Railway overhead line infrared thermal defect identification and early warning method and system

Abstract

The present application is a railway catenary infrared thermal defect identification and early warning method and system, belonging to the field of railway traction power supply system monitoring technology. The method includes the following steps: collecting catenary component images; inputting an improved YOLOv8 deep learning network for target detection, identifying and locating catenary components; based on the boundary box of the catenary component, performing image segmentation on the corresponding region and extracting a temperature matrix to realize quantitative characterization of the catenary component temperature field; based on the sequence of the temperature matrix, using a continuous frame statistical analysis method to distinguish the temperature changes of the same component in multiple images, and distinguishing between persistent thermal defects and transient interference through time series correlation analysis; when a thermal defect is determined, outputting complete early warning information including the defect location, type, temperature and early warning level. The present application realizes end-to-end intelligent processing from the original infrared image to accurate early warning information, significantly improving the accuracy and reliability of railway catenary thermal defect detection.

Description

Technical Field

[0001] This invention relates to a method and system for infrared thermal defect identification and early warning of railway catenary, belonging to the field of railway traction power supply system monitoring technology. Background Technology

[0002] The railway overhead contact system is a crucial component of the traction power supply system, and its operational status directly affects the safety and stability of train power supply. Currently, the main methods for detecting thermal defects in railway systems include manual inspection and automated detection technology. In traditional methods, manual inspection is still widely used for routine equipment checks. However, due to low manual efficiency, strong reliance on the experience of maintenance personnel, and limited visibility in elevated, complex terrain, or nighttime environments, it is difficult to meet the high-frequency, precise inspection requirements of high-speed and heavy-haul railways.

[0003] To improve efficiency, many non-destructive testing technologies have been introduced into the railway inspection field, such as ultrasonic testing, eddy current testing, lidar, and machine vision. Among these, visual inspection has achieved significant breakthroughs in track geometry and surface defect identification. Deep learning has become the mainstream approach in existing automated inspection methods. Most studies often use target detection or classification networks to achieve component identification and defect judgment, but existing methods still have significant shortcomings in the complex application scenarios of railway overhead contact lines: First, in scenarios where components overlap with mountains, viaducts, etc., and during nighttime inspections, visible light images cannot accurately reflect the structure and shape of the equipment, leading to structural identification failures and misjudgments. Second, under high-speed train operation, ordinary uncooled infrared imaging suffers from problems such as ghosting, blurred contours, and feature loss, reducing the accuracy of target positioning and identification, and also affecting the accuracy of temperature acquisition. Furthermore, during the day, under strong sunlight, infrared images are easily affected by reflective interference; for example, various wire clamps and conductors may experience brief bright reflections, producing false high temperatures. Existing algorithms generally remain at the detection and classification stage, lacking systematic post-processing and fusion of detection results. For example, they lack quantitative analysis of temperature fields and do not combine continuous frame information for statistical judgment of anomalies, thus easily generating a large number of false alarms in short-term pseudo-high temperature scenarios. Finally, multimodal fusion algorithms for infrared and visible light are still crude. Current research mostly relies on simple feature splicing or weighted fusion, which is insufficient for extracting target details in complex backgrounds, leading to the easy omission of small-sized components (such as wire clamps and insulators) during detection.

[0004] In summary, existing detection technologies and algorithms generally suffer from problems such as insufficient discrimination accuracy, lack of continuous processing mechanisms, weak anti-interference capabilities, and imperfect post-processing methods. Therefore, they are difficult to meet the actual needs of high-speed railway catenary for thermal defect detection in complex environments. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method and system for identifying and warning of infrared thermal defects in railway overhead contact lines. This system enables accurate identification and positioning of components under complex environments and high-speed operating conditions. It extracts temperature field data and combines it with continuous frame information for fusion judgment, thereby effectively improving the robustness, anti-interference capability, and diagnostic accuracy of detection, while reducing the false alarm rate.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows:

[0007] In a first aspect, the present invention provides a method for infrared thermal defect identification and early warning of railway catenary, comprising the following steps:

[0008] Step S1: Acquire images of the overhead contact system components. The images of the overhead contact system components include infrared images of the components acquired using a cooled infrared imaging device while the train is in operation, and synchronous visible light images acquired using an industrial visible light camera.

[0009] Step S2: Input the collected images of the contact wire components into the improved YOLOv8 deep learning network for target detection, identification and localization of the contact wire components. The improvements to the YOLOv8 deep learning network include introducing a C2f module into the backbone network to enhance feature extraction capabilities, adopting a BiFPN structure in the feature fusion network to achieve cross-scale feature fusion, and embedding a CBAM attention mechanism in the network to improve attention to thermal defect areas.

[0010] Step S3: Based on the identification and positioning of the bounding boxes of the contact wire components, the corresponding regions are segmented and the temperature matrix is ​​extracted to achieve a quantitative characterization of the temperature field of the contact wire components.

[0011] Step S4: Based on the sequence of the temperature matrix, the temperature change of the same component in multiple frames of images is judged by the continuous frame statistical analysis method, and the persistent thermal defects and transient interference are distinguished by the temporal correlation analysis.

[0012] Step S5: When a thermal defect is determined, the component positioning information from step S2, the temperature analysis results from step S3, and the statistical discrimination conclusion from step S4 are combined to output complete early warning information including the defect location, type, temperature, and warning level.

[0013] As one possible implementation of this embodiment, step S1 includes the following steps:

[0014] Step S11: Install a cooled mid-wave infrared thermal imager on the inspection train and set the acquisition parameters of the infrared thermal imager. The acquisition parameters include: spatial resolution set to 640×512 pixels, frame rate set to 100-200Hz, and noise equivalent temperature difference less than 20mK.

[0015] Step S12: During train operation, 16-bit grayscale full radiation thermal image data of the contact wire components are collected in real time, and the collected infrared image data is stored in the on-board storage device in DDT or CSV format.

