ADSS optical cable electro-corrosion defect detection method based on dynamic snakelike convolution

Abstract

The invention discloses an ADSS optical cable electro-corrosion defect detection method based on dynamic snakelike convolution, and the method comprises the following steps: 1) replacing a common convolution kernel with an ADown module, completing the down-sampling of different modes through sub-channels to extract features, and reducing the optical cable information loss generated in the down-sampling process; (2) a Context Anchor Attention (CAA) mechanism is introduced, so that the network can better learn the information of the defect of the optical cable in the deep semantic information; and 3) providing a C2fDSC module for extracting the optical cable information of neck fusion features, and helping the network to improve the adaptability of the network to the tubular structure of the optical cable by introducing the DSConv into the C2fDSC feature extraction module.The method provided by the invention can realize the high-precision detection of the electro-corrosion defect of the ADSS optical cable.

Description

Technical Field

[0001] The invention relates to a method for detecting electrical corrosion defects of an optical cable, and in particular to a method for detecting electrical corrosion conditions of an ADSS optical cable and discovering defects in a timely manner based on a dynamic serpentine convolution technology. Background Art

[0002] ADSS (All Dielectric Self Supporting) optical cable is an all-dielectric self-supporting optical cable. It is often used as a fast and economical transmission channel in power communication systems because of its advantages such as no metal, tension resistance, self-support, high insulation, non-inductive, thin diameter, light, easy construction and economy.

[0003] ADSS optical cables are installed near high-voltage conductors. The spatial electric field around the conductors puts the optical cables at a spatial potential. In dirty weather conditions, such as heavy fog, frost or light rain, moist dirt forms a resistance layer on the surface of the optical cable. Under the action of spatial potential, a current is generated between the surface of the optical cable and the optical cable grounding hardware on the iron tower or steel pipe tower, which is called ground leakage current. The heat generated by the current causes the water to evaporate, forming a small section of dry belt, which blocks the current. When the induced voltage on the dry belt is high enough to exceed the air breakdown field strength, discharge occurs to form an arc, that is, a dry belt arc. The arc discharge occurs repeatedly, resulting in the formation of a dendritic carbonization channel on the surface of the cable sheath, that is, an electric trace, which eventually causes electrical corrosion of the optical cable. For this reason, it is necessary to detect the electrical corrosion of the ADSS optical cable.

[0004] Nowadays, conventional detection methods for electrical corrosion of ADSS optical cables mainly rely on manual inspections and sensor monitoring. Sensors are mostly used in the form of signal monitoring using sensing equipment, such as the "Online Detection System for Electrical Corrosion Faults of ADSS Optical Cables" disclosed in Chinese patent CN201811365795.0, which uses leakage current measurement modules, electrostatic field measurement modules, and signal processing modules to detect dry arcs generated by ADSS optical cables, and transmits information to the ground station through the signal processing module according to the strength of the electric field, thereby determining online whether electrical corrosion occurs in the ADSS optical cable. However, this type of method is susceptible to interference from the external environment on the signal data, resulting in high rates of missed detection and false detection.

[0005] The above situation can be avoided by image detection, and the need for maintenance can be directly judged by the manifestation of electrical corrosion in each stage. For example, the "ADSS optical cable line inspection system and inspection method based on deep learning" disclosed in Chinese patent CN202210198099.5 proposes a method based on graphic detection, but the invention method 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 the present invention is to overcome the above-mentioned shortcomings and provide an ADSS optical cable electrical corrosion defect detection method based on dynamic serpentine convolution with high detection accuracy.

[0007] The object of the present invention is achieved in that:

[0008] A method for detecting electrical corrosion defects of ADSS optical cables based on dynamic serpentine convolution comprises the following steps:

[0009] Step 1: Establish an ADSS optical cable electrical corrosion defect dataset. The dataset is divided into a training set, a validation set, and a test set according to the ratio of 7:2:1. The dataset is divided into five labels according to the manifestation of optical cable defects, namely Pitting, Smallpitting, Break, Breakdown, and Tracking. LabelImg is used to manually annotate rectangular boxes to form an ADSS optical cable electrical corrosion defect dataset and send it to the network for training.

