Electrified railway catenary insulator defect detection method, system, device and medium based on enhanced YOLOv10

Abstract

The application provides an electrified railway catenary insulator defect detection method, system, device and medium based on enhanced YOLOv10, which comprises the following steps: collecting catenary insulator images of a target scene through a shooting device and an automatic inspection device; using a labeling tool to label the catenary insulator and defects in the catenary insulator images, and dividing the labeled images into a training set and a verification set according to a preset ratio; performing model training based on the training set and the verification set, introducing a lightweight EfficientViT module, a Wise-IoU loss function and a knowledge distillation technology during the training process to obtain a catenary insulator defect detection model based on YOLOv10; and inputting the obtained catenary insulator images to be detected into the catenary insulator defect detection model to obtain a defect detection result. The method improves the accuracy and efficiency of insulator defect detection.

Description

Technical Field

[0001] This invention relates to the field of contact wire insulator defect detection technology, and in particular to a method, system, equipment and medium for detecting defects in contact wire insulators of electrified railways based on enhanced YOLOv10. Background Technology

[0002] The pantograph-overhead contact system is crucial for the power supply of modern electrified railways. Insulators, as core components, directly affect the safety of power transmission and the normal operation of the railway. However, with the expansion of the railway network and the increase in operating speed, traditional manual inspections are no longer sufficient to meet the demands for efficient and reliable maintenance. Automated and intelligent detection technologies have emerged to address this need. China State Railway Group Co., Ltd. has developed a high-speed railway power supply safety detection and monitoring system (referred to as the "6C system") based on computer vision for non-contact detection. However, accurate detection of the diverse components and complex defects of the overhead contact system insulators remains a challenge.

[0003] Contact line insulators, due to their long-term exposure to the outdoors, are susceptible to environmental erosion and mechanical damage, resulting in various defects such as breakage, dirt, and ablation. Furthermore, the contact line uses a wide variety of insulator types (different materials, structures, and sizes, such as...). Figure 2 As shown in the figure, images acquired by systems such as 6C are subject to interference such as changes in lighting and motion blur, making the accurate identification and location of various defects (especially early and minute defects) a challenging task.

[0004] In recent years, deep learning-based target detection technology has been widely explored for the automated identification of defects in overhead contact line insulators, aiming to replace traditional, inefficient manual inspections. It is mainly divided into two categories: methods based on convolutional neural networks (CNNs) (such as Faster R-CNN and YOLO series) and methods based on Transformers (such as ViT and DETR). Existing research mainly focuses on improving detection accuracy, handling complex background interference, and adapting to different insulator types. CNN-based methods still face challenges in balancing detection accuracy, processing speed, and generalization ability to diverse insulators and defects; in particular, the computational cost of complex models often becomes a bottleneck for real-time deployment. Transformer-based methods have demonstrated powerful feature extraction capabilities, but their high computational cost is the main obstacle preventing their widespread application in real-time inspection of railway insulators.

[0005] Existing methods for detecting defects in overhead contact line insulators, whether based on CNN or Transformer, often struggle to simultaneously meet the requirements of high accuracy (especially for small and diverse defects), strong generalization (adapting to multiple types of insulators), and high efficiency (meeting the real-time processing needs of edge devices) in practical applications. Summary of the Invention

[0006] In view of this, the present invention provides a method, system, equipment and medium for detecting defects in insulators of electrified railway contact networks based on enhanced YOLOv10, in order to solve the above problems.

[0007] This invention provides a method for detecting defects in overhead contact line insulators of electrified railways based on enhanced YOLOv10, comprising: acquiring images of overhead contact line insulators in a target scene using an imaging device and automated inspection equipment; annotating the overhead contact line insulators and defects in the images using an annotation tool, and dividing the annotated images into a training set and a validation set according to a preset ratio; training a model based on the training set and the validation set, incorporating a lightweight EfficientViT module, a Wise-IoU loss function, and knowledge distillation technology during the training process to obtain an overhead contact line insulator defect detection model based on YOLOv10; and inputting the acquired images of the overhead contact line insulators to be detected into the overhead contact line insulator defect detection model to obtain defect detection results.

[0008] In another implementation of the present invention, the backbone of the contact wire insulator defect detection model includes a convolutional layer, a dual-feature fusion module, an SCDown module, an EfficientViT module, an SPPF module, and a PSA module; image features are extracted step by step through the convolutional layer and the dual-feature fusion module; downsampling is performed through the SCDown module; diverse features and global context information of the insulator are captured through the EfficientViT module; and feature information is further refined through the SPPF module and the PSA module.

[0009] In another implementation of the present invention, the Neck part of the contact wire insulator defect detection model includes a Concat module, an Upsample module, and a C2fCIB module; the Concat module fuses features from different layers; the Upsample module upsamples the feature map; and the C2fCIB module further processes the fused features.

[0010] In another implementation of the present invention, the Head part of the contact wire insulator defect detection model includes a YOLODetect layer; the YOLODetect layer is used to perform the final target classification and bounding box prediction, and outputs the target category and corresponding confidence level.

[0011] In another implementation of the present invention, in the knowledge distillation technique, the teacher model adopts YOLOv10-s and the student model adopts YOLOv10-n-EfficientViT. The layer alignment strategy in the knowledge distillation process is based on aligning feature maps with the same spatial resolution in the feature pyramid. The 16th, 19th and 22nd layer feature maps generated by the teacher model are aligned with the 12th, 15th and 18th layer feature maps of the student model, respectively.

