A high-speed rail power transmission line dual-mode defect intelligent detection method based on improved YOLOv11s

Abstract

The application discloses a high-speed rail power transmission line double-mode defect intelligent detection method based on an improved YOLOv11s, relates to the field of high-speed rail power system equipment fault detection, and solves the problem that the existing method cannot realize synchronous detection of thermal defects and structural defects. The application uses a UAV loaded with an infrared thermal imager and a visible light camera to collect double-source images of insulators and contact line connecting pieces of a high-speed rail power transmission line, constructs a multi-modal sample data set after defect labeling; a weight model of a YOLOv11s network model is obtained after pre-training on the multi-modal sample data set; a wavelet transform convolution, a double-mode gate fusion module and a lightweight attention mechanism are introduced, the weight model is migrated to the improved YOLOv11s network model, and a detection model is trained to realize detection, so that defect changes can be captured in real time, power supply failure rate is reduced, and line operation reliability is improved.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology for high-speed railway power system equipment, and in particular to an intelligent detection method for dual-modal defects in high-speed railway transmission lines based on an improved YOLOv11s. Background Technology

[0002] As the core carrier of power supply for high-speed railways, the operational reliability of high-speed railway power transmission lines directly affects the punctuality and safety of high-speed train operations. High-speed railway power transmission lines are exposed to the outdoor environment for extended periods, and are susceptible to factors such as wind, rain, lightning strikes, interference from train pantograph-catenary friction, and aging. Key components such as insulators and contact wire connectors are prone to localized overheating. If not detected and addressed in time, this can lead to serious faults such as line tripping and equipment burnout, causing high-speed train shutdowns or speed reductions, resulting in huge economic losses and adverse social impacts. Traditionally, the inspection of key components of high-speed railway power transmission lines relies mainly on manual inspections. Workers need to carry infrared thermometers for on-site operations. However, while infrared thermal imagers can capture thermal defects, they cannot determine whether overheating is caused by structural problems. Furthermore, manual inspections have limited coverage and low efficiency, making it difficult to cover lines in complex terrains such as mountainous areas and those crossing rivers, and preventing close-range inspections of lines during high-speed train operation.

[0003] In recent years, with the rapid development of deep learning technology, many excellent machine vision object detection algorithms have emerged and have been widely applied in production and daily life. In the field of high-voltage transmission line inspection, the ability to quickly and accurately identify abnormal temperature rises and structural damage in insulators and contact wire connections has become a core requirement for ensuring the stable operation of the power system. However, traditional image processing methods have significant limitations under the complex operating conditions of high-speed rail transmission lines, and their detection efficiency cannot achieve real-time monitoring.

[0004] The core technology of the YOLOv11 algorithm is to rely on a single deep neural network model to simultaneously predict the position and category of all targets within an image during a single forward pass. However, the existing YOLOv11 algorithm has limitations in practical applications of thermal defect detection in high-voltage transmission lines. It cannot simultaneously detect both thermal and structural defects in key components such as insulators and contact wire connectors of high-speed railway transmission lines. Therefore, it is necessary to propose an intelligent dual-modal defect detection method for high-speed railway transmission lines based on an improved YOLOv11s to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent detection method for dual-modal defects in high-speed railway transmission lines based on an improved YOLOv11s, in order to solve the limitations of the existing YOLOv11 algorithm in the practical application of thermal defect detection in high-voltage transmission lines, which cannot achieve the simultaneous detection of both thermal and structural defects in key components such as insulators and contact wire connectors of high-speed railway transmission lines.

[0006] This invention provides an intelligent detection method for dual-modal defects in high-speed railway transmission lines based on an improved YOLOv11s, comprising:

[0007] Step 1: Use a drone equipped with an infrared thermal imager and a visible light camera to collect dual-source images of the insulators and contact wire connectors of the high-speed railway transmission line.

[0008] Step 2: Preprocess the dual-source images and label each image with defects to construct a multimodal sample dataset with the defective images and corresponding labels.

[0009] Step 3: Obtain the weight model of the YOLOv11s network model after pre-training on the multimodal sample dataset;

[0010] Step 4: Improve the YOLOv11s network model by introducing wavelet transform convolution, dual-modal gating fusion module and lightweight attention mechanism to obtain the improved YOLOv11s network model.

