Railway sectional insulation device arc elimination angle defect identification method

Abstract

The application provides a railway sectional insulation device arc extinction angle defect identification method, belonging to the technical field of intelligent defect identification. The method obtains a sectional insulation device image to be detected; inputs the sectional insulation device image to be detected into a pre-trained arc extinction angle defect identification model, uses a backbone network of the model to extract a local feature map and a global feature map of the sectional insulation device in parallel; uses a bridging fusion module in the backbone network to fuse the global feature into the local feature and output a first feature map, and simultaneously fuse the local feature into the global feature and output a second feature map; inputs the first feature map and the second feature map into a classifier of the arc extinction angle defect identification model for defect identification classification, and outputs an arc extinction angle defect identification result. The application improves the identification effect of the model on the arc extinction angle defect, enhances the local and global perception ability of the model, simultaneously realizes parallel calculation of local feature extraction and global feature extraction, and thus reduces the delay of training and inference.

Description

Technical Field

[0001] This application belongs to the field of intelligent defect identification technology, specifically relating to a method for identifying arc-extinguishing angle defects in railway segmented and sub-item insulators. Background Technology

[0002] Sectional insulators are a crucial component of the overhead contact line suspension system in electrified railways. They divide the contact line of the same phase power supply unit into several independent power supply areas, providing conditions for maintenance work. When the contact line needs repair or malfunctions, the power outage area can be reduced, minimizing the impact on transportation. When an electric locomotive passes over a sectional insulator, an electric arc is generated. The arc-extinguishing angle of the sectional insulator promptly elongates and extinguishes the arc. However, long-term fatigue and repeated arc burning can reduce the arc-extinguishing ability of the arc-extinguishing angle. Furthermore, the arc-extinguishing angle of the sectional insulator is constantly exposed to the external environment, making it susceptible to defects such as detachment, breakage, and bending due to factors like temperature, humidity, and weather.

[0003] Currently, the identification of arc-suppression angle defects in segmented insulators often employs CNNs or Transformer networks. However, even when these two are directly fused into a single network for arc-suppression angle defect identification, the process suffers from low recall and low precision. This is because the fused network fails to leverage the strengths of both CNNs and Transformers, and it also exhibits significant training and inference latency. Therefore, a novel fusion network combining CNNs and Transformers is needed to improve the identification performance of arc-suppression angle defects in segmented insulators. Summary of the Invention

[0004] Therefore, this application provides a method for identifying arc-extinguishing angle defects in railway segmented and sub-item insulators, which helps to solve the problems of poor identification effect and high delay in existing arc-extinguishing angle defect identification technologies.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] This application provides a method for identifying arc-extinguishing angle defects in railway section and item insulators, including:

[0007] Acquire images of the segmented insulator to be inspected;

[0008] The image of the segmented insulator to be detected is input into the pre-trained arc-suppression angle defect recognition model, and the local feature map and global feature map of the segmented insulator are extracted in parallel using the backbone network of the arc-suppression angle defect recognition model.

[0009] The bridging and fusion module in the backbone network is used to fuse global features into local features and output the first feature map, while the local features are fused into global features and output the second feature map.

[0010] The first feature map and the second feature map are input into the classifier of the arc-suppression angle defect recognition model for defect recognition and classification, and the arc-suppression angle defect recognition result is output.

[0011] Furthermore, the arc-suppression angle defect identification model includes an input layer, a backbone network, and a classifier;

[0012] The backbone network comprises four phases, each phase including multiple cross-fusion modules stacked together and a spatial dimensionality reduction module; the spatial dimensionality reduction module of each phase is connected to the first cross-fusion module of the next phase.

[0013] The cross-fusion module consists of Nx stacked convolutional modules, Nx stacked attention modules, and a bridging fusion module; the convolutional modules are used to extract local feature maps of the segmented insulator, and the attention modules are used to extract global feature maps of the segmented insulator; the input of the bridging fusion module is connected to the output of the convolutional modules and the output of the attention modules, respectively.

