Catenary dropper state detection method based on improved YOLOv11s
By improving the YOLOv11s algorithm for overhead contact line dropper status detection, multi-scale feature fusion and occlusion area feature extraction are achieved. This solves the problems of low efficiency in manual detection and missed detection and false detection in existing algorithms, improves the accuracy and robustness of detection, and ensures the safe operation of high-speed railways.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2025-04-22
- Publication Date
- 2026-07-21
AI Technical Summary
The existing overhead contact line dropper condition detection methods suffer from low efficiency due to manual inspection and the existence of algorithms that miss or misdetect, failing to meet the high-efficiency and accurate inspection requirements of high-speed railways.
An improved YOLOv11s algorithm is adopted, which achieves multi-scale feature fusion by adding a small target detection layer to the neck network and combining it with an improved detection head network. The SEAM attention mechanism and depthwise separable convolution are also embedded to improve the small target detection capability.
It improves the accuracy and robustness of small target detection, reduces missed and false detections, ensures the reliability and integrity of fault detection of overhead contact line droppers in high-speed railways, and supports the safe operation of high-speed railways.
Smart Images
Figure CN120451064B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of overhead contact line dropper detection technology, and in particular to an overhead contact line dropper status detection method based on an improved YOLOv11s. Background Technology
[0002] In high-speed railway systems, the overhead contact line is the core of train power transmission, and the droppers, as a key component, bear the responsibility of supporting the power transmission line and buffering vibration and impact. Their condition is crucial to the safe and stable operation of the railway. However, droppers are subjected to the dual loads of vehicles and the environment for extended periods, making them prone to loosening and deformation, threatening power supply stability and train operation safety. Therefore, routine and precise inspection of droppers is essential. In the past, manual inspection was inefficient and prone to errors. While the introduction of 4C devices solved some problems, with the dramatic increase in image data, existing algorithms have experienced missed detections and false detections, failing to meet the demands. Summary of the Invention
[0003] To address the issues of low efficiency in manual inspection and false negatives in existing algorithms for detecting the condition of overhead contact line droppers, this invention provides an improved method for detecting the condition of overhead contact line droppers based on YOLOv11s.
[0004] This invention provides a method for detecting the condition of overhead contact line droppers based on an improved YOLOv11s, the method comprising: A small target dataset was constructed based on images of small target defects in overhead contact line droppers. The small target dataset is input into the constructed improved YOLOv11s algorithm network for training to obtain the trained detection model. By identifying the images of the overhead contact line droppers collected on-site using a trained detection model, small target detection results are obtained, and state detection is completed. The construction process of the improved YOLOv11s algorithm network includes: Establish a YOLOv11s basic network; the YOLOv11s basic network includes a Backbone network, a Neck network, and a Head network; A small object detection layer is constructed and added to the Neck network. A SEAM attention module is introduced to obtain an improved Neck network. The improved Neck network is used to obtain feature images from four different channels of the Backbone network and output feature images of four different resolutions. The Head detection head in the Head network is improved by combining it with the DASFF module to form a four-head DASFFHead detection head, resulting in an improved YOLOv11s algorithm network. The DASFF module is used to fuse four feature images of different resolutions, so that the Head network outputs small target detection results.
[0005] According to a specific implementation, in the above detection method, after the original image is input into the Backbone network, a first channel feature image is output through the first C3k2 module in the Backbone network, a second channel feature image is output through the second C3k2 module in the Backbone network, a third channel feature image is output through the third C3k2 module in the Backbone network, and a fourth channel feature image is output through the C2PSA module in the Backbone network.
[0006] According to a specific implementation, in the above detection method, the small target detection layer includes a first C3k2 module, a second C3k2 module, a first Concat module, a second Concat module, a first Conv module, and a first Upsample module.
[0007] According to a specific implementation, in the above detection method, the improved Neck network is used to upsample the fourth channel feature image, fuse the features with the third channel feature image and extract the features, then upsample again, fuse the features with the second channel feature image, extract the features and upsample them through the first Upsample module, fuse the features with the first channel feature image through the first Concat module, and then extract the features through the first C3k2 module to obtain the first resolution feature image, which is then output to the DASFF module.
[0008] According to a specific implementation, in the above detection method, the improved Neck network is further used to downsample the features extracted by the first C3k2 module through the first Conv module, fuse the features with the input of the first Upsample module through the second Concat module, extract the features through the second C3k2 module, input them to the SEAM attention module to obtain the second resolution feature image, and output them to the DASFF module.
[0009] According to one specific implementation, in the above detection method, the SEAM attention module is used to extract small target features in the occluded region through a depthwise separable convolution with residual connections.
