Enhanced YOLOv10-based electrified railway overhead line system insulator defect detection method, system, equipment and medium

Abstract

The invention provides an enhanced YOLOv10-based electrified railway overhead line system insulator defect detection method, system and equipment and a medium. The method comprises the following steps: acquiring an overhead line system insulator image of a target scene through a shooting device and automatic inspection equipment; marking contact network insulators and defects in the contact network insulator image by using a marking tool, and dividing the marked image into a training set and a verification set according to a preset proportion; performing model training on the basis of the training set and the verification set, and introducing a lightweight OfficientViT module, a Wise-IoU loss function and a knowledge distillation technology in the training process to obtain a contact network insulator defect detection model based on YOLOv10; and inputting the obtained image of the to-be-detected overhead line system insulator into the overhead line system insulator defect detection model to obtain a defect detection result. The method provided by the invention improves the accuracy and efficiency of insulator defect detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of catenary insulator defect detection, and particularly relates to an electrified railway catenary insulator defect detection method, system, device and medium based on enhanced YOLOv10. BACKGROUND

[0002] The pantograph-catenary system is the key to obtaining the operating power of modern electrified railways. Among them, the insulator as the core component, its state is directly related to the safety of power transmission and the normal operation of the railway. However, with the expansion of the railway network and the increase of operating speed, traditional manual inspection has been difficult to meet the efficient and reliable maintenance needs. Automatic and intelligent detection technology has emerged as the times require, and the China State Railway Group has developed a high-speed railway power supply safety detection and monitoring system based on computer vision for non-contact detection (referred to as "6C system"), but it is still a challenge to accurately detect the catenary insulator with various components and complex defects.

[0003] The catenary insulator is easily eroded by the environment and mechanically damaged due to long-term exposure to the outdoors, resulting in a variety of defects such as breakage, dirt, and ablation. In addition, there are many types of insulators in the catenary (different in material, structure, and size, such as Figure 2 As shown), and the images collected by the 6C system are subject to changes in lighting, motion blur, and other disturbances, making it a difficult task to accurately identify and locate various defects, especially early-stage minor defects.

[0004] In recent years, target detection technology based on deep learning has been widely explored for the automatic identification of catenary insulator defects, aiming to replace traditional inefficient manual inspection, and mainly divided into two categories: convolutional neural network (CNN) based methods (such as Faster R-CNN, YOLO series) and Transformer based methods (such as ViT, 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 for diverse insulators and defects, especially the computational overhead of complex models often becomes a bottleneck for real-time deployment. Transformer-based methods exhibit strong feature extraction capabilities, but their high computational cost is a major obstacle to their widespread application in real-time railway insulator detection.

[0005] Existing catenary insulator defect detection methods, whether based on CNN or Transformer, often struggle to simultaneously meet the requirements of high accuracy (especially for minor and diverse defects), strong generalization (adaptation to multiple types of insulators), and high efficiency (to meet the real-time processing needs of edge devices) in practical applications. SUMMARY

[0006] Therefore, the application provides an electrified railway catenary insulator defect detection method, system, device and medium based on enhanced YOLOv10 to solve the above problems.

[0007] The application provides an electrified railway catenary insulator defect detection method based on enhanced YOLOv10, which comprises the following steps: acquiring catenary insulator images of a target scene by a shooting device and an automatic inspection device; labeling catenary insulators and defects in the catenary insulator images by using a labeling tool, and dividing the labeled images into a training set and a verification set according to a preset ratio; training a model based on the training set and the verification set, introducing a lightweight EfficientViT module, a Wise-IoU loss function and a knowledge distillation technology in the training process to obtain a catenary insulator defect detection model based on YOLOv10; and inputting acquired catenary insulator images to be detected into the catenary insulator defect detection model to obtain a defect detection result.

[0008] In another implementation manner of the application, the Backbone part of the catenary insulator defect detection model comprises a convolution layer, a double-feature fusion module, an SCDown module, an Efficient ViT module, an SPPF module and a PSA module; image features are gradually extracted through the convolution layer and the double-feature fusion module; down-sampling is performed through the SCDown module; diversified features and global context information of the insulator are captured through the Efficient ViT module; and feature information is further refined through the SPPF module and the PSA module.

[0009] In another implementation manner of the application, the Neck part of the catenary insulator defect detection model comprises a Concat module, an Upsample module and a C2fCIB module; features from different layers are fused through the Concat module; the feature map is up-sampled through the Upsample module; and the fused features are further processed through the C2fCIB module.

[0010] In another implementation manner of the application, the Head part of the catenary insulator defect detection model comprises a YOLODetect layer; final target classification and bounding box prediction are performed through the YOLODetect layer to output the category of the target and the corresponding confidence.

[0011] In another implementation manner of the present application, in the knowledge distillation technology, the teacher model adopts YOLOv10-s, the student model adopts YOLOv10-n-EfficientViT, and the layer alignment strategy in the knowledge distillation process is to align the 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 respectively aligned with the 12th, 15th and 18th layer feature maps of the student model.

[0012] In another implementation manner of the present application, the contact net insulator defect detection model adopts a Wise-IoU loss function to optimize the boundary box regression, and the loss function is expressed as:

[0013] Wherein, The basic loss function of IoU is expressed as: The distance attention term is expressed as.