[0011] Step 5: Transfer the weight model to the improved YOLOv11s network model; train the improved YOLOv11s network model using the multimodal sample dataset to obtain a detection model suitable for detecting thermal and structural defects in insulators and contact wire connectors of high-speed railway transmission lines.

[0012] Step 6: Use the detection model to detect thermal and structural defects in the high-speed railway transmission line.

[0013] Furthermore, in step two, the preprocessing includes: flipping, rotating, adjusting the brightness of the dual-source image, and expanding the image.

[0014] Furthermore, in step four, the wavelet transform convolution is combined with the C3k2 module of the backbone network of the YOLOv11s network model to form the C3k2_WTConv module.

[0015] Furthermore, the C3k2_WTConv module is used for multi-band feature decomposition, frequency-band feature enhancement, and inverse wavelet transform fusion; wherein, the multi-band feature decomposition includes:

[0016] The initial feature map extracted from the backbone network is decomposed into four frequency bands using the Daubechies-4 wavelet basis: low frequency, horizontal high frequency, vertical high frequency, and diagonal high frequency. The low frequency component is used to preserve global structural information, including the outline of the insulator skirt and the overall geometry of the contact wire connector. The horizontal high frequency component is used to capture the horizontal edge features of the target, including horizontal cracks in the insulator skirt. The vertical high frequency component is used to extract the vertical texture features of the target, including vertical damage marks on the surface of the insulator. The diagonal high frequency component is used to enhance the angular features of the target.

[0017] Furthermore, the frequency band feature enhancement includes:

[0018] For low-frequency components, lightweight feature extraction is performed using 3×3 depthwise separable convolution, which reduces computational cost while preserving the global structure of the target. For the remaining components, subtle feature enhancement is performed using 1×1 grouped convolution, which preserves key details of small targets, including insulator cracks and connector bolts, and suppresses background noise.

[0019] Furthermore, the inverse wavelet transform fusion includes:

[0020] The four enhanced frequency band components are re-aggregated into a complete feature map through inverse wavelet transform, achieving a synergistic expression of global structural information and local detailed features.

[0021] Further, in step four, the dual-modal gating fusion module is used for global modal information extraction, dynamic gating weight calculation, and multi-scale feature filtering and fusion; wherein, the global modal information extraction includes:

[0022] For infrared feature map F ir With visible light feature map F vis Global average pooling is performed separately to compress the spatial dimensionality while preserving the global statistical features of the modality, resulting in a 1×1×C global descriptor G. ir and G vis C represents the number of feature channels; G ir Reflects the overall temperature distribution trend of infrared images; G vis It reflects the structural integrity of visible light images.

[0023] Furthermore, the dynamic gating weight calculation includes:

[0024] global descriptor G ir With G vis Feature concatenation is performed, the channel dimension is compressed to C through 1×1 convolution, and then infrared mode weights w are generated by the Sigmoid activation function. ir As shown in Equation 1, the optical mode weight w vis The calculation is derived from Equation 2;

[0025]

[0026] In the formula, w ir For infrared mode weights, G ir Infrared feature map F ir The obtained global descriptor, G vis Visible light feature map F vis The resulting global descriptor, where Conv represents convolution and σ represents the Sigmoid activation function;

[0027] w vis =1-w ir Formula 2

[0028] In the formula, w vis For visible light mode weights;

[0029] When insulator overheating is detected in the infrared signature, w ir Automatically approaches 1, enhancing infrared mode contribution; when structural information in visible light features is clear, w vis Automatically approaches 1, increasing the weight of the visible light mode; for composite defects, the weight will be in the middle range, realizing dual-mode collaborative judgment.

[0030] Furthermore, the multi-scale feature filtering and fusion includes:

[0031] Dual-modal gated fusion modules are deployed at key scales in the backbone network output to address defects of different sizes. Equation 3 uses dynamic weights to perform element-wise weighting of the dual-modal features to obtain the filtered effective infrared features M. ir Equation 4 yields the effective visible light characteristic M. vis ;

[0032] M ir =w ir ·F ir Formula 3

[0033] In the formula, M ir For effective infrared features, w ir For infrared mode weights, F ir Infrared feature map;

[0034] M vis =w vis ·F vis Formula 4

[0035] In the formula, M vis For visible light effective characteristics, w vis For visible light mode weights, F vis Visible light characteristic map;

[0036] Then, element-wise multiplication is used to capture the spatial correspondence between thermal defects and structural defects, and depthwise separable convolution is used to extract the deep semantics of the associated features. Equation 5 shows that by adding the basic fusion features and the associated enhancement features, the multi-scale fusion feature F is obtained. fusion ;

[0037] F fusion =(M ir +M vis )+DWConv(M ir ⊙M vis Formula 5

[0038] In the formula, F fusion For multi-scale fusion features, M ir M is an effective infrared feature. vis For visible light effective features, DWConv represents depthwise separable convolution.

