A railway track wear line detection system and method

Abstract

The present application relates to railway track detection technology field, specifically a kind of railway track abrasion line detection system and method, system includes multispectral-terahertz composite imaging module, adaptive laser point cloud scanning module, edge intelligent processing unit, space-time synchronous positioning module and track digital twin platform;Multispectral-terahertz composite imaging module synchronously acquires track surface multispectral image and internal terahertz tomographic image, adaptive laser point cloud scanning module dynamically adjusts scanning parameter according to preliminary detection result and adapts curve track, edge intelligent processing unit realizes space-time registration of multimodal data, hierarchical wear identification and three-level data transmission, space-time synchronous positioning module provides sub-centimeter level positioning and microsecond level hardware synchronization, track digital twin platform constructs full life cycle digital model, realizes wear evolution simulation, hierarchical early warning and intelligent maintenance decision-making.The integration of surface wear and internal defect is realized with high precision detection.

Description

Technical Field

[0001] This invention relates to the field of railway track inspection technology, specifically a railway track wear line detection system and method. Background Technology

[0002] As the backbone of the national comprehensive transportation system, railways undertake a large number of passenger and freight transport tasks, and their safe operation is directly related to the development of the national economy and the safety of people's lives and property. The track, as the basic structure of the railway system, directly bears the load of trains and inevitably experiences wear and damage during long-term operation. Track wear mainly includes railhead side wear, top surface wear, and corrugated wear. These types of wear not only reduce the smoothness and comfort of train operation and increase the force between the wheel and rail, accelerating the damage to track and vehicle components, but can also lead to serious safety accidents such as train derailments in severe cases.

[0003] Besides surface wear, internal defects such as fatigue cracks and delamination can also develop within the track. These internal defects are usually caused by a combination of repeated train loading, material fatigue, and environmental factors. They are difficult to detect in the early stages but will continue to expand over time, eventually leading to track breakage and catastrophic consequences. Statistics show that railway accidents caused by internal track defects account for a significant proportion of all track accidents. Therefore, timely detection and assessment of internal track defects are crucial for ensuring railway safety.

[0004] Currently, existing track wear detection technologies mainly include manual inspection, traditional machine vision inspection, and laser scanning inspection. Manual inspection relies on inspectors using simple tools to measure section by section, which is labor-intensive, inefficient, and highly subjective. Moreover, it can only be carried out during maintenance windows, which cannot meet the needs of high-density railway operations. Traditional machine vision inspection uses visible light cameras to capture images of the track surface and identifies wear areas through image processing algorithms. Although this improves efficiency, it is greatly affected by lighting and weather conditions, can only obtain two-dimensional surface information, and cannot measure wear depth or detect internal defects.

[0005] Laser scanning inspection technology can acquire three-dimensional contour information of tracks with high measurement accuracy and is less affected by lighting conditions, but it still has some significant limitations. First, existing laser scanning systems typically use fixed scanning frequencies and point cloud densities. To ensure detection accuracy, high scanning parameters need to be set, which generates a large amount of redundant data, increasing the burden on data processing and transmission. Second, single laser scanning technology cannot distinguish between different types of wear, nor can it identify surface defects such as cracks and spalling, let alone detect internal track defects. Furthermore, most existing inspection systems lack a unified spatiotemporal reference; data collected by different sensors exhibit temporal and spatial deviations, affecting the accuracy of data fusion and the reliability of the inspection results.

[0006] Of particular note are the common shortcomings of existing technologies in curved track detection. Due to the variations in superelevation and curvature of curved tracks, lidar with a fixed scanning angle often fails to completely cover the functional edge area of ​​the rail head, leading to a significant decrease in the accuracy of wear detection on curved sections. These curved sections are precisely the areas where track wear is most severe and where safety accidents are most likely to occur. Furthermore, most existing detection systems use full data uploads, which are prone to data loss and delays in areas with unstable 5G network coverage, failing to meet the requirements for real-time detection.

[0007] In recent years, although some track inspection systems combining multiple sensors have emerged, most of these systems simply overlay data from different sensors without achieving true multimodal data fusion, thus failing to fully leverage the advantages of different sensors. Moreover, most existing inspection systems only focus on the current wear condition, lacking prediction of wear trends and optimization of maintenance plans, and are unable to achieve preventative maintenance and intelligent operation and maintenance.

[0008] Therefore, there is an urgent need to develop a railway track wear line detection system and method that can simultaneously detect track surface wear and internal defects, adapt to curved track inspection, has high data processing efficiency, and has a unified spatiotemporal reference, in order to meet the needs of modern railway safe operation and intelligent maintenance. Summary of the Invention

[0009] To address the problems in existing technologies, this invention provides a railway track wear line detection system and method. This system and method can simultaneously achieve integrated detection of track surface wear and internal defects, solving the problems of traditional technologies being unable to detect internal defects, incomplete scanning of curved tracks, and low data transmission efficiency. This improves detection accuracy and efficiency, and provides strong support for intelligent operation and maintenance of railway tracks.

[0010] The technical solution adopted by this invention to solve its technical problem is: a railway track wear line detection system, comprising a multispectral-terahertz composite imaging module, an adaptive laser point cloud scanning module, an edge intelligent processing unit, a spatiotemporal synchronous positioning module, and a track digital twin platform; the multispectral-terahertz composite imaging module integrates visible light, near-infrared, and short-wave infrared cameras and a terahertz imager for simultaneously acquiring track surface texture images and internal tomographic images; the adaptive laser point cloud scanning module uses a variable frequency line lidar, which can dynamically adjust the scanning frequency and point cloud density according to the feedback signal from the edge intelligent processing unit; the edge intelligent... The processing unit is electrically connected to the multispectral-terahertz composite imaging module, the adaptive laser point cloud scanning module, and the spatiotemporal synchronous positioning module, respectively, to complete multimodal data spatiotemporal registration, preliminary identification of surface wear, screening of internal defects, and hierarchical data transmission. The spatiotemporal synchronous positioning module adopts BeiDou RTK, inertial navigation, and odometer multi-source fusion technology to provide sub-centimeter-level positioning accuracy and microsecond-level time synchronization. The track digital twin platform is connected to the edge intelligent processing unit through a 5G private network to construct a three-dimensional digital twin model of the track, complete accurate calculation of wear, quantitative analysis of internal defects, simulation of wear evolution, and optimization of maintenance schemes.