[0014] In another implementation manner of the present application, the basic loss function of IoU is defined as:

[0015] Wherein, The smaller the value is, the higher the overlap degree of the predicted box and the real box is.

[0016]

[0017] Wherein, And The center point coordinates of the predicted box and the real box are respectively W b And H b The width and height of the minimum closed box covering the predicted box and the real box are represented.

[0018] ​In another aspect of the present application, an electrified railway catenary insulator defect detection system based on enhanced YOLOv10 is provided, comprising: an image acquisition module: acquiring catenary insulator images of a target scene through a shooting device and an automatic inspection device; an image processing module: labeling the catenary insulator and defects in the catenary insulator images using a LabelBee labeling tool, and dividing the labeled images into a training set and a validation set according to a preset ratio; a model training module: training a model based on the training set and the validation 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 a result output module: inputting the obtained catenary insulator image to be detected into the catenary insulator defect detection model to obtain a defect detection result.

[0019] In another aspect of the present application, an electronic device is provided, comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the electrified railway catenary insulator defect detection method based on enhanced YOLOv10 according to any one of the above aspects when executing the computer program. In another aspect of the present application, a computer storage medium is provided, wherein the computer storage medium stores a computer program, and the computer program implements the steps of the electrified railway catenary insulator defect detection method based on enhanced YOLOv10 according to any one of the above aspects when executed by a processor.

[0020] The electrified railway catenary insulator defect detection method based on enhanced YOLOv10 of the present application selects a lightweight YOLOv10-n as a basic framework, integrates a lightweight EfficientViT module in the YOLOv10-n backbone network, enhances the model's ability to capture global features and diversified details of insulators, and controls the calculation overhead at the same time; by introducing a Wise-IoU loss function, the perception loss is used to optimize the bounding box regression, which improves the positioning accuracy of small size and irregular shape defects and accelerates the model convergence; the knowledge distillation technology is used, the pre-trained YOLOv10-s model is used as a teacher to guide the training of the enhanced YOLOv10-n student model, while maintaining the lightweight, effectively improving the detection performance, and making it suitable for resource-limited scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. The advantages and benefits in the solutions will become clear to those skilled in the art by reading the following detailed description of the embodiments. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not considered as limiting the present application. In the drawings: Figure 1 A flowchart of an electric railway catenary insulator defect detection method based on enhanced YOLOv10 according to an embodiment of the present application.

[0022] Figure 2 A schematic diagram of the position distribution of insulators in the catenary according to an embodiment of the present application.

[0023] Figure 3 A schematic diagram of a network structure according to an embodiment of the present application.

[0024] Figure 4 An overview diagram of an EfficientViT enhancement module according to an embodiment of the present application.

[0025] Figure 5 A schematic diagram of the principle of knowledge distillation according to an embodiment of the present application.

[0026] Figure 6 A part of training data images according to an embodiment of the present application.

[0027] Figure 7 A schematic diagram of the data set label and distribution according to an embodiment of the present application.

[0028] Figure 8 A schematic diagram of night visualization contrast analysis according to an embodiment of the present application.

[0029] Figure 9 A schematic diagram of multi-insulator visualization contrast analysis under intermediate light conditions according to an embodiment of the present application.

[0030] Figure 10 A schematic diagram of broken ceramic insulator visualization contrast analysis according to an embodiment of the present application.

[0031] Figure 11 A comparison diagram of the mAP@0.5 of each model in the YOLOv10 series changing with the training round number according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order for those skilled in the art to better understand the improvements made by the present embodiments relative to the prior art, the prior art will be briefly introduced first.

[0033] The reliability of electrified railway catenary insulators is directly related to the operation safety. Automated defect detection is a key technology to ensure its reliability. However, in actual deployment, especially on edge devices with limited computing power, existing methods often struggle to balance high detection accuracy and high operational efficiency.

[0034] For example, advanced YOLOv10 directly applied to the railway inspection edge device with limited computing power faces a significant precision-efficiency dilemma. The lightweight version (such as YOLOv10-n) is fast, but it often lacks precision in the face of the complexity of insulator detection, making it difficult to reliably identify small defects. While the high-performance version (such as YOLOv10-s) has improved accuracy, its high computational cost and model size limit real-time deployment. In addition, the general design of YOLOv10 may not fully consider the specific structure and defect characteristics of insulators, resulting in poor performance in small target and edge detail detection.

[0035] Vision Transformer (ViT) model was proposed in 2020 and published in 2021. It introduced the Transformer architecture originally used in the natural language processing (NLP) field to the computer vision (CV) field, mainly for image classification tasks. Studies have shown that ViT performs better as the size of the training data set increases. Compared with convolutional neural networks (CNN), ViT has stronger global information capture and long-range interaction capabilities, especially when scaling up the model size and training data, it can exhibit better classification accuracy than CNN. However, in high-resolution dynamic scene applications, the performance of ViT is lower than that of CNN.

[0036] In order to make the personnel in the art better understand the technical solutions in the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and in detail below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all embodiments. Based on the embodiments in the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art should be within the scope of protection of the embodiments of the present application.