[0039] Furthermore, in step four, the lightweight attention mechanism is used to automatically generate attention weights for each neuron in the feature map by measuring its differential contribution to surrounding neurons; Equation 6 shows the calculation method of neuron energy parameters, as well as the use of input feature mean and variance, through which the difference between the target neuron and its neighboring neurons is quantified.

[0040]

[0041] In the formula, Represents the neuron's energy parameters. Let λ represent the variance of the input features, λ represent the regularization parameter, and t represent the feature value of the current neuron. This represents the mean of the input features.

[0042] The present invention has the following beneficial effects: The present invention provides an intelligent detection method for dual-modal defects in high-speed railway transmission lines based on an improved YOLOv11s. It uses a drone equipped with an infrared thermal imager and a visible light camera to simultaneously acquire dual-source images of key components such as insulators and contact wire connectors in high-speed railway transmission lines. By improving the YOLOv11s algorithm, wavelet transform convolution, dual-modal gating fusion module and attention mechanism are introduced, thereby enabling real-time capture of defect changes, reducing power supply failure rate and improving line operation reliability. Attached Figure Description

[0043] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0044] Figure 1 Flowchart of the intelligent detection method for dual-modal defects in high-speed railway transmission lines based on the improved YOLOv11s provided by this invention;

[0045] Figure 2 This is a diagram of the improved YOLOv11 architecture;

[0046] Figure 3 This is the original C3K2 structure diagram;

[0047] Figure 4 This is a diagram of the improved C3K2 module of the present invention;

[0048] Figure 5 This is a diagram of the C3K_WTConv module in this invention;

[0049] Figure 6 Generate a graph for the dynamic weighted Mir Mvis. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0051] Please see Figure 1 The present invention provides an intelligent detection method for dual-modal defects in high-speed railway transmission lines based on an improved YOLOv11s, comprising:

[0052] Step 1: Use a drone equipped with an infrared thermal imager and a visible light camera to collect dual-source images of the insulators and contact wire connectors of the high-speed railway transmission line.

[0053] Step 2: Preprocess the dual-source images and label each image with defects to construct a multimodal sample dataset with the defective images and corresponding labels.

[0054] Specifically, the original dual-source images are processed by flipping, rotating, and adjusting brightness to expand the images, and each image is labeled with defects to construct a dataset with defective images and corresponding labels.

[0055] Step 3: Obtain the weight model of the YOLOv11s network model after pre-training on the multimodal sample dataset.

[0056] Specifically, in this embodiment, the YOLOv11s network model is trained and iterated 100 times on a multimodal sample dataset.

[0057] Step four: Improve the YOLOv11s network model by introducing wavelet transform convolution, dual-modal gating fusion module and lightweight attention mechanism to obtain the improved YOLOv11s network model.

[0058] Specifically, wavelet transform convolution is combined with the C3k2 module of the YOLOv11s network model backbone to form the C3k2_WTConv (wavelet transform convolution) module, which enhances the ability to extract features from small targets. In the scenario of UAV inspection of high-speed railway transmission lines, due to limitations in flight altitude and safe distance, it is impossible to take close-up pictures of insulators and contact wire connectors. As a result, the proportion of pixels of these two types of targets in the acquired images is low. Not only do they contain very little feature information such as edges and textures, but they are also easily interfered with by environmental factors such as tower shadows, tree branches and leaves, and sky background. Traditional convolution modules are difficult to effectively capture the features of small targets, resulting in a high false negative rate.