[0014] Furthermore, the bridging and fusion module includes a convolutional unit, an attention unit, and a fusion unit; the fusion unit includes a local fusion subunit and a global fusion subunit.

[0015] The output of the convolution module is connected to the convolution unit, the local fusion sub-unit, and the global fusion sub-unit, respectively; the output of the attention module is connected to the attention unit, the local fusion sub-unit, and the global fusion sub-unit, respectively; the output of the convolution unit is connected to the output of the local fusion sub-unit; the output of the attention unit is connected to the output of the global fusion sub-unit; the local fusion sub-unit is used to fuse the global feature map into the local feature map through a multi-head self-attention mechanism; the global fusion sub-unit is used to fuse the local feature map into the global feature map.

[0016] Furthermore, the spatial dimensionality reduction module includes a downsampling unit and a downsampling unit; the downsampling unit is used to perform spatial dimensionality reduction on the first feature map to obtain a first feature map with a first resolution, and output the first feature map with the first resolution; the downsampling unit is used to perform spatial dimensionality reduction on the second feature map to obtain a second feature map with a second resolution, and output the second feature map with the second resolution; the first resolution is one-quarter of the resolution of the input first feature map, and the second resolution is one-quarter of the resolution of the input second feature map.

[0017] Furthermore, the training process of the arc-suppression angle defect identification model includes:

[0018] Collect arc-suppression angle defect sample data, perform sample augmentation on the arc-suppression angle defect sample data, and divide the augmented arc-suppression angle defect sample data into training set, test set and validation set in a ratio of 85:15:5;

[0019] A corner reduction defect identification model is constructed, and the model is iteratively trained using the training set. The model parameters are optimized using the focal loss function to obtain the initial corner reduction defect identification model.

[0020] The initial arc-angle defect recognition model was tested and validated using a test set and a validation set, respectively. After testing and validation, a pre-trained arc-angle defect recognition model was obtained.

[0021] Furthermore, the number Nx of the convolutional modules and attention modules is specifically 3.

[0022] The application employs the above technical solution and has at least the following beneficial effects:

[0023] The method for identifying arc-suppression angle defects in railway segmented insulators provided in this application involves acquiring an image of the segmented insulator to be inspected; then inputting the image into a pre-trained arc-suppression angle defect identification model; using the backbone network of the arc-suppression angle defect identification model to extract local and global feature maps of the segmented insulator in parallel; then using the bridging and fusion module in the backbone network to fuse the global features into the local features and output a first feature map, and simultaneously fusing the local features into the global features and outputting a second feature map; finally, inputting the first and second feature maps into the classifier of the arc-suppression angle defect identification model for defect identification and classification, and outputting the arc-suppression angle defect identification result. This application first extracts local and global feature maps of segmented insulators in parallel using the backbone network of the arc-suppression angle defect recognition model. Then, it uses a bridging fusion module in the backbone network to fuse the global features into the local features, outputting a first feature map. Simultaneously, the global features fuse the local features into the global features, outputting a second feature map. This ensures that the fused first and second feature maps satisfy both the global features required for image recognition of segmented insulators and the local features required for arc-suppression angle defect recognition, improving the model's defect recognition performance and enhancing its local and global perception capabilities. Furthermore, it enables parallel computation of local and global feature extraction, thereby reducing training and inference latency. In addition, the bridging fusion module provided in this application is easily transferable to other models and can be widely used.

[0024] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0026] Figure 1 This is a flowchart of the railway section and item insulator arc suppression angle defect identification method of the present invention;

[0027] Figure 2 This is an example diagram of the arc suppression angle defect in an existing segmented insulator;

[0028] Figure 3 This is an example diagram for identifying arc-suppression angle defects in segmented insulators in complex scenarios;

[0029] Figure 4 This is a diagram of the backbone network architecture of the present invention;

[0030] Figure 5 This is a comparison diagram of the bridging and fusion architecture of the present invention and existing technologies. Detailed Implementation

[0031] To make the purpose, technical solution and advantages of this application clearer, the technical solution of this application will be described in detail below.