[0012] In another implementation of the present invention, the contact wire insulator defect detection model uses the Wise-IoU loss function to optimize the bounding box regression, and the loss function is expressed as:

[0013]

[0014] in, express IoU The basic loss function, This indicates the distance attention term.

[0015] In another implementation of the present invention, the basic loss function of the IoU Defined as:

[0016]

[0017] in, The smaller the value, the higher the overlap between the predicted bounding box and the ground truth bounding box.

[0018]

[0019] in, and These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. W b and H b This represents the width and height of the smallest closure box that covers both the predicted and actual bounding boxes.

[0020] In another aspect, the present invention provides a defect detection system for overhead contact line insulators of electrified railways based on enhanced YOLOv10, comprising: an image acquisition module for acquiring images of overhead contact line insulators in a target scene using a shooting device and automated inspection equipment; an image processing module for labeling the overhead contact line insulators and defects in the images using the LabelBee annotation tool, and dividing the labeled images into a training set and a validation set according to a preset ratio; a model training module for training a model based on the training set and the validation set, incorporating a lightweight EfficientViT module, a Wise-IoU loss function, and knowledge distillation technology during the training process to obtain an overhead contact line insulator defect detection model based on YOLOv10; and a result output module for inputting the acquired images of the overhead contact line insulators to be detected into the overhead contact line insulator defect detection model to obtain defect detection results.

[0021] In another aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a method for detecting defects in insulators of electrified railway contact networks based on enhanced YOLOv10 as described in any of the preceding claims.

[0022] In another aspect, the present invention provides a computer storage medium storing a computer program that, when executed by a processor, implements the steps of a method for detecting defects in insulators of electrified railway contact networks based on enhanced YOLOv10 as described in any of the preceding claims.

[0023] This invention presents a defect detection method for insulators in electrified railway contact networks based on enhanced YOLOv10. It uses the lightweight YOLOv10-n as the basic framework, integrating a lightweight EfficientViT module into the YOLOv10-n backbone network. This enhances the model's ability to capture global features and diverse details of insulators while controlling computational overhead. By introducing the Wise-IoU loss function, this perceptual loss is used to optimize bounding box regression, improving the localization accuracy for small-sized, irregularly shaped defects and accelerating model convergence. Furthermore, knowledge distillation technology is employed, using a pre-trained YOLOv10-s model as the teacher to guide the training of the enhanced YOLOv10-n student model. This effectively improves the detection performance while maintaining lightweight design, making it suitable for resource-constrained scenarios. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The accompanying drawings are only for illustrating preferred embodiments and are not intended to limit the present invention. In the accompanying drawings:

[0025] Figure 1 This is a schematic diagram of a defect detection method for electrified railway contact wire insulators based on enhanced YOLOv10, according to an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram showing the positional distribution of insulators in a contact network according to an embodiment of the present invention.

[0027] Figure 3 This is a schematic diagram of a network structure according to an embodiment of the present invention.

[0028] Figure 4 This is an overview diagram of the EfficientViT enhancement module according to an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram illustrating the principle of knowledge distillation in one embodiment of the present invention.

[0030] Figure 6 This is a partial training data image of one embodiment of the present invention.

[0031] Figure 7 This is a schematic diagram of the dataset labels and distribution according to an embodiment of the present invention.

[0032] Figure 8 This is a schematic diagram of nighttime visualization comparison analysis according to an embodiment of the present invention.

[0033] Figure 9 This is a schematic diagram illustrating the visual comparison and analysis of multiple insulators under intermittent illumination conditions, according to an embodiment of the present invention.

[0034] Figure 10 This is a schematic diagram illustrating a visual comparison analysis of damaged ceramic insulators according to an embodiment of the present invention.

[0035] Figure 11 This is a schematic diagram showing the variation of mAP@0.5 of various YOLOv10 series models with the number of training rounds, according to an embodiment of the present invention. Detailed Implementation

[0036] To enable those skilled in the art to better understand the improvements made in this embodiment compared to the prior art, the prior art will be briefly described below.

[0037] The reliability of insulators in electrified railway overhead contact lines is directly related to operational safety. Automated defect detection is a key technology to ensure its reliability; however, in actual deployments, especially on edge devices with limited computing power, existing methods often struggle to balance high detection accuracy with high operational efficiency.

[0038] Taking the advanced YOLOv10 as an example, directly applying YOLOv10 to edge devices for railway inspection with limited computing power faces a significant accuracy-efficiency dilemma. While lightweight versions (such as YOLOv10-n) are fast, their accuracy is often insufficient to reliably identify small defects given the complexity of insulator detection. High-performance versions (such as YOLOv10-s), while offering improved accuracy, suffer from high computational costs and model size limitations that restrict real-time deployment. Furthermore, the general design of YOLOv10 may not fully consider the specific structure and defect characteristics of insulators, resulting in poor performance in detecting small targets and edge details.

[0039] The Vision Transformer (ViT) model, proposed in 2020 and published in 2021, introduces the Transformer architecture, originally used in Natural Language Processing (NLP), to Computer Vision (CV), primarily for image classification tasks. Research shows that ViT performs better as the training dataset size increases. Compared to Convolutional Neural Networks (CNNs), ViT has stronger global information capture and long-range interaction capabilities, especially exhibiting superior classification accuracy when scaling the model size and training data. However, in high-resolution dynamic scene applications, ViT's performance degrades compared to CNNs.