[0011] Specifically, the multispectral-terahertz composite imaging module also includes an adaptive composite illumination unit, which integrates a multi-band LED array and a terahertz source, and can automatically adjust the illumination intensity and terahertz emission power according to environmental conditions; the four imaging devices adopt a hardware synchronous triggering method, with a unified trigger signal provided by the spatiotemporal synchronization positioning module, and the imaging time synchronization error is less than 10 microseconds, and are installed on a rigid heat-insulating bracket with a three-level shock absorption device.

[0012] Specifically, the scanning frequency of the adaptive laser point cloud scanning module can be dynamically adjusted between 50Hz and 200Hz, and the point cloud density can be continuously varied between 30 and 120 cross-sections per meter. The module has a built-in track curve detection unit, which can automatically adjust the scanning angle of the laser radar according to the radius of the track curve to ensure that the scanning range completely covers the active and non-active areas of the left and right rail heads.

[0013] Specifically, the edge intelligent processing unit has a built-in Transformer-based multimodal wear recognition model that can process image data and point cloud data simultaneously, and identify wear on the side of the rail head, wear on the top surface, wavy wear, as well as fatigue cracks and delamination defects inside the rail. It adopts a three-level data hierarchical transmission strategy, with different compression ratios and transmission priorities set for different levels of data.

[0014] Specifically, the spatiotemporal synchronization positioning module has a built-in PTP precise time protocol unit, which provides a unified time reference for all sensors, with a time synchronization error of less than 1 microsecond; the external parameter joint calibration adopts a combination of hand-eye calibration and three-dimensional target calibration, and the calibration is verified by standard gauge blocks after calibration, with a spatial calibration error of less than 0.1mm.

[0015] Specifically, the track digital twin platform has a built-in four-level early warning mechanism, and early warning information of different levels is pushed to the corresponding level of management personnel; the wear amount is calculated based on the rail jaw part of the non-functional edge of the rail, and internal defects are divided into three levels: general, relatively heavy and severe according to size, depth and direction of expansion. Severe defects immediately trigger a red warning.

[0016] A method for detecting wear lines on railway tracks, implemented using the aforementioned system, includes the following steps: S1: System initialization, completing multi-sensor intrinsic parameter calibration, extrinsic parameter joint calibration, and spatiotemporal synchronization calibration; S2: Synchronously acquire multispectral images of the orbital surface, internal terahertz images, 3D point cloud data, and corresponding spatiotemporal coordinate information in basic mode; S3: The edge intelligent processing unit performs spatiotemporal registration and preprocessing on multimodal data to extract preliminary features; S4: Use a multimodal wear identification model for rapid screening to mark suspected wear areas and suspected defect areas; S5: Based on the screening results, send adjustment instructions to the adaptive laser point cloud scanning module to perform high-precision scanning of the suspected area; S6: Upload high-precision scan data and multimodal data of suspected areas to the track digital twin platform; S7: The track digital twin platform integrates multi-source data to update the digital twin model, calculates accurate wear and defect parameters, simulates wear evolution trends, and generates the optimal maintenance plan.

[0017] Specifically, in step S1, after the external parameters are jointly calibrated, a standard gauge block is used to verify the accuracy. If the verification error exceeds 0.1 mm, the calibration is repeated. In step S2, the multispectral-terahertz composite imaging module and the adaptive laser point cloud scanning module are synchronously triggered by hardware to ensure that the data collected at the same time correspond to the same physical position on the track.

[0018] Specifically, in step S4, the multimodal wear recognition model adopts a two-level inference mechanism: the first level uses only multispectral images to quickly screen out high-suspection areas with a confidence level higher than 0.4; the second level combines the point cloud data and terahertz images corresponding to the high-suspection areas to perform fine inference and output the final recognition result.

[0019] Specifically, in step S7, the wear amount is calculated based on the rail jaw part of the non-functional edge of the rail, and the wear amount on the side of the rail head, the wear amount on the top surface, and the wear volume are calculated; the defect level is classified according to the internal defect parameters, and a red warning is immediately generated for severe defects and pushed to the highest level of management personnel.

[0020] The beneficial effects of this invention are: This invention employs multispectral-terahertz composite imaging technology, which can simultaneously acquire multispectral images of the track surface and terahertz tomographic images of the interior, achieving integrated detection of surface wear and internal defects, and solving the problem that traditional technologies cannot detect hidden defects inside the track.

[0021] This invention employs adaptive laser point cloud scanning technology, which can dynamically adjust scanning parameters based on preliminary detection results and has a built-in track curve detection unit to automatically adjust the scanning angle. This solves the problem of incomplete track curve scanning in traditional technologies and greatly reduces redundant data while ensuring detection accuracy.

[0022] This invention employs a three-level data hierarchical transmission strategy, setting different compression ratios and transmission priorities according to the importance of the data, which significantly reduces network bandwidth requirements and improves the reliability and real-time performance of data transmission.

[0023] This invention employs a multimodal wear recognition model with a two-level inference mechanism, which balances detection speed and recognition accuracy. It can quickly filter normal area data and concentrate resources to process suspected areas, thereby improving the overall detection efficiency of the system.

[0024] This invention uses the rail jaws, which are not the active edge of the rail, as the benchmark for wear calculation. It adopts a four-level early warning mechanism and a three-level defect classification standard, which improves the accuracy and comparability of the detection results and facilitates standardized operation and maintenance by railway departments. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Figure 1 An architecture diagram of a railway track wear line detection system provided by the present invention; Figure 2 A flowchart of a railway track wear line detection method provided by the present invention. Detailed Implementation

[0027] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0028] like Figure 1As shown, the railway track wear line detection system of the present invention includes a multispectral-terahertz composite imaging module, an adaptive laser point cloud scanning module, an edge intelligent processing unit, a spatiotemporal synchronous positioning module, and a track digital twin platform.

