Electrified railway cotter pin anomaly detection method
Patent Information
- Application Number
- CN202510767188.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-10
- Publication Date
- 2025-09-23
Smart Images
Figure CN120689291A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of machine vision technology, and in particular to a method for detecting abnormalities in cotter pins of electrified railways. Background Art
[0002] The overhead catenary system, installed above the railway, plays a vital role in the railway system's power transmission. It primarily consists of a positioning device, a support base, a support mechanism, and a suspension mechanism. Cotter pins, primarily located on the catenary's suspension mechanism, prevent the nut and bolt from rotating relative to each other, thereby preventing components from loosening or falling off. Because the catenary is typically installed outdoors and experiences severe vibrations when trains pass through, the health and quality of the cotter pins gradually deteriorate over the long-term operation of the railway system. Cotter pins are prone to abnormal conditions such as missing, loosening, and insufficient opening angles. These can cause the catenary to fall off or have insufficient opening angles, impacting power supply and even causing safety accidents. Therefore, real-time online detection of cotter pin abnormalities is extremely urgent and important.
[0003] As an emerging detection technology, machine vision-based methods offer the advantages of being non-contact and highly efficient. In recent years, they have been introduced to detect abnormalities in cotter pins on contact networks. Some researchers have used Hough transform and SIFT (Scale Invariant Feature Transform) algorithms to locate cotter pins at oblique and non-oblique wrist locations. During the detection phase, cotter pin anomalies are detected by statistically analyzing the distribution patterns of non-connected areas within a circular ring. Other researchers have used a normalized cross-correlation template matching algorithm to locate the base of the contact network locator, while also employing an unsupervised learning autoencoder to extract dimensionality reduction features from the image. A support vector machine is then used to identify missing cotter pins. This traditional method requires manual design of features for the object being detected and relies on expert-designed features, resulting in insufficient robustness and generalization.
[0004] Deep learning is rapidly developing, and convolutional neural networks (CNNs), with their ability to automatically learn features from large numbers of images, have found applications in the detection of catenary cotter pins. Wang Xinyu et al. proposed a three-stage cascade architecture based on a deep convolutional generative adversarial network. After expanding the defect sample set, a convolutional neural network (CNN) was trained to detect cotter pin defects. Wang Jian, Luo Longfu, Zou Jinhai et al. first used the SSD algorithm to classify cotter pins, then employed Deeplabv3+ for semantic segmentation. The SURF algorithm was then used to extract features from the segmented images. ERF was then used to train these features to generate a classification model, enabling cotter pin defect detection. Min Feng, Lang Da, and Wu Tao trained a Deeplabv3+ model using cotter pin data labeled with head and tail information, obtaining semantic information about the cotter pin's head and tail, enabling cotter pin status detection. Li Jian and Wang Zhengtong first used the YOLO (You Only Look Once) algorithm to locate cotter pins at the arm base, and then used a convolutional neural network to classify and identify missing lateral and vertical cotter pins. Qi Donglian, Qian Jiaying, and Yan Yunfeng used the RefineDet detection network to locate the dropper string components in the contact network image, and combined the traditional Hough transform algorithm and deep learning network to establish a multi-scale dropper string state detection model to achieve accurate identification of the dropper string state.
[0005] Shi Qiang, Liu Guangwen, Liu Limin, and others built an attention mechanism module and a multi-scale fusion module based on YOLOX to accurately locate the cotter pins of U-shaped clamps. They then used the ResNet50 network to classify and detect missing cotter pins. Gu Guimei and Wang Xiaoliang used a lightweight neural network, MobileNetV3, for front-end feature extraction and improved the SSD algorithm by replacing the inverse residual structure with the CA attention mechanism to optimize the identification of small defects in contact lines.