[0040] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0041] Figure 1 A schematic flowchart of a defect detection method for electrified railway contact wire insulators based on enhanced YOLOv10 is provided for an embodiment of the present invention, as shown below. Figure 1 As shown, this embodiment mainly includes:

[0042] S101. Collect images of contact network insulators in the target scene using a shooting device and automated inspection equipment.

[0043] S102. Use a labeling tool (such as LabelBee labeling tool) to label the contact wire insulators and defects in the contact wire insulator image, and divide the labeled image into a training set and a validation set according to a preset ratio.

[0044] S103. Based on the training set and validation set, the model is trained. During the training process, the lightweight EfficientViT module, Wise-IoU loss function and knowledge distillation technology are introduced to obtain the contact wire insulator defect detection model based on YOLOv10.

[0045] S104. Input the acquired image of the contact wire insulator to be detected into the contact wire insulator defect detection model to obtain the defect detection result.

[0046] This invention presents a defect detection method for insulators in electrified railway contact networks based on enhanced YOLOv10. It uses the lightweight YOLOv10-n as the basic framework, integrating a lightweight EfficientViT module into the YOLOv10-n backbone network. This enhances the model's ability to capture global features and diverse details of insulators while controlling computational overhead. By introducing the Wise-IoU loss function, this perceptual loss is used to optimize bounding box regression, improving the localization accuracy for small-sized, irregularly shaped defects and accelerating model convergence. Furthermore, knowledge distillation technology is employed, using a pre-trained YOLOv10-s model as the teacher to guide the training of the enhanced YOLOv10-n student model. This effectively improves the detection performance while maintaining lightweight design, making it suitable for resource-constrained scenarios.

[0047] In another implementation of the present invention, the backbone of the contact wire insulator defect detection model includes a convolutional layer, a dual-feature fusion module, an SCDown module, an EfficientViT module, an SPPF module, and a PSA module; image features are extracted step by step through the convolutional layer and the dual-feature fusion module; downsampling is performed through the SCDown module; diverse features and global context information of the insulator are captured through the EfficientViT module; and feature information is further refined through the SPPF module and the PSA module.

[0048] For example, such as Figure 3As shown, based on the YOLOv10-n framework, the network structure was designed and optimized. In the backbone part, the network progressively extracts image features through convolutional layers (Conv) and C2f modules, and introduces the SCDown module for downsampling. In the application of contact network defect detection, although YOLOv10 has advanced target detection capabilities, it requires more computational resources and complex feature extraction layers to achieve higher detection accuracy. While this design improves detection accuracy, for the detection of small defects or subtle features (such as cracks or slight wear), the model may require more refined feature processing and a deeper network structure. However, deeper network layers bring more computation, leading to a decrease in model processing speed and affecting the efficiency of real-time detection. Therefore, the Efficient ViT (Vision Transformer) module was introduced, such as... Figure 4 As shown, this module enhances the network's ability to capture long-range dependent features, making it particularly suitable for feature extraction in complex scenes. Subsequently, through the SPPF (Fast Spatial Pyramid Pooling) and PSA modules, the network further refines the feature information, providing rich multi-scale features for subsequent detection.

[0049] It should be understood that by using the EfficientViT module to optimize the YOLOv10 backbone, the EfficientViT model improves computational efficiency while maintaining the global feature extraction advantages of ViT. EfficientViT addresses the computational bottleneck of ViT by replacing the traditional softmax attention mechanism with linear attention, and enhances the model's local feature extraction capabilities by introducing depthwise convolution, thereby reducing the complexity of linear computation.

[0050] This invention employs hierarchical local attention, performing attention calculations only on local information of the input image, reducing complexity while preserving the ability to capture global information. Furthermore, different methods are used for feature extraction at different network levels. Specifically, convolutional networks are used for low-level feature extraction, while Transformers are used for high-level global feature representation. This hybrid architecture fully leverages the locality of convolutional networks and the globality of Transformers, making the model more lightweight and efficient. Considering that the contact wire insulator defect detection model needs to be deployed on edge devices with limited computing resources, it is necessary to reduce the number of model parameters and computational cost to lower deployment costs. However, while lightweighting the network structure, a loss of accuracy is inevitable. To address this issue, a sparse computation strategy is adopted, i.e., precise calculations are performed only in the required regions, reducing invalid calculations on the entire image, significantly accelerating detection speed while ensuring prediction accuracy.

[0051] In another implementation of the present invention, the Neck part of the contact wire insulator defect detection model includes a Concat module, an Upsample module, and a C2fCIB module; the Concat module fuses features from different layers; the Upsample module upsamples the feature map; and the C2fCIB module further processes the fused features.

[0052] For example, in the Neck section, the network employs a Concat operation to fuse features from different layers, ensuring that features at different scales are considered simultaneously, thereby improving the detection capability for targets of different sizes. Furthermore, an Upsample operation is used to upsample the feature maps, helping to retain more detailed features. The Neck also introduces a C2fCIB module to further process these fused features, ensuring that the features are fully expressed before the detection layer.

[0053] In another implementation of the present invention, the Head part of the contact wire insulator defect detection model includes a YOLODetect layer; the YOLODetect layer is used to perform the final target classification and bounding box prediction, and outputs the target category and corresponding confidence level.

