Lightweight detection method for electro-corrosion defect of ADSS (all dielectric self-supporting) optical cable based on hierarchical network structure

Abstract

According to the ADSS optical cable electro-corrosion defect lightweight detection method based on the layered network structure, the LS-FPN is adopted to replace a traditional neck structure, favorable position information in channel dimensions is reserved, and the defect positioning capability is improved; the idea of deformable convolution is introduced, and the convolution of the original backbone network is replaced, so that the network can pay more attention to surrounding defect information in the feature extraction process; finally, a Focus-MPDIOU loss function is used for replacing an original Compatible Intersection over unit (CIOU), the Focus-MPDIOU is excellent in performance in processing the electrocorrosion boundary condition, the loss gradient caused by excessive bias can be avoided, and therefore the method has the advantages of being high in detection efficiency, accurate in detection, low in cost and easy to popularize.

Description

Technical Field

[0001] The present invention relates to a lightweight detection method for electric corrosion defects of optical cables, and in particular to a method for detecting the electric corrosion of ADSS optical cables based on a hierarchical network structure and discovering defects in a timely manner. Background Art

[0002] At present, in the existing technology, the detection methods of electrical corrosion defects of ADSS optical cables (All Dielectric Self Supporting) mainly include manual inspection, traditional electrical testing, and optical fiber sensor detection. These methods can identify and locate electrical corrosion defects in optical cables to a certain extent, but they all have some obvious shortcomings.

[0003] (1) Manual inspection relies on the experience of the inspectors and on-site conditions, and is highly subjective and unpredictable, making it difficult to achieve large-scale, comprehensive and accurate inspection of optical cables. In addition, manual inspection also has the problems of heavy workload and low efficiency.

[0004] (II) Although traditional electrical testing methods can detect electrical corrosion defects in optical cables, they are less sensitive to defects, especially small corrosion defects. Such methods can usually only detect obvious changes in resistance, and are difficult to detect early defects that have not yet caused serious damage. They lack sufficient detection accuracy and cannot provide early warning and intervention, resulting in safety hazards during the use of optical cables.

[0005] (III) Fiber optic sensing detection technology is a relatively advanced detection method. It can monitor the changes of temperature, stress and other parameters inside the optical cable through optical fiber sensors, and indirectly infer possible electrical corrosion defects. Although it has high detection accuracy, this method also has the following limitations: fiber optic sensors are complex to install and costly, and are easily interfered by external environmental factors in actual applications, resulting in unstable detection results. In addition, for long-distance optical cables, the attenuation of signal transmission will also affect the accuracy of detection.

[0006] In summary, the existing ADSS optical cable electrical corrosion defect detection technology has deficiencies in sensitivity, stability, efficiency and cost. There is an urgent need for a more efficient, accurate and economical detection method to improve the safety and reliability of optical cables. Summary of the invention

[0007] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure, which has high detection efficiency, accurate detection, low cost and is easy to promote.

[0008] The object of the present invention is achieved in that:

[0009] A lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure comprises the following steps:

[0010] 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 a ratio of 7:2:1. The dataset is divided into five labels according to the manifestation of optical cable defects, namely Smallpitting, Pitting, Break, Breakdown, and Tracking. LabelImg is used to annotate the dataset, thus forming an ADSS optical cable electrical corrosion defect dataset and sending it to the network for training.

[0011] Step 2: In the backbone network, a hierarchical scale feature pyramid network LS-FPN is used for feature recognition, which is completed by the hierarchical scale feature pyramid network LS-FPN in two steps: (i) feature self-filtering and screening of feature maps of different scales in the data set in the feature selection module; (ii) synergistic integration of high-level and low-level information in the filtered feature maps through the selective feature fusion SFF mechanism;

[0012] Step 3: Perform convolution operation on the feature map identified in step 2 through a deformable convolutional network DCN;

[0013] Step 4: Use the Focus-MPDIoU loss function to process the electrical corrosion boundary condition of the feature map processed in step 3.