[0006] Although the above deep learning-based method has improved the detection effect of cotter pins in complex scenarios to a certain extent, it still faces the following difficulties in practical applications:
[0007] 1) Complex inspection scenarios. Because high-speed rail lines are in operation most of the time, overhead line image acquisition can only be performed within a very limited "window" of time. In practice, various inspection scenarios arise, including varying weather conditions and lighting conditions. Furthermore, the characteristics of the contact pins are diverse.
[0008] 2) Complex image background. Catenary images contain not only the sky but also complex backgrounds such as towers, catenary, and transmission lines. The contrast between the background and foreground objects is low, making feature extraction difficult.
[0009] 3) Small objects: The cotter pin occupies fewer pixels in the contact network image and is a typical small object.
[0010] 4) Limited computing resources. Due to cost and volume constraints, the computing resources configured on trains are insufficient. It is necessary to balance detection accuracy, model parameter size, and computational complexity to facilitate field application deployment.
[0011] To this end, a method for detecting abnormalities in cotter pins of electrified railways is designed to provide a technical solution to the above technical problems. Summary of the Invention
[0012] Based on this, it is necessary to provide a method for detecting abnormalities in cotter pins of electrified railways in order to solve the technical problems raised in the above-mentioned background technology.
[0013] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0014] A method for detecting abnormalities in cotter pins of an electrified railway, comprising the following steps:
[0015] S1: Collect on-site images of the cotter pins of the electrified railway to be inspected;
[0016] S2: Input the collected image into a pre-trained electrified railway cotter pin anomaly detection model to obtain a detection result of whether the cotter pin being detected is abnormal.
[0017] As a preferred embodiment of the method for detecting abnormality of cotter pins in electrified railways provided by the present invention, in step S1, the steps are as follows:
[0018] Collect images of contact network cotter pins;
[0019] Use LabelImg to label the contact network cotter pin image;
[0020] Divide the dataset into training validation set and test set;
[0021] The training validation set is divided into a training set and a validation set.
[0022] As a preferred embodiment of the method for detecting abnormality of cotter pins in electrified railways provided by the present invention, in step S2, the steps are as follows:
[0023] An abnormality detection model for cotter pins in electrified railways is constructed based on the backbone feature extraction network Backbone, the neck network Neck and the detection head network Head.
[0024] As a preferred embodiment of the method for detecting abnormalities in cotter pins of electrified railways provided by the present invention, the backbone feature extraction network Backbone is constructed in the following steps:
[0025] A ten-layer structure is used to build the backbone feature extraction network Backbone;
[0026] The first layer is the Focus module;
[0027] The second, fourth, sixth, and eighth layers are Conv_BN_SiLU modules;
[0028] The third, fifth, seventh, and ninth layers are lightweight feature enhancement extraction modules;
[0029] The tenth layer is the auxiliary enhanced feature network.
[0030] As a preferred embodiment of the method for detecting anomaly of split pins in electrified railways provided by the present invention, the Conv_BN_SiLU module is a 2D convolution layer Conv, BatchNorm2d and activation function SiLU cascaded in sequence;
[0031] The lightweight feature enhancement extraction module adopts four feature processing branches;
[0032] The first feature processing branch is a cascade of 1×1 convolution modules, SimAM attention, and Ghost modules;
[0033] The second feature processing branch is a 1×1 convolution block;
[0034] The third feature processing branch consists of a cascade of 1×1 convolutional blocks and two Ghost modules;
[0035] The fourth feature processing branch consists of a cascade of 1×1 convolutional blocks and four Ghost modules;
[0036] The outputs of the four feature processing branches are concatenated as the output of the lightweight feature enhancement extraction module.
[0037] As a preferred implementation of the method for detecting abnormalities in open pins of electrified railways provided by the present invention, the feature maps extracted by the third, fifth, seventh and ninth layers of the backbone feature extraction network Backbone are respectively recorded as C2, C3, C4 and C5, and the input of the auxiliary enhanced feature network is the feature maps C2, C3, C4 and C5 extracted by the backbone feature extraction network, and the output is the feature map F1.