[0032] Due to installation process issues, frequent arcing occurs when electric locomotives pass over the sectional insulators, adversely affecting the operation of the arc-extinguishing angle. Furthermore, the arc-extinguishing angles of the sectional insulators are constantly exposed to the external environment, affected by factors such as temperature, humidity, and weather. These adverse factors lead to two serious problems with the arc-extinguishing angles: 1) long-term fatigue of the arc-extinguishing angles, resulting in a decrease in arc-extinguishing capacity; 2) prolonged arcing causes severe melting, cracking, and vaporization of the arc-extinguishing angles, leading to a complete loss of arc-extinguishing capacity. When the arc-extinguishing angle loses its arc-extinguishing capacity or its arc-extinguishing capacity decreases, it will further lead to arcing burns and damage to the sectional insulators, and may even cause the contact network to fail. In recent years, arc-extinguishing problems in the sectional insulators of high-speed railways in China have gradually emerged and are increasing exponentially, seriously threatening the safe operation of the contact network. To facilitate real-time detection of arc-extinguishing angle defects, arc-extinguishing angle defects are classified into three types according to the severity of the damage: arc-extinguishing angle detachment, arc-extinguishing angle breakage, and arc-extinguishing angle bending. Among them, arc-extinguishing angle detachment defects are as follows... Figure 2 As shown in the attached diagrams (a), (b), (c), and (d), the arc-extinguishing angle is broken as follows. Figure 2 As shown in the attached diagrams (e), (f), (g), and (h), the arc-extinguishing angle bends as follows: Figure 2 The (i), (j), (k), and (l) parts are shown in the attached diagram. The segmented insulator has four arc-suppression angles, and these three types of defects can occur at each arc-suppression angle.

[0033] In real-world operating scenarios, the identification of arc-suppression angle defects is also affected by complex environments and identification algorithms, as detailed below:

[0034] First, the long mileage and wide coverage of high-speed railways in China bring about diverse and complex practical application environments, which greatly increases the difficulty of identifying arc-suppression angle defects. Figure 3 The attached figures (a), (b), (c), (d), (e), and (f) illustrate various scenarios encountered when identifying arc-suppression angle defects in segmented insulators from acquired images. Figure 3 Part (a) of the attached diagram shows a blurry image; Figure 3 Part (b) shows that the image acquisition lens is dirty and part of the image is obscured; Part (c) of Image 3 shows that the image encounters a complex background, with the target object and the background intertwined, making it difficult to distinguish the segmented insulator. Figure 3 The attached figures (d) and (e) show the segmented insulator in the image being blocked by the pantograph and the arcing. Figure 3 The attached diagram (f) shows the segmented insulator being truncated in the image; Figure 3 The attached figures (g), (h), (i), and (j) illustrate situations where the arc suppression angle image of the segmented insulator is blurred and difficult to distinguish due to adverse weather conditions such as heavy fog, nighttime, heavy rain, and strong sunlight. In real-world applications, the various complex conditions encountered by the segmented insulator image make arc suppression angle defect identification exceptionally difficult.

[0035] II. Identification of arc-extinguishing angle defects on segmented insulators differs from other defect identification methods. Due to the small size of the arc-extinguishing angle target and its high similarity to other targets, directly identifying it using existing algorithms leads to low recall and precision. Inspired by license plate recognition (which requires first identifying the vehicle and then the license plate), achieving high precision and recall for arc-extinguishing angle defect detection typically involves two steps: first, globally identifying the segmented insulator; then, based on the results of the first step, identifying whether the arc-extinguishing angle is abnormal. This requires establishing two identification models: a global model and a local model. The global model identifies the segmented insulator, and the local model identifies the arc-extinguishing angle. However, cascading these two models results in significant inference latency, and the detection performance of a single model affects the overall effectiveness in identifying arc-extinguishing angle defects.