[0037] Figure 1 A flowchart of an electrified railway catenary insulator defect detection method based on enhanced YOLOv10 is provided in the embodiments of the present application, as shown in Figure 1 The present embodiment mainly includes: S101, acquiring the catenary insulator image of the target scene through the shooting device and the automated inspection equipment.

[0038] S102, the contact net insulator image and defects in the contact net insulator image are labeled by using a labeling tool (for example, a LabelBee labeling tool), and the labeled image is divided into a training set and a verification set according to a preset ratio.

[0039] S103, model training is performed based on the training set and the verification set, a lightweight EfficientViT module, a Wise-IoU loss function and a knowledge distillation technology are introduced in the training process, and a contact net insulator defect detection model based on YOLOv10 is obtained.

[0040] S104, the obtained contact net insulator image to be detected is input into the contact net insulator defect detection model, and a defect detection result is obtained.

[0041] The electrical railway contact net insulator defect detection method based on the enhanced YOLOv10 of the present application selects a lightweight YOLOv10-n as a basic framework, integrates a lightweight EfficientViT module in the YOLOv10-n backbone network, enhances the model's ability to capture global features and diversified details of the insulator, and controls the calculation overhead at the same time; by introducing a Wise-IoU loss function, the perception loss is used to optimize the boundary box regression, the positioning accuracy of small size and irregular shape defects is improved, and the model convergence is accelerated; the knowledge distillation technology is used, the pre-trained YOLOv10-s model is used as a teacher to guide the training of the enhanced YOLOv10-n student model, while keeping the lightweight, the detection performance is effectively improved, and it is suitable for resource limited scenes.

[0042] In another implementation manner of the present application, the Backbone part of the contact net insulator defect detection model includes a convolution layer, a double feature fusion module, an SCDown module, an EfficientViT module, an SPPF module and a PSA module; the image features are gradually extracted through the convolution layer and the double feature fusion module; the SCDown module is used for down-sampling; the EfficientViT module is used for capturing diversified features and global context information of the insulator; and the SPPF module and the PSA module are used for further refining the feature information.

[0043] Exemplarily, as Figure 3As shown, based on the YOLOv10-n framework, the network structure is designed and optimized. In the Backbone part, the network gradually extracts image features through convolution layers (Conv) and C2f modules, and introduces SCDown modules for down-sampling. In the catenary defect detection application, YOLOv10 has advanced target detection capabilities, but in order to achieve high detection accuracy, YOLOv10 needs more computing resources and complex feature extraction layers. Although this design improves the detection accuracy, for small defects or subtle features (such as cracks, slight wear), the model may need more fine-grained feature processing and deeper network structure. However, more network layers result in more computational complexity, leading to a decrease in model processing speed and affecting the efficiency of real-time detection. Therefore, the Efficient ViT (Visual Transformer) module is introduced, as shown. Figure 4 This module enhances the network's ability to capture long-range dependency features, especially suitable for feature extraction in complex scenarios. Subsequently, through SPPF (Fast Spatial Pyramid Pooling) and PSA modules, the network further refines the feature information, providing rich multi-scale features for subsequent detection.

[0044] It should be understood that the adoption of the EfficientViT module optimizes the YOLOv10 backbone, and the EfficientViT model improves the computational efficiency while maintaining the advantages of ViT global feature extraction. EfficientViT replaces the traditional softmax attention mechanism with linear attention, solving the computational bottleneck of ViT, and enhancing the local feature extraction capability of the model by introducing deep convolution, thereby reducing the complexity of linear computation.

[0045] In this invention, hierarchical local attention is adopted, which only performs attention computation on the local information of the input image, reducing the complexity while retaining the ability to capture global information. In addition, different methods are used for different network levels in feature extraction. Specifically, convolutional networks are used for low-level feature extraction, while Transformers are used for high-level global feature representation. This hybrid architecture takes full advantage of the locality of convolutional networks and the globality of Transformers, making the model more lightweight and efficient. Considering that the catenary insulator defect detection model needs to deploy the detection algorithm on edge devices with limited computing resources, it is necessary to reduce the parameter quantity and computational complexity of the model to reduce the deployment cost. However, reducing the network structure inevitably leads to a loss of accuracy. To solve the above problems, a sparse computing strategy is adopted, which only performs accurate calculation in the required area, reducing the invalid calculation of the entire image, significantly speeding up the detection speed while ensuring the prediction accuracy.

[0046] In another implementation manner of the present application, the Neck part of the catenary insulator defect detection model comprises a Concat module, an Upsample module and a C2fCIB module; the features from different layers are fused through the Concat module; the feature maps are upsampled through the Upsample module; and the fused features are further processed through the C2fCIB module.

[0047] Exemplarily, in the Neck part, the network adopts the Concat operation to fuse the features from different layers, ensures that the features of different scales are considered at the same time, and thus improves the detection capability for targets of different sizes. In addition, the Upsample operation is adopted to upsample the feature maps, which is helpful to retain more detailed features. The C2fCIB module is also introduced in the Neck part to further process the fused features, so as to ensure that the features are fully expressed before the detection layer.

[0048] In another implementation manner of the present application, the Head part of the catenary insulator defect detection model comprises a YOLODetect layer; the final target classification and bounding box prediction are performed through the YOLODetect layer, and the category and corresponding confidence of the target are output.