[0038] As a preferred embodiment of the method for detecting abnormalities in cotter pins of electrified railways provided by the present invention, the neck network Neck is used to fully fuse the feature maps of three scales, F3, F4 and F6, extracted by the backbone feature extraction network to obtain feature maps P3, P4 and P5;
[0039] The detection head network Head is used to detect features P3, P4 and P5 to obtain detection results of the helmet wearing image.
[0040] As a preferred embodiment of the method for detecting abnormalities in cotter pins of electrified railways provided by the present invention, the auxiliary enhanced feature network is divided into a detail feature processing module and an advanced feature enhancement module;
[0041] A. The detail feature processing module takes the feature maps C2 and C3 extracted by the feature extraction network as input and is divided into three feature processing branches;
[0042] B. The level feature enhancement module takes the feature maps C4 and C5 extracted by the feature extraction network as input and includes three different convolutional type branches;
[0043] C. In the output stage of the auxiliary enhanced feature network, the small target positioning information extracted by the detail feature processing module is concatenated and aggregated with the semantic information optimized by the advanced feature enhancement module, and the final output of the auxiliary enhanced feature network is obtained through a 1×1 convolution adjustment channel.
[0044] As a preferred embodiment of the method for detecting abnormalities in cotter pins of electrified railways provided by the present invention, in step A, three feature processing branches are performed as follows:
[0045] The feature processing branch optimizes a pair of input features C2 and sequentially undergoes 7×1 and 1×7 spatially separable convolutions, with a convolution kernel size of 3×3 and a dilation rate of 7;
[0046] Feature processing branch 2 optimizes the input feature C3 by sequentially performing 5×1 and 1×5 spatially separable convolutions, and adds a 3×3 dilated convolution with a dilation rate of 5 at the end of the branch.
[0047] The input feature processing branch 3 uses CA attention to encode channel relationships and long-range dependencies through precise position information, allowing the network to focus on the details of the target with a low computational cost. The features are further processed using a 3×3 dilated convolution with a dilation rate of 2.
[0048] The outputs of the three feature processing branches are concatenated and passed through a 1×1 convolution adjustment channel as the output of the detail feature processing module.
[0049] As a preferred embodiment of the method for detecting abnormalities in cotter pins of electrified railways provided by the present invention, in step B, the three branches of different convolution types are as follows:
[0050] The input of branch 1 is feature map C4, which is processed by a series of 3×1 and 1×3 spatially separable convolutions, with a convolution kernel size of 3×3 and a dilation rate of 3.
[0051] Branch 2 includes an adaptive average pooling layer and a dilated convolution with a dilation rate of 2 and a convolution kernel size of 3×3;
[0052] Branch three consists of a 1×1 convolution, a 3×3 Ghost module, and a 3×3 dilated convolution with a dilation rate of 1, all connected in series.
[0053] The input features of branches 2 and 3 are both feature maps C5;
[0054] The output features of the three branches are spliced and aggregated, the positioning information and semantic information are fully integrated, the positioning information is supplemented to the deep features, and the final output feature map is obtained through the 1×1 convolution adjustment channel.
[0055] It can be seen without a doubt that the above-mentioned technical solution of this application can definitely solve the technical problem to be solved by this application.
[0056] At the same time, through the above technical solutions, the present invention has at least the following beneficial effects:
[0057] 1. The present invention provides a method for detecting abnormalities in cotter pins on electrified railways. It adopts a lightweight feature enhancement extraction module. On the one hand, it adopts a multi-branch structure to extract rich information. On the other hand, it adopts parameter-free attention SimAM attention, which can enhance the learning of fine-grained features of cotter pin targets without increasing the number of model parameters.
[0058] 2. The present invention has the advantage of approaching the standard convolution performance with fewer parameters and computational complexity through the Ghost module. The Ghost module is used to further process the features, so that the network can learn more features and reduce the number of model parameters.