[0029] The multispectral-terahertz composite imaging module integrates a visible light camera, a near-infrared camera, a short-wave infrared camera, and a terahertz imager. The visible light camera clearly captures the texture and color information of the track surface; the near-infrared camera is sensitive to moisture and oil on the track surface; and the short-wave infrared camera can penetrate rust and dust on the track surface to reveal wear beneath the surface. The terahertz imager emits terahertz waves and receives reflected signals. Terahertz waves have strong penetrating power, allowing them to penetrate the track material and obtain tomographic images of the track's interior, thereby detecting latent defects such as fatigue cracks, delamination, and inclusions. By fusing imaging information from these four different modes, the surface and internal condition of the track can be comprehensively and accurately reflected.

[0030] The multispectral-terahertz composite imaging module also includes an adaptive composite illumination unit, integrating a multi-band LED array and a terahertz source. It can automatically adjust the illumination intensity and terahertz emission power of different bands according to ambient light intensity, temperature, and humidity. The four imaging devices employ a hardware synchronous triggering method, with a unified trigger signal provided by a spatiotemporal synchronization positioning module. The imaging time synchronization error is less than 10 microseconds, ensuring that images acquired at the same time correspond to the same physical position on the track. The imaging devices are mounted on a rigid, heat-insulated bracket with a three-stage vibration damping system, effectively isolating vibrations and engine heat during vehicle movement and ensuring stable image quality.

[0031] The adaptive laser point cloud scanning module employs a variable frequency linear lidar, which dynamically adjusts the scanning frequency and point cloud density based on feedback signals from the edge intelligent processing unit. On normal road sections, the system performs a basic scan with a lower scanning frequency and point cloud density to reduce data volume. When a suspected wear or defect area is detected, the system automatically increases the scanning frequency and point cloud density to perform a high-precision scan of the area, acquiring more detailed three-dimensional contour information. The module incorporates a track curve detection unit, which automatically adjusts the lidar's scanning angle according to the track curve radius, ensuring complete coverage of both the active and inactive rail edges on curved sections, thus solving the problem of incomplete scanning of curved tracks using traditional techniques.

[0032] The edge intelligent processing unit is electrically connected to the multispectral-terahertz composite imaging module, the adaptive laser point cloud scanning module, and the spatiotemporal synchronous positioning module, and is the core processing unit of the entire system. The edge intelligent processing unit has a built-in high-performance computing chip, capable of locally performing spatiotemporal registration, preprocessing, feature extraction, and preliminary recognition of multimodal data. The edge intelligent processing unit incorporates a Transformer-based multimodal wear recognition model. Employing a multimodal Transformer architecture, it can simultaneously process image data and point cloud data, automatically learn the correlations between different modalities, and extract more comprehensive and effective features.

[0033] The edge intelligent processing unit employs a three-tiered data transmission strategy: Level 1 data consists of wear statistics and location information for normal areas, transmitted using the highest compression ratio; Level 2 data consists of low-resolution multispectral images and sparse point cloud data for suspected areas, transmitted using a medium compression ratio; and Level 3 data consists of high-precision multispectral images, terahertz images, and dense point cloud data for suspected areas, transmitted using lossless compression. Different transmission priorities are assigned to different levels of data, with Level 3 data having the highest priority, ensuring that critical data can be transmitted to the cloud in a timely and reliable manner.

[0034] The spatiotemporal synchronization positioning module employs a multi-source fusion technology combining BeiDou RTK, inertial navigation, and odometer, providing sub-centimeter-level positioning accuracy and microsecond-level time synchronization. BeiDou RTK technology provides high-precision absolute position information, inertial navigation provides continuous position and attitude information even when satellite signals are blocked, and odometer provides accurate relative distance information. By fusing these three types of information using an extended Kalman filter algorithm, problems such as satellite signal blockage and multipath effects can be effectively overcome, ensuring positioning accuracy in complex environments such as tunnels, mountainous areas, and railway stations.

[0035] The spatiotemporal synchronization positioning module has a built-in PTP precise time protocol unit, which can synchronize the clocks of all sensors to the BeiDou satellite clock via the network, with a time synchronization error of less than 1 microsecond. The external parameter joint calibration adopts a combination of hand-eye calibration and three-dimensional target calibration. After calibration, standard gauge blocks are used for accuracy verification to ensure that the spatial calibration error between all sensors is less than 0.1mm, providing an accurate spatial reference for multimodal data fusion.

[0036] The track digital twin platform, connected to the edge intelligent processing unit via a 5G private network, serves as the cloud-based brain of the entire system. Based on high-precision geographic information and track design data, the platform constructs a three-dimensional digital twin model covering the entire railway line, achieving millimeter-level accuracy. The platform receives detection data uploaded by the edge intelligent processing unit, updates the digital twin model in real time, and achieves precise mapping between the physical and digital tracks.

[0037] The rail digital twin platform incorporates a four-level early warning mechanism: a blue warning corresponds to wear reaching 60% of the safety threshold, a yellow warning to 80%, an orange warning to 100%, and a red warning to exceeding the safety threshold. Different levels of warning information are pushed to different levels of management personnel, triggering corresponding response processes. Wear calculation uses the non-operating edge of the rail jaw as a benchmark. This part generates almost no wear during operation, providing a stable and reliable reference benchmark, improving the accuracy and comparability of wear calculations. Internal defects are categorized into three levels—general, moderate, and severe—based on size, depth, and direction of expansion. For severe defects, a red warning is immediately generated and pushed to the highest-level management personnel.

