional layers, and is used to obtain a feature map having the same resolution as the color stripe feature map according to the color semantic feature map; The input ends of the first residual connection and normalization layer are respectively connected to the output ends of the first four convolutional layers and the output ends of the first upsampling layer, and are used to obtain a residual connection output feature map containing color stripe features according to the color stripe feature map and a feature map having the same resolution as the color stripe feature map; The input end of the first second upsampling layer is connected to the output end of the first first residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the color spot feature map according to the residual connection output feature map containing the color stripe feature; The input ends of the first and second residual connections and the normalization layer are respectively connected to the output ends of the first and third convolutional layers and the output ends of the first and second upsampling layers, and are used to obtain a residual connection output feature map containing color blob features according to the color blob feature map and a feature map having the same resolution as the color blob feature map; The input end of the first three upsampling layers is connected to the output end of the first two residual connections and the normalization layer, and is used to obtain a feature map having the same resolution as the color texture feature map according to the residual connection output feature map containing the color spot feature; The input ends of the first three residual connections and the normalization layer are respectively connected to the output ends of the first two convolutional layers and the output ends of the first three upsampling layers, and are used to obtain a residual connection output feature map containing color texture features according to the color texture feature map and a feature map having the same resolution as the color texture feature map; The input end of the first four upsampling layers is connected to the output end of the first three residual connections and normalization layers, and is used to obtain a feature map having the same resolution as the color edge feature map according to the residual connection output feature map containing color texture features; The input ends of the first four residual connections and the normalization layer are respectively connected to the output ends of the first four convolutional layers and the output ends of the first four upsampling layers, and are used to obtain a residual connection output feature map containing color edge features according to the color edge feature map and a feature map having the same resolution as the color edge feature map; The input end of the first five upsampling layers is connected to the output end of the first four residual connections and normalization layers, and is used to obtain a feature map having the same resolution as the color input feature map according to the residual connection output feature map containing the color edge feature; The input ends of the first five residual connections and the normalization layer are respectively connected to the output ends of the input layer and the first five upsampling layers, and are used to obtain a color-dominant branch output feature map according to the ADSS optical fiber cable image after ice coating and a feature map having the same resolution as the color input feature map; The output ends of the first five residual connections and the normalization layer are connected to the GE attention mechanism module.

4. The method for detecting ice coating on ADSS optical fiber cables based on GBRNet network according to claim 2 is characterized in that: The depth-dominant branch includes a 21st convolutional layer, a 22nd convolutional layer, a 23rd convolutional layer, a 24th convolutional layer, a 25th convolutional layer, a 21st upsampling layer, a 21st residual connection and a normalization layer, a 22nd upsampling layer, a 22nd residual connection and a normalization layer, a 23rd upsampling layer, a 23rd residual and upsampling layer, a 24th upsampling layer, a 24th residual connection and a normalization layer, a 25th upsampling layer, a 25th residual connection and a normalization layer; The input end of the second convolutional layer is connected to the input layer, and is used to obtain a depth edge feature map according to the ADSS optical fiber cable image after ice coating; The input end of the second-second convolutional layer is connected to the output end of the second-first convolutional layer, and is used to obtain a depth texture feature map according to the depth edge feature map; The input end of the second-third convolutional layer is connected to the output end of the second-second convolutional layer, and is used to obtain a depth spot feature map according to the depth texture feature map; The input end of the second fourth convolutional layer is connected to the output end of the second third convolutional layer, and is used to obtain a depth stripe feature map according to the depth spot feature map; The input end of the second fifth convolutional layer is connected to the output end of the second fourth convolutional layer, and is used to obtain a depth semantic feature map according to the depth stripe feature map; The input end of the second upsampling layer is respectively connected to the output end of the second fifth convolutional layer and the output end of the feature interaction module, so as to obtain a feature map having the same resolution as the deep stripe feature map according to the deep semantic feature map and the shallow features of the color branch; The input end of the second residual connection and normalization layer is connected to the output end of the second fourth convolutional layer and the output end of the second upsampling layer respectively, and is used to obtain a residual connection output feature map containing depth stripe features according to the depth stripe feature map and a feature map having the same resolution as the depth stripe feature map; The input end of the second upsampling layer is connected to the output end of the second residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth spot feature map according to the residual connection output feature map containing the depth stripe feature; The input ends of the second-second residual connection and normalization layer are respectively connected to the output ends of the second-third convolutional layer and the output end of the second-second upsampling layer, and are used to obtain a residual connection output feature map containing deep speckle features according to the deep speckle feature map and a feature map having the same resolution as the deep speckle feature map; The input end of the second third upsampling layer is connected to the output end of the second second residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth texture feature map according to the residual connection output feature map containing the depth spot feature; The input ends of the second third residual connection and normalization layer are respectively connected to the output ends of the second second convolutional layer and the output ends of the second third upsampling layer, and are used to obtain a residual connection output feature map containing deep texture features according to the deep texture feature map and a feature map having the same resolution as the deep texture feature map; The input end of the second fourth upsampling layer is connected to the output end of the second third residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth edge feature map according to the residual connection output feature map containing the depth texture feature; The input ends of the second four residual connections and normalization layers are respectively connected to the output ends of the second first convolutional layer and the output ends of the second four upsampling layers, for obtaining a residual connection output feature map containing a depth edge feature according to the depth edge feature map and a feature map having the same resolution as the depth edge feature map; The input end of the second fifth upsampling layer is connected to the output end of the second fourth residual connection and normalization layer, and is used to obtain a feature map having the same resolution as the depth input feature map according to the residual connection output feature map containing the depth edge feature; The input ends of the second-fifth residual connection and the normalization layer are respectively connected to the output ends of the input layer and the second-fifth upsampling layer, for obtaining a depth-dominant branch output feature map according to the ADSS optical fiber cable image after ice coating and a feature map having the same resolution as the depth input feature map; The output ends of the second five residual connections and normalization layers are connected to the GE attention mechanism module.