[0016] Step S13: Use an industrial visible light camera to synchronously acquire visible light images of the contact wire components, and keep the visible light images and infrared images synchronized in time and space;

[0017] Step S14: Register the infrared image and the visible light image to ensure the consistency of the position of the same target in different modal images;

[0018] Step S15: Establish pixel-level correspondence between infrared and visible light images to achieve precise alignment of multimodal images;

[0019] Step S16: Normalize the registered image to a uniform image size of 640×640 pixels.

[0020] As one possible implementation of this embodiment, the registration processing of the infrared image and the visible light image adopts a feature point-based image registration method, which includes the following steps:

[0021] (a) Extract feature points from infrared and visible light images using SIFT, SURF, or ORB algorithms, respectively;

[0022] (b) Calculate the matching relationship between feature points and estimate the affine transformation parameters between the two images;

[0023] (c) Estimate the geometric transformation parameters between the two images;

[0024] (d) Transform the visible light image according to the geometric transformation parameters to align it with the infrared image.

[0025] As one possible implementation of this embodiment, step S2 includes the following steps:

[0026] Step S21: Input the collected images of the contact wire components into the backbone of the improved YOLOv8 network, and extract multi-scale features through the C2f module;

[0027] Step S22: In the feature fusion stage, a BiFPN structure is used to perform bidirectional cross-scale feature fusion;

[0028] Step S23: Embed the CBAM attention mechanism at key locations in the network to enhance attention to thermal defect areas;

[0029] Step S24: Output the bounding box coordinates, category confidence, and category label of the contact wire component through the detection head.

[0030] As one possible implementation of this embodiment, the C2f module performs feature extraction and gradient flow optimization in the YOLOv8 backbone network through a parallelized Bottleneck path, and its specific structure includes:

[0031] Adjust the number of channels in a 1×1 convolutional layer;

[0032] 3×3 depth separable convolutional layers extract features;

[0033] Residual connectivity maintains gradient flow;

[0034] Channel splicing and fusion of multi-scale features.

[0035] As one possible implementation of this embodiment, the CBAM attention mechanism includes:

[0036] Channel attention module, calculated as ,

[0037] Where σ is the Sigmoid function;

[0038] Spatial attention module, calculated as ;

[0039] Insert the CBAM module into the middle layer of the backbone network and key positions in the feature fusion network.

[0040] As one possible implementation of this embodiment, step S3 includes the following steps:

[0041] Step S31: Based on the bounding box of the contact wire component obtained in step S2, extract the corresponding region of interest from the original infrared image;

[0042] Step S32: Perform pixel-level semantic segmentation on the region of interest to accurately define the component outline;

[0043] Step S33: Extract the temperature value of each pixel based on the segmentation results to form a two-dimensional temperature matrix;

[0044] Step S34: Perform statistical analysis on the temperature matrix to calculate the maximum temperature, minimum temperature, average temperature, and temperature standard deviation.

[0045] As one possible implementation of this embodiment, the semantic segmentation adopts:

[0046] DeepLabv3+ segmentation network, combined with encoder-decoder structure;

[0047] A void space pyramid pooling module captures multi-scale contextual information;

[0048] Boundary-aware loss function improves contour segmentation accuracy.

[0049] As one possible implementation of this embodiment, step S4 includes the following steps:

[0050] Step S41: Establish a continuous frame temperature sequence to track the temperature change of the same component in multiple frames of images;

[0051] Step S42: Perform statistical analysis on the temperature series using the sliding window method;

[0052] Step S43: Calculate the mean, variance, and trend index of temperature changes;

[0053] Step S44: Identify persistent temperature anomalies based on statistical features and filter out transient interference.

[0054] As one possible implementation of this embodiment, the establishment of the continuous frame temperature sequence includes:

[0055] The same component in consecutive frames is associated using a target tracking algorithm;

[0056] Extract the average temperature value of the component in each frame;

[0057] Construct time-temperature series data.

[0058] As one possible implementation of this embodiment, step S5 includes the following steps:

[0059] Step S51: Generate early warning information based on the temperature anomaly identification result. The early warning information includes: precise location information of the defective component, description of the type and degree of temperature anomaly, detection timestamp and GPS coordinates, and relevant image and data evidence.

[0060] Step S52: Determine the warning level and urgency level;

[0061] Step S53: Organize the early warning data packet, which contains complete detection information;

[0062] Step S54: Send early warning information through multiple communication channels.

[0063] Secondly, an embodiment of the present invention provides a railway catenary infrared thermal defect identification and early warning system, comprising:

[0064] The overhead contact line component image acquisition module is used to acquire images of overhead contact line components. The images of the overhead contact line components include infrared images of the overhead contact line components acquired using a cooled infrared imaging device while the train is in operation, and synchronous visible light images acquired using an industrial visible light camera.

[0065] An improved YOLOv8 target detection module is used to input the acquired images of contact wire components into the improved YOLOv8 deep learning network for target detection, identification, and localization of contact wire components. The improvements to the YOLOv8 deep learning network include introducing a C2f module into the backbone network to enhance feature extraction capabilities, adopting a BiFPN structure in the feature fusion network to achieve cross-scale feature fusion, and embedding a CBAM attention mechanism in the network to improve attention to thermal defect areas.

[0066] The temperature matrix analysis module is used to segment the corresponding area and extract the temperature matrix based on the identification and location of the bounding box of the contact wire component, so as to realize the quantitative characterization of the temperature field of the contact wire component.

[0067] The continuous frame statistical discrimination module is used to discriminate the temperature changes of the same component in multiple frames of images based on the sequence of the temperature matrix using the continuous frame statistical analysis method, and to distinguish between persistent thermal defects and transient interference through temporal correlation analysis.