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

[0011] Step 2: Between the backbone network and the neck network, the CAA attention mechanism is used to complete the prominent extraction of deep electrical corrosion features; then PAFPN is used to complete the interactive fusion of different feature levels of the backbone network. On the basis of the upward fusion of deep semantic features completed by the feature pyramid network, the strong positioning information of shallow features is fused to the deep layer through the path aggregation network to enhance the positioning information of the features. In addition, the C2f_DSC module is added to the PAFPN, and the improved C2f_DSC module is used. Its specific structure is as follows: In the bottleneck structure Neck, DSC_block is added, and the A set of convolution operations is used to adjust the number of channels of DSC_block; in DSC_block, the input is sent to the ordinary convolution kernel with a convolution kernel size of 3 and the dynamic snake convolution in the X-axis direction and Y-axis direction with a convolution kernel size of 3 for processing, and finally the three different features are spliced ​​and sent to the subsequent process for processing; DSConv is combined with the C2f feature extraction module, and the C2f feature extraction module is used to generate rich gradient information flow, and the characteristics of dynamic snake convolution are used to perform better linear adaptation extraction of optical cable features, improve the accuracy of network detection of optical cable electrical corrosion, and optimize the gradient information flow of the network;

[0012] Step three, the features extracted and fused by the above two networks will be sent to the head network to complete the detection. An advanced decoupled head structure is adopted, and the Anchor-Free strategy is introduced, which abandons the concept of anchor frame, 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 electrical corrosion defect dataset uses ADSS electrical corrosion defect image photos obtained by ADSS optical cable line inspection as the basic dataset; at the same time, since ADSS electrical corrosion defect images are difficult to obtain, resulting in an insufficient number of dataset samples, in order to improve the algorithm accuracy and accelerate network convergence, the dataset is enhanced and expanded by randomly adding Gaussian noise or salt and pepper noise, changing brightness, mirroring, scaling, and random flipping. On this basis, in order to improve the robustness of the algorithm, normal ADSS optical cable pictures are finally added as background for training.

[0014] Preferably, in step one, three groups of strided convolution modules and two ADown downsampling modules are used to complete five downsampling operations on the feature map of size 640×640, and the samples are sampled to 320×320, 160×160, 80×80, 40×40 and 20×20 respectively; the middle and deep features of 40×40 and 20×20 are generated by the ADown module, and the strategy used is: one part of the channels is directly downsampled through the traditional strided convolution kernel, and the other part is downsampled through the maximum pooling operation with a pooling kernel size of 3, a step size of 2, and a padding of 1, and then the features are extracted through a convolution operation with a convolution kernel size of 1, a step size and a padding of 0, and finally the features obtained by the two downsampling parts are spliced ​​in the spatial dimension to complete the entire downsampling process.

[0015] Preferably, the operation flow of the CAA attention mechanism in step 2 is: smooth the feature information by performing an average pooling operation with a pooling kernel size of 7 on the feature map, then extract the features using a 1×1 convolution kernel, and then complete the feature information extraction and connection in the width and height directions through two large-kernel depth-separable convolutions, and finally use a 1×1 convolution kernel to extract the features again and use the Sigmoid function to output the corresponding weight value.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] The present invention can better avoid the occurrence of sensor interference through image detection means, and directly judge whether maintenance is needed through the manifestation of electrical corrosion in each stage. However, there is currently no deep learning image detection method optimized for ADSS optical cable electrical corrosion.

[0018] The present invention adopts image deep learning algorithm to perform defect detection on ADSS optical cable images, which has higher accuracy than traditional sensor-based detection methods. At the same time, it adopts targeted optimized deep learning network algorithm to enhance the detection accuracy of ADSS optical cable electrical corrosion defects.