[0054] For example, in the Head section, the network performs the final target classification and bounding box prediction through the YOLODetect layer, outputting the target category and corresponding confidence score.

[0055] In another implementation of the present invention, in the knowledge distillation technique, the teacher model adopts YOLOv10-s and the student model adopts YOLOv10-n-EfficientViT. The layer alignment strategy in the knowledge distillation process is based on aligning feature maps with the same spatial resolution in the feature pyramid. The 16th, 19th and 22nd layer feature maps generated by the teacher model are aligned with the 12th, 15th and 18th layer feature maps of the student model, respectively.

[0056] For example, knowledge distillation is employed, using YOLOv10-s, a model with larger parameters and higher computational complexity, as the teacher model to guide the student model YOLOv10-n-EfficientViT, which has smaller parameters and higher computational efficiency. This allows the student model to learn richer and more detailed features than YOLOv10-n. By learning the output of the teacher model (such as soft labels or intermediate layer features), the student model retains its own architectural advantages while absorbing beneficial information from the teacher model, thereby effectively improving its predictive ability. Figure 5As shown, logistic distillation is performed on the Detect layer of YOLOv10-s and the Detect layer of YOLOv10-n-EfficientViT using the BCKD (Bidirectional Contrastive Knowledge Distillation) method. This method combines the advantages of knowledge distillation and contrastive learning, which can improve model performance and train small-scale student models more efficiently, making it particularly suitable for resource-constrained application environments.

[0057] In the knowledge distillation process, the layer alignment strategy between the teacher model (YOLOv10-s) and the student model (YOLOv10-n-EfficientViT) is based on aligning feature maps with the same spatial resolution in the feature pyramid. Specifically, the 16th, 19th, and 22nd layer feature maps generated by the teacher model are aligned with the 12th, 15th, and 18th layer feature maps of the student model, respectively. This alignment method fully utilizes the similarity of the feature pyramid structures of the two models, ensuring that the student model can learn matching spatial feature representations from the corresponding layers of the teacher model, thereby improving the performance of the student model in multi-scale object detection tasks. The cross-feature mapping of this invention adjusts the number of channels of the teacher model's feature map to be consistent with that of the student model through 1x1 convolution before calculating the loss. This method can flexibly handle differences in feature dimensions, ensuring effective alignment and learning of feature representations during the distillation process.

[0058] In another implementation of the present invention, the contact wire insulator defect detection model uses the Wise-IoU loss function to optimize the bounding box regression, and the loss function is expressed as:

[0059]

[0060] in, express IoU The basic loss function, This indicates the distance attention term.

[0061] In another implementation of the present invention, the basic loss function of the IoU Defined as:

[0062]

[0063] in, The smaller the value, the higher the overlap between the predicted bounding box and the ground truth bounding box.

[0064]

[0065] in, and These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively.W b and H b This represents the width and height of the smallest closure box that covers both the predicted and actual bounding boxes.

[0066] For example, the loss function of the entire network uses a combination of L_IoU and WIoU, where WIoU (weighted IoU) can more accurately handle the size, position and shape of the target bounding box, thereby improving the accuracy of target detection.

[0067] The Wise-IoU (WIoU) loss function plays a crucial role in object detection tasks, especially in scenarios involving small-sized or irregularly shaped defects, such as suspension insulator detection, where its design demonstrates unique advantages. The WIoU loss mainly consists of two parts: an IoU-based loss term and a distance attention mechanism.

[0068] IoU base loss term ( It is constructed based on the Intersection over Union (IoU). IoU quantifies the degree of overlap between the predicted bounding box and the ground truth box, and is calculated using the following formula:

[0069]

[0070] in, W i and H i These represent the width and height of the intersection region between the predicted bounding box and the ground truth bounding box, respectively. S u It is the area of ​​the region where the two are joined.

[0071] The basic loss function of IoU Defined as:

[0072]

[0073] in, A smaller value indicates a higher degree of overlap between the predicted and ground truth boxes, and vice versa. This loss term provides the basic geometric constraints for model optimization.

[0074] To make up for To address the shortcomings of focusing only on overlapping regions while ignoring the geometric relationships of non-overlapping regions, Wise-IoU introduces a distance attention term. Its definition is as follows:

[0075]

[0076] in, and These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. W b and H b This represents the width and height of the smallest enclosing box that covers both the predicted and ground truth boxes. Its superscript indicates that it is used when calculating gradients during backpropagation. W b and H b The values ​​of are treated as constants (i.e., separated from the computation graph) to prevent them from generating gradients that could interfere with training convergence. By calculating an exponential function of the distance to the center point, a larger loss weight is assigned to the distant, low-quality anchor boxes, prompting the model to prioritize predictions with better geometric alignment. This mechanism is particularly important for improving the model's localization accuracy and robustness when detecting small or irregularly shaped defects.

[0077] The complete loss function of Wise-IoU-v1 This combines the two parts mentioned above:

[0078]

[0079] This form utilizes distance attention items. For the basics Weighting was applied to improve the model's localization performance.

[0080] In another implementation of the invention, Wise-IoU also introduces a dynamic non-monotonic focusing mechanism, which assesses the "outlierness" of the anchor frame. This allows for dynamic adjustment of its contribution to the loss function. The definition is as follows:

[0081]

[0082] in, This also indicates separation from the computation graph. Value, and yes The exponential moving average reflects the average IoU loss level of recent samples. The smaller the value, the higher the quality of the anchor frame (relative to the recent average level).