[0049] Exemplarily, in the Head part, the network performs the final target classification and bounding box prediction through the YOLODetect layer, and outputs the category and corresponding confidence of the target.

[0050] In another implementation manner of the present application, in the knowledge distillation technology, the teacher model adopts YOLOv10-s, the student model adopts YOLOv10-n-EfficientViT, and the layer alignment strategy in the knowledge distillation process is to align the 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.

[0051] Exemplarily, the knowledge distillation technology is adopted, the YOLOv10-s with large parameters and complex calculation is taken as the teacher model to guide the student model YOLOv10-n-EfficientViT with small parameters and efficient calculation, so that the student model can learn more rich and detailed features than YOLOv10-n. Through learning the output (such as soft label or intermediate layer feature) of the teacher model, the student model can absorb the beneficial information in the teacher model while retaining the advantages of its own architecture, so as to effectively improve its prediction capability. Figure 5As shown, the Detect layer of YOLOv10-s and the Detect layer of YOLOv10-n-EfficientViT are implemented with logical distillation using the BCKD (Bidirectional Contrastive Knowledge Distillation) method, which combines the advantages of knowledge distillation and contrastive learning, can more efficiently train small-scale student models while improving model performance, and is particularly suitable for resource-constrained application environments.

[0052] In the knowledge distillation process, the layer alignment strategy of the teacher model (YOLOv10-s) and the student model (YOLOv10-n-EfficientViT) is based on the feature maps of the same spatial resolution in the feature pyramid. Specifically, the 16, 19 and 22 layer feature maps generated by the teacher model are aligned with the 12, 15 and 18 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 matched 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 the present application adjusts the channel number of the feature maps of the teacher model to be consistent with the student model through 1x1 convolution, and then calculates the loss. This method can flexibly handle the difference in feature dimensions and ensure effective alignment and learning of feature representations during the distillation process.

[0053] In another implementation manner of the present application, the overhead line insulator defect detection model optimizes the boundary box regression using a Wise-IoU loss function, which is expressed as:

[0054] wherein, represents IoU the basic loss function of IoU, represents the distance attention term.

[0055] In another implementation manner of the present application, the basic loss function of IoU is defined as:

[0056] wherein, The smaller the value, the higher the overlap between the predicted box and the real box.

[0057]

[0058] wherein, and are the center point coordinates of the predicted box and the real box, respectively; W b andH b The width and height of the minimum closed box representing the prediction box and the real box.

[0059] Exemplarily, the loss function of the entire network adopts a combination of L_IoU and WIoU, wherein the WIoU (Weighted IoU) can more accurately process the size, position and shape of the target bounding box, thereby improving the accuracy of target detection.

[0060] The Wise-IoU (WIoU) loss function plays a key role in the target detection task, especially in scenarios involving small-sized or irregularly shaped defects such as suspension insulator detection, and its design exhibits unique advantages. The WIoU loss mainly consists of two parts: an IoU-based loss term and a distance attention mechanism.

[0061] The IoU-based loss term ( ) is constructed based on the Intersection over Union (IoU). IoU is used to quantify the degree of overlap between the predicted bounding box and the ground truth box, and the calculation formula is as follows:

[0062] wherein, W i and H i represent the width and height of the intersection area of the prediction box and the real box, S u and the area of the union area of the two is

[0063] The IoU-based loss function is defined as:

[0064] wherein, The smaller the value, the higher the degree of overlap between the prediction box and the real box, and vice versa. This loss term provides a basic geometric constraint for model optimization.

[0065] To make up for the deficiency of only focusing on the overlapping area and ignoring the geometric relationship of the non-overlapping area, the Wise-IoU introduces a distance attention term , which is defined as follows:

[0066] wherein, and are the center point coordinates of the prediction box and the real box, respectively;W b and H b represent the width and height of the smallest enclosing box of the predicted and ground truth boxes. The superscript indicates that it is used when backpropagating the gradient; W b and H b are treated as constants (i.e., detached from the computation graph), which is done to prevent them from producing gradients that can interfere with the training convergence. By computing an exponential function of the center point distance, a larger loss weight is assigned to low-quality anchor boxes that are far away, encouraging the model to pay more attention to predictions with better geometric alignment. This mechanism is particularly important for improving the positioning accuracy and robustness of the model when detecting small or irregularly shaped defects.

[0067] The complete loss function of Wise-IoU-v1 combines the two parts above:

[0068] This form weights the basic by the distance attention term , which synergistically improves the positioning performance of the model.

[0069] In another implementation of the present application, Wise-IoU also introduces a dynamic non-monotonic focusing mechanism, which dynamically adjusts the contribution of the anchor box to the loss function by evaluating its "outlierness" . The definition of is as follows:

[0070] where also represents the value detached from the computation graph, while is the exponential moving average of , reflecting the average IoU loss level of recent samples. The smaller the value, the higher the quality of the anchor box (relative to the recent average level).

[0071] Based on the dynamic focusing coefficient , the is defined as:

[0072] wherein, and are hyperparameters that control the shape of the focusing curve.

[0073] The core idea of this mechanism is: 1) For high-quality anchor boxes (small ), assign a small gradient weight to prevent the model from overfitting to high-quality samples.

