An ADSS optical cable electric corrosion defect detection method based on dynamic serpentine convolution

Abstract

The application discloses an ADSS optical cable electric corrosion defect detection method based on a dynamic snake convolution, which comprises the following steps: 1) replacing a common convolution kernel with an ADOWN module, extracting features by completing different mode down-sampling through a split channel, and reducing optical cable information loss generated in the down-sampling process; 2) introducing a context anchor attention mechanism (CAA), so that the network can better learn information of an optical cable defect in deep semantic information; and 3) proposing a C2f_DSC module for neck fusion feature optical cable information extraction, and introducing a dynamic snake convolution (DSConv) in the C2f feature extraction module, so as to help the network improve adaptability to a tubular structure of the optical cable. Through the method, high-precision detection of ADSS optical cable electric corrosion defects can be realized.

Description

Technical Field

[0001] This invention relates to a method for detecting electrical corrosion defects in optical cables, and more particularly to a method for detecting the electrical corrosion of ADSS optical cables and timely identifying defects based on dynamic serpentine convolution technology. Background Technology

[0002] ADSS (All Dielectric Self Supporting) optical cable is a type of all-dielectric self-supporting optical cable. Due to its advantages such as being metal-free, tensile-resistant, self-supporting, highly insulated, non-inductive, thin-diameter, lightweight, easy to install, and economical, it is often used as a fast and economical transmission channel in power communication systems.

[0003] ADSS optical cables are laid near high-voltage power lines, and the electric field around the power lines causes the cable to be at a space potential. In polluted weather conditions, such as heavy fog, frost, or light rain, the damp contaminants form a resistive layer on the cable surface. Under the influence of this space potential, a current, known as grounding leakage current, is generated between the cable surface and the grounding hardware on the tower or steel pipe tower. This current generates heat, causing moisture to evaporate and forming short dry bands that block the current. When the induced voltage on these dry bands is high enough to exceed the air breakdown field strength, a discharge occurs, forming an electric arc, known as a dry band arc. Repeated arcing leads to the formation of dendritic carbonized channels on the cable sheath surface, known as electrical tracking, ultimately causing electro-corrosion of the cable. Therefore, it is necessary to inspect the electro-corrosion status of ADSS optical cables.

[0004] Currently, conventional methods for detecting electro-corrosion in ADSS optical cables mainly rely on manual inspection and sensor monitoring. Sensors are primarily used for signal monitoring, as exemplified by the "Online Detection System for Electro-corrosion Faults in ADSS Optical Cables" disclosed in Chinese Patent CN201811365795.0. This system uses a leakage current measurement module, an electrostatic field measurement module, and a signal processing module to detect the dry-charge arc generated in the ADSS optical cable. Based on the electric field strength, the signal processing module transmits the information to a ground station, thereby determining online whether electro-corrosion has occurred in the ADSS optical cable. However, this type of method is susceptible to interference from the external environment, leading to high rates of missed and false detections.

[0005] Image detection methods can effectively avoid the above-mentioned situations, directly determining whether maintenance is needed by observing the manifestations of electro-corrosion at each stage. For example, Chinese patent CN202210198099.5 discloses "An ADSS Optical Cable Line Inspection System and Inspection Method Based on Deep Learning," which proposes an image-based detection method. However, this invention does not propose a targeted deep learning algorithm for the characteristics of ADSS optical cables, resulting in insufficient inspection accuracy. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a high-precision ADSS optical cable electrical corrosion defect detection method based on dynamic serpentine convolution.

[0007] The objective of this invention is achieved as follows:

[0008] A method for detecting electrical corrosion defects in optical cables based on dynamic serpentine convolution (ADSS) includes the following steps:

[0009] Step 1: Establish the ADSS optical cable electrical corrosion defect dataset. The dataset is divided into training set, validation set and test set in a 7:2:1 ratio. The dataset is divided into five labels according to the manifestation of optical cable defects: Pitting, Smallpitting, Break, Breakdown and Tracking. Rectangular boxes are manually labeled using LabelImg to form the ADSS optical cable electrical corrosion defect dataset, which is then fed into the network for training.