[0010] According to a specific implementation, in the above detection method, the improved Neck network is further used to upsample the fourth channel feature image, fuse the features with the third channel feature image and extract the features, then fuse the features with the output of the downsampled SEAM attention module, extract the features to obtain a third resolution feature image, and output it to the DASFF module; then downsample the third resolution feature image, fuse the features with the fourth channel feature image, extract the features to obtain a fourth resolution feature image, and output it to the DASFF module.
[0011] According to one specific implementation, in the above detection method, the DASFFHead detection head uses a DASFF module to receive inputs at four different scales, and respectively fuses the first resolution feature image to the fourth resolution feature image.
[0012] According to one specific implementation, in the above detection method, constructing a small target dataset based on small target defect images of overhead contact line hangers specifically includes: Random Poisson noise is introduced into the small target defect image, and data augmentation and data expansion are performed to obtain the dataset; The dataset is divided into a training set, a validation set, and a test set to generate a small target dataset.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, based on an improved YOLOv11s detection model, achieves multi-scale feature fusion by adding a small target detection layer to the neck network and combining it with an improved detection head network. This enables more accurate detection of small targets and reduces false negatives and missed detections. By embedding the SEAM attention mechanism and utilizing depthwise separable convolution, this invention can effectively extract small target features in occluded areas, improving the robustness, reliability, and completeness of detection in complex environments. Although the average detection time increases slightly, it significantly improves the small target detection capability without affecting the detection accuracy of other categories. The overall detection performance is superior, providing reliable technical support for the detection of catenary dropper faults in high-speed railways and ensuring the safe operation of high-speed railways. Attached Figure Description
[0014] Figure 1 A flowchart illustrating a catenary dropper status detection method based on an improved YOLOv11s provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the improved YOLOv11s algorithm network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the SEAM attention module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the DASFFHead detection head module provided in an embodiment of the present invention; Figure 5 This diagram illustrates the comparison of average detection accuracy of different algorithms provided in embodiments of the present invention. Detailed Implementation
[0015] The present invention will now be described in further detail with reference to specific embodiments. However, this should not be construed as limiting the scope of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.
[0016] Unless otherwise specified, the use of terms such as "first," "second," and "third" in the description of specific embodiments of the present invention is merely for distinguishing descriptions of identical or similar components and should not be construed as emphasizing or implying the relative importance of a particular component.
[0017] The purpose of this invention is to achieve accurate and efficient identification of overhead contact line dropper status and detection of small targets, thereby improving the accuracy and efficiency of overhead contact line detection and ensuring the stable operation of the high-speed railway power supply system. Therefore, it is necessary to achieve multi-scale feature fusion to allow the model to adapt to the detection of dropper targets of different sizes; and to extract small target features in occluded areas to overcome detection obstacles caused by occlusion.
[0018] The technical solution provided by the present invention will be described and explained in detail below with reference to specific implementation methods.
[0019] Please refer to Figure 1 This document illustrates a flowchart of a catenary dropper status detection method based on an improved YOLOv11s according to an embodiment of the present invention. The method includes: Step 1: Construct a small target dataset based on small target defect images of overhead contact line droppers.
[0020] The construction of a small-target dataset specifically includes: Random Poisson noise is introduced into the small target defect image, and data augmentation and data expansion are performed to obtain the dataset; The dataset is divided into a training set, a validation set, and a test set to generate a small target dataset.
[0021] In one possible implementation, random Poisson noise is introduced due to the complexity of the actual operating environment and the special conditions of nighttime image acquisition, which allows the training images to conform to low-light characteristics. Simultaneously, various data augmentation techniques such as brightness enhancement, rotation, translation, cropping, mirroring, and occlusion are used to augment the training images, resulting in a dataset.
[0022] Specifically, the dataset includes four defect types: normal droppers, bent droppers, current-carrying ring faults, and minor defects. For example, there are 800 images of normal droppers, and 400 images each of bent droppers, current-carrying ring faults, and minor defects, all with a resolution of 5120×5120 pixels, totaling 2000 images. This 2000-image dataset is then divided into training, validation, and test sets in an 8:1:1 ratio, completing the construction of the small-objective dataset.
[0023] Step 2: Input the small target dataset into the constructed improved YOLOv11s algorithm network for training to obtain the trained detection model.
[0024] In one possible implementation, a basic YOLOv11s network is first built. Then, a small object detection layer is constructed and added to the Neck network, with a SEAM attention module introduced to obtain an improved Neck network. This improved Neck network is used to acquire feature images from four different channels of the Backbone network and output four feature images at different resolutions. Finally, the Head detection head in the Head network is combined with a DASFF module to improve it into a four-head DASFFHead detection head, resulting in the improved YOLOv11s algorithm network. The DASFF module is used to fuse the four feature images at different resolutions, enabling the Head network to output small object detection results. Please refer to [reference needed]. Figure 2 This diagram illustrates the structure of the improved YOLOv11s algorithm network provided in an embodiment of the present invention.