[0068] The early warning output module is used to output complete early warning information, including the defect location, type, temperature and warning level, when a thermal defect is determined, by integrating component location information, step temperature analysis results and statistical judgment conclusions.

[0069] Thirdly, an electronic device provided by an embodiment of the present invention includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of any of the above-described methods for identifying and warning of infrared thermal defects in railway contact wires.

[0070] Fourthly, an embodiment of the present invention provides a storage medium storing a computer program, which, when run by a processor, executes the steps of any of the above-described methods for identifying and warning of infrared thermal defects in railway contact wires.

[0071] One of the above technical solutions has the following advantages or beneficial effects:

[0072] This invention achieves end-to-end intelligent processing from raw infrared images to precise early warning information through the close logical connection of five steps: high-quality data acquisition, intelligent component identification, quantitative temperature analysis, time-series statistical discrimination, and comprehensive early warning output. This significantly improves the accuracy and reliability of thermal defect detection in railway catenary.

[0073] This invention enables accurate identification and positioning of components under complex environments and high-speed operating conditions. It extracts temperature field data and combines it with continuous frame information for fusion judgment, thereby effectively improving the robustness, anti-interference ability and diagnostic accuracy of detection, reducing false alarm rate, and meeting the actual detection needs of railway catenary thermal defects in high-frequency and complex environments.

[0074] This invention, by combining high-quality input from cooled infrared imaging with an improved deep learning detection algorithm, achieves high-precision detection, visibility of small targets, strong anti-interference capabilities, and more reliable temperature discrimination. Compared with existing methods, this invention significantly improves the defect detection capability of railway catenary under complex environments and high-speed operating conditions, providing strong support for ensuring safe train operation and preventive maintenance of equipment. Attached Figure Description

[0075] Figure 1 This is a flowchart illustrating an infrared thermal defect identification and early warning method for railway catenary according to an exemplary embodiment;

[0076] Figure 2 This is a schematic diagram of the structure of an infrared thermal defect identification and early warning system for railway catenary according to an exemplary embodiment;

[0077] Figure 3 This is a pooling schematic diagram illustrated according to an exemplary embodiment;

[0078] Figure 4 This is a visualization of the ReLU function according to an exemplary embodiment;

[0079] Figure 5 This is a visualization of the SiLU function according to an exemplary embodiment;

[0080] Figure 6 This is a schematic diagram of a residual structure according to an exemplary embodiment;

[0081] Figure 7 This is a schematic diagram of an existing YOLOv8 network structure according to an exemplary embodiment;

[0082] Figure 8 This is a schematic diagram of an optimized YOLOv8 network structure according to an exemplary embodiment. Detailed Implementation

[0083] To more clearly illustrate the technical features of the present invention, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0084] like Figure 1As shown in the figure, an embodiment of the present invention provides a method for infrared thermal defect identification and early warning of railway catenary, which includes the following steps:

[0085] Step S1: Acquire images of the overhead contact system components. The images of the overhead contact system components include infrared images of the components acquired using a cooled infrared imaging device while the train is in operation, and synchronous visible light images acquired using an industrial visible light camera.

[0086] Step S2: Input the collected images of the contact wire components into the improved YOLOv8 deep learning network for target detection, identification and localization of the contact wire components. The improvements to the YOLOv8 deep learning network include introducing a C2f module into the backbone network to enhance feature extraction capabilities, adopting a BiFPN structure in the feature fusion network to achieve cross-scale feature fusion, and embedding a CBAM attention mechanism in the network to improve attention to thermal defect areas.

[0087] Step S3: Based on the identification and positioning of the bounding boxes of the contact wire components, the corresponding regions are segmented and the temperature matrix is ​​extracted to achieve a quantitative characterization of the temperature field of the contact wire components.

[0088] Step S4: Based on the sequence of the temperature matrix, the temperature change of the same component in multiple frames of images is judged by the continuous frame statistical analysis method, and the persistent thermal defects and transient interference are distinguished by the temporal correlation analysis.

[0089] Step S5: When a thermal defect is determined, the component positioning information from step S2, the temperature analysis results from step S3, and the statistical discrimination conclusion from step S4 are combined to output complete early warning information including the defect location, type, temperature, and warning level.

[0090] As one possible implementation of this embodiment, step S1 includes the following steps:

[0091] Step S11: Install a cooled mid-wave infrared thermal imager on the inspection train and set the acquisition parameters of the infrared thermal imager. The acquisition parameters include: spatial resolution set to 640×512 pixels, frame rate set to 100-200Hz, and noise equivalent temperature difference less than 20mK.

[0092] Step S12: During train operation, 16-bit grayscale full radiation thermal image data of the contact wire components are collected in real time, and the collected infrared image data is stored in the on-board storage device in DDT or CSV format.

[0093] Step S13: Use an industrial visible light camera to synchronously acquire visible light images of the contact wire components, and keep the visible light images and infrared images synchronized in time and space;

[0094] Step S14: Register the infrared image and the visible light image to ensure the consistency of the position of the same target in different modal images;

[0095] Step S15: Establish pixel-level correspondence between infrared and visible light images to achieve precise alignment of multimodal images;

[0096] Step S16: Normalize the registered image to a uniform image size of 640×640 pixels.

[0097] As one possible implementation of this embodiment, the registration processing of the infrared image and the visible light image adopts a feature point-based image registration method, which includes the following steps:

[0098] (a) Extract feature points from infrared and visible light images using SIFT, SURF, or ORB algorithms, respectively;

[0099] (b) Calculate the matching relationship between feature points and estimate the affine transformation parameters between the two images;

[0100] (c) Estimate the geometric transformation parameters between the two images;

[0101] (d) Transform the visible light image according to the geometric transformation parameters to align it with the infrared image.

