Communication optical cable surface defect detection method based on adaptive feature extraction

Abstract

According to the communication optical cable surface defect detection method based on adaptive feature extraction, a common convolution kernel in a backbone network is replaced by a deformable convolution DCNv3, and the shape of the optical cable surface defect is adaptively learned through a deformable convolution C2fDCN module, so that defect feature information of the optical cable is better extracted. Through the C2fDCN module, the optical cable information of the neck fusion feature is extracted, and the neck network can be helped to better extract the optical cable fusion feature. Meanwhile, a large-scale full-dimension fusion attention LAF-Block attention module is introduced, and important features of deep features of the network are re-extracted, so that background features are diluted or ignored, and the detection precision of the model is effectively improved.

Description

Technical Field

[0001] The present invention relates to a method for detecting surface defects of a communication optical cable, and in particular to a method for detecting surface defects of a communication optical cable by image processing and based on adaptive feature extraction. Background Art

[0002] Power communication optical cables (ADSS, All Dielectric Self Supporting) are installed near high-voltage conductors. The spatial electric field around the high-voltage conductors puts the optical cables in a spatial potential position. In dirty weather conditions (such as heavy fog, frost, etc.), moist dirt will form a resistance layer on the surface of the optical cable. Under the action of spatial potential, a ground leakage current is generated between the surface of the optical cable and the optical cable grounding hardware on the iron tower or steel pipe tower. This ground leakage current generates heat and causes water to evaporate, forming a small section of dry belt, blocking 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, thereby causing electrical corrosion of the optical cable. For this reason, it is necessary to detect the electrical corrosion of power communication optical cables.

[0003] 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.

[0004] 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 Chinese patent CN202210198099.5 discloses a method based on graphic detection, but the invention method does not propose a targeted deep learning algorithm for the characteristics of power communication optical cables, resulting in insufficient detection accuracy. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a method for detecting surface defects of communication optical cables based on adaptive feature extraction, which realizes the identification of surface defects of optical cables based on image processing and realizes adaptive detection through a deformable convolution module. Therefore, compared with manual methods or sensor methods, the detection is more convenient and reliable, and compared with conventional graphic detection, the detection accuracy is higher.

[0006] The object of the present invention is achieved in that: A method for detecting surface defects of communication optical cables based on adaptive feature extraction comprises the following steps: Step 1: Establish a dataset of electrical corrosion defects in communication optical cable lines. Import the basic dataset of defect photos found during the inspection of communication optical cable lines, and annotate the communication optical cable lines with five labels: Pitting, Smallpitting, Break, Breakdown, and Tracking according to the electrical corrosion defects. Then, enhance and expand the dataset through data enhancement. Then, import multiple photos of normal communication optical cable lines collected from different angles or remotely into the dataset. Finally, divide the dataset into training set, validation set, and test set according to the ratio of 7:2:1. Step 2: In the backbone network, the features of the electrical corrosion defect images in the dataset are adaptively downsampled and extracted based on the deformable convolution module DCNv3; Step 3: After sampling in step 2, between the backbone network and the neck network, a LAF-Block module is used to perform attention enhancement extraction to strengthen the information expression of the deep network, and the feature information is smoothed by performing an average pooling operation with a pooling kernel size of 7 on the feature map. Then, a 1×1 convolution kernel is used to extract features, and then 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; Step 4: After the features extracted and fused by the backbone network and the neck network in step 3, they will be sent to the head network to complete the detection.

[0007] Preferably, the specific process of step 2 is: first use 3 groups of strided convolution modules and 2 DCNv3 downsampling modules to complete 5 downsampling operations on the feature map of size 640×640 in the data set, and sample it to 320×320, 160×160, 80×80, 40×40 and 20×20 respectively; among which the middle and deep features of 40×40 and 20×20 are generated by 2 DCNv3; then use SPPF at the deepest feature layer to complete the pooling fusion of deep features to enhance feature expression.

[0008] Preferably, the process of DCNv3 in step 2 is as follows: DCNv3 combines the characteristics of deep separable convolution to divide the convolution weights into depth-wise and point-wise parts; secondly, through the multi-head attention mechanism, the spatial aggregation process is split into G groups, so that each group has an independent sampling offset and modulation factor, so that different groups on each convolution layer have different spatial aggregation modes to enhance the feature expression ability of deformable convolution; finally, Softmax normalization along the sample points is used to constrain the sum of the modulation factors between 0 and 1 to enhance the network stability.

[0009] Preferably, the process of the attention LAF-Block module in step three is: divide the deep feature P5 in the backbone network into two paths, and extract important features of the deep feature P5 through large separable convolutional attention and triple attention respectively.

[0010] Preferably, the convolution operation adopts a deformable convolution module C2f_DCN module, and its specific structure is: in the bottleneck structure Neck, a DCN_block is added, and a group of convolution operations are added to adjust the number of channels of the DCN_block.