[0036] Therefore, a specific algorithm is needed to identify arc-suppression angle defects in segmented insulators. This algorithm should simultaneously satisfy the requirements of global perception of segmented insulators and local recognition of arc-suppression angle defects, and should also be able to accurately identify arc-suppression angle defects in complex natural scenes. Based on this, we designed a novel algorithm that fuses CNN and Transformer bridge mode to specifically solve the problem of arc-suppression angle defect identification in real-world scenarios, achieving high accuracy, high recall, and low latency.

[0037] Corner reduction defect recognition requires locating the defect position and type in an image, which is essentially object detection in the field of machine vision. Currently, object detection is mainly divided into one-stage and two-stage detection. One-stage detection, represented by the YOLO series and its improvements, constructs an end-to-end object detector; two-stage detection, represented by Faster R-CNN and its improvements, constructs high-precision, high-recall object detectors. Attempts to apply these two types of object detection algorithms to corner reduction defect recognition in complex natural scenes resulted in low precision and recall. The reason for this is that both types of object detection are based on convolutional neural networks (CNNs), which excel at localized features but lack the ability to recognize global long-term dependencies.

[0038] To address the above deficiencies, this application provides a method for identifying arc-extinguishing angle defects in railway segmented insulators, thereby improving the defect identification effect of the target detection model, enhancing the model's local and global perception capabilities, and enabling parallel computation of local and global feature extraction, thus reducing the latency of training and inference.

[0039] Please see Figure 1 , Figure 1 This is based on a method for identifying arc-suppression angle defects in railway segmented and component-specific insulators, such as... Figure 1 As shown, the method includes:

[0040] S1: Acquire the image of the segmented insulator to be inspected;

[0041] S2: Input the image of the segmented insulator to be detected into the pre-trained arc-suppression angle defect recognition model, and use the backbone network of the arc-suppression angle defect recognition model to extract the local feature map and global feature map of the segmented insulator in parallel;

[0042] S3: Use the bridging and fusion module in the backbone network to fuse global features into local features and output the first feature map, and at the same time fuse local features into global features and output the second feature map;

[0043] S4: Input the first feature map and the second feature map into the classifier of the arc-suppression angle defect recognition model to perform defect recognition and classification, and output the arc-suppression angle defect recognition result.

[0044] Furthermore, in one embodiment, the arc-suppression angle defect identification model of this application includes an input layer, a backbone network, and a classifier. Wherein,

[0045] The backbone network consists of four phases, each of which includes multiple cross-fusion modules stacked together and a spatial dimensionality reduction module; the spatial dimensionality reduction module of each phase is connected to the first cross-fusion module of the next phase.

[0046] The cross-fusion module consists of Nx stacked convolutional modules, Nx stacked attention modules, and a bridging fusion module. The convolutional modules are used to extract local feature maps of the segmented insulator, and the attention modules are used to extract global feature maps of the segmented insulator. The input of the bridging fusion module is connected to the output of the convolutional modules and the output of the attention modules, respectively.

[0047] The backbone network of this application is a multi-scale feature pyramid backbone network architecture, referred to as the CTBF-AH2D backbone network (hereinafter, the backbone network refers to the CTBF-AH2D backbone network). The backbone network has two branches: CNN and Transformer. The CNN branch extracts local features of the segmented insulator, while the Transformer branch extracts global features. The extracted local and global features are then cross-fused by a bridging and fusion module within the backbone network. The backbone network consists of four stages: the first stage, the second stage, the third stage, and the fourth stage. The resolutions of the feature maps from the first to the fourth stage are (H / 4xW / 4, H / 8xW / 8, H / 16xW / 16, H / 32xW / 32), respectively, and the resolution of the feature map in the next stage is one-quarter of that in the previous stage.