[0059] Wavelet transform convolution, through its unique mechanism of multi-band feature decomposition, targeted enhancement, and inverse transform fusion, can accurately extract subtle features of small targets and filter background interference. Combining it with the C3k2 module of YOLOv11s to form the C3k2_WTConv module significantly improves the network's feature perception capability for insulators and contact wire connectors. Its working principle is as follows:

[0060] Multi-band feature decomposition: First, the initial feature map extracted from the backbone network is subjected to a two-level wavelet decomposition using the Daubechies-4 wavelet basis, splitting the feature information into four frequency band components: low frequency, horizontal high frequency, vertical high frequency, and diagonal high frequency. The low frequency component is used to preserve global structural information such as the outline of the insulator skirts and the overall geometry of the contact wire connectors; the horizontal high frequency component is used to capture the horizontal edge features of the target, such as horizontal cracks in the insulator skirts; the vertical high frequency component is used to extract the vertical texture features of the target, such as vertical damage marks on the surface of the insulator; and the diagonal high frequency component is used to enhance the angular features of the target, aiding in the accurate identification of small targets.

[0061] Frequency-band feature enhancement: For low-frequency components, lightweight feature extraction is performed using 3×3 depthwise separable convolutions, which reduces computational cost while preserving the global structure of the target. For the remaining components, subtle features are enhanced using 1×1 grouped convolutions, focusing on preserving key details of small targets such as insulator cracks and connector bolts, while suppressing background noise.

[0062] Inverse wavelet transform fusion: The four frequency band components that have undergone enhancement processing are re-aggregated into a complete feature map through inverse wavelet transform, so as to achieve the coordinated expression of global structural information and local detailed features.

[0063] A dual-modal gated fusion module (BMGFM) is embedded between the backbone and the neck module to achieve dynamic adaptation and fusion of dual modalities. Through global information perception, dynamic weight adjustment, and multi-scale feature adaptation, the BMGFM achieves adaptive fusion of dual-modal features without significantly increasing computational burden. Defect detection in high-speed rail transmission lines relies on both infrared and visible light dual-modal information. However, the value of these two modalities dynamically changes with the scene. For example, local overheating of insulators requires infrared temperature gradient features, loose bolts in contact wire connectors require visible light edge texture features, and combined overheating and crack defects require collaborative dual-modal judgment. Traditional fixed-weight fusion methods cannot adapt to this dynamic requirement. For instance, failing to increase infrared weights when visible light is blurred in cloudy weather can easily lead to missed thermal defects; failing to strengthen visible light weights when infrared noise interferes in bright sunlight can easily cause misjudgments of structural defects. BMGFM's core logic is based on generating dynamic weights from bimodal global statistical information to achieve precise matching of scene, modality, and defect. This adapts to the hardware limitations of drones with limited computing power and accurately matches the modal requirements of different defect types. Specifically, feature fusion is achieved through three steps:

[0064] Global modal information extraction: extraction of infrared feature map F ir With visible light feature map F vis Global average pooling is performed separately to compress the spatial dimensionality while preserving the global statistical features of the modality, resulting in a 1×1×C global descriptor G. ir and G vis C represents the number of feature channels; G ir Reflects the overall temperature distribution trend of infrared images; G vis It reflects the structural integrity of visible light images.

[0065] Dynamic gating weight calculation: The global descriptor G... ir With G vis Feature concatenation is performed, the channel dimension is compressed to C through 1×1 convolution, and then infrared mode weights w are generated by the Sigmoid activation function. ir As shown in Equation 1, the optical mode weight w vis The calculation is derived from Equation 2;

[0066]

[0067] In the formula, w ir For infrared mode weights, G ir Infrared feature map F ir The obtained global descriptor, G vis Visible light feature map F vis The resulting global descriptor, where Conv represents convolution and σ represents the Sigmoid activation function;

[0068] w vis =1-w ir Formula 2

[0069] In the formula, w vis For visible light mode weights;

[0070] When insulator overheating is detected in the infrared signature, w ir Automatically approaches 1, enhancing infrared mode contribution; when structural information in visible light features is clear, w vis Automatically approaches 1, increasing the weight of the visible light mode; for composite defects, the weight will be in the middle range, realizing dual-mode collaborative judgment.