5. The method for detecting ice coating on ADSS optical fiber cables based on GBRNet network according to claim 2 is characterized in that: The BiPANet feature fusion network includes: a first upsampling layer, a first convolutional layer, a second upsampling layer, a second convolutional layer, a first feature interaction layer, a second feature interaction layer, a third feature interaction layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first residual connection and normalization layer, a second residual connection and normalization layer, and an output layer; The input end of the first upsampling layer is connected to the GENet feature extraction network, and is used to obtain a fused depth output feature map having the same resolution as the depth dominant branch output feature map according to the fused depth output feature map; The input end of the first feature interaction layer is respectively connected to the output end of the first upsampling layer and the GENet feature extraction network, and is used to obtain a shallow fusion feature map according to the depth-dominant branch output feature map and a fused depth output feature map having the same resolution as the depth-dominant branch output feature map; The input end of the first convolutional layer is connected to the output end of the first feature interaction layer, and is used to obtain shallow feature information according to the shallow fusion feature map; The input end of the second upsampling layer is connected to the output end of the first convolutional layer, and is used to obtain a shallow feature map having the same resolution as the color dominant branch output feature map according to the shallow feature information; The input end of the second feature interaction layer is respectively connected to the output end of the second upsampling layer and the GENet feature extraction network, and is used to obtain a deep fusion feature map according to the color-dominant branch output feature map and a shallow feature map having the same resolution as the color-dominant branch output feature map; The input end of the second convolutional layer is connected to the output end of the second feature interaction layer, and is used to obtain deep feature information according to the deep fusion feature map; The input end of the third feature interaction layer is respectively connected to the output end of the second convolutional layer and the GENet feature extraction network, so as to obtain deep color feature information according to the feature map and deep feature information output by the color dominant branch; The input end of the third convolutional layer is connected to the output end of the third feature interaction layer, and is used to obtain a first size object detection head according to the deep color feature information; The input ends of the first residual connection and the normalization layer are respectively connected to the output end of the third convolutional layer and the GENet feature extraction network, so as to obtain a feature map having original depth-dominant branch output feature information according to the depth-dominant branch output feature map and the first-size object detection head; The input end of the fourth convolutional layer is connected to the output end of the first residual connection and the normalization layer, and is used to obtain a second-size object detection head according to the feature map having the original depth-dominant branch output feature information; The input ends of the second residual connection and the normalization layer are respectively connected to the output end of the fourth convolutional layer and the GENet feature extraction network, so as to obtain a feature map having original fusion depth output information according to the fusion depth output feature map and the second size object detection head; The input end of the fifth convolutional layer is connected to the output end of the second residual connection and the normalization layer, and is used to obtain a third-size object detection head according to the feature map with original fusion depth output information; The output layer is used to obtain and output the optical fiber cable image with the ice-covered area identification frame added according to the first size target detection head, the second size target detection head, and the third size target detection head; Among them, the relationship between the first size target detection head, the second size target detection head and the third size target detection head is: The size of the second-size target detection head=the size of the first-size target detection head*2, and the size of the third-size target detection head=the size of the second-size target detection head*2.

6. The method for detecting ice coating on ADSS optical fiber cables based on GBRNet network according to claim 1, characterized in that: The formula used to obtain the ice thickness of ADSS optical fiber cable is as follows: Wherein: y represents the average thickness of ice; D represents the pixel measurement value of the outer diameter after ice coating, that is, the pixel thickness of the optical cable in the ADSS optical fiber cable image after ice coating; d represents the pixel measurement value of the outer diameter without ice coating, that is, the pixel thickness of the optical cable in the ADSS optical fiber cable image before ice coating; h represents the standard diameter of the optical cable.

Images