[0048] Reference Figure 4 As shown, in the CTBF-AH2D backbone network, each stage mainly consists of multiple stacked cross-fusion modules (FEFB) and a spatial dimensionality reduction module (CTRB). The cross-fusion modules perform deep fusion of local and global features through a bridge pattern. The fused CNN and Transformer branches possess both local and global features, meaning they have both global perception capabilities for segmented insulators and arc-suppression angle defect recognition capabilities. In the four stages of the CTBF-AH2D backbone network, the number of stacked cross-fusion modules is L1, L2, L3, and L4, respectively. The CTRB module is only used at the end of each stage to fuse the features of the CNN and Transformer branches and reduce the dimensionality of the feature maps. The dimensionality-reduced feature maps serve as the input for the next stage.

[0049] To improve model training speed and reduce inference latency, the convolutional module, or CNN block (hereinafter referred to as CB), employs depthwise separable convolution. The attention module, or Transformer block (hereinafter referred to as TB), also employs an efficient multi-head self-attention mechanism (MHSA). These optimizations further improve detection efficiency while ensuring defect detection performance.

[0050] Each stage of the CTBF-AH2D backbone network is composed of multiple cross-fusion modules FEFB stacked together. The purpose is to extract global and local features, and then cross-fuse these features to enhance the model's representation ability.

[0051] Reference Figure 4 As shown, the cross-fusion module consists of Nx convolutional modules (CB), attention modules (TB), and a bridging fusion module (CTFB). CB and TB are primarily responsible for feature extraction, while CTFB fuses the extracted features. Specifically, CB uses a convolutional neural network (CNN) to extract local features, TB uses a transformer to extract long-dependent global features, and CTFB performs deep fusion of global and local features. The resulting CNN and Transformer branches contain both local and global features.

[0052] During feature extraction, Nx block boundaries (CBs) and block boundaries (TBs) are extracted in parallel to improve model efficiency. In FEFB, the number of TBs and CBs is Nx, and the number of CTFB blocks is 1. This application fixes the number of FEFB modules in stages 1, 2, and 4 to be 3, 4, and 3 respectively, and finds the optimal Nx value by configuring the two parameters Nx and L3. Extensive experimental comparisons are shown in Table 1, demonstrating that an Nx value of 3 achieves optimal performance in terms of network latency and accuracy.

[0053] The cross-fusion module of this application stacks multiple parallel CNN blocks and Transformer blocks to extract local and global features respectively, and then performs feature fusion through CTFB. The fused network has both global and local perception capabilities, richer features, and stronger representation capabilities. Furthermore, the cross-fused features can satisfy both the global features required for image recognition of segmented insulators and the local features required for arc-suppression angle defect recognition.

[0054] The CNN block employs a depthwise separable convolution approach that combines a convolutional generator with separable convolutions, reducing parameters and computational cost while improving network efficiency.

[0055] The Transformer block uses separable convolutions to reduce the spatial dimensions of K and V by 1 / s. 2 This significantly reduces the computational load and parameters required for subsequent multi-head attention calculations.

[0056] The CNN block and transformer block in this application ensure high efficiency and effective extraction of local and global features.

[0057] In this application, reference is made to Figure 5As shown, existing technologies often employ two-way bridging fusion when fusing global and local features, using Mobile-Former

[37] ( Figure 5 (a) and Conformer

[19] Figure 5 Taking (d) as an example, the two-way bridging fusion method provides an innovative approach to fusing convolutional neural network (CNN) and Transformer branches. This method employs local-to-global fusion, embedding local features from the CNN branch into the Transformer branch, enhancing the Transformer branch's ability to capture image texture and contour details. Simultaneously, global-to-local fusion injects global long-range dependencies from the Transformer branch into the CNN branch, enabling the CNN branch to acquire global features and cognitive capabilities. However, the two-way bridging fusion method used by Mobile-Former and Conformer still faces several challenges:

[0058] 1. In Mobile-Former, the Mobile→Former fusion module ( Figure 5 Part (b) uses a multi-head self-attention mechanism (MHSA) for local-to-global fusion, enabling the Transformer branches to perceive local details. However, the MHSA used in this module weakens local features and impairs the Transformer branches' ability to accurately capture local cognition.