[0074] 2) For very poor quality anchor boxes (large ), also appropriately reduce their gradient weight to avoid the negative impact of these outliers on model training, and focus the main optimization focus on anchor boxes of "ordinary quality".

[0075] The final loss function that applies the dynamic focusing mechanism achieves gradient adjustment by multiplying as multiplier (or otherwise combining). Experiments show that this dynamic focusing mechanism significantly improves the defect localization accuracy of the model and enhances the model's robustness to inconsistent or low-quality labeled samples.

[0076] Embodiment 1 The present application studies the contact net insulator and defects, and the main detection targets are two categories of insulators (Insulator) and insulator defects (Defect). The data set used comes from various shooting means under the railway 6C detection system, as shown in Figure 6 , including high-resolution images shot by 3C and 4C devices and images shot by other automated inspection equipment. These images cover scenes of different light, angle and clarity, ensuring the diversity of the data set.

[0077] The contact net insulator and defects are labeled using the LabelBee labeling tool, and the data set is randomly divided according to the ratio of training set:validation set=8:2, and the proportion distribution of each type of insulator in the data set is shown in Table 1, and the data set label distribution is shown in Figure 7 .

[0078] Table 1 Mixed distribution of each type of insulator in the data set

[0079] The configuration environment used in the experiment is shown in Table 2.

[0080] Table 2 Experiment configuration table

[0081] To verify the effect of the algorithm, the F1 score, AP, accuracy, precision (P) and recall (R) are used as evaluation indexes of the catenary insulator defect detection algorithm, and the formulas are as follows:

[0082] Wherein, TP is true positive (number of detected defects), FP is false positive (number of false detected defects).

[0083]

[0084] Wherein, FN is false negative (number of missed defects).

[0085]

[0086] The F1 score is the harmonic mean of precision and recall, and is used to comprehensively evaluate the accuracy and integrity of the model.

[0087] AP is usually obtained by calculating the precision and recall at different thresholds, and taking the area, p(r) The function representing the accuracy.

[0088] Experimental results and analysis: To visually evaluate the actual performance of different models in the catenary insulator defect detection task, three representative scenes are selected for visual result comparison. Figure 8 The detection of defective wrist and needle insulators under complex night lighting conditions is shown; Figure 9 Then focus on the dense scene of multiple target normal insulators at night; Figure 10 Focus on the recognition of damaged ceramic insulators during the day. Each scene compares FasterNet-t0, YOLOv8-n / s, YOLOv10-n / s, RT-DETR-l and the enhanced YOLOv10-n model proposed in the application. In the figure, (a) is the original image, (b) to (h) correspond to the inference results of each model in turn.

[0089] In Figure 8The performance of lightweight models such as FasterNet-t0 (b) and YOLOv8-n (c) is significantly affected by the complex night light and multi-type defect scene, with problems such as defect omission or extremely low confidence, and inaccurate positioning of the bounding box. The performance of YOLOv10-s (f) and RT-DETR-l (g) is usually good for defect detection, and the bounding box is relatively accurate. In contrast, the method (h) proposed in the application can not only stably detect the defects of the insulator under this challenging condition, but also has a high confidence score, which is no worse than the advanced RT-DETR model, indicating that it has good adaptability to complex lighting and diversified targets.

[0090] Figure 9 The performance of the model in the night dense normal insulator scene was investigated, which requires the model to have consistent recognition and false alarm suppression capabilities. The lightweight model (b) missed some targets in the dense target scene, and the confidence score of all targets was low. It is worth noting that models (d) and (g) misjudged the light spots or other components in the background as defects. The method (h) of the application and the stable-performing baseline (such as YOLOv10-s (f)) can reliably detect all insulator targets in the field of view, and do not introduce obvious false alarms, showing good target recognition stability and background suppression ability.

[0091] Finally, Figure 10 The generalization detection ability of the model for damaged ceramic insulators in daytime scenes was evaluated. Some comparative models (b, c, e) have limited recognition ability for this specific damage mode, resulting in low confidence, and even the YOLOv10-n model has obvious repeated anchor boxes. The method (h) of the application also performs well in this scene, accurately identifying and framing the damaged area of the ceramic insulator, with accurate positioning and high confidence. Its performance is very close to that of RT-DETR-l (g) (although this is only the smallest parameter model of RT-DETR), and significantly better than the other several comparative models, which confirms that the proposed model has good generalization ability across materials and defect morphologies.

[0092] In summary, through visual analysis of different lighting, target types, defect conditions, and material, etc. in various practical scenes, the enhanced YOLOv10-n model proposed in the application exhibits consistent and reliable detection performance. In terms of key defect positioning accuracy, confidence reliability, and complex environment robustness, the model usually reaches the level comparable to YOLOv10-s, which consumes significantly more computing resources, and consistently outperforms other lightweight baseline models.

[0093] To comprehensively evaluate the performance of the invented method (Ours), it was quantitatively compared with a series of representative baseline models on the reserved test set, with the results shown in Table 3 and Figure 11 Fig. 3. 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 an input resolution of 640x640. Key indicators of interest include mAP@0.5, average precision for insulator bodies ) and defects , precision (P), recall (R), and maximum F1 score at the optimal threshold, as well as the parameter quantity (Params), computational complexity (FLOPs), and inference speed (FPS) to measure model efficiency.