[0059] 3. The present invention uses convolutions of different sizes to capture target features of different granularities for different levels of features through auxiliary enhanced feature networks, thereby enriching feature expression. On the other hand, different types of convolutions are used to fully mine spatial information, expand the receptive field, and retain rich semantic information in the deep layer. The shallow positioning features are used to supplement the small target positioning information that is lacking in the deep layer, thereby further optimizing the feature map. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0061] Figure 1 It is a structural diagram of the detection method of the present invention;
[0062] Figure 2 This is a structural diagram of the lightweight feature enhancement extraction module of the present invention;
[0063] Figure 3 This is a diagram of the auxiliary enhancement feature network structure of the present invention;
[0064] Figure 4 Schematic diagram of the appearance of three typical catenary split pins used to collect data for this invention. DETAILED DESCRIPTION
[0065] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0066] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0067] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions therein may be combined with each other.
[0068] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not require further definition or explanation in subsequent drawings.
[0069] Example 1
[0070] Reference Figure 1-Figure 4 A method for detecting abnormalities in cotter pins of electrified railways comprises the following steps:
[0071] Collect on-site images of the electrified railway cotter pins to be inspected;
[0072] The dataset contains 4745 images, which are divided into three types of contact network cotter pin images: Figure 4 a Flat arm base - longitudinal cotter pin (CP-V), Figure 4 b Flat arm base - horizontal split pin (CP-H) and Figure 4 c is a cotter pin (CP-U).
[0073] First, LabelImg is used to annotate the contact network cotter pin image. Then, the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:3.
[0074] The training validation set is divided into a training set and a validation set in a ratio of 8:2.
[0075] Finally, we obtained 2657 training sets, 664 validation sets, and 1423 test sets.
[0076] The collected image is input into a pre-trained electric railway cotter pin anomaly detection model to obtain a detection result of whether the cotter pin being detected is abnormal;
[0077] The steps to build an anomaly detection model for cotter pins in electrified railways include:
[0078] An abnormality detection model for cotter pins in electrified railways is constructed based on the backbone feature extraction network, the neck network, and the detection head network.
[0079] (1) The backbone feature extraction network Backbone adopts a ten-layer structure. The first layer is the Focus module, the second, fourth, sixth, and eighth layers are Conv_BN_SiLU, the third, fifth, seventh, and ninth layers are lightweight feature enhancement extraction modules, and the tenth layer is the auxiliary enhancement feature network;
[0080] 1) The Conv_BN_SiLU module is a cascade of 2D convolutional layers Conv, BatchNorm2d, and SiLU activation function.
[0081] 2) The lightweight feature enhancement extraction module uses four feature processing branches. The first feature processing branch consists of a cascade of a 1×1 convolutional module, SimAM attention, and a Ghost module. The second feature processing branch consists of a 1×1 convolutional block. The third feature processing branch consists of a cascade of a 1×1 convolutional block and two Ghost modules. The fourth feature processing branch consists of a cascade of a 1×1 convolutional block and four Ghost modules. The outputs of the four feature processing branches are concatenated as the output of the lightweight feature enhancement extraction module.
[0082] 3) The feature maps extracted from the third, fifth, seventh, and ninth layers of the backbone feature extraction network are denoted as C2, C3, C4, and C5 respectively;
[0083] 4) The input of the auxiliary enhancement feature network is the feature maps C2, C3, C4 and C5 extracted by the backbone feature extraction network, and the output is the feature map F1;
[0084] (2) Neck network, which is used to fully fuse the feature maps of three scales, F3, F4 and F6, extracted by the backbone feature extraction network to obtain feature maps P3, P4 and P5;
[0085] (3) The detection head network Head is used to detect features P3, P4 and P5 to obtain the detection results of the helmet wearing image.
[0086] The auxiliary enhanced feature network is divided into a detail feature processing module and a high-level feature enhancement module;
[0087] (1) Detail feature processing module: This module takes the feature maps C2 and C3 extracted by the feature extraction network as input and is divided into three feature processing branches.