[0010] In the backbone network, CSPDarkNet is used to extract features from the ADSS optical cable electrical corrosion defect images in the ADSS optical cable electrical corrosion defect dataset. It includes 3 convolution-batch normalization-activation layers, 4 C2f feature extraction modules, 1 SPPF feature pooling and aggregation module, and 2 ADown downsampling modules. After downsampling, the 4 C2f feature extraction modules are used to perform further feature extraction, which enhances the quality of the network gradient flow information and strengthens feature representation. Then, SPPF is used to perform pooling and fusion of deep features at the deepest feature layer to enhance feature representation.

[0011] Step two: A CAA attention mechanism is used between the backbone network and the neck network to extract prominent features of deep electro-corrosion. Subsequently, PAFPN is used to perform interactive fusion of different feature levels in the backbone network. Building upon the upward fusion of deep semantic features by the feature pyramid network, a path aggregation network is used to fuse strong localization information of shallow features to deeper layers, enhancing the localization information of the features. Furthermore, a C2f_DSC module is added to PAFPN, and an improved C2f_DSC module is adopted. Its specific structure is as follows: A DSC_block is added to the bottleneck structure (Neck), and... A set of convolution operations is used to adjust the number of channels in the DSC_block. Within the DSC_block, the input is processed through a regular convolution kernel of size 3 and dynamic serpentine convolutions of size 3 in the X and Y axes. Finally, the three different features are concatenated and fed into subsequent processing steps. By combining DSConv with the C2f feature extraction module, and utilizing the rich gradient information flow generated by the C2f module, along with the characteristics of dynamic serpentine convolution, better linear adaptation extraction of optical cable features is achieved, improving the accuracy of the network in detecting optical cable electrical corrosion and optimizing the network's gradient information flow.

[0012] Step 3: The features extracted and fused by the two networks above will be fed into the head network to complete the detection. An advanced decoupled head structure is adopted and an Anchor-Free strategy is introduced, which abandons the concept of anchor boxes, simplifies the way of matching positive and negative samples, and improves the network's detection effect on ADSS optical cable electrical corrosion defects.

[0013] Preferably, in step one, the ADSS optical cable electro-corrosion defect dataset uses ADSS optical cable line inspection images of ADSS electro-corrosion defects as the basic dataset. Meanwhile, since ADSS electro-corrosion defect images are difficult to obtain, resulting in insufficient dataset samples, in order to improve algorithm accuracy and accelerate network convergence, random addition of Gaussian noise or salt-and-pepper noise, changes in brightness, mirroring, scaling, and random flipping are used to enhance and expand the dataset. On this basis, to improve the robustness of the algorithm, normal ADSS optical cable images are finally added as background for training.

[0014] Preferably, in step one, three sets of strided convolutional modules and two ADown downsampling modules are used to perform five downsampling operations on the 640×640 feature map, sampling to 320×320, 160×160, 80×80, 40×40 and 20×20 respectively. Among them, the mid-level and deep features of 40×40 and 20×20 are generated by the ADown module. The strategy used is as follows: one part of the channels is downsampled directly through the traditional strided convolutional kernel, and the other part is downsampled through max pooling with a kernel size of 3, a stride of 2 and padding of 1. Then, features are extracted through a convolutional operation with a kernel size of 1, a stride and padding of 0. Finally, the features obtained from the two downsampling parts are concatenated in the spatial dimension to complete the entire downsampling process.

[0015] Preferably, the operation flow of the CAA attention mechanism in step two is as follows: smooth the feature information by performing an average pooling operation with a kernel size of 7 on the feature map, then extract the features using a 1×1 convolution kernel, then complete the extraction and connection of feature information in the width and height directions through two large kernel depthwise separable convolutions, and finally extract the features again using a 1×1 convolution kernel and output the corresponding weight values ​​using the Sigmoid function.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] This invention can effectively avoid sensor interference through image detection. It can directly determine whether maintenance is needed by observing the manifestation of electro-corrosion at each stage. Currently, there is no deep learning image detection method optimized for the electro-corrosion of ADSS optical cables.

[0018] This invention employs an image deep learning algorithm to detect defects in ADSS optical cable images, achieving higher accuracy compared to traditional sensor-based detection methods. Furthermore, it utilizes a specially optimized deep learning network algorithm to enhance the detection accuracy of electrical corrosion defects in ADSS optical cables.