[0094] The experimental results clearly demonstrate the excellent balance between precision and efficiency achieved by the invented method. In terms of the core mAP@0.5 indicator, the model of the invention achieved 82.0%, not only significantly surpassing all lightweight opponents (such as YOLOv10-n with 79.4%), but also being highly competitive compared to high-performance YOLOv10-s (82.4%) and YOLOv8-s (81.15%), only slightly lower than the computationally resource-intensive RT-DETR-l (83.3%).

[0095] Table 3 Results of quantitative experimental comparison on the dataset

[0096] Bold indicates statistically significant improvements (p<0.05).

[0097] Notably, compared to the YOLOv10-n framework, the invented method achieved a 2.6% mAP improvement, directly demonstrating the effectiveness of the introduced improvements such as EfficientViT, Wise-IoU, and knowledge distillation. In fine classification categories, the model performed particularly well in recognizing insulator bodies, reaching 95.1%, the highest value among all comparison models. For more challenging defect detection, it also reached 68.9%, maintaining a similar level to top models.

[0098] Further analysis of precision-recall performance (P / R), the model of the present application achieved a recall rate of 0.87, which is on par with YOLOv8-s, indicating that it can effectively capture most of the targets. At the same time, the precision rate remains at a high level of 0.963, which means a lower false positive rate. In summary, the model has achieved a very ideal balance between accurate identification and false positive avoidance, which is crucial for practical applications.

[0099] However, the most attractive advantage of the method of the present application lies in its outstanding computing efficiency. The model parameter amount is only 4.1M, and the calculation amount is 9.9G FLOPs. Compared with YOLOv10-s with similar performance, this means that the parameter amount is reduced by nearly half, and the calculation amount is reduced by more than 60%. This significant efficiency improvement is of decisive significance for resource-constrained edge deployment scenarios. Compared with RT-DETR-l with the highest accuracy, the computing cost is even lower by an order of magnitude. Although compared with the original YOLOv10-n, the calculation amount increases slightly (about 17.9%), but this part of the overhead brings significant accuracy returns, showing a very high cost-effectiveness ratio. In terms of actual inference speed, the performance of 47.7 FPS is also better than YOLOv10-s and RT-DETR-l, which can meet the needs of many real-time detection.

[0100] In summary, the quantitative evaluation results strongly prove that the enhanced YOLOv10-n model proposed in the present application successfully addresses the challenge of balancing accuracy and efficiency in catenary insulator detection. It achieves detection accuracy comparable to one of the current best models (YOLOv10-s) with significantly reduced computing and storage requirements, and maintains an advantage in inference speed. This careful trade-off between performance and efficiency makes the model an ideal candidate for deploying advanced defect detection capabilities on actual railway inspection systems, especially on edge computing platforms.

[0101] Example 2 Ablation experiment: To further demonstrate the feasibility of the algorithmic innovations mentioned in the study, an ablation experiment was conducted to analyze the effect of each module. The ablation experiment results are shown in Table 4.

[0102] Table 4 Ablation experiment comparison results, bold represents the optimal result

[0103] The present application verifies the contribution of each module to the model performance through the ablation experiment, especially the impact on balancing accuracy and efficiency. Four different model configurations are designed, and their performances in multiple performance indicators are compared.

[0104] Firstly, the model without knowledge distillation (w / o Distillation) shows a significant performance drop. Although the precision (P) remains high (0.927), the F1 score drops to 0.71, and the recall (R) drops to 0.73, with mAP@0.5 being 80.4. This indicates that knowledge distillation plays a crucial role in improving feature extraction capabilities and enhancing model generalization performance. By transferring the knowledge of the YOLOv10-s model to the YOLOv10-n, the student model can better learn fine-grained features with fewer parameters, thus improving precision and recall.

[0105] Secondly, after removing EfficientViT (w / o EfficientViT), the model's performance decreases. Although the precision (0.938) slightly improves, the F1 score drops to 0.69, and the recall drops to 0.72, with mAP@0.5 being 79.8. EfficientViT significantly optimizes the efficiency of feature extraction, reduces computational cost, and maintains high precision. After removing this module, the model still achieves good precision, but the overall performance decreases, indicating that EfficientViT plays a key role in balancing feature extraction and computational efficiency.

[0106] Further, after removing Wise-IoU (w / o WIoU), although the precision (0.958) and recall (0.77) remain high, mAP@0.5 is 81.3, but it leads to a decrease in F1 score (to 0.78). This indicates that Wise-IoU, as a perception loss function, plays an important role in optimizing the positioning accuracy in the target detection task. After removing this module, the model shows a decrease in precision when dealing with complex defect regions, especially in the detection of small defects, affecting the improvement of recall and F1 score.

[0107] The comprehensive configuration (Ours) combines all the modules: knowledge distillation, EfficientViT and Wise-IoU, and shows the best performance. Specifically, the F1 value is 0.78, the precision is 0.963, the recall is 0.78, and the mAP@0.5 is 82. This shows that after reasonably combining these modules, the model achieves a balance between precision and efficiency. By strengthening feature extraction and optimizing the perception loss function, the model proposed in the application can achieve higher precision with fewer parameters, especially in the balance between detection precision and recall. In addition, the synergy of knowledge distillation and EfficientViT effectively improves the computational efficiency, so that the model can still obtain relatively excellent performance in the case of limited resources. In summary, the results of the ablation experiment verify the necessity and contribution of each module in improving the performance of the model, especially in the balance optimization between the number of parameters and the precision of the model, which provides strong support for subsequent practical applications.