[0019] At the same time, in order to prevent missed detection and false detection, the present invention introduces a specific downsampling module, an attention mechanism and a targeted convolution kernel to the network. Specifically, compared with the downsampling using traditional convolution, the ADown module uses average pooling to help the backbone network retain more optical cable feature information, and by downsampling the feature map of the two-segment channel connection in different ways, the maximum pooling kernel is used to retain more texture and contour information of the optical cable defects. In summary, the ADown downsampling module can make up for the information loss problem of the traditional strided convolution in the downsampling process, and retain more optical cable defect feature information. Adding an attention mechanism can solve the problem of overly complex background information of the optical cable image, and effectively improve the sensitivity of the network to the electrical corrosion defects of the optical cable. Since optical cable defects are often local small area defects in the image, whether the network can effectively learn its accurate position in space will affect the detection accuracy, so the present invention introduces a context anchor attention mechanism to solve this problem. At the same time, by introducing the concept of offset in the convolution kernel, a linear constraint is made on the direction and position generated by the offset, so that the convolution kernel is more in line with the characteristics of the linear continuity structure. This feature can improve the accuracy and robustness of the network in detecting electrical corrosion defects in optical cables. Dynamic snake 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. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0022] Figure 3 This is a CAA structure diagram in the present invention.

[0023] Figure 4 This is the C2f_DSC structure diagram of the present invention.

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

[0025] Figure 6 This is a diagram of the experimental environment configuration of an embodiment of the present invention. DETAILED DESCRIPTION

[0026] See also Figures 1 to 6 The present invention relates to a method for detecting electrical corrosion defects of ADSS optical cables based on dynamic serpentine convolution, and its overall structure is as follows Figure 1 shown.

[0027] Step 1, establish a data set containing ADSS electrical corrosion defects. In this embodiment, the inspection defect photos of 21 ADSS optical cable lines completed by Jiangyin Power Supply Bureau in the second quarter of 2022 are used as the basic data set, with a total of 1025 pictures; due to the difficulty in obtaining optical cable defect images, the number of data set samples is insufficient. Therefore, in order to improve the algorithm accuracy and accelerate network convergence, this paper adopts traditional data enhancement methods such as randomly adding Gaussian noise or salt and pepper noise, changing brightness, mirroring, scaling, and random flipping to enhance and expand the data set; on this basis, in order to improve the robustness of the algorithm, 257 normal ADSS optical cable pictures are finally added as background for training; after the combination, there are a total of 4357 pictures, and the data set is divided into training set, verification set, and test set according to 7:2:1.

[0028] The dataset is divided into five labels according to the manifestation of optical cable defects, namely Pitting, Smallpitting, Break, Breakdown and Tracking. LabelImg is used to manually annotate rectangular boxes, and finally the ADSS optical cable electrical corrosion defect dataset is formed, which is then sent to the network for training.

[0029] First, in the backbone network, the improved CSPDarkNet is used in this embodiment to complete the feature downsampling extraction of the ADSS electrocorrosion defect image, which includes 3 convolutional-batch normalization-activation layers, 4 C2f feature extraction modules, 1 SPPF feature pooling aggregation module, and 2 ADown downsampling modules. The specific structure of ADown is as follows: Figure 2As shown. In this embodiment, three groups of strided convolution modules and two ADown downsampling modules are used to complete five downsampling operations on the feature map of size 640×640, and the samples are sampled to 320×320, 160×160, 80×80, 40×40 and 20×20 respectively. The middle and deep features of 40×40 and 20×20 are generated by the ADown module. The strategy used is: one part of the channels is directly downsampled through the traditional strided convolution kernel, and the other part is downsampled through the maximum pooling operation with a pooling kernel size of 3, a stride of 2, and a padding of 1. Then, a convolution operation with a convolution kernel size of 1, a stride and padding of 0 is used to extract features. Finally, the features obtained by the two parts of downsampling are spliced ​​in the spatial dimension to complete the entire downsampling process. Compared with downsampling using traditional convolution, the ADown module uses average pooling to help the backbone network retain more optical cable feature information, and downsamples the feature maps of the two channel connections in different ways, using the maximum pooling kernel to retain more texture and contour information of optical cable defects. In summary, the ADown downsampling module can make up for the information loss problem of traditional strided convolution during the downsampling process and retain more optical cable defect feature information.

[0030] After downsampling, four C2f feature extraction modules are used to perform further feature extraction, enhance the quality of network gradient flow information, and enhance feature expression. Then, SPPF is used at the deepest feature layer to complete the pooling fusion of deep features and enhance feature expression.