[0071] Multi-scale feature screening and fusion: Considering the differences in defect size, such as overall insulator overheating being a large-area defect, and bolt loosening and micro-cracks being small-target defects, BMGFM is deployed at the key scales of the backbone network output to specifically address defects of different sizes. Equation 3 uses dynamic weights to perform element-wise weighting of the dual-modal features to obtain the screened effective infrared features M. ir Equation 4 yields the effective visible light characteristic M. vis ;

[0072] M ir =w ir ·F ir Formula 3

[0073] In the formula, M ir For effective infrared features, w ir For infrared mode weights, F ir Infrared feature map;

[0074] M vis =w vis ·F vis Formula 4

[0075] In the formula, M vis For visible light effective characteristics, w vis For visible light mode weights, F vis Visible light characteristic map;

[0076] Then, element-wise multiplication is used to capture the spatial correspondence between thermal defects and structural defects, and depthwise separable convolution is used to extract the deep semantics of the associated features. Equation 5 shows that by adding the basic fusion features and the associated enhancement features, the multi-scale fusion feature F is obtained. fusion ;

[0077] F fusion =(M ir +M vis )+DWConv(Mir ⊙M vis Formula 5

[0078] In the formula, F fusion For multi-scale fusion features, M ir M is an effective infrared feature. vis For visible light effective features, DWConv represents depthwise separable convolution.

[0079] BMGFM's design fully considers the limited computing power of drones. It controls computational costs by using 1×1 convolutions for weight calculations, eliminating complex fully connected layers or multi-head attention, and employing depthwise separable convolutions, which have a computational cost that is only 1 / C of that of ordinary convolutions.

[0080] A lightweight attention module (SimAM) is added to the Neck layer to focus on key defect regions and suppress background interference. This lightweight attention mechanism requires no additional parameters and is well-suited for drones with limited computing power that need to enhance small target features. In high-speed rail transmission line inspection images, key features such as insulators and contact wire connectors often occupy only 5-10 pixels, easily obscured by background information such as tower shadows and vegetation textures. Traditional feature extraction models often fail to adequately focus on small target features, leading to a high false negative rate. SimAM, by simulating the activation mechanism of biological neurons, adaptively enhances the response intensity of key features without increasing computational burden, while suppressing redundant background information, thus improving small target detection accuracy with lightweight design. The core principle of SimAM is to calculate feature importance based on neuron activity. For each neuron in the feature map, attention weights are automatically generated by measuring its differential contribution compared to surrounding neurons. Equation 6 shows the calculation method of neuron energy parameters and the application of the mean and variance of input features, using these statistics to quantify the difference between the target neuron and its neighboring neurons.

[0081]

[0082] In the formula, Represents the neuron's energy parameters. Let λ represent the variance of the input features, λ represent the regularization parameter, and t represent the feature value of the current neuron. This represents the mean of the input features.

[0083] Step 5: Transfer the weight model to the improved YOLOv11s network model; train the improved YOLOv11s network model using the multimodal sample dataset to obtain a detection model suitable for detecting thermal and structural defects in insulators and contact wire connectors of high-speed railway transmission lines.

[0084] Step 6: Use the detection model to detect thermal and structural defects in the high-speed railway transmission line.

[0085] For thermal defects, the first step is to simultaneously acquire infrared thermal images and visible light images of the transmission line using a drone. Infrared images are used to capture the temperature distribution of components to detect thermal anomalies, while visible light images are used to obtain structural details of components as an aid to thermal defect analysis. The improved YOLOv11s model first performs feature fusion and target detection on the dual-modal images to identify key components such as insulators and contact wire connectors. Second, through the enhanced thermal feature extraction module in the model, it accurately captures abnormal temperature rise areas in local areas of insulators and contact wire connectors in the infrared images. Finally, by combining the structural state of the components in the visible light images, the type and severity of the thermal defect are comprehensively judged, because structural defects are often the cause of thermal defects, and dual-modal information fusion can more accurately locate the root cause of thermal defects.

[0086] For structural defects, after completing the dataset construction, the improved YOLOv11s model first identifies components such as insulators and contact wire connectors in transmission lines. Secondly, the SimAM attention module in the model focuses on key structural areas to accurately detect structures such as insulator skirts and bolts. For insulator appearance damage, after detecting the damaged area, its area, shape, and other characteristics are analyzed to determine whether the damage will affect the insulation performance. For bolt issues, by judging the relative positional relationship between the bolt and the connector, it can be determined whether the bolt is loose or missing.

[0087] In this scheme, the following improvements were made to enhance the model's detection accuracy of target objects:

[0088] The backbone network's C3k2 module is replaced with the C3k2_WTConv module to enhance small target feature extraction capabilities: To address the issue of low pixel ratios, few features, and susceptibility to background interference in high-speed rail transmission line insulators and contact wire connectors during UAV inspections, where traditional convolutional methods struggle to capture small target features, this invention replaces the C3k2 module of the YOLOv11s backbone network with the C3k2_WTConv module to strengthen small target feature extraction capabilities. Its core logic is achieved through three steps: multi-band feature decomposition, sub-band feature enhancement, and inverse wavelet transform fusion.