[0059] 2. In Conformer, FCU2 ( Figure 5 Part (f) is designed to use CNNs for global-to-local fusion, enhancing the ability of CNN branches to perceive global information. However, CNNs are not well-suited for effectively fusing global features into local features, which may lead to the loss of global information and weaken the ability of CNN branches to perceive global features. FCU1 ( Figure 5 Part (e) was designed to use Transformer for local-to-global fusion, enhancing the ability of Transformer branches to capture local features; however, the actual extraction effect of local features is not ideal.

[0060] In addition to the issues mentioned above, the high coupling between the two fusion methods in Mobile-Former and Conformer makes parallel fusion difficult and introduces significant training and inference latency. Furthermore, the specific fusion modules required by the method (such as the FCU1 and FCU2 modules in Conformer) increase the implementation complexity of integrating it into other models, thus limiting its potential for widespread application.

[0061] Specifically, refer to Figure 5As shown in the figures for parts (g) and (h), the bridging fusion module includes a convolutional unit, an attention unit, and a fusion unit; the fusion unit includes a local fusion subunit and a global fusion subunit;

[0062] The output of the convolution module is connected to the convolution unit, the local fusion sub-unit, and the global fusion sub-unit, respectively; the output of the attention module is connected to the attention unit, the local fusion sub-unit, and the global fusion sub-unit, respectively; the output of the convolution unit is connected to the output of the local fusion sub-unit; the output of the attention unit is connected to the output of the global fusion sub-unit; the local fusion sub-unit is used to fuse the global feature map into the local feature map through a multi-head self-attention mechanism; the global fusion sub-unit is used to fuse the local feature map into the global feature map.

[0063] This application provides a novel bridging fusion paradigm that uses a multi-head self-attention mechanism (MHSA) for global-to-local fusion, which is more advantageous in enriching local feature fusion, while using a convolutional neural network (CNN) for local-to-global fusion is beneficial for advancing global feature fusion.

[0064] To address the shortcomings of the aforementioned two-way bridging and fusion methods, this application proposes a novel and effective convolutional neural network and Transformer fusion block (CTFB), such as... Figure 5 As shown in sections (g) and (h), it enhances the fusion of global and local features without affecting their effectiveness. Here, the fusion unit (FB) represents the fusion process between the CNN and the Transformer block, as... Figure 5 As shown in section (h), it comprises two sub-units. The first sub-unit is the local fusion sub-unit for global-to-local fusion, denoted by f(X). i Z i →X i+1 This indicates that it uses MHSA to extract local feature maps X from CNN branches. i and the global feature map Z from the Transformer branch i Fused into the output feature map X i+1 In the middle. For example Figure 5 The (h) part is shown on the left side of the attached figure. The fusion process is as shown in formula (1):

[0065]

[0066] Where X i and Z i It is used as the input parameter for the MHSA operation. Reshape represents the local feature map X. i Perform a transformation operation to convert it into an input parameter x in (B,N,D) format. i x iShrinkOP will use vector K i sum vector V i The dimensionality is reduced to provide efficient computation and memory savings for subsequent MHSA operations. Here, vector K... i sum vector V i Through global feature map Z i Obtained through transformation. The Softmax function is used to calculate the weights, where d represents the vector dimension and W... q W represents the query learning matrix. o This represents the output learning matrix. i +x i 'Perform the superposition operation on two vectors and assign the result to x' i BatchNorm represents the summed parameter x. i Perform batch standardization operations. i× W q x i× W o respectively with the parameter x i With query learning matrix W q and output learning matrix W o Perform cross product operation.

[0067] The second subunit is the global fusion subunit from local to global, denoted by H(X). i Z i →Z i+1 Indicates, such as Figure 5 The (h) part is shown on the right side of the attached diagram, Z i+1 Using CNN on X i and Z i The fusion process is obtained by performing a convolution operation, as shown in formula (2):

[0068]

[0069] Where Reshape represents the global feature map Z i Perform a reshaping operation to convert the input parameter z into a (B,H,W,C) style. i Conv1×1 represents a 1×1 convolution operation, DyReLU is a dynamic ReLU activation function, and z i +X i Perform a superposition operation on two vectors and assign the result to z. i LayerNorm represents the layer normalization operation.