[0088] The feature processing branch optimizes the input feature C2 by sequentially passing through 7×1 and 1×7 spatially separable convolutions, and a dilated convolution with a kernel size of 3×3 and a dilation ratio of 7. Ordinary convolution only processes channel feature information, while spatially separable convolution processes features from the spatial dimension, which can further capture spatial positioning information and reduce the amount of computation. Dilated convolution introduces spacing between convolution kernel elements to expand the receptive field without increasing the number of parameters. Feature map C2 is a shallow feature that contains rich detailed information and has a relatively high resolution. The convolution operation using a large convolution kernel captures more comprehensive feature information of the target.
[0089] The feature processing branch 2 optimizes the input feature C3, and sequentially performs 5×1 and 1×5 spatially separable convolutions. At the end of the branch, a 3×3 dilated convolution with a dilation rate of 5 is added. Since the feature map C3 belongs to the middle-level feature, the image resolution is reduced after downsampling, so the convolution kernel is reduced accordingly. After downsampling the feature map C2 and splicing it with the feature map C3,
[0090] The input feature processing branch 3 uses CA attention to encode channel relationships and long-range dependencies through precise position information, allowing the network to focus on the details of the target with a low computational cost. The features are further processed using a 3×3 dilated convolution with a dilation rate of 2.
[0091] The outputs of the three feature processing branches are concatenated and passed through a 1×1 convolution adjustment channel as the output of the detail feature processing module.
[0092] (2) Advanced feature enhancement module: This module takes the feature maps C4 and C5 extracted by the feature extraction network as input. The module also contains three different convolutional branches:
[0093] The input of branch 1 is feature map C4, which is processed by a series of 3×1 and 1×3 spatially separable convolutions, with a convolution kernel size of 3×3 and a dilation rate of 3.
[0094] Branch 2 includes an adaptive average pooling layer and a dilated convolution with a dilation rate of 2 and a convolution kernel size of 3×3. The adaptive average pooling layer is used to remove redundant semantic information in deep features.
[0095] Branch three consists of a 1×1 convolution, a 3×3 Ghost module, and a 3×3 dilated convolution with a dilation rate of 1, all connected in series.
[0096] The input features of branches two and three are both feature maps C5. Dilated convolutions at the end of each of the three branches ensure that the information flowing through different branches has the same receptive field, maintaining semantic consistency. The output features of the three branches are concatenated and aggregated, fully integrating positioning and semantic information. This information is then added to the deep features, enriching the feature representation. Finally, a 1×1 convolution is performed to adjust the channels to obtain the final output feature map.
[0097] In the output stage of the auxiliary enhanced feature network, the small target positioning information extracted by the detail feature processing module and the semantic information optimized by the advanced feature enhancement module are spliced and aggregated, and the final output of the auxiliary enhanced feature network is obtained through a 1×1 convolution adjustment channel.
[0098] Example 2
[0099] A verification method is disclosed based on the above-mentioned embodiment 1.
[0100] 1. Dataset: The present invention is verified on the dataset.
[0101] 2. This paper compares the effectiveness and advancement of this paper with mainstream deep learning object detection models, including two-stage object detection networks Faster-RCNN and EfficientDet-D1, one-stage object detection networks YOLOv5-s, YOLOX-s, YOLOv7, YOLOv8, CenterNet, and ConvNeXt, anchor-free detection network FCOS, and Transformer-based object detection networks DETR and Swin, to verify the effectiveness and advancement of this paper. Refer to Table 1, where the optimal and suboptimal results are marked in bold and underlined, respectively.