[0102] As one possible implementation of this embodiment, step S2 includes the following steps:

[0103] Step S21: Input the collected images of the contact wire components into the backbone of the improved YOLOv8 network, and extract multi-scale features through the C2f module;

[0104] Step S22: In the feature fusion stage, a BiFPN structure is used to perform bidirectional cross-scale feature fusion;

[0105] Step S23: Embed the CBAM attention mechanism at key locations in the network to enhance attention to thermal defect areas;

[0106] Step S24: Output the bounding box coordinates, category confidence, and category label of the contact wire component through the detection head.

[0107] As one possible implementation of this embodiment, the C2f module performs feature extraction and gradient flow optimization in the YOLOv8 backbone network through a parallelized Bottleneck path, and its specific structure includes:

[0108] Adjust the number of channels in a 1×1 convolutional layer;

[0109] 3×3 depth separable convolutional layers extract features;

[0110] Residual connectivity maintains gradient flow;

[0111] Channel splicing and fusion of multi-scale features.

[0112] As one possible implementation of this embodiment, the CBAM attention mechanism includes:

[0113] Channel attention module, calculated as ,

[0114] Where σ is the Sigmoid function;

[0115] Spatial attention module, calculated as ;

[0116] Insert the CBAM module into the middle layer of the backbone network and key positions in the feature fusion network.

[0117] As one possible implementation of this embodiment, step S3 includes the following steps:

[0118] Step S31: Based on the bounding box of the contact wire component obtained in step S2, extract the corresponding region of interest from the original infrared image;

[0119] Step S32: Perform pixel-level semantic segmentation on the region of interest to accurately define the component outline;

[0120] Step S33: Extract the temperature value of each pixel based on the segmentation results to form a two-dimensional temperature matrix;

[0121] Step S34: Perform statistical analysis on the temperature matrix to calculate the maximum temperature, minimum temperature, average temperature, and temperature standard deviation.

[0122] As one possible implementation of this embodiment, the semantic segmentation adopts:

[0123] DeepLabv3+ segmentation network, combined with encoder-decoder structure;

[0124] A void space pyramid pooling module captures multi-scale contextual information;

[0125] Boundary-aware loss function improves contour segmentation accuracy.

[0126] As one possible implementation of this embodiment, step S4 includes the following steps:

[0127] Step S41: Establish a continuous frame temperature sequence to track the temperature change of the same component in multiple frames of images;

[0128] Step S42: Perform statistical analysis on the temperature series using the sliding window method;

[0129] Step S43: Calculate the mean, variance, and trend index of temperature changes;

[0130] Step S44: Identify persistent temperature anomalies based on statistical features and filter out transient interference.

[0131] As one possible implementation of this embodiment, the establishment of the continuous frame temperature sequence includes:

[0132] The same component in consecutive frames is associated using a target tracking algorithm;

[0133] Extract the average temperature value of the component in each frame;

[0134] Construct time-temperature series data.

[0135] As one possible implementation of this embodiment, step S5 includes the following steps:

[0136] Step S51: Generate early warning information based on the temperature anomaly identification result. The early warning information includes: precise location information of the defective component, description of the type and degree of temperature anomaly, detection timestamp and GPS coordinates, and relevant image and data evidence.

[0137] Step S52: Determine the warning level and urgency level;

[0138] Step S53: Organize the early warning data packet, which contains complete detection information;

[0139] Step S54: Send early warning information through multiple communication channels.

[0140] like Figure 2 As shown in the figure, an infrared thermal defect identification and early warning system for railway catenary provided by an embodiment of the present invention includes:

[0141] The overhead contact line component image acquisition module is used to acquire images of overhead contact line components. The images of the overhead contact line components include infrared images of the overhead contact line components acquired using a cooled infrared imaging device while the train is in operation, and synchronous visible light images acquired using an industrial visible light camera.

[0142] An improved YOLOv8 target detection module is used to input the acquired images of contact wire components into the improved YOLOv8 deep learning network for target detection, identification, and localization of contact wire components. The improvements to the YOLOv8 deep learning network include introducing a C2f module into the backbone network to enhance feature extraction capabilities, adopting a BiFPN structure in the feature fusion network to achieve cross-scale feature fusion, and embedding a CBAM attention mechanism in the network to improve attention to thermal defect areas.

[0143] The temperature matrix analysis module is used to segment the corresponding area and extract the temperature matrix based on the identification and location of the bounding box of the contact wire component, so as to realize the quantitative characterization of the temperature field of the contact wire component.

[0144] The continuous frame statistical discrimination module is used to discriminate the temperature changes of the same component in multiple frames of images based on the sequence of the temperature matrix using the continuous frame statistical analysis method, and to distinguish between persistent thermal defects and transient interference through temporal correlation analysis.

[0145] The early warning output module is used to output complete early warning information, including the defect location, type, temperature and warning level, when a thermal defect is determined, by integrating component location information, step temperature analysis results and statistical judgment conclusions.

[0146] To overcome the problems of insufficient discrimination accuracy, weak anti-interference ability, and imperfect post-processing in existing railway catenary thermal defect detection technologies, this embodiment proposes a railway catenary infrared thermal defect identification and early warning system based on cooled infrared imaging and deep learning. By combining hardware imaging devices and improved deep learning algorithms, high-precision detection and discrimination of thermal defects in catenary equipment is achieved.

[0147] The railway catenary infrared thermal defect identification and early warning system of the present invention mainly consists of the following parts:

[0148] 1. Imaging Acquisition Module: This module includes a cooled infrared camera and a visible light camera, which acquire the thermal field distribution and structural texture information of the equipment, respectively. The infrared images contain temperature matrix data, which is used for subsequent defect identification.

[0149] 2. Data labeling and storage module: The collected images are labeled and divided according to a unified format. The dataset is divided into training set, validation set and test set in a ratio of 7:1:2 and stored in a high-speed data storage device.