[0014] Preferably, the feature self-filtering method of sub-step (a) in step 2 is: the feature map is processed according to the channel attention mechanism, the processed features are subjected to two-level pooling by the dimension matching module, and the attention weight is determined by the Sigmoid activation function.

[0015] Preferably, the selective feature fusion method of sub-step (ii) in step 2 is as follows: given a high-level feature and a low-level feature in the feature map, the high-level feature is first expanded using a transposed convolution with a stride of 2 and a convolution kernel of 3×3, and after obtaining the expanded feature size, the high-level feature in sub-step (ii) is upsampled or downsampled using bilinear interpolation to obtain the sampled feature; finally, the attention weight in sub-step (i) is used to transform and filter the low-level features of the high-level features, and after obtaining features of consistent size, the features are fused and the feature values ​​are output.

[0016] Preferably, the operation method of the deformable convolutional network DCN in step three is: the input feature map is processed by standard convolution to generate a bias domain with a channel number of 2N; each pixel point in the bias domain corresponds to a matrix encoding offset information, and the deformable convolutional network DCN adjusts the offset of the input feature map at each point according to the value of the matrix, and the deformable convolutional network DCN adds a weight coefficient after the calculation result.

[0017] Preferably, the weight coefficient Δm in step 3 is n It must meet formula (2):

[0018]

[0019] in:

[0020] ω represents the weight of the sampling point,

[0021] ω(p n ) is at position p n The convolution kernel weight value,

[0022] x is the feature map to be detected,

[0023] p0 is the position on the output feature map,

[0024] p n is the position element in R,

[0025] Δp n is the offset,

[0026] p0+p n +Δp n Represents any location in the region.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The verification evaluation index mAP50 of the model of the present invention is improved to 87.8%, which is significantly improved compared with the YOLOv8 model of the same series, and has obvious advantages in accuracy even compared with the RT-DETR model.

[0029] At the same time, the improved model parameters of the present invention are reduced to 1.96M, which is 34.9% lower than the baseline model, so that the model algorithm can run smoothly on low-end and mid-end devices and is suitable for deployment in resource-constrained environments.

[0030] In addition, the model has high computational efficiency. While the accuracy is improved, the amount of calculation (i.e., GFLOPs) is reduced by 16%, which means that the algorithm can meet the real-time requirements while maintaining high accuracy.

[0031] Finally, the present invention introduces the LS-FPN structure and the deformable convolution module DCN, which makes the present invention perform better in processing complex backgrounds and multi-scale defects, and can better adapt to diverse detection environments than traditional technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a network structure diagram of the present invention.

[0033] Figure 2 Schematic diagram of the deformable convolutional network DCN module in the present invention.

[0034] Figure 3 Schematic diagram of the convolution operation of the deformable convolutional network DCN module in the present invention.

[0035] Figure 4 Schematic diagram of the hierarchical scale feature pyramid network LS-FPN in the present invention.

[0036] Figure 5 Schematic diagram of the selective feature fusion module in the present invention.

[0037] Figure 6 It is a thermal comparison diagram of the electrical corrosion defects of the ADSS optical cable in the present invention.

[0038] Figure 7 It is a comparison chart of the detection results in the present invention.

[0039] Figure 8 It is a schematic diagram of the implementation process of the present invention.

[0040] Fig. 9 It is a comparison chart of comparative experimental results of the present invention. DETAILED DESCRIPTION

[0041] See also Figures 1 to 9 The present invention relates to a lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure, the steps of which are:

[0042] Step one, establish a data set containing ADSS electrical corrosion defects. The data sets in this embodiment are all from ADSS optical cable line inspection defect photos collected by the State Grid Jiangsu Jiangyin Power Supply Bureau in each quarter of 2022 and 2023, as well as laboratory simulation data. It includes high-definition images obtained during the maintenance and inspection process of the power company, covering a variety of typical optical cable defect types, such as wear, cracks, peeling, etc.; at the same time, in order to supplement and expand the diversity of the data set, various types of optical cable damage were simulated and recorded through high-precision imaging under laboratory conditions, with a total of 1,374 pictures; and in order to accelerate model convergence and improve the robustness of the model in different environments, in the data preprocessing stage, the data set is randomly cropped, horizontally or vertically flipped, rotated, scaled, color jittered, noise added, etc., with a total of 4,979 pictures after enhancement, and divided into training set, verification set and test set in a ratio of 7:2:1.