[0038] like Figure 2 As shown, the railway track wear line detection method of the present invention includes the following steps: S1: System Initialization. Before testing begins, a comprehensive initialization calibration of the system is performed. The calibration process includes intrinsic parameter calibration of each sensor, joint extrinsic parameter calibration of multiple sensors, and spatiotemporal synchronization calibration. Intrinsic parameter calibration is performed on each camera and lidar separately to obtain their internal parameters. The joint extrinsic parameter calibration uses a combination of hand-eye calibration and 3D target calibration, simultaneously collecting data from all sensors relative to the calibration target, calculating the rotation matrix and translation vector between each sensor, and establishing a unified spatial coordinate system. After calibration, accuracy is verified using standard gauge blocks; if the verification error exceeds 0.1 mm, recalibration is performed. Spatiotemporal synchronization calibration synchronizes the clocks of all sensors to the BeiDou satellite clock via the PTP precise time protocol, with a time synchronization error of less than 1 microsecond.

[0039] S2: Basic Mode Data Acquisition. After system calibration, the inspection vehicle begins to travel along the track, and all sensors operate synchronously in basic mode. The multispectral-terahertz composite imaging module and the adaptive laser point cloud scanning module are hardware-synchronized and triggered by the spatiotemporal synchronization positioning module, ensuring that the images and point cloud data acquired at the same time correspond to the same physical location on the track. The multispectral-terahertz composite imaging module acquires multispectral images of the track surface and internal terahertz images at a fixed frame rate; the adaptive laser point cloud scanning module performs a basic scan at a lower scanning frequency and point cloud density to obtain the three-dimensional contour point cloud data of the track; the spatiotemporal synchronization positioning module outputs position information, attitude information, and timestamps at a high frequency for each acquisition moment.

[0040] S3: Multimodal Data Preprocessing and Spatiotemporal Registration. After receiving the acquired data, the edge intelligent processing unit first preprocesses the data from different modalities. Image data undergoes denoising, enhancement, and distortion correction; point cloud data undergoes filtering, redundancy removal, and coordinate transformation. Then, a joint registration method based on spatiotemporal feature points is used to unify the data from all sensors into the same spatiotemporal coordinate system. This method utilizes both the spatial and temporal characteristics of the data, effectively improving registration accuracy and resolving the problem of spatiotemporal inconsistency between data from different sensors.

[0041] S4: Rapid Screening of Multimodal Data. After preprocessing and spatiotemporal registration, the edge intelligent processing unit inputs the multimodal data into the trained Transformer-based multimodal wear recognition model. The model employs a two-level inference mechanism: the first level uses only multispectral images for rapid inference, filtering out high-suspective regions with a confidence level higher than 0.4 and discarding most normal region data; the second level combines the point cloud data corresponding to the high-suspective regions with terahertz images for refined inference, outputting the wear type, defect type, and final confidence level for each region. Regions with a confidence level below a preset threshold are determined to be normal regions and no further processing is performed; regions with a confidence level above the preset threshold are marked as suspected wear regions or suspected defect regions.

[0042] S5: Adaptive High-Precision Scanning. Based on the results of the rapid screening, the edge intelligent processing unit sends adjustment commands to the adaptive laser point cloud scanning module. When a suspected area is detected, the system calculates the time period and scanning range for which the LiDAR needs to increase its scanning frequency based on the area's location and size. It then sends a command to the LiDAR to increase the scanning frequency from 50Hz to 200Hz for high-precision scanning of the area. Simultaneously, the track curve detection unit monitors the track curve radius in real time and automatically adjusts the LiDAR's scanning angle to ensure complete coverage of the track head area. After the high-precision scan is completed, the system automatically reverts to the basic scanning mode and continues detecting subsequent road sections.

[0043] S6: Tiered Data Upload. The edge intelligent processing unit performs tiered processing and transmission of data. For normal areas, only the average wear value, maximum wear value, and location information of that road segment are uploaded; for suspected areas, low-resolution data or high-precision data is uploaded according to the confidence level. Data upload uses the LZ4 lossless compression algorithm for compression, with a compression ratio of approximately 3:1. Through tiered upload and data compression, the amount of data to be uploaded is only about 8% of the original data volume, greatly saving network bandwidth and transmission time.

[0044] S7: Precise Analysis and Decision-Making via Digital Twin Platform. Upon receiving uploaded data, the track digital twin platform first decompresses and verifies the data to ensure its integrity and accuracy. Then, it integrates the data into the track digital twin model, updating the model's state. Using the rail jaws (non-operating edges) as a reference, the platform registers high-precision point cloud data with the standard track profile, calculating precise railhead side wear, top surface wear, and wear volume. Combining terahertz images and point cloud data, the platform analyzes the location, size, depth, and expansion direction of internal defects, classifying defect levels. Then, the platform utilizes a wear evolution simulation engine to simulate the development trends of wear and defects, assessing the track's remaining service life. Finally, the platform comprehensively considers multiple factors to generate the optimal maintenance plan and pushes it to relevant management and maintenance personnel.

[0045] Example 1: This example provides a railway track wear line detection system, such as... Figure 1 As shown, the system includes a multispectral-terahertz composite imaging module, an adaptive laser point cloud scanning module, an edge intelligent processing unit, a spatiotemporal synchronous positioning module, and an orbital digital twin platform.

[0046] The multispectral-terahertz composite imaging module integrates a visible light camera, a near-infrared camera, a short-wave infrared camera, and a terahertz imager. The visible light camera has a resolution of 2448×2048, a frame rate of 30fps, and a spectral response range of 400-700nm; the near-infrared camera has a resolution of 1920×1080, a frame rate of 30fps, and a spectral response range of 700-1100nm; the short-wave infrared camera has a resolution of 640×512, a frame rate of 30fps, and a spectral response range of 1000-2500nm; the terahertz imager operates at a frequency of 0.3THz, has a resolution of 320×240, a frame rate of 15fps, and a penetration depth of up to 5mm.