[0150] 3. Improved YOLOv8 feature extraction and detection module: Based on the YOLOv8 backbone network, the C2f module, BiFPN bidirectional feature fusion path and CBAM attention mechanism are introduced to enhance the detection capability of small target hot spots and improve anti-interference performance.

[0151] 4. Post-processing module: Performs temperature field analysis and continuous frame statistics on the detection results, eliminates false high temperatures such as solar reflection, and improves the reliability of the final defect identification.

[0152] The system acquires infrared and visible light images in real time during train operation → inputs them into an improved YOLOv8 network for target detection (identification + localization) → segments the target area and extracts the temperature matrix → performs continuous frame analysis and temperature threshold discrimination through a post-processing module → outputs the defect type and location.

[0153] The specific process of using the railway catenary infrared thermal defect identification and early warning system of the present invention to identify and warn of railway catenary infrared thermal defects is as follows.

[0154] 1. Data Acquisition:

[0155] During the data acquisition phase, a cooled mid-wave infrared thermal imager with a resolution of 640×512 was used. This imager digitized the infrared signal using a 16-bit ADC, acquiring 16-bit grayscale total radiation thermal image data, which was stored in a custom DDT or CSV format. Simultaneously, an industrial visible light camera was used to acquire auxiliary images. The infrared images not only contain spatial structure information but also pixel-level temperature field data, providing accurate temperature distribution. The acquired data was formatted and stored in an onboard storage device, supporting temperature field data backtracking and retraining.

[0156] 2. Data preprocessing and labeling:

[0157] In the data annotation and integration stage, VOC format (XML annotation) was used to annotate the target boxes of infrared and visible light images. The annotation categories covered key components of the overhead contact system (insulators, clamps, locators, droppers, etc.) and were divided into training set, validation set and test set in a ratio of 7:1:2 to construct a thermal defect dataset for railway overhead contact system equipment.

[0158] 3. Improved YOLOv8 object detection:

[0159] C2f module: Optimizes shallow feature extraction path, preserves edge and texture details, and improves small target detection capability;

[0160] BiFPN bidirectional fusion: replaces the traditional PAN-FPN, enhances feature interaction at different scales, and improves contextual representation and multi-scale detection capabilities;

[0161] CBAM attention mechanism: guides the network to focus on real hotspot areas, suppresses background noise such as mountains, sky and reflections, and reduces false alarm rate.

[0162] The improved YOLOv8 network can maintain stable detection performance in high-speed driving, complex backgrounds, and day and night environments.

[0163] 4. Temperature matrix extraction and defect identification:

[0164] The detected target region is segmented into an image, and the corresponding temperature matrix is ​​extracted.

[0165] Based on the structural characteristics of the components, a dynamic temperature threshold is set to conduct preliminary defect identification;

[0166] By comparing consecutive frames and using statistical methods, short-term false high temperatures (such as solar reflections) are filtered out.

[0167] 5. Outputs and Alarms:

[0168] The final output includes the location, type, and temperature information of defects in the overhead contact line equipment.

[0169] If the temperature exceeds the warning threshold, an alarm will be triggered and the data will be recorded for maintenance personnel to refer to.

[0170] Compared with the prior art, the present invention has the following significant advantages:

[0171] 1. High-precision detection: Based on the improved YOLOv8 multi-level optimization structure, detection accuracy and recall are significantly improved;

[0172] 2. Strong small target recognition capability: The C2f module enhances the feature extraction effect of small components such as insulators and wire clamps;

[0173] 3. Stable multi-scale representation: BiFPN enhances cross-scale feature interaction capabilities;

[0174] 4. Strong anti-interference capability: CBAM effectively suppresses background noise and reduces false alarms;

[0175] 5. More reliable temperature judgment: Combining temperature matrix with continuous frame analysis avoids misjudgment caused by single-frame anomalies;

[0176] 6. Strong applicability: Stable detection can be achieved under high-speed operation, day and night changes, and complex backgrounds.

[0177] The improvement process of the YOLOv8 deep learning network is as follows.

[0178] (a) Basic Units of Convolutional Neural Networks:

[0179] The core of deep learning in computer vision is the Convolutional Neural Network (CNN), whose main function is to extract features from input images layer by layer, obtaining semantic information at different levels from shallow to deep. In this invention, YOLOv8 and its optimized structure are built on the foundation of CNN.

[0180] 1. Convolutional Layer:

[0181] Convolutional layers extract features from the input image by sliding the convolutional kernel (filter). The formula is as follows:

[0182] ,

[0183] in, For input pixels, is the kernel weight, and b is the bias term.

[0184] In infrared images, convolutional layers can extract edge features of hot spots and temperature gradient changes.

[0185] Different layers of convolutional kernels focus on different features: shallow layers focus on texture and edges, while deep layers focus on the overall shape of the target and its contextual semantics.

[0186] 2. Pooling Layer:

[0187] like Figure 3 As shown, pooling layers commonly use max pooling and average pooling. Their main functions are: reducing feature map size and computational load; enhancing the translation invariance of features; and in infrared target detection, pooling helps eliminate noise and highlight the target region.

[0188] 3. Activation Function:

[0189] YOLOv8 uses the SiLU (Sigmoid LinearUnit) activation function:

[0190] .

[0191] compared to Figure 4 The ReLU shown Figure 5 The SiLU algorithm, as shown, retains some gradients even when the gradient is less than zero, thus improving the network's expressive power and convergence performance. In infrared scenes, SiLU performs better in handling weak-contrast hotspots.

[0192] 4. Residual Block:

[0193] To avoid the vanishing gradient problem during deep network training, residual connections are introduced:

[0194]

[0195] Where F(x) is the convolution and nonlinear transformation, and x is the identity mapping.

[0196] like Figure 6 As shown, the residual structure ensures that the network can learn "incremental features" without degrading due to excessive depth. The C2f module in the YOLOv8 backbone is an improvement based on this idea.