[0019] Meanwhile, to prevent missed detections and false detections, this invention introduces a specific downsampling module, attention mechanism, and targeted convolutional kernels into the network. Specifically, compared to using traditional convolution for downsampling, the ADown module uses average pooling to help the backbone network retain more fiber optic cable feature information. Furthermore, by downsampling the feature maps connecting two channels in different ways, it uses max pooling kernels to retain more texture and contour information of fiber optic cable defects. In summary, the ADown downsampling module can compensate for the information loss problem in the downsampling process of traditional strided convolution, retaining more feature information of fiber optic cable defects. The addition of an attention mechanism can solve the problem of overly complex background information in fiber optic cable images, effectively improving the network's sensitivity to fiber optic cable electrical corrosion defects. Since fiber optic cable defects are often localized and small-area defects in images, the network's ability to effectively learn their accurate spatial location affects detection accuracy. Therefore, this invention introduces a context anchor attention mechanism to address this issue. Simultaneously, by introducing the concept of offset into the convolutional kernel, the direction and position of the offset are constrained to be linear, making the convolutional kernel more consistent with the characteristics of a linear continuous structure. This characteristic improves the accuracy and robustness of the network in detecting electrical corrosion defects in optical cables. Dynamic serpentine convolution linearizes each element of the traditional convolution kernel in the X and Y directions by learning the offset, and gradually adapts to the tubular structure of the ADSS optical cable during the learning process. Attached Figure Description

[0020] Figure 1 This is a diagram of the overall network structure of the improved model of the present invention.

[0021] Figure 2 This is a structural diagram of the Adown module in this invention.

[0022] Figure 3 This is a structural diagram of the CAA in this invention.

[0023] Figure 4 This is a diagram of the C2f_DSC structure in this invention.

[0024] Figure 5 This is a comparison chart of the optical cable defect detection effects in this invention.

[0025] Figure 6 This is a diagram showing the experimental environment configuration of an embodiment of the present invention. Detailed Implementation

[0026] See Figures 1-6 This invention relates to an ADSS optical cable electrical corrosion defect detection method based on dynamic serpentine convolution, the overall structure of which is as follows: Figure 1 As shown.

[0027] Step 1: Establish a dataset containing ADSS electrical corrosion defects. In this embodiment, the defect photos of 21 ADSS optical cable lines inspected by Jiangyin Power Supply Bureau in the second quarter of 2022 are used as the basic dataset, totaling 1025 images. Due to the difficulty in obtaining optical cable defect images, resulting in insufficient dataset samples, this paper uses traditional data augmentation methods such as randomly adding Gaussian noise or salt-and-pepper noise, changing brightness, mirroring, scaling, and random flipping to enhance and expand the dataset in order to improve algorithm accuracy and accelerate network convergence. On this basis, to improve the robustness of the algorithm, 257 normal ADSS optical cable images are added as background for training. The final dataset contains 4357 images, which are divided into training, validation, and test sets in a 7:2:1 ratio.

[0028] The dataset is divided into five labels based on the manifestation of optical cable defects: Pitting, Smallpitting, Break, Breakdown, and Tracking. Rectangular bounding boxes were manually labeled using LabelImg to create the ADSS optical cable electrical corrosion defect dataset, which was then fed into the network for training.

[0029] Firstly, in the backbone network, this embodiment uses an improved CSPDarkNet to perform feature downsampling extraction for ADSS electrical corrosion defect images. This includes three convolutional-batch normalization-activation layers, four C2f feature extraction modules, one SPPF feature pooling aggregation module, and two ADown downsampling modules. The specific structure of ADown is as follows... Figure 2As shown in the figure. In this embodiment, three sets of strided convolutional modules and two ADown downsampling modules are used to complete five downsampling operations on a 640×640 feature map, sampling to 320×320, 160×160, 80×80, 40×40, and 20×20 respectively. The mid- and deep features of 40×40 and 20×20 are generated by the ADown module. The strategy is as follows: one part of the channels is downsampled directly through a traditional strided convolutional kernel, and the other part is downsampled through a max pooling operation with a kernel size of 3, a stride of 2, and padding of 1. Then, a convolution operation with a kernel size of 1, a stride, and padding of 0 is used to extract features. Finally, the features obtained from the two downsampled parts are concatenated in the spatial dimension to complete the entire downsampling process. Compared to using traditional convolutions for downsampling, the ADown module utilizes average pooling to help the backbone network retain more fiber optic cable feature information. Furthermore, by applying different downsampling methods to the feature maps connecting the two channels, it uses max pooling kernels to preserve more texture and contour information of fiber optic cable defects. In summary, the ADown downsampling module can compensate for the information loss problem in the downsampling process of traditional strided convolutions, preserving more feature information about fiber optic cable defects.