[0083] based on Dynamic focusing coefficient Defined as:

[0084]

[0085] in, and It is a hyperparameter that controls the shape of the focusing curve.

[0086] The core idea of ​​this mechanism is:

[0087] 1) For high-quality anchor frames ( Smaller), assign smaller gradient weights This prevents the model from overfitting to high-quality samples.

[0088] 2) For anchor frames of extremely poor quality ( (Very large), so appropriately reduce its gradient weights. To avoid these outliers having too much negative impact on model training, the main optimization focus is placed on anchor boxes of "normal quality".

[0089] The loss function that ultimately applies the dynamic focusing mechanism will be obtained by... As Gradient adjustment is achieved through a multiplier (or in combination in other ways). Experiments show that this dynamic focusing mechanism significantly improves the model's defect localization accuracy and enhances its robustness to inconsistent or low-quality labeled samples.

[0090] Example 1

[0091] This invention studies overhead contact line insulators and defects, primarily targeting two categories: insulators and insulator defects. The datasets used are derived from various imaging methods employed by the railway 6C inspection system, such as... Figure 6 As shown, the dataset includes high-resolution images captured by 3C and 4C devices, as well as images taken by other automated inspection equipment. These images cover scenes with varying lighting, angles, and sharpness, ensuring the diversity of the dataset.

[0092] The LabelBee annotation tool was used to annotate the contact wire insulators and defects. The dataset was randomly divided according to a training set:validation set ratio of 8:2. The distribution of the proportion of each type of insulator in the dataset is shown in Table 1, and the distribution of the dataset labels is shown in Table 2. Figure 7 As shown.

[0093] Table 1. Mixed distribution of insulator types in the dataset

[0094]

[0095] The configuration environment used in the experiment is shown in Table 2.

[0096] Table 2 Experiment Configuration Table

[0097]

[0098] To verify the effectiveness of the algorithm, this invention uses F1 score, AP, precision (P), and recall (R) as evaluation metrics for the contact wire insulator defect detection algorithm, as shown in the following formulas:

[0099]

[0100] TP stands for true positive (the number of defects detected), and FP stands for false positive (the number of defects incorrectly detected).

[0101]

[0102] FN stands for false negative (the number of defects that were missed).

[0103]

[0104] The F1 score is the harmonic mean of precision and recall, used to comprehensively evaluate the accuracy and completeness of a model.

[0105]

[0106] AP is typically obtained by calculating precision and recall at different thresholds and then taking the area between them. p(r) A function representing accuracy.

[0107] Experimental Results and Analysis:

[0108] To intuitively evaluate the actual performance of different models in the task of detecting defects in overhead contact line insulators, three representative scenarios were selected for visualization and comparison of the results. Figure 8 The test results of defective brace arms and pin insulators under complex nighttime lighting conditions were demonstrated. Figure 9 The focus is on dense scenarios involving multiple normal insulators at night. Figure 10 The focus is on the identification of daytime damaged ceramic insulators. The FasterNet-t0, YOLOv8-n / s, YOLOv10-n / s, RT-DETR-l, and the enhanced YOLOv10-n model proposed in this invention were compared for each scenario. In the figure, (a) is the original image, and (b) to (h) correspond to the inference results of each model, respectively.

[0109] exist Figure 8In the nighttime complex lighting and multi-type defect scenarios shown, the performance of lightweight models such as FasterNet-t0(b) and YOLOv8-n(c) is significantly affected, with problems such as missed defect detection or extremely low confidence scores, and the localization bounding boxes are not accurate enough. The higher-performance YOLOv10-s(f) and RT-DETR-l(g) can generally detect defects well, and the bounding boxes are relatively accurate. In contrast, the method proposed in this invention (h) under these challenging conditions not only can stably detect insulator defects, but its localization boxes also fit the defect area perfectly and give a high confidence score. Its performance is comparable to the advanced RT-DETR model, indicating its good adaptability to complex lighting and diverse targets.

[0110] Figure 9 The model's performance was examined in a nighttime scenario with dense, normal insulators, which places demands on the model's recognition consistency and false alarm suppression capabilities. The lightweight model (b) exhibited missed detections under dense target conditions and had low confidence scores for all targets. Notably, models (d) and (g) misclassified light spots or other components in the background as defects. The proposed method (h), similar to robust baselines (such as YOLOv10-s (f)), reliably detected all insulator targets within the field of view without introducing significant false alarms, demonstrating good target recognition stability and background suppression capabilities.

[0111] at last, Figure 10 The model's generalization detection capability for damaged ceramic insulators in daytime scenarios was evaluated. Some comparative models (b, c, e) showed limited ability to identify this specific damage mode, resulting in low confidence scores, and even the YOLOv10-n model exhibited obvious duplicate anchor frames. The proposed method (h) also performed excellently in this scenario, accurately identifying and defining the damaged area of ​​the ceramic insulator with precise localization and high confidence. Its performance was very close to RT-DETR-l(g) (although this is the model with the minimum parameters of RT-DETR) and significantly better than several other comparative models, which confirms that the proposed model has good generalization ability across materials and defect morphologies.