[0197] (II) YOLOv8 Basic Structure:

[0198] The YOLO (You Only Look Once) series is a typical end-to-end object detection network. Building upon the principles of the YOLO series, YOLOv8 adopts an anchor-free approach and introduces a decoupled head and lightweight design, achieving a balance between efficiency and high accuracy. Its overall structure typically consists of three parts: Input Layer, Backbone, and Head.

[0199] In this study, the acquired cooled infrared images were single-channel grayscale images. To ensure compatibility with the official YOLOv8 pre-trained weights, we copied the grayscale images into a 3-channel pseudo-color format and input them into the YOLOv8 model. The advantage of this approach is that it allows for direct reuse of large-scale pre-trained models without modifying the network structure, thereby improving convergence speed and detection accuracy.

[0200] 1. Input Layer:

[0201] The input end primarily performs image preprocessing to ensure it meets the requirements of network computation. The main steps are as follows:

[0202] 1.1. Scaling and Filling (Letterbox):

[0203] YOLOv8's default input size is 640×640. Our acquired infrared image is 640×512, so the system will perform a 1.25x magnification in the width direction: (640,512) → (640,640).

[0204] This process does not produce black borders, but it does cause some stretching in the width direction, which may cause slight distortion of the target shape.

[0205] 1.2. Normalization:

[0206] The input pixel values ​​are linearly scaled from [0,255] to [0,1] to reduce the impact of the numerical range on training stability.

[0207] 1.3. Data Augmentation (Training Phase Only):

[0208] Methods such as random flipping, rotation, translation, scaling, and mosaic enhancement are used to improve the robustness and generalization ability of the model.

[0209] 1.4. Batch Processing:

[0210] After preprocessing, multiple images are combined into a batch, forming a tensor with dimensions (B, 3, 640, 640), which is then fed into the subsequent network structure for feature extraction and detection.

[0211] 2. YOLOv8 network structure and advantages:

[0212] YOLOv8 is a single-stage object detection algorithm, such as... Figure 7 As shown, it adopts the classic three-section structure of Backbone–Neck–Head.

[0213] (1) Backbone:

[0214] For feature extraction from the input image, YOLOv8 replaces the earlier YOLO CSPDarknet structure with an improved C2f module. The C2f module optimizes feature extraction and gradient flow through a parallelized Bottleneck path, improving feature representation while maintaining lightweight architecture. At the end of the backbone, YOLOv8 employs SPPF (Spatial Pyramid Pooling Fast), which expands the receptive field and integrates contextual information through multi-scale pooling operations.

[0215] The basic formula for convolution is:

[0216] ,

[0217] in For the input feature map, For convolution kernel, These represent the core size and output, respectively. This represents the extracted features.

[0218] CBS: Basic Convolutional Module, usually a combination of "Conv (convolution) + BN (batch normalization) + activation function (such as SiLU)," is the basic unit for feature extraction;

[0219] C2F: A feature extraction module commonly used in YOLO series (such as YOLOv8) (improved from CSPNet), which extracts richer features through a multi-branch residual structure;

[0220] SPPF: Fast Spatial Pyramid Pooling Module, which can fuse spatial features of different scales while reducing computational load.

[0221] (2) Neck:

[0222] YOLOv8 uses PAN-FPN (Path Aggregation Network – Feature Pyramid Network) for multi-scale feature fusion. FPN passes high-level semantic information from top to bottom, while PAN passes localization information from bottom to top; the combination of the two enhances the detection capability of targets at different scales. Upsample: Upsampling (such as bilinear interpolation) amplifies deep, low-resolution, high-semantic features to match shallow, high-resolution features;

[0223] Concat: Channel-level concatenation combines "shallow high-resolution detail features" and "deep low-resolution semantic features," allowing subsequent detection to utilize both types of information simultaneously.

[0224] (3) Head:

[0225] An anchor-free detection head is used to directly predict the center point and size of the target bounding box, reducing the complexity of anchor box matching. The output layer is divided into three scales (small, medium, and large) to correspond to the detection of targets of different sizes.

[0226] (4) Summary of advantages:

[0227] Lightweighting: C2f improves gradient flow efficiency;

[0228] Multi-scale fusion: PAN-FPN takes into account targets at different scales;

[0229] Accurate detection: Anchor-free design reduces false matches;

[0230] High real-time performance: Suitable for real-time detection in high-speed scenarios.

[0231] 3. Shortcomings of YOLOv8 in railway infrared detection:

[0232] Although YOLOv8 has strong target detection performance, its direct application in infrared thermal imaging inspection of railway overhead contact lines still has the following shortcomings:

[0233] 3.1. Lack of attention:

[0234] In complex backgrounds (such as mountains, buildings, and solar reflection interference), YOLOv8 is susceptible to false hotspots.

[0235] The original C2f and SPPF modules lack an explicit attention mechanism, and the network does not focus enough on thermal defect areas.

[0236] 3.2. Limitations of Feature Fusion:

[0237] The feature fusion method of PAN-FPN uses fixed weights and lacks dynamic adjustment capabilities, which makes it easy for information of small targets (such as wire clamps and insulators) to be submerged in multi-scale fusion.

[0238] Especially in high-speed train environments, infrared imaging ghosting leads to a loss of detail, and fixed fusion methods are unable to recover effective features.

[0239] 3.3. Insufficient subsequent feature extraction:

[0240] The PAN-FPN output is directly input into the detection head, lacking further compression and feature reconstruction, resulting in redundancy and noise in the features of the final input head.

[0241] To address the aforementioned problems, this invention proposes an improved network structure for detecting thermal defects in railway overhead contact lines based on YOLOv8. The main optimizations include the CBAM attention mechanism, BiFPN bidirectional feature fusion, and C2f refinement. Its overall structure is as follows: Figure 8 As shown.