[0089] Dual-Modal Gated Fusion Module (BMGFM): In response to the problem that the infrared and visible light modal values ​​change dynamically with the scene in the detection of high-speed railway transmission lines, and that traditional fixed-weight fusion is prone to missed detections and misjudgments, BMGFM improves the utilization rate of dual-modal features under the premise of lightweight design through global information perception, dynamic weight adjustment and multi-scale fusion.

[0090] To address the issue that small targets (only 5-10 pixels) such as insulator thermal defects and bolts in high-speed rail inspection images are easily obscured by the background, resulting in insufficient attention from traditional models, SimAM employs a lightweight design with no additional parameters to enhance key feature responses, suppress redundant backgrounds, and improve the detection accuracy of small targets. Based on neuron activity, feature importance is calculated by quantifying the importance of each neuron in the feature map by calculating the deviation from the global mean, allowing the network to autonomously focus on defect regions. Without pre-setting a fixed attention structure, it automatically assigns high weights to key areas such as insulator overheating points and bolt edges, while filtering out interference from tower shadows and vegetation. This significantly enhances the feature response intensity of small targets, increasing the concentration of attention weights by 25%, effectively reducing the false negative rate of small targets and the false positive rate of background detection. The improved YOLOv11s network achieves targeted performance improvements through a three-layer optimization system, as detailed below:

[0091] Enhance the ability to extract features from small targets: By fusing wavelet transform convolution, it can not only solve the problem that the local overheating features of insulators in infrared features are easily submerged by the background, but also enhance the texture details such as cracks in visible light features.

[0092] To solve the problem of modal information mismatch and achieve complementary enhancement of dual-modal information: relying on the dual-modal gating fusion module, dynamic weights are calculated based on global descriptors of infrared and visible light, replacing fixed weight fusion.

[0093] Suppressing background interference and adapting to low-computing-power hardware: The SimAM attention module is introduced, which automatically assigns high weights to the target area by quantifying the deviation between neurons and the global mean, effectively suppressing background interference such as pole shadows and vegetation, and perfectly adapting to low-computing-power hardware of drones.

[0094] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention.

Claims

1. A dual-modal intelligent detection method for high-speed railway transmission lines based on an improved YOLOv11s, characterized in that, include: Step 1: Use a drone equipped with an infrared thermal imager and a visible light camera to collect dual-source images of the insulators and contact wire connectors of the high-speed railway transmission line. Step 2: Preprocess the dual-source images and label each image with defects to construct a multimodal sample dataset with the defective images and corresponding labels. Step 3: Obtain the weight model of the YOLOv11s network model after pre-training on the multimodal sample dataset; Step 4: Improve the YOLOv11s network model by introducing wavelet transform convolution, dual-modal gating fusion module and lightweight attention mechanism to obtain the improved YOLOv11s network model. Specifically, wavelet transform convolution is combined with the C3k2 module of the backbone network of the YOLOv11s network model to form the C3k2_WTConv module; the C3k2_WTConv module is used for multi-band feature decomposition, frequency-band feature enhancement and inverse wavelet transform fusion. The multi-band feature decomposition includes: performing a second-level wavelet decomposition on the initial feature map extracted from the backbone network using the Daubechies-4 wavelet basis, splitting the feature information into four frequency band components: low frequency, horizontal high frequency, vertical high frequency, and diagonal high frequency. The low frequency component is used to retain global structural information, including the outline of the insulator skirts and the overall geometry of the contact wire connectors. The horizontal high frequency component is used to capture the horizontal edge features of the target, including horizontal cracks in the insulator skirts. The vertical high frequency component is used to extract the vertical texture features of the target, including vertical damage marks on the surface of the insulator. The diagonal high frequency component is used to enhance the angular features of the target. The frequency band feature enhancement includes: using 3×3 depthwise separable convolution for lightweight feature extraction of low-frequency components, which reduces computational consumption while preserving the global structure of the target; and using 1×1 grouped convolution for subtle feature enhancement of the remaining components, which preserves key details of small targets, including insulator cracks and connector bolts, and suppresses background noise. Step 5: Transfer the weight model to the improved YOLOv11s network model; train the improved YOLOv11s network model using the multimodal sample dataset to obtain a detection model suitable for detecting thermal and structural defects in insulators and contact wire connectors of high-speed railway transmission lines. Step 6: Use the detection model to detect thermal and structural defects in the high-speed railway transmission line.