[0112] Based on the comprehensive qualitative results, through visualization analysis of various real-world scenarios including different lighting conditions, target types, defect conditions, and materials, the enhanced YOLOv10-n model proposed in this invention demonstrates consistently reliable detection performance. In terms of key defect localization accuracy, confidence reliability, and robustness to complex environments, this model typically achieves levels comparable to YOLOv10-s, which consumes significantly more computational resources, and consistently outperforms other lightweight baseline models.

[0113] To comprehensively evaluate the performance of the method of this invention (Ours), it was quantitatively compared with a series of representative baseline models on a reserved test set. The results are shown in Table 3 and Figure 11 As shown. These baselines cover different scales and architectures, including lightweight YOLOv8-n, YOLOv10-n, FasterNet-t0, and more powerful YOLOv8-s, YOLOv10-s, and RT-DETR-l. All experiments were conducted at a 640x640 input resolution. Key metrics of interest include mAP@0.5, performance against insulator body ( ) and defects ( The average precision, precision (P), recall (R), and maximum F1 score at the optimal threshold are calculated as follows: ), as well as the number of parameters, computational cost (FLOPs), and inference speed (FPS) used to measure model efficiency.

[0114] The experimental results clearly demonstrate the excellent balance between accuracy and efficiency achieved by the method of this invention. On the core mAP@0.5 metric, the model of this invention achieves 82.0%, significantly outperforming all lightweight competitors (such as YOLOv10-n at 79.4%), and is highly competitive with high-performance YOLOv10-s (82.4%) and YOLOv8-s (81.15%), only slightly lower than the computationally resource-intensive RT-DETR-l (83.3%).

[0115] Table 3 shows the results of quantitative experimental comparisons on the dataset.

[0116]

[0117] Bold text indicates statistically significant improvement (p<0.05).

[0118] It is worth noting that, compared to the YOLOv10-n framework, the method of this invention achieves a 2.6% improvement in mAP, which directly proves the effectiveness of the introduced improvements such as EfficientViT, Wise-IoU, and knowledge distillation. In terms of subcategories, the model excels particularly in identifying the insulator ontology. It achieved a success rate of 95.1%, the highest among all comparative models. For the more challenging defect detection... It also reached 68.9%, maintaining a level similar to that of top-tier models.

[0119] Further analysis of the precision-recall performance (P / R) revealed that the model of this invention achieved a recall of 0.87, on par with YOLOv8-s, indicating its ability to effectively capture the vast majority of targets. Simultaneously, the precision remained high at 0.963, implying a low false positive rate. Overall, the model achieves a very good balance between accurate identification and avoiding false positives, which is crucial for practical applications.

[0120] However, the most striking advantage of this invention lies in its superior computational efficiency. The model has only 4.1M parameters and requires 9.9 G FLOPs of computation. Compared to the similarly performing YOLOv10-s, this means nearly half the number of parameters and over 60% less computation. This significant efficiency improvement is crucial for resource-constrained edge deployment scenarios. Compared to the highest-accuracy RT-DETR-l, its computational cost is an order of magnitude lower. Although the computational cost is slightly increased (approximately 17.9%) compared to the original YOLOv10-n, this overhead is compensated by a significant increase in accuracy, demonstrating an extremely high cost-effectiveness ratio. In terms of actual inference speed, 47.7 FPS also outperforms YOLOv10-s and RT-DETR-l, meeting the needs of many real-time detection applications.

[0121] In summary, the quantitative evaluation results strongly demonstrate that the enhanced YOLOv10-n model proposed in this invention successfully addresses the challenge of balancing accuracy and efficiency in overhead contact line insulator inspection. It achieves detection accuracy nearly equivalent to one of the current best models (YOLOv10-s) with significantly reduced computational and storage requirements, while maintaining an advantage in inference speed. This careful trade-off between performance and efficiency makes this model an ideal candidate for deploying advanced defect detection capabilities in practical railway inspection systems, particularly on edge computing platforms.

[0122] Example 2

[0123] Ablation experiment:

[0124] To further demonstrate the feasibility of the algorithmic innovations mentioned in the study, ablation experiments were conducted to analyze the effect of each module. The results of the ablation experiments are shown in Table 4.

[0125] Table 4 Comparison of ablation experiment results; bold indicates the best result.

[0126]

[0127] This invention verifies the contribution of each module to model performance through ablation experiments, particularly its impact on balancing accuracy and efficiency. Four different model configurations were designed, and their performance on multiple performance metrics was compared.

[0128] First, the model without knowledge distillation exhibits a significant performance degradation. While the precision (P) remains high (0.927), the F1 score drops to 0.71, the recall (R) drops to 0.73, and the mAP@0.5 is 80.4. This indicates that knowledge distillation plays a crucial role in improving feature extraction capabilities and enhancing model generalization performance. By transferring knowledge from the YOLOv10-s model to YOLOv10-n, the student model is able to learn finer features better with fewer parameters, thereby improving precision and recall.

[0129] Secondly, removing EfficientViT (w / o EfficientViT) resulted in a decrease in model performance. Although the precision (0.938) improved slightly, the F1 score dropped to 0.69, the recall to 0.72, and the mAP@0.5 to 79.8. EfficientViT significantly optimized the efficiency of feature extraction, reduced computational costs, and maintained high precision. While the model still achieved good precision after removing this module, the overall performance declined, indicating that EfficientViT played a crucial role in balancing feature extraction and computational efficiency.