[0031] Step 2: In this embodiment, the CAA attention mechanism is used between the backbone network and the neck network to complete the prominent extraction of deep electrical corrosion features. CAA, namely the context anchor attention mechanism (Context Anchor Attention), is as follows: Figure 3 The figure shows the specific structure of the CAA attention mechanism. The feature information is smoothed by performing an average pooling operation with a pooling kernel size of 7 on the feature map, and then a 1×1 convolution kernel is used to extract features. Subsequently, two large-kernel depth-separable convolutions are used to complete the feature information extraction and connection in the width and height directions. Finally, a 1×1 convolution kernel is used to extract features again and the corresponding weight value is output using the Sigmoid function.

[0032] Then, this embodiment uses PAFPN to complete the interactive fusion of different feature levels of the backbone network. On the basis of the feature pyramid network completing the upward fusion of deep semantic features, the path aggregation network is used to fuse the strong positioning information of shallow features to the deep layer to enhance the positioning information of features. In addition, the C2f_DSC module is added to PAFPN. In formulas (1) and (2), the output of the specific position of each weight grid of the X and Y axes is respectively represented. 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. i±c The selection of is a cumulative process. From the center position K i Initially, the position of the grid away from the center depends on the position of the previous grid: K i+1 Relative to K i Added offset Δ.

[0033]

[0034] like Figure 4 The specific structure of the improved C2f_DSC is shown in Figure 1. In the bottleneck structure Neck, DSC_block is added, and a set of convolution operations is added to adjust the number of channels of DSC_block. In DSC_block, the input is sent to the ordinary convolution kernel with a convolution kernel size of 3 and the dynamic snake convolution in the X-axis direction and Y-axis direction with a convolution kernel size of 3 for processing. Finally, the three different features are spliced ​​and sent to the subsequent process for processing. By combining DSConv with the C2f feature extraction module, the C2f feature extraction module can be used to generate rich gradient information flow, and the characteristics of dynamic snake convolution can be used to extract better linear adaptation of optical cable features, improve the accuracy of network detection of optical cable electrical corrosion, and optimize the gradient information flow of the network. The C2f_DSC module is used for the extraction of optical cable information of the neck fusion feature, and the dynamic snake convolution (DSConv) is introduced in the C2f feature extraction module to help the network improve its adaptability to the tubular structure of the optical cable.

[0035] Step 3: After the features extracted and fused by the above two networks, they are sent to the head network to complete the detection. The advanced decoupled head structure is adopted, and the Anchor-Free strategy is introduced, which abandons the concept of anchor frame and simplifies the way of matching positive and negative samples, thereby further improving the network's detection effect on ADSS optical cable electrical corrosion defects. Figure 5 shown.

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

[0037] In addition: It should be noted that the above specific implementation is only an optimization scheme of this patent. Any changes or improvements made by technicians in this field based on the above concept are within the scope of protection of this patent.

Claims

1. A method for detecting electrical corrosion defects of ADSS optical cables based on dynamic serpentine convolution, characterized in that: The following steps are involved: Step 1: Establish an ADSS optical cable electrical corrosion defect dataset. The dataset is divided into a training set, a validation set, and a test set according to the ratio of 7:2:

1. The dataset is divided into five labels according to the manifestation of optical cable defects, namely Pitting, Smallpitting, Break, Breakdown, and Tracking. LabelImg is used to manually annotate rectangular boxes to form an ADSS optical cable electrical corrosion defect dataset and send it to the network for training. In the backbone network, CSPDarkNet is used to complete the feature sampling and extraction of ADSS electrical corrosion defect images in the ADSS optical cable electrical corrosion defect dataset, which includes 3 convolution-batch normalization-activation layers, 4 C2f feature extraction modules, 1 SPPF feature pooling aggregation module and 2 ADown downsampling modules. After the downsampling is completed, the 4 C2f feature extraction modules are used to complete further feature extraction, enhance the quality of network gradient flow information, and enhance feature expression. Subsequently, SPPF is used at the deepest feature layer to complete the pooling fusion of deep features and enhance feature expression. Step 2: Between the backbone network and the neck network, the CAA attention mechanism is used to highlight the extraction of deep electrical corrosion features; then PAFPN is used to complete the interactive fusion of different feature levels of the backbone network. The specific process of PAFPN is as follows: on the basis of the upward fusion of deep semantic features completed by the feature pyramid network, the strong positioning information of shallow features is integrated into the deep layer through the path aggregation network to enhance the positioning information of the features; then the improved C2f_DSC module is added to PAFPN. The specific structure of the C2f_DSC module is as follows: in the bottleneck structure Neck, DSC_block is added, and a set of convolution operations is added to adjust the number of channels of DSC_block; in DSC_block, the input is sent to the ordinary convolution kernel with a convolution kernel size of 3 and the dynamic snake convolution in the X-axis and Y-axis directions with a convolution kernel size of 3 for processing, and finally the three different features are spliced ​​and sent to the subsequent process for processing; Step 3: After the features extracted and fused by the above two networks, they are sent to the head network to complete the detection. The advanced decoupled head structure is adopted, and the Anchor-Free strategy is introduced to abandon the concept of anchor frame.

2. According to claim 1, a method for detecting electrical corrosion defects of ADSS optical cables based on dynamic serpentine convolution is characterized in that: In step 1, the ADSS optical cable electrical corrosion defect dataset uses ADSS electrical corrosion defect image photos obtained by ADSS optical cable line inspection as the basic dataset; at the same time, randomly adding Gaussian noise or salt and pepper noise (Salt and Pepper Noise is a common type of noise in digital image processing. It is named after the black and white spots on the image, which look like rough pepper and salt. Salt and pepper noise is actually black and white light and dark point noise generated by image sensors, transmission channels, decoding processing, etc. Specifically, salt and pepper noise may appear as black pixels in bright areas, that is, "salt" noise, or white pixels in dark areas, that is, "pepper" noise, or even both at the same time), changing brightness, mirroring, scaling, and random flipping to enhance and expand the dataset. On this basis, in order to improve the robustness of the algorithm, normal ADSS optical cable pictures are finally added as background for training.

3. According to claim 1, a method for detecting electrical corrosion defects of ADSS optical cables based on dynamic serpentine convolution is characterized in that: In step 1, three groups of strided convolution modules and two ADown downsampling modules are used to complete five downsampling operations on the feature map of size 640×640, sampling to 320×320, 160×160, 80×80, 40×40 and 20×20 respectively; The 40×40 and 20×20 middle and deep features are generated by the ADown module. The strategy used is: one part of the channels is directly downsampled through the traditional strided convolution kernel, and the other part is downsampled through the maximum pooling operation with a pooling kernel size of 3, a stride of 2, and a padding of 1. Then, features are extracted through a convolution operation with a convolution kernel size of 1, a stride and padding of 0. Finally, the features obtained from the two downsamplings are spliced ​​in the spatial dimension to complete the entire downsampling process.

4. According to claim 1, a method for detecting electrical corrosion defects of ADSS optical cables based on dynamic serpentine convolution is characterized in that: The specific operation process of the CAA attention mechanism in step 2 is as follows: smooth the feature information by performing an average pooling operation with a pooling kernel size of 7 on the feature map, then extract the features using a 1×1 convolution kernel, and then complete the feature information extraction and connection in the width and height directions through two large-kernel depth-separable convolutions, and finally use a 1×1 convolution kernel to extract the features again and use the Sigmoid function to output the corresponding weight values.

5. According to claim 1, a method for detecting electrical corrosion defects of ADSS optical cables based on dynamic serpentine convolution is characterized in that: In step 2, the specific position outputs of each weight grid in the X and Y axes of the DSConv convolution kernel introduced in the C2f_DSC module are expressed by formulas (1) and (2), respectively. Taking the X axis direction as an example, the specific position of each grid in the convolution kernel K is expressed as: K i±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; each grid position K in the convolution kernel K i±c The selection of is a cumulative process; from the central position K i Initially, the position of the grid away from the center depends on the position of the previous grid: K i+1 Relative to K i Added offset Δ;

Images