[0047] Four imaging devices are mounted on the same rigid, heat-insulating bracket made of aluminum alloy with a heat-insulating coating to effectively isolate the heat generated by the vehicle's engine. A three-stage vibration damping system is installed between the bracket and the vehicle body: a rubber damping pad, a spring damper, and an air damper, effectively attenuating vibrations of different frequencies and ensuring the stability of the imaging devices. The optical axes of the four imaging devices are parallel to each other and perpendicular to the track surface, enabling simultaneous image acquisition of the same area along the track.

[0048] The multispectral-terahertz composite imaging module also includes an adaptive composite illumination unit, which consists of 36 multi-band LEDs and 2 terahertz sources, evenly distributed around the imaging device. The illumination unit incorporates light, temperature, and humidity sensors, enabling real-time monitoring of environmental parameters and automatic adjustment of the brightness of different LED bands and the emission power of the terahertz sources based on the detection results. Under strong light, the illumination unit automatically reduces LED brightness to prevent image overexposure; at night or in low-light conditions, the illumination unit automatically activates LED illumination; and when detecting internal defects, the illumination unit automatically increases the emission power of the terahertz sources to enhance penetration.

[0049] The adaptive laser point cloud scanning module employs a variable-frequency linear lidar, whose scanning frequency can be continuously adjusted between 50Hz and 200Hz, with a ranging range of 0.1-20m and a ranging accuracy of ±1cm. The lidar is mounted below the multispectral-terahertz composite imaging module, with its scanning plane perpendicular to the longitudinal direction of the track. The module incorporates a track curve detection unit, which can calculate the track curve radius in real time based on the track position and attitude information provided by the spatiotemporal synchronization positioning module, and automatically adjust the lidar's scanning angle. When the track curve radius is less than 1000m, the lidar's scanning angle shifts inwards by a certain angle to ensure that the scanning range completely covers the track head action edge area of ​​both left and right tracks.

[0050] The edge intelligent processing unit utilizes the NVIDIA Jetson AGX Orin development board, featuring a 12-core ARM CPU and a 2048-core NVIDIA Ampere GPU, delivering an AI computing power of up to 275 TOPS. Connected to the multispectral-terahertz composite imaging module, adaptive laser point cloud scanning module, and spatiotemporal synchronous positioning module via 10 Gigabit Ethernet, the edge intelligent processing unit can receive and process large amounts of data collected by various sensors in real time. The edge intelligent processing unit incorporates image preprocessing algorithms, point cloud processing algorithms, spatiotemporal registration algorithms, and a Transformer-based multimodal wear recognition model.

[0051] The edge intelligent processing unit employs a three-tiered data transmission strategy: Level 1 data consists of wear statistics and location information for normal areas, with a data volume of approximately 10KB per kilometer, transmitted using a 10:1 compression ratio; Level 2 data comprises low-resolution multispectral images and sparse point cloud data for suspected areas, with a data volume of approximately 1MB per location, transmitted using a 5:1 compression ratio; Level 3 data includes high-precision multispectral images, terahertz images, and dense point cloud data for suspected areas, with a data volume of approximately 10MB per location, transmitted using lossless compression. Level 3 data has the highest transmission priority, allowing it to occupy network bandwidth preferentially and ensuring timely transmission of critical data.

[0052] The spatiotemporal synchronous positioning module employs a multi-source fusion scheme, utilizing a BeiDou RTK receiver, an inertial measurement unit (IMU), and a wheeled odometer. The BeiDou RTK receiver can simultaneously receive signals from four satellite systems: BeiDou, GPS, GLONASS, and Galileo, achieving a static positioning accuracy of ±0.5 cm and a dynamic positioning accuracy of ±1 cm. The inertial measurement unit's accelerometer has an accuracy of 0.001 g, and its gyroscope has an accuracy of 0.01° / h, providing high-precision attitude and acceleration information. The wheeled odometer, mounted on the wheels of the testing vehicle, accurately measures the number of wheel rotations to calculate the vehicle's travel distance.

[0053] The spatiotemporal synchronization positioning module fuses three types of information using an extended Kalman filter algorithm, effectively overcoming problems such as satellite signal obstruction and multipath effects, ensuring positioning accuracy in complex environments such as tunnels, mountainous areas, and railway stations. The module has a built-in PTP (Precise Time Protocol) unit, which can synchronize the clocks of all sensors to the BeiDou satellite clock via the network, with a time synchronization error of less than 1 microsecond. External parameter joint calibration employs a combination of hand-eye calibration and 3D target calibration. After calibration, verification is performed using standard gauge blocks with an accuracy of 0.01mm to ensure that the spatial calibration error between all sensors is less than 0.1mm.

[0054] The railway digital twin platform is deployed on a dedicated cloud server cluster, boasting powerful computing and storage capabilities. Based on high-precision geographic information and track design data, the platform constructs a 3D digital twin model covering the entire railway line, achieving millimeter-level accuracy. Connected to edge intelligent processing units via a 5G private network, the platform can receive real-time detection data uploaded from the edge. The platform incorporates algorithms for calculating wear, defect analysis, wear evolution simulation, and maintenance scheme optimization.

[0055] The platform has a built-in four-level early warning mechanism: a blue warning corresponds to a wear level reaching 60% of the safety threshold and is pushed to the work area maintenance personnel; a yellow warning corresponds to 80% and is pushed to workshop management personnel; an orange warning corresponds to 100% and is pushed to section-level management personnel; and a red warning corresponds to exceeding the safety threshold and is pushed to bureau-level management personnel. Internal defects are classified into three levels—general, relatively serious, and severe—based on size, depth, and direction of expansion. For severe defects, a red warning is immediately generated and pushed to bureau-level management personnel, and it is recommended to suspend train operations on that section.

[0056] Comparative Example 1: This comparative example uses a conventional laser scanning detection system and is compared with Example 1. The conventional detection system only includes a fixed-frequency lidar and an industrial computer, lacking a multispectral-terahertz imaging module, adaptive scanning function, curved trajectory detection unit, and three-level data transmission strategy. The lidar's scanning frequency is fixed at 100Hz, and the scanning angle remains constant.