[0130] Furthermore, while removing Wise-IoU (w / o WIoU) still resulted in high precision (0.958) and recall (0.77), with an mAP@0.5 of 81.3, it led to a decrease in the F1 score (down to 0.78). This indicates that Wise-IoU, as a perceptual loss function, plays a crucial role in optimizing localization accuracy in object detection tasks. Removing this module resulted in a decrease in precision when handling complex defect regions, particularly in the detection of small defects, impacting the improvement of recall and F1 score.

[0131] The comprehensive configuration (Ours) combines all modules: knowledge distillation, EfficientViT, and Wise-IoU, demonstrating the best performance. Specifically, it achieves an F1 score of 0.78, precision of 0.963, recall of 0.78, and mAP@0.5 of 82. This indicates that the model achieves a balance between accuracy and efficiency by reasonably combining these modules. By enhancing feature extraction and optimizing the perceptual loss function, the proposed model can achieve higher accuracy with fewer parameters, especially demonstrating excellent performance in balancing detection precision and recall. Furthermore, the synergistic effect of knowledge distillation and EfficientViT effectively improves computational efficiency, enabling the model to achieve relatively good performance even with limited resources. In summary, the ablation experiments verify the necessity and contribution of each module in improving model performance, particularly in optimizing the balance between parameter count and model accuracy, providing strong support for subsequent practical applications.

[0132] This invention validated its method on a large-scale, non-public real-world dataset collected by the Railway 6C system. Results show that, compared to the YOLOv10-n baseline, the proposed model achieves a 2.6% improvement in mAP@0.5 while only increasing computational complexity (FLOPs) by 17.9%. Compared to the superior-performing YOLOv10-s, with only a 0.5% difference in mAP@0.5, the computational complexity and parameter count are significantly reduced by 60.1% and nearly half, respectively. This demonstrates that the proposed method effectively alleviates the trade-off between accuracy and efficiency, providing a more practical and efficient solution for real-time edge detection of defects in overhead contact line insulators.

[0133] This invention combines Efficient-ViT and WIoU in its overall design, enabling the network to not only possess powerful feature extraction capabilities but also improve the accuracy of detection results through more precise bounding box prediction. Multi-scale feature fusion and a refined loss function design further optimize the network's performance, making it outstanding in handling complex scenes and multi-scale targets, achieving a good balance between speed and accuracy.

[0134] (1) A YOLOv10-based algorithm for detecting defects in contact network insulators is proposed. By introducing the EfficientViT attention module, the algorithm can capture the diverse features and global context information of insulators more effectively, thereby improving the defect detection effect while reducing the computational cost and detection time of the model.

[0135] (2) Wise-IoU was introduced as the perceptual loss function to optimize bounding box regression, improving the accuracy of locating complex defects. Its effectiveness in detecting defects in contact network insulators was verified through ablation experiments. Experimental results show that Wise-IoU significantly improves the model's localization accuracy, especially in cases of inconsistent labeling or ambiguous defect boundaries. Specifically, after using Wise-IoU, the model achieved a score of 82% on mAP@0.5, a 0.7% improvement compared to the model without Wise-IoU. This result demonstrates the significant value of Wise-IoU in improving defect detection accuracy and enhancing model robustness.

[0136] (3) Knowledge distillation is employed to further enhance the performance ceiling of the lightweight student model, transferring the knowledge learned by the model with a larger number of parameters to the model with a smaller number of parameters, thereby improving the model performance while achieving model lightweighting. The model has 4.1M parameters, 9.9G FLOPs, and an mAP@0.5 value of 0.82, which is close to 82.4% of the YOLOv10-s model, but the computational cost is reduced by 60%. The backbone network is enhanced by introducing the lightweight EfficientViT module; the Wise-IoU loss function is adopted; and knowledge distillation technology is used (with YOLOv10-s as the teacher model).

[0137] This invention develops a detection model that achieves a better balance between accuracy, efficiency, and model complexity, making it more suitable for deployment and application on edge devices in real-world scenarios such as the railway 6C system. This improves the accuracy and efficiency of insulator defect detection, thereby providing strong support for the sustainable development of the electrified railway industry.

[0138] Another aspect of the present invention provides a defect detection system for insulators of electrified railway contact networks based on enhanced YOLOv10, comprising:

[0139] Image acquisition module: Acquires images of contact wire insulators in the target scene through imaging devices and automated inspection equipment;

[0140] Image processing module: Uses LabelBee annotation tool to annotate the contact wire insulators and defects in the contact wire insulator image, and divides the annotated image into training set and validation set according to a preset ratio;

[0141] Model training module: The model is trained based on the training set and validation set. During the training process, the lightweight EfficientViT module, Wise-IoU loss function and knowledge distillation technology are introduced to obtain a contact wire insulator defect detection model based on YOLOv10.

[0142] Result output module: Input the acquired image of the contact wire insulator to be detected into the contact wire insulator defect detection model to obtain the defect detection result.

[0143] This invention presents an electrified railway contact network insulator defect detection system based on enhanced YOLOv10. It uses the lightweight YOLOv10-n as its basic framework, integrating a lightweight EfficientViT module into the YOLOv10-n backbone network. This enhances the model's ability to capture global features and diverse details of insulators while controlling computational overhead. By introducing the Wise-IoU loss function, this perceptual loss is used to optimize bounding box regression, improving the localization accuracy of small-sized, irregularly shaped defects and accelerating model convergence. Furthermore, knowledge distillation technology is employed, using a pre-trained YOLOv10-s model as a teacher to guide the training of the enhanced YOLOv10-n student model. This effectively improves its detection performance while maintaining a lightweight design, making it suitable for resource-constrained scenarios.