[0030] After downsampling, four C2f feature extraction modules are used to further extract features, enhancing the quality of the network gradient flow information and improving feature representation. Subsequently, SPPF is used at the deepest feature layer to perform pooling and fusion of deep features, further enhancing feature representation.

[0031] Step two: In this embodiment, a CAA attention mechanism is used between the backbone network and the neck network to extract prominent features of deep electro-erosion. CAA stands for Context Anchor Attention. Figure 3 The diagram shows the specific structure of the CAA attention mechanism. It smooths the feature information by performing an average pooling operation with a kernel size of 7 on the feature map. Then, it extracts features using a 1×1 convolution kernel. Subsequently, it extracts and connects feature information in the width and height directions using two large-kernel depthwise separable convolutions. Finally, it extracts features again using a 1×1 convolution kernel and outputs the corresponding weight values ​​using the Sigmoid function.

[0032] Subsequently, this embodiment uses PAFPN to complete the interactive fusion of different feature levels of the backbone network. Based on the feature pyramid network's upward fusion of deep semantic features, the path aggregation network fuses the strong localization information of shallow features into deeper layers, enhancing the feature localization information. Furthermore, a C2f_DSC module is added to PAFPN. In formulas (1) and (2), the specific positions of each weight grid along the X and Y axes are represented, respectively. Taking the X-axis direction as an example, the specific position of each grid in the convolution kernel K is represented as: Ki±c =(x i±c ,y i±c ), where c represents the horizontal distance from the center grid; i represents the i-th grid in the X-axis direction. The position of each grid in the convolution kernel K is K. i±c The selection is a cumulative process. From the central position K... i Initially, the position furthest from the center grid depends on the position of the previous grid: K i+1 Relative to K i An offset Δ has been added.

[0033]

[0034] like Figure 4 The improved C2f_DSC structure incorporates a DSC_block within the neck bottleneck structure, with an additional set of convolutional operations to adjust the number of channels. Within the DSC_block, the input is processed through a standard convolutional kernel of size 3 and dynamic snake-like convolutions of size 3 in both the X and Y axes. Finally, the three distinct features are concatenated and fed into subsequent processing steps. By combining DSConv with the C2f feature extraction module, the rich gradient information flow generated by C2f is utilized, along with the characteristics of dynamic snake-like convolution, to achieve better linear adaptation in the extraction of optical cable features, improving the network's accuracy in detecting optical cable electrical corrosion and optimizing the network's gradient information flow. The C2f_DSC module is used for extracting optical cable information from the neck fusion features. Introducing dynamic snake-like convolution (DSConv) into the C2f feature extraction module enhances the network's adaptability to the tubular structure of optical cables.

[0035] Step three: The features extracted and fused by the two networks above are fed into the head network for detection. An advanced decoupled head structure is adopted, and an anchor-free strategy is introduced, eliminating the concept of anchor boxes and simplifying the matching of positive and negative samples, thereby further improving the network's detection performance for ADSS optical cable electrical corrosion defects. Detection performance comparison: Figure 5 As shown.

[0036] The experimental environment in this embodiment is as follows: Figure 6 As shown: The system is Ubuntu 22.04, the graphics card is NVIDIA RTX 3090, and the deep learning framework is PyTorch 2.0.1. The accuracy of the model proposed in this embodiment is higher than that of any mainstream model.

[0037] Additionally, it should be noted that the above-described specific implementation is merely an optimized solution of this patent, and any modifications or improvements made by those skilled in the art based on the above concept are within the scope of protection of this patent.