[0057] Under the same test conditions, the systems of Example 1 and Comparative Example 1 were used to inspect a 10km long railway track, including a 2km curved section (curve radius 800m). Test results showed that on straight sections, the surface wear identification accuracy of the system of Example 1 was 99.2%, while that of the system of Comparative Example 1 was 95.3%; on curved sections, the surface wear identification accuracy of the system of Example 1 was 98.7%, while that of the system of Comparative Example 1 was only 72.5%. The system of Example 1 could detect internal defects with an accuracy of 96.8%, while the system of Comparative Example 1 could not detect internal defects. In terms of data volume, the system of Example 1 generated approximately 18GB of total data, while the system of Comparative Example 1 generated approximately 50GB. In terms of data transmission time, the system of Example 1 took approximately 5 minutes, while the system of Comparative Example 1 took approximately 25 minutes.

[0058] Example 2: This example provides a method for detecting wear lines on railway tracks, such as... Figure 2 As shown, this method is implemented based on the detection system described in Example 1, and the specific steps are as follows: S1: System Initialization. Before testing begins, a comprehensive system initialization and calibration are performed. Intrinsic parameter calibration uses the Zhang calibration method to calibrate the three cameras separately. Images at different angles and distances are acquired using a checkerboard calibration board, and the camera's intrinsic parameter matrix and distortion coefficients are calculated. Intrinsic parameter calibration of the lidar is performed by acquiring point cloud data from a standard plane to calculate the lidar's internal parameters.

[0059] The external parameter joint calibration employs a combination of hand-eye calibration and 3D target calibration. First, hand-eye calibration is used to obtain preliminary external parameters between the camera and the LiDAR. Then, a 3D calibration target with multiple spherical markers is used for precise calibration. The calibration target is placed at 10 different positions on the track, and data from all sensors relative to the target is collected simultaneously. A nonlinear optimization algorithm is then used to calculate the rotation matrix and translation vector between the sensors, establishing a unified spatial coordinate system. After calibration, accuracy is verified using standard gauge blocks with a precision of 0.01 mm. The length and width of the gauge blocks are measured; if the measurement error exceeds 0.1 mm, recalibration is performed.

[0060] Spatiotemporal synchronization calibration is achieved using the PTP (Precise Time Protocol). The BeiDou RTK receiver is used as the PTP master clock, and all other sensors are used as PTP slave clocks, synchronized via 10 Gigabit Ethernet. After synchronization, the trigger signals of each sensor are measured using an oscilloscope to ensure that the time synchronization error is less than 1 microsecond.

[0061] S2: Basic Mode Data Acquisition. After system calibration, the inspection vehicle travels along the track at a speed of 120 km / h, and all sensors operate synchronously in basic mode. The multispectral-terahertz composite imaging module and the adaptive laser point cloud scanning module are hardware synchronously triggered, with a 1 kHz trigger signal provided by the spatiotemporal synchronization positioning module, ensuring that the images and point cloud data acquired at the same time correspond to the same physical position on the track.

[0062] The multispectral-terahertz composite imaging module acquires visible light, near-infrared, and short-wave infrared images at a frame rate of 30 fps, and terahertz images at a frame rate of 15 fps. The adaptive laser point cloud scanning module performs a basic scan at a scanning frequency of 50 Hz, acquiring point cloud data from approximately 15 cross-sections per meter of track. The spatiotemporal synchronous positioning module outputs position information, attitude information, and a timestamp at a frequency of 200 Hz. Each acquired image and point cloud data set carries a corresponding timestamp and position information, ensuring the spatiotemporal consistency of the data.

[0063] S3: Multimodal Data Preprocessing and Spatiotemporal Registration. After receiving the collected data, the edge intelligent processing unit first preprocesses the data of different modalities. For image data, 3×3 median filtering is used to remove salt-and-pepper noise, 5×5 Gaussian filtering is used to remove Gaussian noise, adaptive histogram equalization is used to enhance image contrast, and camera intrinsic parameters are used for distortion correction. For point cloud data, a combination of statistical filtering and radius filtering is used to remove outliers. The parameters for statistical filtering are set as follows: neighborhood number k=50, standard deviation threshold=1.0; the parameters for radius filtering are set as follows: search radius r=0.1m, neighborhood number threshold=5. Voxel downsampling is used to reduce the amount of point cloud data, with the voxel size set to 0.01m×0.01m×0.01m.

[0064] Then, a joint registration method based on spatiotemporal feature points is used for spatiotemporal registration. This method first extracts points with obvious spatiotemporal features from image data and point cloud data, such as track joints, fasteners, and bolts. Then, based on the timestamps and spatial coordinates of these feature points, the spatiotemporal transformation relationship between different sensor data is calculated, unifying all data into the same spatiotemporal coordinate system. After registration, the registration accuracy is verified by checking the overlap of feature points, ensuring that the registration error is less than 0.1 mm.

[0065] S4: Rapid Screening of Multimodal Data. After preprocessing and spatiotemporal registration, the edge intelligent processing unit inputs the multimodal data into the trained Transformer-based multimodal wear recognition model. The model employs a two-level inference mechanism: the first level uses only multispectral images for rapid inference at a speed of approximately 100fps, quickly identifying high-suspection areas with a confidence level higher than 0.4 and filtering out approximately 95% of normal area data; the second level combines point cloud data corresponding to high-suspection areas with terahertz images for refined inference at a speed of approximately 10fps, outputting the wear type, defect type, and final confidence level for each area.

[0066] The final confidence threshold was set to 0.6. Areas with a confidence level below 0.6 were classified as normal areas and no further processing was performed; areas with a confidence level above 0.6 were marked as suspected wear areas or suspected defect areas. The model can simultaneously identify three types of surface wear: railhead side wear, top surface wear, and wavy wear, as well as two types of internal defects: fatigue cracks and delamination.