[0102] Table 1 Comparative experiments of the present invention and other networks
[0103]
[0104]
[0105] The detection accuracy results of the present invention and the mainstream deep learning target detection model are shown in Table 1. Overall, the detection accuracy of the present invention is higher than that of other methods, with an mAP50 of 96.59%. The mAP50 of the present invention is 1.42% higher than that of YOLOv7, which ranks second in detection accuracy. YOLOv5-s is the baseline of the present invention, and the mAP50 of the present invention exceeds that of YOLOv5-s by 4.66%. In the three specific categories, the detection accuracy of the present invention is also the best, at 96.64%, 97.52% and 95.6% respectively. For the detection accuracy of CP-V, the present invention is 0.33% higher than that of YOLOv8, which ranks second in this category. For the detection accuracy of CP-H and CP-U, the present invention is 1.54% and 1.52% higher than that of YOLOv7, which ranks second in these two categories. Compared with YOLOv5-s as the baseline, the detection accuracy of the proposed method for CP-V, CP-H and CP-U exceeds YOLOv5-s by 5.39%, 3.61% and 4.96% respectively.
[0106] Because the present invention modifies some modules on the baseline, the number of parameters and computational complexity of the present invention have increased, while the inference speed is still slower than that of YOLOv5-s. Compared with YOLOv7, which ranks second in detection accuracy, the number of parameters and computational complexity of the present invention are reduced by 36.57% and 9.3%, respectively. Although the inference speed of the present invention is only half that of YOLOv7, it still reaches 32.2 frames per second, meeting the requirements of real-time detection.
[0107] 3. Achieve results
[0108] (1) Lightweight feature enhancement extraction module
[0109] Due to the diversity of detection scenarios and complex image backgrounds, existing feature extraction networks are insufficiently capable of capturing image features, resulting in the loss of many detailed features and the generation of redundant information. To extract important detailed features from contact network images and suppress invalid noise features, the lightweight feature enhancement extraction module employs a multi-branch structure to extract rich information and employs parameter-free attention (SimAM) to enhance the learning of fine-grained features of cotter pin targets without increasing the number of model parameters. Furthermore, as a new convolutional operation, the Ghost module approaches the performance of standard convolution with fewer parameters and computational complexity. Further feature processing using the Ghost module enables the network to learn more features, reducing the number of model parameters.
[0110] (2) Auxiliary Enhancement Feature Network
[0111] Split pins occupy relatively few pixels in contact network images and are typical small targets. Detecting small targets has always been a challenge in target detection. The auxiliary enhanced feature network is divided into a detail feature processing module and a high-level feature enhancement module. The auxiliary enhanced feature network uses convolutions of varying sizes for different levels of features to capture target features of varying granularity and enrich feature expression. Furthermore, different types of convolutions are used to fully exploit spatial information, expand the receptive field, and retain rich semantic information in the deep layers. By supplementing the small target positioning information lacking in the deep layers with shallow positioning features, the feature map is further optimized.
[0112] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A method for detecting abnormalities in cotter pins of electrified railways, characterized in that: Here are the steps: S1: Collect on-site images of the cotter pins of the electrified railway to be inspected; S2: Input the collected image into a pre-trained electrified railway cotter pin anomaly detection model to obtain a detection result of whether the cotter pin being detected is abnormal.
2. The method for detecting abnormality of split pins in electrified railways according to claim 1, characterized in that: In step S1, the steps are as follows: Collect images of contact network cotter pins; Use LabelImg to label the contact network cotter pin image; Divide the dataset into training validation set and test set; The training validation set is divided into a training set and a validation set.
3. The method for detecting abnormality of cotter pins in electrified railways according to claim 2, characterized in that: In step S2, the steps are as follows: An abnormality detection model for cotter pins in electrified railways is constructed based on the backbone feature extraction network Backbone, the neck network Neck and the detection head network Head.
4. The method for detecting abnormality of split pins in electrified railways according to claim 3, characterized in that: The backbone feature extraction network Backbone is constructed as follows: A ten-layer structure is used to build the backbone feature extraction network Backbone; The first layer is the Focus module; The second, fourth, sixth, and eighth layers are Conv_BN_SiLU modules; The third, fifth, seventh, and ninth layers are lightweight feature enhancement extraction modules; The tenth layer is the auxiliary enhanced feature network.