2. The intelligent detection method for dual-mode defects in high-speed railway transmission lines based on improved YOLOv11s as described in claim 1, characterized in that, In step two, the preprocessing includes: flipping, rotating, adjusting the brightness of the dual-source image, and expanding the image.

3. The intelligent detection method for dual-mode defects in high-speed railway transmission lines based on improved YOLOv11s as described in claim 1, characterized in that, The inverse wavelet transform fusion includes: The four enhanced frequency band components are re-aggregated into a complete feature map through inverse wavelet transform, achieving a synergistic expression of global structural information and local detailed features.

4. The intelligent detection method for dual-mode defects in high-speed railway transmission lines based on improved YOLOv11s as described in claim 1, characterized in that, In step four, the dual-modal gating fusion module is used for global modal information extraction, dynamic gating weight calculation, and multi-scale feature filtering and fusion; wherein, the global modal information extraction includes: For infrared feature map F ir With visible light feature map F vis Global average pooling is performed separately to compress the spatial dimensionality while preserving the global statistical features of the modality, resulting in a 1×1×C global descriptor G. ir and G vis C represents the number of feature channels; G ir Reflects the overall temperature distribution trend of infrared images; G vis It reflects the structural integrity of visible light images.

5. The intelligent detection method for dual-mode defects in high-speed railway transmission lines based on improved YOLOv11s as described in claim 4, characterized in that, The dynamic gating weight calculation includes: global descriptor G ir With G vis Feature concatenation is performed, the channel dimension is compressed to C through 1×1 convolution, and then infrared mode weights w are generated by the Sigmoid activation function. ir As shown in Equation 1, the optical mode weight w vis The calculation is derived from Equation 2; In the formula, w ir For infrared mode weights, G ir Infrared feature map F ir The obtained global descriptor, G vis Visible light feature map F vis The resulting global descriptor, where Conv represents convolution, This represents the Sigmoid activation function; In the formula, w vis For visible light mode weights; When insulator overheating is detected in the infrared signature, w ir Automatically approaches 1, enhancing infrared mode contribution; when structural information in visible light features is clear, w vis Automatically approaches 1, increasing the weight of the visible light mode; for composite defects, the weight will be in the middle range, realizing dual-mode collaborative judgment.

6. The intelligent detection method for dual-mode defects in high-speed railway transmission lines based on improved YOLOv11s as described in claim 4, characterized in that, The multi-scale feature filtering and fusion includes: Dual-modal gated fusion modules are deployed at key scales in the backbone network output to address defects of different sizes. Equation 3 uses dynamic weights to perform element-wise weighting of the dual-modal features to obtain the filtered effective infrared features M. ir Equation 4 yields the effective visible light characteristic M. vis ; In the formula, M ir For effective infrared features, w ir For infrared mode weights, F ir Infrared feature map; In the formula, M vis For visible light effective characteristics, w vis For visible light mode weights, F vis Visible light characteristic map; Then, element-wise multiplication is used to capture the spatial correspondence between thermal defects and structural defects, and depthwise separable convolution is used to extract the deep semantics of the associated features. Equation 5 shows that by adding the basic fusion features and the associated enhancement features, the multi-scale fusion feature F is obtained. fusion ; In the formula, F fusion For multi-scale fusion features, M ir M is an effective infrared feature. vis For visible light effective features, DWConv represents depthwise separable convolution.

7. The intelligent detection method for dual-mode defects in high-speed railway transmission lines based on improved YOLOv11s as described in claim 1, characterized in that, In step four, the lightweight attention mechanism is used to automatically generate attention weights for each neuron in the feature map by measuring its differential contribution to surrounding neurons; Equation 6 shows the calculation method of neuron energy parameters and the use of input feature mean and variance, which are used to quantify the difference between the target neuron and its neighboring neurons. In the formula, Represents the neuron's energy parameters. Represents the variance of the input features. Represents the regularization parameter. This represents the feature value of the current neuron. This represents the mean of the input features.

Images