[0067] S5: Adaptive High-Precision Scanning. Based on the results of rapid screening, the edge intelligent processing unit sends adjustment commands to the adaptive laser point cloud scanning module. When a suspected area is detected, the system calculates the time period and scanning range for which the LiDAR needs to increase its scanning frequency based on the area's location and size. It then sends a command to the LiDAR to increase the scanning frequency from 50Hz to 200Hz, performing a high-precision scan of the area, collecting point cloud data from approximately 60 cross-sections per meter of track.

[0068] Simultaneously, the track curve detection unit calculates the radius of the track curve in real time based on the track position and attitude information provided by the spatiotemporal synchronization positioning module. When the radius of the track curve is less than 1000m, the scanning angle of the lidar shifts 3° inwards towards the curve to ensure that the scanning range completely covers the rail head action edge area of ​​the left and right rails. After the high-precision scan is completed, the system automatically restores the scanning frequency to 50Hz and the scanning angle to the initial position, and continues to detect subsequent road sections.

[0069] S6: Data Hierarchical Upload. The edge intelligent processing unit performs hierarchical processing and transmission of data. For normal areas, only the average wear value, maximum wear value, and location information of the road segment are uploaded, resulting in a very small data volume. For suspected areas with a confidence level between 0.6 and 0.8, low-resolution multispectral images and sparse point cloud data are uploaded. For suspected areas with a confidence level higher than 0.8, complete high-precision multispectral images, terahertz images, and dense point cloud data are uploaded.

[0070] All uploaded data is compressed using the LZ4 lossless compression algorithm, with a compression ratio of approximately 3:1. Data transmission utilizes a 5G private network, with different transmission priorities assigned to different data levels. Level 3 data has the highest priority and is given priority in utilizing network bandwidth. When network signal is weak, the system automatically suspends the transmission of Level 1 and Level 2 data, prioritizing the transmission of Level 3 data to ensure that critical data is not lost.

[0071] S7: Precise Analysis and Decision-Making via Digital Twin Platform. After receiving uploaded data, the track digital twin platform first decompresses and verifies the data, checking its integrity and accuracy. Then, it integrates the data into the track digital twin model, updates the model's state, and achieves a precise mapping between the physical and digital tracks.

[0072] The platform uses the rail jaw section (non-functional edge) as a reference, registers high-precision point cloud data with the standard rail profile, and employs the Iterative Closest Point (ICP) algorithm for precise alignment. It then calculates the deviation between the actual profile and the standard profile to obtain the wear amount on the rail head side, top surface, and wear volume. Combining terahertz images and point cloud data, the platform analyzes the location, size, depth, and expansion direction of internal defects, classifying them into three levels: general, moderate, and severe, based on defect parameters.

[0073] Then, the platform invokes a wear evolution simulation engine, inputting historical inspection data, train pass counts, train axle loads, track type, ambient temperature, and humidity, to simulate the development trend of wear and defects over the next 3-12 months and assess the remaining service life of the track. Finally, considering factors such as wear level, defect severity, train operation plan, number of maintenance personnel, and maintenance equipment, the platform uses a genetic algorithm to generate the optimal maintenance plan, including the best maintenance time, location, content, and required personnel, equipment, and materials. The maintenance plan is then pushed to relevant managers and maintenance personnel via web interface, SMS, and email.

[0074] Comparative Example 2: This comparative example uses a traditional fixed-scan detection method and is compared with Example 2. The traditional method uses a fixed-frequency lidar for scanning, with the same scanning parameters used in all areas. All collected data is uploaded to the cloud for processing, and wear calculation is based on the rail base.

[0075] Under the same test conditions, a 10km long railway track was inspected using the methods of Example 2 and Comparative Example 2, respectively. The track contained 100 known surface abrasions and 20 known internal defects. Test results showed that the method of Example 2 achieved a surface abrasion identification accuracy of 99.2% and an internal defect identification accuracy of 96.8%, with an average absolute error of 0.08mm for abrasion measurement. The method of Comparative Example 2 achieved a surface abrasion identification accuracy of 95.3%, failed to identify internal defects, and had an average absolute error of 0.32mm for abrasion measurement. Regarding data transmission volume, the method of Example 2 uploaded approximately 1.5GB of data, while the method of Comparative Example 2 uploaded approximately 50GB. In terms of processing time, the method of Example 2 took approximately 25 minutes from data acquisition to report generation, while the method of Comparative Example 2 took approximately 2 hours.

[0076] Example 3: This example verifies the application effect of the detection system and method described in this invention on an actual high-speed railway line. The test line is a section of a busy high-speed railway, 80km long, with a 60kg / m seamless track, a design speed of 350km / h, an actual operating speed of 310km / h, and approximately 120 trains passing through daily. This line has a long operating history and includes 12 curved sections (curve radii 800-2000m), with some sections exhibiting varying degrees of wear and internal defects.

[0077] The testing period was from November 2025 to February 2026, with a total of 5 tests conducted, each three weeks apart. The test vehicle traveled at a speed of 120 km / h, and each test lasted approximately 40 minutes. During the testing, the system operated stably and was able to adapt to various environmental conditions, including sunny days, cloudy days, nighttime, and light rain. It also maintained good positioning accuracy in areas with satellite signal obstruction, such as tunnels and mountainous areas.

[0078] The test results show that in five tests, the system identified a total of 218 surface wear areas, including 142 straight sections and 76 curved sections; and identified 37 internal defects, including 25 fatigue cracks and 12 delaminations. Verified by manual drilling and ultrasonic testing, the system's surface wear identification accuracy was 98.6%, with a false positive rate of 1.2% and a false negative rate of 0.2%; the internal defect identification accuracy was 94.6%, with a false positive rate of 3.8% and a false negative rate of 1.6%. The average absolute error for wear measurement was 0.09 mm, and the average absolute error for internal defect depth measurement was 0.28 mm, fully meeting the accuracy requirements for railway track inspection.