5. The method for detecting abnormality of cotter pins in electrified railways according to claim 4, characterized in that: The Conv_BN_SiLU module is a cascade of 2D convolutional layers Conv, BatchNorm2d and activation function SiLU; The lightweight feature enhancement extraction module adopts four feature processing branches; The first feature processing branch is a cascade of 1×1 convolution modules, SimAM attention, and Ghost modules; The second feature processing branch is a 1×1 convolution block; The third feature processing branch consists of a cascade of 1×1 convolutional blocks and two Ghost modules; The fourth feature processing branch consists of a cascade of 1×1 convolutional blocks and four Ghost modules; The outputs of the four feature processing branches are concatenated as the output of the lightweight feature enhancement extraction module.
6. The method for detecting abnormality of cotter pins in electrified railways according to claim 4, characterized in that: The feature maps extracted by the third, fifth, seventh and ninth layers of the backbone feature extraction network Backbone are denoted as C2, C3, C4 and C5 respectively. The input of the auxiliary enhancement feature network is the feature maps C2, C3, C4 and C5 extracted by the backbone feature extraction network, and the output is the feature map F1.
7. The method for detecting abnormality of cotter pins in electrified railways according to claim 6, characterized in that: The neck network Neck is used to fully fuse the feature maps of three scales F3, F4 and F6 extracted by the backbone feature extraction network to obtain feature maps P3, P4 and P5; The detection head network Head is used to detect features P3, P4 and P5 to obtain detection results of the helmet wearing image.
8. The method for detecting abnormality of split pins in electrified railways according to claim 6, characterized in that: The auxiliary enhanced feature network is divided into a detail feature processing module and a high-level feature enhancement module; A. The detail feature processing module takes the feature maps C2 and C3 extracted by the feature extraction network as input and is divided into three feature processing branches; B. The level feature enhancement module takes the feature maps C4 and C5 extracted by the feature extraction network as input and includes three different convolutional type branches; C. In the output stage of the auxiliary enhanced feature network, the small target positioning information extracted by the detail feature processing module is concatenated and aggregated with the semantic information optimized by the advanced feature enhancement module, and the final output of the auxiliary enhanced feature network is obtained through a 1×1 convolution adjustment channel.
9. The method for detecting abnormality of cotter pins in electrified railways according to claim 8, characterized in that: In step A, there are three feature processing branches, and the steps are as follows: The feature processing branch optimizes a pair of input features C2 and sequentially undergoes 7×1 and 1×7 spatially separable convolutions, with a convolution kernel size of 3×3 and a dilation rate of 7; Feature processing branch 2 optimizes the input feature C3 by sequentially performing 5×1 and 1×5 spatially separable convolutions, and adds a 3×3 dilated convolution with a dilation rate of 5 at the end of the branch. The input feature processing branch 3 uses CA attention to encode channel relationships and long-range dependencies through precise position information, allowing the network to focus on the details of the target with a low computational cost. The features are further processed using a 3×3 dilated convolution with a dilation rate of 2. The outputs of the three feature processing branches are concatenated and passed through a 1×1 convolution adjustment channel as the output of the detail feature processing module.
10. The method for detecting abnormality of cotter pins in electrified railways according to claim 8, characterized in that: In step B, there are three different convolutional branches, and the steps are as follows: The input of branch 1 is feature map C4, which is processed by a series of 3×1 and 1×3 spatially separable convolutions, with a convolution kernel size of 3×3 and a dilation rate of 3. Branch 2 includes an adaptive average pooling layer and a dilated convolution with a dilation rate of 2 and a convolution kernel size of 3×3; Branch three consists of a 1×1 convolution, a 3×3 Ghost module, and a 3×3 dilated convolution with a dilation rate of 1, all connected in series. The input features of branches 2 and 3 are both feature maps C5; The output features of the three branches are spliced and aggregated, the positioning information and semantic information are fully integrated, the positioning information is supplemented to the deep features, and the final output feature map is obtained through the 1×1 convolution adjustment channel.