Claims

1. A method for detecting electrical corrosion defects in optical cables based on dynamic serpentine convolution (ADSS), characterized in that: Includes the following steps: Step 1: Establish the ADSS optical cable electrical corrosion defect dataset. The dataset is divided into training set, validation set and test set in a 7:2:1 ratio. The dataset is divided into five labels according to the manifestation of optical cable defects: Pitting, Smallpitting, Break, Breakdown and Tracking. Rectangular boxes are manually labeled using LabelImg to form the ADSS optical cable electrical corrosion defect dataset, which is then fed into the network for training. In the backbone network, CSPDarkNet is used to extract features from the ADSS optical cable electrical corrosion defect images in the ADSS optical cable electrical corrosion defect dataset. It includes 3 convolution-batch normalization-activation layers, 4 C2f feature extraction modules, 1 SPPF feature pooling and aggregation module, and 2 ADown downsampling modules. After downsampling, the 4 C2f feature extraction modules are used to perform further feature extraction, which enhances the quality of the network gradient flow information and strengthens feature representation. Then, SPPF is used to perform pooling and fusion of deep features at the deepest feature layer to enhance feature representation. Step 2: A CAA attention mechanism is used between the backbone network and the neck network to prominently extract deep electro-erosion features. Then, PAFPN is used to complete the interactive fusion of different feature levels in the backbone network. The specific process of PAFPN is as follows: based on the upward fusion of deep semantic features completed by the feature pyramid network, the path aggregation network fuses the strong localization information of shallow features into deeper layers, enhancing the localization information of the features. Subsequently, an improved C2f_DSC module is added to PAFPN. The specific structure of the C2f_DSC module is as follows: a DSC_block is added to the bottleneck structure (Neck), and a set of convolution operations is added to adjust the number of channels in the DSC_block. In the DSC_block, the input is fed into a regular convolutional kernel with a kernel size of 3, as well as dynamic serpentine convolutions with kernel sizes of 3 in the X-axis and Y-axis directions for processing. Finally, the three different features are concatenated and fed into subsequent processes. Step 3: The features extracted and fused by the feature pyramid network and path aggregation network are fed into the head network to complete the detection. A decoupled head structure is adopted and an Anchor-Free strategy is introduced.

2. The ADSS optical cable electrical corrosion defect detection method based on dynamic serpentine convolution according to claim 1, characterized in that: In step one, the ADSS optical cable electro-corrosion defect dataset uses ADSS optical cable line inspection images of ADSS electro-corrosion defects as the base dataset. At the same time, the dataset is enhanced and expanded by randomly adding Gaussian noise or salt and pepper noise, changing brightness, mirroring, scaling, and random flipping. Finally, normal ADSS optical cable images are added as background for training.

3. The ADSS optical cable electrical corrosion defect detection method based on dynamic serpentine convolution according to claim 1, characterized in that: In step one, three sets of strided convolutional modules and two ADown downsampling modules are used to complete five downsampling operations on the 640×640 feature map, which are sampled to 320×320, 160×160, 80×80, 40×40 and 20×20 respectively. The 40×40 and 20×20 mid- and deep features are generated by the ADown module. The strategy is as follows: one part of the channels is downsampled directly through a traditional strided convolution kernel, and the other part is downsampled through max pooling with a kernel size of 3, a stride of 2, and padding of 1. Then, features are extracted through a convolution operation with a kernel size of 1, a stride, and padding of 0. Finally, the features obtained from the two downsampled parts are concatenated in the spatial dimension to complete the entire downsampling process.

4. The ADSS optical cable electrical corrosion defect detection method based on dynamic serpentine convolution according to claim 1, characterized in that: The specific operation process of the CAA attention mechanism in step two is as follows: smooth the feature information by performing an average pooling operation with a kernel size of 7 on the feature map, then extract the features using a 1×1 convolution kernel, then complete the extraction and connection of feature information in the width and height directions through two large kernel depthwise separable convolutions, and finally extract the features again using a 1×1 convolution kernel and output the corresponding weight values ​​using the Sigmoid function.

5. The ADSS optical cable electrical corrosion defect detection method based on dynamic serpentine convolution according to claim 1, characterized in that: In step two, the specific positions of the weight grids along the X and Y axes in the DSConv convolution kernel introduced in the C2f_DSC module are represented by formulas (1) and (2) respectively. Taking the X-axis direction as an example, the convolution kernel... The specific location of each grid cell is represented as follows: ,in Indicates the horizontal distance from the center grid; Represents the first in the X-axis direction One grid; convolution kernel Each grid position The selection process is cumulative; starting from the central position Initially, the position away from the center grid depends on the position of the previous grid: Compared to Increase offset ; 。

Images