[0043] The dataset is divided into five labels according to the manifestation of optical cable defects, namely Smallpitting, Pitting, Break, Breakdown and Tracking. The LabelImg library is used to annotate the data, and the ADSS optical cable electrical corrosion defect dataset is finally formed, which is then sent to the network for training.

[0044] Step 2: Considering the large scale span of defects on the surface of the optical cable, the layered scale feature pyramid network LS-FPN (Layered scales-FPN, which is an improved structure of the feature pyramid network Feature Pyramid Network) is used to replace the traditional neck structure, retain the favorable position information in the channel dimension, and improve the ability to locate defects; and realize multi-scale feature fusion, so that the model can more accurately locate the electrical corrosion position information on the optical cable;

[0045] LS-FPN is completed in two stages: (i) feature maps of different scales in the dataset are self-filtered in the feature selection module; (ii) high-level and low-level information in these feature maps are synergistically integrated through the selective feature fusion (SFF) mechanism. Selective feature fusion produces semantically rich features, which is crucial for detecting subtle features in optical cable electrical corrosion images, thereby enhancing the detection capability of the model. Specifically:

[0046] (I) Feature self-filtering: The filter features are processed by the CA module and the Dimensionality Matching module; the CA module initially processes the input feature map, and its processed weight f in ∈R C×H×W, where C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map; after the feature map passes through the two pooling layers of global average pooling and global maximum pooling, the features generated after pooling are combined, and the Sigmoid activation function is used to determine the weight value of each channel, and finally the weight f of each channel is obtained CA ∈R C×1×1 .

[0047] (ii) Selective feature fusion: Although the multi-scale feature maps in the data set in the backbone network provide rich high-level semantic information features, the positioning of the target features is relatively vague. In contrast, low-level features provide accurate target positioning information, but the semantic information they carry is very limited. A common way to resolve this contradiction is to directly add the upsampled high-level features and low-level features pixel by pixel so that each layer is full of semantic information; however, this method simply sums the pixel values ​​of multiple feature layers without making any feature selection. In order to address the limitations of the fusion process, this embodiment uses selective feature fusion (SFF) instead of the original fusion method, and uses a feature fusion method that uses high-level features as weights to filter important semantic information embedded in low-level features.

[0048] Specifically: given an input high-level feature f for the feature map high ∈R C×H×W and an input low-level feature First, we use the transposed convolution (T-Conv) with a stride of 2 and a convolution kernel of 3×3 to expand the high-level features, and the expanded feature size is Afterwards, in order to unify the size of high-level features and low-level features, bilinear interpolation (BI) is used to upsample or downsample the high-level features in step (2) to obtain the sampled features Finally, the channel attention module is used to convert the high-level features into corresponding attention weights (from step (a)) to filter the low-level features, and fuse them after obtaining features of consistent size; finally, the filtered low-level features are fused with the high-level features to enhance the feature representation of the model and obtain the output features

[0049] Step 3: Introduce the idea of ​​deformable convolution to replace the convolution of the original backbone network, so that the network can pay more attention to the surrounding defect information during feature extraction;