[0242] 1. Backbone optimization: Introducing the CBAM attention module;

[0243] CBAM (Convolutional Block Attention Module) comprises two sub-modules: channel attention and spatial attention.

[0244] Channel attention:

[0245] ,

[0246] Where σ is the Sigmoid function;

[0247] Spatial attention:

[0248] ,

[0249] Spatial attention maps are generated through convolution operations to guide the model to focus on key regions.

[0250] Inserting CBAM into the middle layer of the Backbone effectively improves the model's sensitivity to thermal defect areas in infrared images and reduces false hotspots caused by reflective interference.

[0251] 2. Neck optimization:

[0252] BiFPN replaces PAN-FPN. BiFPN (Bidirectional Feature Pyramid Network) is an improvement on FPN, employing bidirectional cross-layer connections and learnable weight fusion.

[0253] ,

[0254] in Input features for different layers, For learnable weights, To prevent the denominator from being zero.

[0255] BiFPN can automatically learn the importance of features at different scales, ensuring that information about small targets is not masked during fusion, thereby improving the detection accuracy of small components such as wire clamps and insulators. Adding a C2f module after BiFPN fusion further refines the fused features, enhances nonlinear expressive power, and suppresses redundant information.

[0256] C2f's Bottleneck structure ensures smooth gradient flow and prevents gradient vanishing in deep networks.

[0257] 5. Compared with the original YOLOv8, the optimized YOLOv8 of this invention has the following advantages:

[0258] Significantly reduces false alarm rate: CBAM enables the network to focus on thermal defect areas, suppressing false hotspots caused by solar reflection;

[0259] Improving the accuracy of small target detection: BiFPN provides bidirectional feature interaction and improves the ability to recognize small-sized parts through learnable weights;

[0260] Enhanced feature representation and robustness: C2f refinement ensures that the Head receives cleaner and more compact features, improving overall detection robustness;

[0261] Adaptable to complex scenarios and high-speed conditions: The improved network can still maintain stable detection performance and temperature analysis capabilities under complex conditions such as nighttime, mountainous areas, and high-speed train operation.

[0262] This invention, by introducing an improved deep learning detection algorithm based on cooled infrared imaging, and supplementing it with post-processing and multi-frame statistical discrimination, can achieve the following technical effects in the detection of thermal defects in railway catenary equipment:

[0263] First, this invention utilizes a cooled infrared imaging device to acquire images. Compared to ordinary uncooled infrared devices, its detector operates in a low-temperature environment, resulting in higher sensitivity and lower noise levels. Even under conditions of high-speed train operation, low-light nighttime conditions, and strong light interference, it can still output infrared images with clear boundaries and accurate temperature distribution, thus providing high-quality input for subsequent algorithm processing.

[0264] Secondly, in the feature extraction and target detection stages, this invention has made several optimizations based on the YOLOv8 backbone structure:

[0265] The introduction of the C2F module improves the ability to extract shallow features, allowing the complete preservation of device edges and texture details, and significantly enhancing the ability to recognize small-sized components (such as wire clamps and insulators); the C2F module is present in YOLOv8 itself.

[0266] The BiFPN bidirectional feature fusion path improves the multi-scale feature interaction mechanism, enabling the model to have stable detection performance when processing both large-scale wires and small-scale components simultaneously.

[0267] The CBAM attention mechanism further enhances the network's ability to focus on key hotspots, effectively suppresses interference from complex backgrounds such as mountains, buildings, sky, and solar reflections, and significantly reduces the false alarm rate of false hotspots.

[0268] Furthermore, in the defect identification stage, this invention goes beyond traditional detection and classification. It achieves quantitative defect analysis by segmenting the detection area and extracting a temperature matrix. Combined with dynamic temperature threshold settings, it can accurately distinguish between the normal operating state and abnormal temperature rise state of different components. Simultaneously, the continuous frame comparison and statistical post-processing method proposed in this invention can effectively eliminate single-frame interference (such as short-term pseudo-high temperatures caused by solar reflection), ensuring the continuity and stability of the identification results.

[0269] Finally, this invention provides real-time and reliable defect location and alarm functions at the output and application levels. The detection results include not only the component category and spatial location, but also the corresponding temperature parameters, thus providing a quantifiable basis for the condition assessment and maintenance decisions of the railway catenary.

[0270] In summary, by combining the high-quality input of cooled infrared imaging with an improved deep learning detection algorithm, this invention achieves high-precision detection, visibility of small targets, strong anti-interference capabilities, and more reliable temperature discrimination. Compared with existing methods, this invention significantly improves the defect detection capability of railway catenary under complex environments and high-speed operating conditions, providing strong support for ensuring safe train operation and preventive maintenance of equipment.

[0271] An electronic device provided by an embodiment of the present invention includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of any of the above-described methods for identifying and warning of infrared thermal defects in railway contact wires.

[0272] Specifically, the aforementioned memory and processor can be general-purpose memory and processor, without any specific limitations. When the processor runs the computer program stored in the memory, it can execute the aforementioned method for identifying and warning of infrared thermal defects in railway catenary.

[0273] Corresponding to the above application startup method, this embodiment of the invention also provides a storage medium storing a computer program, which, when run by a processor, executes the steps of any of the above railway catenary infrared thermal defect identification and early warning methods.