[0070] This novel fusion paradigm not only enhances the local and global awareness capabilities of the fused model but also enables parallel computation, thereby reducing training and inference latency. Furthermore, the fusion module is easily transferable to other models and can be widely applied.

[0071] Furthermore, in one embodiment, reference is made to... Figure 4 As shown, the spatial dimensionality reduction module includes a downsampling unit and a downsampling unit; the downsampling unit is used to perform spatial dimensionality reduction on the first feature map to obtain a first feature map with a first resolution, and outputs the first feature map with the first resolution; the downsampling unit is used to perform spatial dimensionality reduction on the second feature map to obtain a second feature map with a second resolution, and outputs the second feature map with the second resolution; the first resolution is one-quarter of the resolution of the input first feature map, and the second resolution is one-quarter of the resolution of the input second feature map.

[0072] The spatial dimensionality reduction module CTRB performs spatial dimensionality reduction, making the output X... i+1 and Z i+1 The resolution is the input X i and Z i The resolution is reduced to 1 / 4, reducing the output X dimension. i+1 and Z i+1 These are used as inputs to the next-level CNN branch and Transformer branch, respectively. The specific implementation method for dimensionality reduction is: the CNN branch uses a DownSampling module to downsample and reduce the resolution, outputting X. i+1 The resolution is reduced to the input X i One-quarter of the resolution. In the Transformer branch, the PatchMerging module (downsampling) is used to reduce Z. i+1 Output dimension.

[0073] Furthermore, in one embodiment, the training process of the arc-suppression angle defect identification model includes:

[0074] Collect arc-suppression angle defect sample data, perform sample augmentation on the arc-suppression angle defect sample data, and divide the augmented arc-suppression angle defect sample data into training set, test set and validation set in a ratio of 85:15:5;

[0075] A corner reduction defect identification model is constructed, and the model is iteratively trained using the training set. The model parameters are optimized using the focal loss function to obtain the initial corner reduction defect identification model.

[0076] The initial arc-angle defect recognition model was tested and validated using a test set and a validation set, respectively. After testing and validation, a pre-trained arc-angle defect recognition model was obtained.

[0077] Specifically, this application uses over 2000 AHSI (Arcing Horns in Sectional Insulators) defect samples, i.e., arc horn samples from segmented insulators. These samples are categorized into three different types: AHSI detachment, AHSI fracture, and AHSI deformation. This dataset, referred to as the AHSI defect dataset, is used for training and evaluation, as detailed in Table 1 below. The sample distribution for each defect type can be found in Table 1. The dataset is divided into three subsets: a training subset (80%), an evaluation subset (5%), and a test subset (15%).

[0078] To train the CTBM-DAHD network, this application employs a two-step approach. First, the model was pre-trained for 300 epochs on the Image-22K dataset using the AdamW optimizer with a batch size of 256 and a weight decay of 0.05. Subsequently, the model was fine-tuned for 12 epochs using the AHSI defect dataset. During fine-tuning, a learning rate decay of 0.02 was applied at 67% and 89% of the training epochs. Momentum of 0.9 and a weight decay of 0.0004 were used, and all networks were trained on four Titan RTX GPUs with 24GB of memory each. This training method was consistently applied to all subsequent experimental networks.

[0079] To address the challenges posed by imbalanced and difficult samples, this application employs the focus loss algorithm as the loss function for the model.

[0080] Table 1. Number of Defect Samples

[0081] type Sample size AHSI Departure 942 AHSI fracture 473 AHSI Deformation 640

[0082] The samples were divided into three subsets: 80% for training, 5% for evaluation, and 15% for test. The CTBM-AH2D network was trained using the training set for 300 iterations, followed by 12 fine-tuning iterations using samples of arc-suppression angle defects in segmented insulators. The model's loss algorithm employed focal loss to adjust model parameters and address the issues of difficult and imbalanced samples.