[0144] In another aspect of the present invention, the electronic device includes: a processor, a memory, a communication bus, and a communication interface.

[0145] in:

[0146] The processor, memory, and communication interface communicate with each other via a communication bus.

[0147] A communication interface is used to communicate with other electronic devices or servers.

[0148] The processor is used to execute programs, specifically, to perform any of the steps of the electrified railway contact wire insulator defect detection method based on enhanced YOLOv10 in the above embodiments.

[0149] Specifically, the program may include program code, which includes computer operation instructions.

[0150] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0151] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0152] Specifically, the program can be used to cause the processor to execute the steps of any of the electrified railway contact network insulator defect detection methods based on enhanced YOLOv10 described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed in any of the above-described electrified railway contact network insulator defect detection methods based on enhanced YOLOv10, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment and modules can be referred to the corresponding process descriptions in the foregoing method embodiments.

[0153] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods of various embodiments of this application.

[0154] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0155] Specific embodiments of the present invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result.

[0156] It should be noted that all directional indications (such as up, down, left, right, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship between the components in a certain order (as shown in the figure). If the specific order changes, the directional indication will also change accordingly.

[0157] In the description of this invention, the terms "first" and "second" are used only for convenience in describing different components or names, and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include at least one of that feature.

[0158] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0159] It should be noted that although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of the present invention.

[0160] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting defects in insulators of electrified railway contact networks based on enhanced YOLOv10, characterized in that, include: Images of overhead contact line insulators in the target scene are collected using imaging devices and automated inspection equipment; The contact wire insulators and defects in the contact wire insulator images are labeled using a labeling tool, and the labeled images are divided into training set and verification set according to a preset ratio; The model is trained based on the training set and the validation set. During the training process, a lightweight EfficientViT module, a Wise-IoU loss function, and knowledge distillation technology are introduced to obtain a YOLOv10-based contact network insulator defect detection model. In the knowledge distillation technology, the teacher model uses YOLOv10-s, and the student model uses YOLOv10-n-EfficientViT. The layer alignment strategy in the knowledge distillation process is to align feature maps with the same spatial resolution in the feature pyramid. The 16th, 19th, and 22nd layer feature maps generated by the teacher model are aligned with the 12th, 15th, and 18th layer feature maps of the student model, respectively. The acquired image of the contact wire insulator to be inspected is input into the contact wire insulator defect detection model to obtain the defect detection result.

2. The method according to claim 1, characterized in that, The backbone of the contact wire insulator defect detection model includes a convolutional layer, a dual-feature fusion module, an SCDown module, an Efficient ViT module, an SPPF module, and a PSA module. Image features are extracted step by step using convolutional layers and a dual-feature fusion module; Downsampling is performed using the SCDown module; Capture diverse features and global contextual information of insulators using the Efficient ViT module; The SPPF and PSA modules are used to further refine the feature information.

3. The method according to claim 2, characterized in that, The Neck section of the contact wire insulator defect detection model includes the Concat module, the Upsample module, and the C2fCIB module. The Concat module is used to fuse features from different layers; The feature map is upsampled using the Upsample module; The C2fCIB module is used to further process the fused features.

4. The method according to claim 3, characterized in that, The Head section of the contact wire insulator defect detection model includes the YOLODetect layer; The YOLODetect layer performs the final target classification and bounding box prediction, outputting the target category and corresponding confidence score.

5. The method according to claim 1, characterized in that, The contact wire insulator defect detection model uses the Wise-IoU loss function to optimize bounding box regression. The loss function is expressed as: in, express IoU The basic loss function, This indicates the distance attention term.

6. The method according to claim 5, characterized in that, The IoU-based basic loss function Defined as: in, The smaller the value, the higher the overlap between the predicted bounding box and the ground truth bounding box; in, and These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. W b and H b This represents the width and height of the smallest closure box that covers both the predicted and actual bounding boxes.

7. A defect detection system for insulators of electrified railway contact networks based on enhanced YOLOv10, characterized in that, include: Image acquisition module: Acquires images of contact wire insulators in the target scene through imaging devices and automated inspection equipment; Image processing module: Uses annotation tools to annotate the contact wire insulators and defects in the contact wire insulator images, and divides the annotated images into training set and validation set according to a preset ratio; Model training module: Model training is performed based on the training and validation sets. During training, a lightweight EfficientViT module, Wise-IoU loss function, and knowledge distillation technique are introduced to obtain a YOLOv10-based contact network insulator defect detection model. In the knowledge distillation technique, the teacher model uses YOLOv10-s, and the student model uses YOLOv10-n-EfficientViT. The layer alignment strategy in the knowledge distillation process is based on aligning feature maps with the same spatial resolution in the feature pyramid. The 16th, 19th, and 22nd layer feature maps generated by the teacher model are aligned with the 12th, 15th, and 18th layer feature maps of the student model, respectively. Result output module: Input the acquired image of the contact wire insulator to be detected into the contact wire insulator defect detection model to obtain the defect detection result.

8. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the electrified railway contact wire insulator defect detection method based on enhanced YOLOv10 as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps in the method for detecting defects in insulators of electrified railway contact networks based on enhanced YOLOv10 as described in any one of claims 1 to 6.

Images