[0011] Preferably, the detection in step 4 adopts a decoupled head structure and introduces an Anchor-Free strategy.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention realizes the detection of surface defects of communication optical cables based on image analysis. Compared with conventional manual detection or sensor detection, it is not easily interfered by external environmental factors, thereby effectively avoiding missed detection (i.e., failure to detect actual electrical corrosion) and false detection (i.e., misjudging non-electrical corrosion phenomena as electrical corrosion).

[0013] At the same time, the algorithm of the present invention detects the images of electrical corrosion of ADSS optical cables based on deep learning, and achieves adaptation through variable convolution, which can improve the accuracy and efficiency of detection and provide more reliable technical support for the maintenance of optical cables.

[0014] Finally, in order to improve the network's pertinence to optical cable defects, the present invention introduces a specific convolution form and a hybrid attention mechanism to the network, which helps to further improve the detection accuracy and avoid missed detections. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is the overall network structure diagram of the present invention.

[0016] Figure 2 This is a schematic diagram of the DCNv3 module in the present invention.

[0017] Figure 3 It is a structural diagram of the LAF-Block module in the present invention.

[0018] Figure 4 This is a structural diagram of the C2f_DCN module in the present invention.

[0019] Figure 5 It is a comparison chart of comparative experimental results of the present invention. DETAILED DESCRIPTION

[0020] See also Figures 1 to 5 The present invention relates to a method for detecting surface defects of a communication optical cable based on adaptive feature extraction, comprising the following steps: Step 1: Establish a data set of electrical corrosion defects in communication optical cable lines (ADSS). To verify the validity of this patent, this patent uses the inspection defect photos of 21 ADSS optical cable lines completed by Jiangyin Power Supply Bureau in the second quarter of 2022 as the basic data set. There are a total of 1,025 pictures, of which, except for the defects of broken cable type, all defect images were collected by manual pole climbing inspection; broken cable defects are all optical cables that cannot be operated, so they are collected and entered in a laboratory environment; due to the influence of weather conditions, some defect photos in the basic data set will cause highlights and shadows on the surface of the optical cable, which increases the difficulty of data set detection, so the collection area is mostly from the outlet of the pre-twisted wire of the optical cable clamp to 1-2 meters away from the anti-vibration whip.

[0021] Since it is difficult to obtain defective images of optical cables, the number of data set samples is insufficient. Therefore, in order to improve the accuracy of the algorithm and accelerate network convergence, this embodiment uses traditional data enhancement methods such as randomly adding Gaussian or salt and pepper noise, changing brightness, mirroring, scaling, and random flipping to enhance and expand the data set (the number of pictures reaches 1025*4=4100), and on this basis, improve the robustness of the algorithm. Finally, since the photos in the basic data set are manually collected samples with limitations in collection position and angle, 257 normal ADSS optical cable pictures at multiple angles and farther positions are added as background negative samples for training, thereby further improving the richness of the data set samples.

[0022] Therefore, there are a total of 4357 mixed images, and the data set is divided into training set, validation set and test set according to the ratio of 7:2:1. The data set refers to the IEEE P1222 standard to divide the degree of optical cable electrical corrosion and the changes in optical cable morphology caused by it into five defect type labels. The five defect type labels are Pitting, Smallpitting, Break, Breakdown and Tracking, and LabelImg is used to manually annotate rectangular boxes, finally forming the ADSS optical cable electrical corrosion defect data set, which is then sent to the network for training.

[0023] Step 2: In the backbone network, the improved CSPDarkNet is used in this embodiment to complete the feature downsampling extraction of the ADSS electrical corrosion defect image. In this embodiment, three groups of strided convolution modules and two DCNv3 downsampling modules are used to complete five downsampling operations on the feature map of size 640×640 in the data set, which 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 two DCNv3s; then, the SPPF is used at the deepest feature to complete the pooling fusion of the deep features and enhance the feature expression; the reason for using the most advanced version of DCNv3 is that the electrical corrosion defect features of ADSS optical cables are usually irregular in shape, and the traditional convolution kernel is not strong in the extraction ability of unconventional geometric shape features such as electrical corrosion defects. Therefore, DCNv3 is used in this step to improve the network's adaptive matching and extraction ability for irregular shape features. DCNv3 has made the following upgrades based on the previous known version DCNv2: First, DCNv3 combines the characteristics of deep separable convolution and divides the convolution weights into depth-wise and point-wise parts to improve network efficiency; secondly, by adopting a multi-head attention mechanism similar to the Transformer structure, the spatial aggregation process is split into G groups, so that each group has an independent sampling offset and modulation factor, so that different groups on each convolutional layer have different spatial aggregation modes, thereby enhancing the feature expression ability of DCN; finally, the element-by-element Sigmoid normalization strategy of DCNv2 is improved to Softmax normalization along the sample points, so that the sum of the modulation factors is constrained between 0 and 1, enhancing the stability of the network.