[0050] Conventional convolution operations with fixed shapes and sizes have limited capabilities in dealing with ADSS communication optical cables with complex backgrounds, large noises, and strong interference. In order to enhance the ability to extract features of electrical corrosion defects, the idea of ​​Deformable Convolutional Networks (DCN) is introduced into the backbone network. The operation method of Deformable Convolutional Networks (DCN) is as follows: the input feature map is processed by standard convolution to generate a bias domain with 2N channels; each pixel point in the bias domain corresponds to a matrix encoding offset information, and the Deformable Convolutional Network (DCN) adjusts the offset of the input feature map at each point according to the value of the matrix, thereby realizing the deformable convolution operation and enhancing the feature extraction capability; and the Deformable Convolutional Network (DCN) adds a weight coefficient after the calculation result, and performs weighted calculation on the offset of the sampling points on the convolution input feature map, so that the model can distinguish between relevant and irrelevant information, and reduce the negative impact of irrelevant background content in the electrical corrosion defect dataset caused by feature extraction.

[0051] Specifically, the convolution operation of the deformable convolutional network DCN adds an offset to the sampling position in the standard convolution operation. First, the input feature map is processed by standard convolution to generate a bias domain with 2N channels; each pixel in the bias domain corresponds to an ω matrix that encodes the offset information, and the deformable convolution is based on Δp n The value of the matrix adjusts the offset of the input feature map at each point, thereby expanding the receptive field of the convolution kernel, realizing deformable convolution operations, and enhancing feature extraction capabilities.

[0052] Among them: Based on the standard convolution, the deformable convolution adds an offset to each sampling point, and the calculation formula of the output eigenvalue is shown in formula (1).

[0053]

[0054] Where R = {(-1,-1), (-1,0), …, (0,1), (1,1)},

[0055] ω represents the weight of the sampling point,

[0056] ω(p n ) is at position p n The convolution kernel weight value,

[0057] x is the feature map to be detected,

[0058] p0 is the position on the output feature map,

[0059] p n is the position element in R,

[0060] Δp n is the offset,

[0061] p0+p n +Δp n Represents any location in the region.

[0062] Based on the above DCN, the deformable convolution in this embodiment adds a weight coefficient Δm after the calculation result. n , see formula (2).

[0063]

[0064] This enables it to perform weighted calculations on the offsets of the sampling points on the convolution input feature map, allowing the model to distinguish between relevant and irrelevant information and reduce the negative impact of irrelevant background content in the electrical corrosion defect dataset caused by feature extraction.

[0065] Step 4: Use the Focus-MPDIoU loss function to replace the original CIoU (Compatible intersection over union). Focus-MPDIoU performs well in dealing with electrical corrosion boundary conditions and can avoid causing excessively biased loss gradients. The grid division method of CIoU (Compatible intersection over union) in the YOLOv8 model will cause the electrical corrosion defects of small-sized and long boundary boxes in the data set to span multiple grids, causing the annotation boxes to overlap multiple times, and their aspect ratios and height-width values ​​to change, which in turn makes it difficult for the original loss function to meet the expected results. The loss function Focus-MPDIoU in this embodiment can calculate IoU by measuring the minimum point distance between the predicted box and the true box, making up for the original loss function when the predicted boundary has the same aspect ratio as the true boundary but different width and height values. The problem of slow convergence speed and inaccurate boundary box prediction.

[0066] The Focus-MPDIoU loss function is:

[0067]

[0068]

[0069]

[0070]

[0071] In the above four formulas, define:

[0072] A is the predicted bounding box,

[0073] B is the ground-truth bounding box,

[0074] and These are the coordinates of the diagonal points of the A and B boxes.

[0075] and is the square of the diagonal distance between box A and box B,

[0076] ω and h represent the width and height of the A box and the A box respectively.

[0077] In the MPDIoU formula, the first term is the conventional intersection-over-union ratio, minus the ratio of the Euclidean distance between the two key points of the predicted box and the true box in the horizontal and vertical directions relative to the diagonal length of the bounding box.

[0078] u and d in Focus MPDIoU represent the upper and lower bounds of the current IoU, respectively. This formula linearly scales the original IoU value from the (d,u) range to the (0,1) range. This adjustment method will not treat highly overlapping boundary conditions (such as IoU values ​​close to 1) as completely accurate predictions when calculating losses, which helps improve the robustness of the model, avoid overly biased loss gradients, and make the model more flexible during training.