[0108] The application verifies the method on a large-scale real dataset collected by a railway 6C system. The results show that compared with the YOLOv10-n baseline, the mAP@0.5 of the model of the application is improved by 2.6%, and the computational amount (FLOPs) is only increased by 17.9%. Compared with the YOLOv10-s with better performance, the computational amount and the parameter amount are significantly reduced by 60.1% and nearly half respectively, with only a difference of 0.5% in mAP@0.5. This proves that the method of the application effectively alleviates the contradiction between precision and efficiency, and provides a more practical and efficient solution for edge real-time detection of catenary insulator defects.

[0109] The overall design of the application combines Efficient-ViT and WIoU, so that the network not only has strong feature extraction capability, but also can improve the accuracy of the detection result through more accurate bounding box prediction. Multi-scale feature fusion and fine loss function design further optimize the performance of the network, so that it performs well in processing complex scenes and multi-scale targets, achieving a good balance between speed and accuracy.

[0110] (1) A catenary insulator defect detection algorithm based on YOLOv10 is proposed, which introduces an EfficientViT attention module to more effectively capture the diversified features and global context information of the insulator, while reducing the computational cost and detection time of the model.

[0111] (2) Introducing Wise-IoU as a perception loss function to optimize boundary box regression, improving the positioning accuracy of complex defects, and verifying its effectiveness in catenary insulator defect detection through ablation experiments. The experimental results show that Wise-IoU significantly improves the positioning accuracy of the model, especially in the case of inconsistent labeling or ambiguous defect boundaries. Specifically, after using Wise-IoU, the model's score on mAP@0.5 reached 82%, an increase of 0.7% compared to the model without using WIoU. This result shows that Wise-IoU has important value in improving defect detection accuracy and enhancing model robustness.

[0112] (3) Using knowledge distillation to further improve the performance of lightweight student models, transferring the knowledge learned by larger parameter models to smaller parameter models, achieving model lightweight while improving model effectiveness. The model parameter is 4.1M, FLOPs is 9.9G, and mAP@0.5 value reaches 0.82, which is similar to the 82.4% of YOLOv10-s model, but the calculation amount is reduced by 60%. By introducing the lightweight EfficientViT module to enhance the backbone network; using the Wise-IoU loss function; and using knowledge distillation technology (with YOLOv10-s as the teacher model).

[0113] The present application develops a detection model that achieves a better balance between precision, efficiency and model complexity, making it more suitable for deployment and application on edge devices in actual scenarios such as railway 6C systems, improving the accuracy and efficiency of insulator defect detection, and providing strong support for the sustainable development of the electrified railway industry.

[0114] In another aspect of the present application, an electrified railway catenary insulator defect detection system based on enhanced YOLOv10 is provided, comprising: An image acquisition module: acquiring catenary insulator images of the target scene through a shooting device and an automatic inspection device; An image processing module: labeling the catenary insulator and defects in the catenary insulator images using the LabelBee labeling tool, and dividing the labeled images into a training set and a validation set according to a predetermined ratio; A model training module: training the model based on the training set and the validation set, introducing a lightweight EfficientViT module, a Wise-IoU loss function and a knowledge distillation technology during the training process, and obtaining a catenary insulator defect detection model based on YOLOv10; A result output module: inputting the acquired catenary insulator image to be detected into the catenary insulator defect detection model to obtain the defect detection result.

[0115] The electrified railway catenary insulator defect detection system based on the enhanced YOLOv10 of the application selects a lightweight YOLOv10-n as a basic framework, integrates a lightweight EfficientViT module in the YOLOv10-n backbone network, enhances the model's ability to capture global features and diversified details of insulators, and controls the calculation overhead; by introducing a Wise-IoU loss function, the perception loss is used to optimize the boundary box regression, improve the positioning accuracy of small size and irregular shape defects, and accelerate the model convergence; the knowledge distillation technology is used, the pre-trained YOLOv10-s model is used as a teacher to guide the training of the enhanced YOLOv10-n student model, while keeping the lightweight, effectively improving the detection performance, and making it suitable for resource-limited scenarios.

[0116] In another aspect of the application, the electronic device includes a processor, a memory, and a communication bus, a communication interface.

[0117] Among them: The processor, the memory and the communication interface complete the communication with each other through the communication bus.

[0118] The communication interface is used for communication with other electronic devices or servers.

[0119] The processor is used for executing programs, and specifically can execute the steps of the electrified railway catenary insulator defect detection method based on the enhanced YOLOv10 in any of the above embodiments.

[0120] Specifically, the program can include program code including computer operation instructions.

[0121] The processor can be a central processing unit CPU, or an application specific integrated circuit ASIC, or one or more integrated circuits configured to implement one or more embodiments of the application. One or more processors included in the smart device can be the same type of processor, such as one or more CPUs; or can be different types of processors, such as one or more CPUs and one or more ASICs.

[0122] The memory is used to store programs. The memory can include a high-speed RAM memory, and can also include a non-volatile memory, such as at least one disk memory.