[0025] Specifically, the YOLOv11s basic network includes a Backbone network, a Neck network, and a Head network. The Backbone network consists of two Conv modules, a C3k2 module, a Conv module, a C3k2 module, a Conv module, a C3k2 module, a Conv module, a C3k2 module, a Conv module, a C3k2 module, an SPPF module, and a C2PSA module connected sequentially. When an image is input into the Backbone network, the first C3k2 module outputs a first-channel feature image, the second C3k2 module outputs a second-channel feature image, the third C3k2 module outputs a third-channel feature image, and the C2PSA module outputs a fourth-channel feature image.
[0026] Further, the small target detection layer includes a first C3k2 module, a second C3k2 module, a first Concat module, a second Concat module, a first Conv module, and a first Upsample module. The improved Neck network is used to upsample the fourth channel feature image, fuse features with the third channel feature image, extract features, upsample again, fuse features with the second channel feature image, extract features, upsample again through the first Upsample module, fuse features with the first channel feature image through the first Concat module, extract features again through the first C3k2 module, obtain a first resolution feature image, and output it to the DASFF module. Based on this, the improved Neck network is also used to downsample the features extracted by the first C3k2 module through the first Conv module, fuse features with the input of the first Upsample module through the second Concat module, extract features again through the second C3k2 module, input it to the SEAM attention module, obtain a second resolution feature image, and output it to the DASFF module. In one possible implementation, please refer to... Figure 3 The diagram illustrates the structure of the SEAM attention module provided in an embodiment of the present invention. The SEAM attention module is used to extract small target features from occluded regions through depthwise separable convolutions with residual connections.
[0027] Furthermore, the improved Neck network is also used to upsample the fourth channel feature image, fuse the features with the third channel feature image and extract the features, then fuse the features with the output of the downsampled SEAM attention module, extract the features to obtain a third resolution feature image, and output it to the DASFF module; then downsample the third resolution feature image, fuse the features with the fourth channel feature image, extract the features to obtain a fourth resolution feature image, and output it to the DASFF module.
[0028] Furthermore, the DASFFHead detection head uses a DASFF module to receive inputs at four different scales, fusing the first resolution feature image to a fourth resolution feature image. Please refer to [link / reference]. Figure 4 This diagram illustrates the structure of the DASFFHead detection head provided in an embodiment of the present invention.
[0029] Understandably, DASFF enables the network to adaptively learn how to filter features from other layers in the spatial domain, retaining only useful information for combination. At each spatial location, features from different layers are adaptively fused, meaning some features may be filtered out because they carry contradictory information, while others may dominate the detection, providing more discriminative cues.
[0030] Understandably, the head detection heads in the YOLOv11s basic network typically employ only three different scale head modules. This invention, however, adds a higher-resolution feature map output to the small target detection layer, enabling the detection of minute targets. Specifically, the DASFF module employs dynamic adaptive spatial feature fusion technology, achieving pixel-wise feature fusion through a dynamic spatial weight mechanism (1x1 convolution + Softmax). Combined with lightweight channel compression (8 / 16 dimensions) and cross-layer adaptive sampling strategies (high-level convolutional pooling downsampling / low-level interpolation upsampling), a four-head detection structure is constructed to enhance multi-scale feature interaction, significantly improving small target detection performance while maintaining computational efficiency. For example, the three head detection heads in the YOLOv11s basic network have strides of 8, 16, and 32, respectively. This invention, based on this, sets the strides of the four DASFF head detection heads to 4, 8, 16, and 32, respectively. An additional DASFF module with a stride of 4 is added, which can extract information from the early feature maps of the backbone network, preserving more spatial details. Small targets may only occupy a few pixels in a low-resolution feature map, making it easy to lose details. By introducing a high-resolution detection layer, the model can capture more shallow features (such as edges and textures), thereby improving the detection capability of small targets.
[0031] In one possible implementation, this embodiment of the invention provides a training example of an improved YOLOv11s algorithm network, wherein the training parameters include a batch size of 16, an initial learning rate of 0.01, a final learning rate of 0.01, a momentum parameter of 0.937, a weight decay of 0.0005, input images uniformly converted to a size of 640x640, and a stochastic gradient descent optimizer.
[0032] Step 3: Use the trained detection model to identify the contact wire images collected on site, obtain the small target detection results, and complete the state detection.
[0033] For details, please refer to Figure 5 The diagram illustrates a comparison of the average detection accuracy of different algorithms provided in this embodiment of the invention. Specifically, the improved YOLOv11s algorithm network provided in this embodiment of the invention improves the overall dropper detection accuracy mAP@0.5 by 1.3%, with the AP value for detecting droppers with small defects improving by 8.7%.