[0274] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for infrared thermal defect identification and early warning of railway overhead contact lines, characterized in that, Includes the following steps: Step S1: Acquire images of the overhead contact system components. The images of the overhead contact system components include infrared images of the components acquired using a cooled infrared imaging device while the train is in operation, and synchronous visible light images acquired using an industrial visible light camera. Step S2: Input the collected images of the contact wire components into the improved YOLOv8 deep learning network for target detection, identification and localization of the contact wire components. The improvements to the YOLOv8 deep learning network include introducing a C2f module into the backbone network to enhance feature extraction capabilities, adopting a BiFPN structure in the feature fusion network to achieve cross-scale feature fusion, and embedding a CBAM attention mechanism in the network to improve attention to thermal defect areas. Step S3: Based on the identification and positioning of the bounding boxes of the contact wire components, the corresponding regions are segmented and the temperature matrix is ​​extracted to achieve a quantitative characterization of the temperature field of the contact wire components. Step S4: Based on the sequence of the temperature matrix, the temperature change of the same component in multiple frames of images is judged by the continuous frame statistical analysis method, and the persistent thermal defects and transient interference are distinguished by the temporal correlation analysis. Step S5: When a thermal defect is determined, the component positioning information from step S2, the temperature analysis results from step S3, and the statistical discrimination conclusion from step S4 are combined to output complete early warning information including the defect location, type, temperature, and warning level.

2. The method for infrared thermal defect identification and early warning of railway catenary according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Install a cooled mid-wave infrared thermal imager on the inspection train and set the acquisition parameters of the infrared thermal imager. The acquisition parameters include: spatial resolution set to 640×512 pixels, frame rate set to 100-200Hz, and noise equivalent temperature difference less than 20mK. Step S12: During train operation, 16-bit grayscale full radiation thermal image data of the contact wire components are collected in real time, and the collected infrared image data is stored in the on-board storage device. Step S13: Use an industrial visible light camera to synchronously acquire visible light images of the contact wire components, and keep the visible light images and infrared images synchronized in time and space; Step S14: Register the infrared image and the visible light image to ensure the consistency of the position of the same target in different modal images; Step S15: Establish pixel-level correspondence between infrared and visible light images to achieve precise alignment of multimodal images; Step S16: Normalize the registered image to a uniform image size of 640×640 pixels.

3. The method for infrared thermal defect identification and early warning of railway catenary according to claim 2, characterized in that, The registration process for infrared and visible light images employs a feature-point-based image registration method, including the following steps: (a) Extract feature points from infrared and visible light images using SIFT, SURF, or ORB algorithms, respectively; (b) Calculate the matching relationship between feature points and estimate the affine transformation parameters between the two images; (c) Estimate the geometric transformation parameters between the two images; (d) Transform the visible light image according to the geometric transformation parameters to align it with the infrared image.

4. The method for infrared thermal defect identification and early warning of railway catenary according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Input the collected images of the contact wire components into the backbone of the improved YOLOv8 network, and extract multi-scale features through the C2f module; Step S22: In the feature fusion stage, a BiFPN structure is used to perform bidirectional cross-scale feature fusion; Step S23: Embed the CBAM attention mechanism at key locations in the network to enhance attention to thermal defect areas; Step S24: Output the bounding box coordinates, category confidence, and category label of the contact wire component through the detection head.

5. The method for infrared thermal defect identification and early warning of railway catenary according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Based on the bounding box of the contact wire component obtained in step S2, extract the corresponding region of interest from the original infrared image; Step S32: Perform pixel-level semantic segmentation on the region of interest to accurately define the component outline; Step S33: Extract the temperature value of each pixel based on the segmentation results to form a two-dimensional temperature matrix; Step S34: Perform statistical analysis on the temperature matrix to calculate the maximum temperature, minimum temperature, average temperature, and temperature standard deviation.

6. The method for infrared thermal defect identification and early warning of railway catenary according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Establish a continuous frame temperature sequence to track the temperature change of the same component in multiple frames of images; Step S42: Perform statistical analysis on the temperature series using the sliding window method; Step S43: Calculate the mean, variance, and trend index of temperature changes; Step S44: Identify persistent temperature anomalies based on statistical features and filter out transient interference.

7. The method for infrared thermal defect identification and early warning of railway catenary according to any one of claims 1-6, characterized in that, Step S5 includes the following steps: Step S51: Generate early warning information based on the temperature anomaly identification result. The early warning information includes: precise location information of the defective component, description of the type and degree of temperature anomaly, detection timestamp and GPS coordinates, and relevant image and data evidence. Step S52: Determine the warning level and urgency level; Step S53: Organize the early warning data packet, which contains complete detection information; Step S54: Send early warning information through multiple communication channels.

8. A railway overhead contact line infrared thermal defect identification and early warning system, characterized in that, include: The overhead contact line component image acquisition module is used to acquire images of overhead contact line components. The images of the overhead contact line components include infrared images of the overhead contact line components acquired using a cooled infrared imaging device while the train is in operation, and synchronous visible light images acquired using an industrial visible light camera. An improved YOLOv8 target detection module is used to input the acquired images of contact wire components into the improved YOLOv8 deep learning network for target detection, identification, and localization of contact wire components. The improvements to the YOLOv8 deep learning network include introducing a C2f module into the backbone network to enhance feature extraction capabilities, adopting a BiFPN structure in the feature fusion network to achieve cross-scale feature fusion, and embedding a CBAM attention mechanism in the network to improve attention to thermal defect areas. The temperature matrix analysis module is used to segment the corresponding area and extract the temperature matrix based on the identification and location of the bounding box of the contact wire component, so as to realize the quantitative characterization of the temperature field of the contact wire component. The continuous frame statistical discrimination module is used to discriminate the temperature changes of the same component in multiple frames of images based on the sequence of the temperature matrix using the continuous frame statistical analysis method, and to distinguish between persistent thermal defects and transient interference through temporal correlation analysis. The early warning output module is used to output complete early warning information, including the defect location, type, temperature and warning level, when a thermal defect is determined, by integrating component location information, step temperature analysis results and statistical judgment conclusions.

9. An electronic device, characterized in that, The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions that the processor can execute. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the railway catenary infrared thermal defect identification and early warning method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, performs the steps of the railway catenary infrared thermal defect identification and early warning method as described in any one of claims 1-7.

Images