[0123] The program can be specifically used for enabling the processor to execute to realize the steps of any one of the electric railway catenary insulator defect detection methods based on enhanced YOLOv10 described in the embodiments. The specific implementation of each step in the program can refer to the corresponding description in the steps and units executed by any one of the electric railway catenary insulator defect detection methods based on enhanced YOLOv10 described above, and details are not described here. It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, the specific working processes of the devices and modules described above can refer to the corresponding process description in the foregoing method embodiments.

[0124] The exemplary embodiments of the present application also provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the method of the embodiments of the present application.

[0125] The method according to the embodiments of the present application described above 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, a RAM, a floppy disk, a hard disk or a magneto-optical disk, or be downloaded through a network originally stored in a remote recording medium or a non-transitory machine-readable medium and then stored in a local recording medium, so that the method described herein can be processed by such software using a general computer, a special purpose processor or programmable or special hardware (such as ASIC or FPGA). It can be understood that the computer, processor, microprocessor controller or programmable hardware includes a storage component (for example, RAM, ROM, flash memory, etc.) that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, processor or hardware, the method described herein is implemented. In addition, when a general computer accesses the code for implementing the method shown herein, the execution of the code will convert the general computer into a special computer for executing the method shown herein.

[0126] So far, specific embodiments of the present application have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve desirable results. In addition, the processes depicted in the figures do not necessarily require the particular order shown, or sequential order, to achieve the desired results.

[0127] It should be noted that all directional indications, such as upper, lower, left, right, back, etc., in the embodiments of the present application are only used to explain the relative positional relationship between components, etc. in a certain specific order (as shown in the drawings), and if the specific order changes, the directional indications also change accordingly.

[0128] In the description of the present application, the terms "first", "second" are only used for the convenience of describing different components or names, and cannot be understood as indicating or implying the order relationship, relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can be explicitly or implicitly included at least one of the features.

[0129] 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 application belongs. The terminology used in the description herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0130] It should be noted that, although the specific embodiments of the present application are described in detail with reference to the accompanying drawings, it should not be understood as limiting the scope of protection of the present application. Various modifications and variations made by those skilled in the art within the scope described in the claims are still within the scope of protection of the present application.

[0131] The examples of the embodiments of the present application are intended to simply illustrate the technical features of the embodiments of the present application, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present application, and are not improper limitations on the embodiments of the present application.

[0132] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting defects in electrified railway contact network insulators based on enhanced YOLOv10, characterized in that: include: Capture images of contact network insulators in target scenes using a camera and automated inspection equipment; Annotating the contact network insulators and defects in the contact network insulator image using an annotation tool, and dividing the annotated image into a training set and a validation set according to a preset ratio; Model training is performed 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. The acquired image of the contact network insulator to be inspected is input into the contact network insulator defect detection model to obtain a defect detection result.

2. The method according to claim 1, wherein The backbone part of the contact network insulator defect detection model includes a convolution layer, a dual feature fusion module, an SCDown module, an Efficient ViT module, an SPPF module and a PSA module; Image features are gradually extracted through convolutional layers and dual-feature fusion modules; Downsampling is performed through the SCDown module; Capture the diverse features and global context information of insulators through the Efficient ViT module; The feature information is further refined through the SPPF module and PSA module.

3. The method according to claim 2, characterized in that The Neck part of the contact network insulator defect detection model includes a Concat module, an Upsample module, and a C2fCIB module; The features from different layers are fused through the Concat module; Upsample the feature map through the Upsample module; The fused features are further processed through the C2fCIB module.

4. The method according to claim 3, characterized in that The head part of the contact network insulator defect detection model includes a YOLODetect layer; The final target classification and bounding box prediction are performed through the YOLODetect layer, and the target category and corresponding confidence level are output.

5. The method according to claim 1, wherein 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 based on aligning feature maps with the same spatial resolution in the feature pyramid. The 16-, 19-, and 22-layer feature maps generated by the teacher model are aligned with the 12-, 15-, and 18-layer feature maps of the student model, respectively.

6. The method according to claim 1, characterized in that The contact network 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 is Represents the distance attention item.

7. The method according to claim 6, characterized in that The IoU basic loss function Defined as: in, The smaller the value, the higher the overlap between the predicted box and the true box; in, and are the center point coordinates of the predicted box and the real box respectively; W b and H b Represents the width and height of the minimum enclosing box covering the predicted box and the true box.

8. An electrified railway contact network insulator defect detection system based on enhanced YOLOv10, characterized in that: include: Image acquisition module: collects images of contact network insulators in the target scene through shooting devices and automated inspection equipment; Image processing module: using a labeling tool to label the contact network insulator and defects in the contact network insulator image, and dividing the labeled image into a training set and a validation set according to a preset ratio; Model training module: Model training is performed 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 YOLOv10-based contact network insulator defect detection model; Result output module: inputs the acquired catenary insulator image to be detected into the catenary insulator defect detection model to obtain defect detection results.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method for detecting defects in insulators of an electrified railway contact network based on enhanced YOLOv10 are implemented as described in any one of claims 1 to 7.

10. 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 electrified railway contact network insulators based on enhanced YOLOv10 according to any one of claims 1 to 7.

Images