[0034] Based on the above technical solutions, this invention, using an improved YOLOv11s detection model, adds a small target detection layer to the neck network and combines it with an improved detection head network to achieve multi-scale feature fusion, enabling more accurate detection of small targets and reducing missed detections and false detections. By embedding the SEAM attention mechanism and utilizing depthwise separable convolution, this invention can effectively extract small target features in occluded areas, improving the robustness, reliability, and completeness of detection in complex environments. Although the average detection time increases slightly, it significantly improves the small target detection capability without affecting the detection accuracy of other categories, resulting in superior overall detection performance. This provides reliable technical support for fault detection of overhead contact line droppers in high-speed railways, ensuring the safe operation of high-speed railways.
[0035] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.
[0036] Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner for ease of understanding.
[0037] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for detecting the condition of overhead contact line droppers based on an improved YOLOv11s, characterized in that, The method includes: A small target dataset was constructed based on images of small target defects in overhead contact line droppers. The small target dataset is input into the constructed improved YOLOv11s algorithm network for training to obtain the trained detection model. By identifying the images of the overhead contact line droppers collected on-site using a trained detection model, small target detection results are obtained, and state detection is completed. The construction process of the improved YOLOv11s algorithm network includes: Establish a YOLOv11s basic network; the YOLOv11s basic network includes a Backbone network, a Neck network, and a Head network; A small object detection layer is constructed and added to the Neck network. A SEAM attention module is introduced to obtain an improved Neck network. The improved Neck network is used to obtain feature images from four different channels of the Backbone network and output feature images of four different resolutions. The Head detection head in the Head network is improved by combining it with the DASFF module to form a four-head DASFFHead detection head, resulting in an improved YOLOv11s algorithm network. The DASFF module is used to fuse four feature images of different resolutions, so that the Head network outputs small target detection results.
2. The contact wire dropper status detection method based on the improved YOLOv11s according to claim 1, characterized in that, After the original image is input into the Backbone network, the first channel feature image is output through the first C3k2 module in the Backbone network, the second channel feature image is output through the second C3k2 module in the Backbone network, the third channel feature image is output through the third C3k2 module in the Backbone network, and the fourth channel feature image is output through the C2PSA module in the Backbone network.
3. The contact wire dropper status detection method based on the improved YOLOv11s according to claim 2, characterized in that, The small target detection layer includes a first C3k2 module, a second C3k2 module, a first Concat module, a second Concat module, a first Conv module, and a first Upsample module.
4. The contact wire dropper status detection method based on the improved YOLOv11s according to claim 3, characterized in that, The improved Neck network is used to upsample the fourth channel feature image, fuse the features with the third channel feature image, extract the features, upsample again, fuse the features with the second channel feature image, extract the features, and upsample them through the first Upsample module. Then, it fuses the features with the first channel feature image through the first Concat module, and extracts the features through the first C3k2 module to obtain the first resolution feature image, which is then output to the DASFF module.
5. The contact wire dropper status detection method based on the improved YOLOv11s according to claim 4, characterized in that, The improved Neck network is also used to downsample the features extracted by the first C3k2 module through the first Conv module, fuse the features with the input of the first Upsample module through the second Concat module, extract the features through the second C3k2 module, input them to the SEAM attention module to obtain the second resolution feature image, and output them to the DASFF module.
6. The contact wire dropper status detection method based on the improved YOLOv11s according to claim 5, characterized in that, The SEAM attention module is used to extract small target features in occluded regions through depthwise separable convolutions with residual connections.
7. The contact wire dropper status detection method based on the improved YOLOv11s according to claim 5, characterized in that, The improved Neck network is also used to upsample the fourth channel feature image, fuse the features with the third channel feature image and extract the features, then fuse the features with the output of the downsampled SEAM attention module to extract the features and obtain the third resolution feature image, which is then output to the DASFFHEAD detection head; then downsample the third resolution feature image, fuse the features with the fourth channel feature image, extract the features and obtain the fourth resolution feature image, which is then output to the DASFF module.
8. The contact wire dropper state detection method based on the improved YOLOv11s according to claim 7, characterized in that, The DASFFHead detection head uses a DASFF module to receive inputs at four different scales, and fuses the first resolution feature image to the fourth resolution feature image respectively.
9. The contact wire dropper status detection method based on the improved YOLOv11s according to claim 1, characterized in that, A small target dataset was constructed based on images of small target defects in overhead contact line droppers, specifically including: Random Poisson noise is introduced into the small target defect image, and data augmentation and data expansion are performed to obtain the dataset; The dataset is divided into a training set, a validation set, and a test set to generate a small target dataset.