[0079] Step 5: After the boundary is processed by the Focus-MPDIoU loss function, the final features are input into the backbone network to compare and identify the electrical corrosion of the ADSS optical cable.

[0080] Based on the above-mentioned lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure, the ADSS communication optical cable surface defect data set is preprocessed, trained and improved. Finally, the optimization effect of the model is verified from multiple angles, and a comparison chart of the actual prediction effect is given. In the training phase, the optimization effect of the model is explored by retraining after modifying the model to determine the final optimization solution. In the phase of verifying the optimization effect of the model, ablation experiments are used to verify the effectiveness of the improvement points to ensure that the improved modules do not affect each other; and the classic target detection model and the same series detection model are compared horizontally to ensure the effectiveness and reliability of the model in the improvement process, providing theoretical support for the experimental results.

[0081] The experimental model training and testing are carried out under the Linux operating system. The CPU model of the experimental configuration is Intel(R) Xeon(R) Platinum 8375C, the GPU model is NVIDIA RTX 3090GPU, and the video memory size is 24G.

[0082] 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 lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure, 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 a ratio of 7:2:

1. The dataset is divided into five labels according to the manifestation of optical cable defects, namely Smallpitting, Pitting, Break, Breakdown, and Tracking. LabelImg is used to annotate the dataset, thus forming an ADSS optical cable electrical corrosion defect dataset and sending it to the network for training. Step 2: In the backbone network, a hierarchical scale feature pyramid network LS-FPN is used for feature recognition, which is completed by the hierarchical scale feature pyramid network LS-FPN in two steps: (i) feature self-filtering and screening of feature maps of different scales in the data set in the feature selection module; (ii) synergistic integration of high-level and low-level information in the filtered feature maps through the selective feature fusion SFF mechanism; Step 3: Perform convolution operation on the feature map identified in step 2 through a deformable convolutional network DCN; Step 4: Use the Focus-MPDIoU loss function to process the electrical corrosion boundary condition of the feature map processed in step 3.

2. According to claim 1, a lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure is characterized by: The feature self-filtering method of sub-step (I) in step 2 is: the feature map is processed according to the channel attention mechanism, the processed features are subjected to two-level pooling by the dimension matching module, and the attention weight is determined by the Sigmoid activation function.

3. According to claim 2, a lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure is characterized by: The selective feature fusion method of sub-step (ii) in step 2 is as follows: given a high-level feature and a low-level feature in the feature map, the high-level feature is first expanded using a transposed convolution with a stride of 2 and a convolution kernel of 3×3. After obtaining the expanded feature size, the high-level feature in sub-step (ii) is upsampled or downsampled using bilinear interpolation to obtain a sampled feature; Finally, the attention weights in step (a) are used to transform the high-level features to filter the low-level features, and after obtaining features of consistent size, they are fused and the feature values ​​are output.

4. According to claim 1, a lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure is characterized by: The operation method of the deformable convolutional network DCN in step 3 is: processing the input feature map through standard convolution to generate a bias domain with a channel number of 2N; Each pixel point in the bias domain corresponds to a matrix that encodes the offset information. The deformable convolutional network DCN adjusts the offset of the input feature map at each point according to the value of the matrix, and the deformable convolutional network DCN adds a weight coefficient after the calculation result.

5. According to claim 4, a lightweight detection method for ADSS optical cable electrical corrosion defects based on a hierarchical network structure is characterized by: The weight coefficient Δm in step 3 n It must meet formula (2): in: ω represents the weight of the sampling point, ω(p n ) is at position p n The convolution kernel weight value, x is the feature map to be detected, p0 is the position on the output feature map, p n is the position element in R, Δp n is the offset, p0+p n +Δp n Represents any location in the region.

Images