[0024] Step three, between the backbone network and the neck network, after downsampling is completed, a LAF-Block module based on a large-scale full-dimensional fusion attention mechanism is used to complete further attention enhancement extraction, strengthen the information expression of the deep network, and smooth the feature information by performing an average pooling operation on the feature map with a pooling kernel size of 7. Then, a 1×1 convolution kernel is used to extract features, and then 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.

[0025] In the attention LAF-Block module, the deep feature P5 in the backbone network is divided into two paths, and the important features of the deep feature P5 are extracted through large separable convolution attention and triple attention respectively. This is because the core of the conventional LSKA is still the deep separation convolution operation, so it is still impossible to avoid the inherent disadvantage that the deep separation convolution cannot make good use of the information at the same spatial position of different channels for learning, resulting in a certain degree of information loss. Therefore, in view of this shortcoming of LSKA, triple attention is introduced to compensate for the information loss caused by LSKA.

[0026] In addition, the above convolution operation adopts a deformable convolution module C2f_DCN module, and its specific structure is as follows: in the bottleneck structure Neck, a DCN_block is added, and a group of convolution operations are added to adjust the number of channels of the DCN_block.

[0027] Therefore, in step 2 and step 3, by combining DCNv3 with the C2f_DCN module, better linear adaptive extraction of optical cable features can be performed, the accuracy of network detection of optical cable electrical corrosion can be improved, and the gradient information flow of the network can be optimized.

[0028] Step 4: After the features extracted and fused by the backbone network and the neck network in step 3, they will be sent to the head network to complete the detection. In this step, an 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.

[0029] The experimental environment of this embodiment: system Ubuntu22.04, graphics card NVIDIA RTX 3090, deep learning framework Pytorch2.0.1. The accuracy of the model proposed in this embodiment is higher than the accuracy of any mainstream model, such as Figure 5 The control experiment data are shown in the graph.

[0030] 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 surface defects of communication optical cables based on adaptive feature extraction, characterized in that: The following steps are involved: Step 1: Establish a dataset of electrical corrosion defects in communication optical cable lines. Import the basic dataset of defect photos found during the inspection of communication optical cable lines, and annotate the communication optical cable lines with five labels: Pitting, Smallpitting, Break, Breakdown, and Tracking according to the electrical corrosion defects. Then, enhance and expand the dataset through data enhancement. Then, import multiple photos of normal communication optical cable lines collected from different angles or remotely into the dataset. Finally, divide the dataset into training set, validation set, and test set according to the ratio of 7:2:

1. Step 2: In the backbone network, the features of the electrical corrosion defect images in the dataset are adaptively downsampled and extracted based on the deformable convolution module DCNv3; Step 3: After sampling in step 2, between the backbone network and the neck network, a LAF-Block module is used to perform attention enhancement extraction to strengthen the information expression of the deep network, and the feature information is smoothed by performing an average pooling operation with a pooling kernel size of 7 on the feature map. Then, a 1×1 convolution kernel is used to extract features, and then 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; Step 4: After the features extracted and fused by the backbone network and the neck network in step 3, they will be sent to the head network to complete the detection.

2. A method for detecting surface defects of communication optical cables based on adaptive feature extraction according to claim 1, characterized in that: The specific process of step 2 is as follows: first, use three groups of strided convolution modules and two DCNv3 downsampling modules to complete five downsampling operations on the feature map of size 640×640 in the dataset, 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 two DCNv3s; then SPPF is used at the deepest feature layer to complete the pooling fusion of the deep features to enhance the feature expression.

3. A method for detecting surface defects of communication optical cables based on adaptive feature extraction according to claim 2, characterized in that: The process of DCNv3 in step 2 is as follows: DCNv3 combines the characteristics of deep separable convolution to divide the convolution weights into depth-wise and point-wise parts; secondly, through the multi-head attention mechanism, the spatial aggregation process is split into G groups, so that each group has an independent sampling offset and modulation factor, so that different groups on each convolution layer have different spatial aggregation modes to enhance the feature expression ability of deformable convolution; Finally, Softmax normalization is used along the sample points so that the sum of the modulation factors is constrained between 0 and 1 to enhance the network stability.

4. The method for detecting surface defects of communication optical cables based on adaptive feature extraction according to claim 1, characterized in that: The process of the attention LAF-Block module in step three is: divide the deep feature P5 in the backbone network into two paths, and extract important features of the deep feature P5 through large separable convolutional attention and triple attention respectively.

5. A method for detecting surface defects of communication optical cables based on adaptive feature extraction according to claim 4, characterized in that: The convolution operation adopts a deformable convolution module C2f_DCN module, and its specific structure is: in the bottleneck structure Neck, a DCN_block is added, and a group of convolution operations are added to adjust the number of channels of the DCN_block.

6. The method for detecting surface defects of communication optical cables based on adaptive feature extraction according to claim 1, characterized in that: The detection in step 4 adopts a decoupled head structure and introduces an Anchor-Free strategy.

Images