[0083] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0084] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for identifying arc-suppression angle defects in railway segmented and component-specific insulators, characterized in that, include: Acquire images of the segmented insulator to be inspected; The image of the segmented insulator to be detected is input into the pre-trained arc-suppression angle defect recognition model, and the local feature map and global feature map of the segmented insulator are extracted in parallel using the backbone network of the arc-suppression angle defect recognition model. The backbone network utilizes a bridging fusion module to fuse global feature maps into local feature maps and output a first feature map, while simultaneously fusing local feature maps into the global feature map and outputting a second feature map. Specifically, the first feature map is obtained by the bridging fusion module fusing the global feature map into the local feature map using a multi-head self-attention mechanism. The second feature map is obtained by the bridging fusion module performing convolution operations on the global and local feature maps using a convolutional neural network. The bridging fusion module includes a convolutional unit, an attention unit, and a fusion unit. The fusion unit includes a local fusion subunit and a global fusion subunit. The output of the convolutional module is connected to the convolutional unit, the local fusion subunit, and the global fusion subunit, respectively. The output of the attention module is connected to the attention unit, the local fusion subunit, and the global fusion subunit, respectively. The output of the convolutional unit is connected to the output of the local fusion subunit. The output of the attention unit is connected to the output of the global fusion subunit. The local fusion subunit is used to fuse the global feature map into the local feature map using a multi-head self-attention mechanism. The global fusion subunit is used to fuse the local feature map into the global feature map. The first feature map and the second feature map are input into the classifier of the arc-suppression angle defect recognition model for defect recognition and classification, and the arc-suppression angle defect recognition result is output.

2. The method for identifying arc-suppression angle defects in railway segmented and sub-item insulators according to claim 1, characterized in that, The arc-suppression angle defect identification model includes an input layer, a backbone network, and a classifier; The backbone network comprises four phases, each phase including multiple cross-fusion modules stacked together and a spatial dimensionality reduction module; the spatial dimensionality reduction module of each phase is connected to the first cross-fusion module of the next phase. The cross-fusion module consists of Nx stacked convolutional modules, Nx stacked attention modules, and a bridging fusion module; the convolutional modules are used to extract local feature maps of the segmented insulator, and the attention modules are used to extract global feature maps of the segmented insulator; the input of the bridging fusion module is connected to the output of the convolutional modules and the output of the attention modules, respectively.

3. The method for identifying arc-suppression angle defects in railway segmented and sub-item insulators according to claim 2, characterized in that, The spatial dimensionality reduction module includes a downsampling unit and a downsampling unit; the downsampling unit is used to perform spatial dimensionality reduction on the first feature map to obtain a first feature map with a first resolution, and outputs the first feature map with the first resolution; the downsampling unit is used to perform spatial dimensionality reduction on the second feature map to obtain a second feature map with a second resolution, and outputs the second feature map with the second resolution; the first resolution is one-quarter of the resolution of the input first feature map, and the second resolution is one-quarter of the resolution of the input second feature map.

4. The method for identifying arc-suppression angle defects in railway segmented and sub-item insulators according to claim 1, characterized in that, The training process of the arc-suppression angle defect identification model includes: Collect arc-suppression angle defect sample data, perform sample augmentation on the arc-suppression angle defect sample data, and divide the augmented arc-suppression angle defect sample data into training set, test set and validation set in a ratio of 85:15:5; A corner reduction defect identification model is constructed, and the model is iteratively trained using a training set. The model parameters are then optimized using the focal loss function to obtain the initial corner reduction defect identification model. The initial arc-angle defect recognition model was tested and validated using a test set and a validation set, respectively. After testing and validation, a pre-trained arc-angle defect recognition model was obtained.

5. The method for identifying arc-suppression angle defects in railway segmented and sub-item insulators according to claim 2, characterized in that, The number of convolutional modules and attention modules, Nx, is specifically 3.

Images