[0079] Of particular note is the superior performance of the system of this invention in curve segment detection. Traditional detection systems typically achieve a wear identification accuracy of less than 80% in curve segments, while the system of this invention achieves a wear identification accuracy of 97.8% in curve segments, successfully detecting three severe lateral wear points that were missed by traditional systems, thus preventing potential safety accidents.

[0080] Based on historical data from five inspections and train operation data, the track digital twin platform simulated and predicted the development trend of track wear and defects. The predictions showed that five surface wear areas and two internal defects would reach or exceed safety thresholds within the next six months, requiring maintenance. The platform generated detailed maintenance plans based on the predictions, recommending grinding and welding repairs to these areas during the next two maintenance windows. Maintenance personnel promptly carried out maintenance based on the precise location information and maintenance plans provided by the platform, preventing further deterioration and ensuring the safe operation of the railway.

[0081] Compared with traditional manual inspection methods, the inspection system and method of this invention have significant advantages. Traditional manual inspection requires 30 inspectors and takes 3 days to complete the inspection of 80km of track, and can only detect surface wear, not internal defects. The accuracy is even lower on curved sections. In contrast, the system of this invention requires only 2 operators and can complete the inspection of the same length in 40 minutes. It can simultaneously detect surface wear and internal defects, maintaining high accuracy even on curved sections, increasing inspection efficiency by nearly 100 times. Furthermore, the system of this invention enables real-time online inspection without affecting normal railway operations, while traditional manual inspection can only be carried out during maintenance windows, severely impacting railway transport efficiency.

[0082] Comparative Example 3: This comparative example uses a domestically produced laser track inspection vehicle for comparison with the system of the present invention. This inspection vehicle uses a single laser scanning technology, costs approximately 3 million yuan, and has an annual maintenance cost of approximately 200,000 yuan. The system of the present invention costs approximately 2 million yuan, and has an annual maintenance cost of approximately 120,000 yuan.

[0083] Under the same test conditions, the system of this invention and the inspection vehicle of Comparative Example 3 were used to inspect the same 80km long railway track. Test results showed that the system of this invention achieved a surface wear identification accuracy of 98.6% and a curve segment identification accuracy of 97.8%, and could detect internal defects with an internal defect identification accuracy of 94.6%. In contrast, the inspection vehicle of Comparative Example 3 achieved a surface wear identification accuracy of 96.1% and a curve segment identification accuracy of 78.3%, but could not detect internal defects. The system of this invention can provide wear evolution prediction, graded early warning, and maintenance scheme optimization functions, while the inspection vehicle of Comparative Example 3 can only provide current wear data. Furthermore, the system of this invention has higher data processing and transmission efficiency, enabling real-time online inspection, while the inspection vehicle of Comparative Example 3 requires offline processing after inspection, with a report generation time of approximately 24 hours.

[0084] In summary, the railway track wear line detection system and method of the present invention are superior to traditional detection technologies in terms of detection function, detection accuracy, detection efficiency and intelligence level. Moreover, they are reasonably priced and have low maintenance costs, which can meet the needs of modern railway safe operation and intelligent maintenance, and have broad application prospects.

[0085] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A railway track wear line detection system, characterized in that, It includes a multispectral-terahertz composite imaging module, an adaptive laser point cloud scanning module, an edge intelligent processing unit, a spatiotemporal synchronous positioning module, and an orbital digital twin platform; The multispectral-terahertz composite imaging module integrates visible light, near-infrared, and short-wave infrared cameras and a terahertz imager to simultaneously acquire orbital surface texture images and internal tomographic images. The adaptive laser point cloud scanning module uses a variable frequency linear lidar, which can dynamically adjust the scanning frequency and point cloud density according to the feedback signal from the edge intelligent processing unit. The edge intelligent processing unit is electrically connected to the multispectral-terahertz composite imaging module, the adaptive laser point cloud scanning module, and the spatiotemporal synchronous positioning module, respectively, and is used to complete multimodal data spatiotemporal registration, preliminary identification of surface wear, screening of internal defects, and hierarchical data transmission. The spatiotemporal synchronization positioning module adopts BeiDou RTK, inertial navigation and odometer multi-source fusion technology to provide sub-centimeter positioning accuracy and microsecond-level time synchronization; The track digital twin platform is connected to the edge intelligent processing unit via a 5G private network to construct a three-dimensional digital twin model of the track, and to complete the accurate calculation of wear amount, quantitative analysis of internal defects, wear evolution simulation and maintenance scheme optimization.

2. The railway track wear line detection system according to claim 1, characterized in that: The multispectral-terahertz composite imaging module also includes an adaptive composite supplementary lighting unit, which integrates a multi-band LED array and a terahertz source, and can automatically adjust the supplementary lighting intensity and terahertz emission power according to environmental conditions. The four imaging devices adopt a hardware synchronous triggering method, with a unified trigger signal provided by the spatiotemporal synchronization positioning module. The imaging time synchronization error is less than 10 microseconds, and they are installed on a rigid heat-insulating bracket with a three-level shock absorption device.

3. The railway track wear line detection system according to claim 1, characterized in that: The scanning frequency of the adaptive laser point cloud scanning module can be dynamically adjusted between 50Hz and 200Hz, and the point cloud density can be continuously varied between 30 and 120 cross-sections per meter. The module has a built-in track curve detection unit, which can automatically adjust the scanning angle of the laser radar according to the radius of the track curve to ensure that the scanning range completely covers the active and non-active areas of the left and right rail heads.

4. The railway track wear line detection system according to claim 1, characterized in that: The edge intelligent processing unit has a built-in Transformer-based multimodal wear recognition model, which can process image data and point cloud data simultaneously, and identify wear on the side and top surfaces of the rail head, wavy wear, as well as fatigue cracks and delamination defects inside the rail. It adopts a three-level data hierarchical transmission strategy, with different compression ratios and transmission priorities set for different levels of data.

5. The railway track wear line detection system according to claim 1, characterized in that: The spatiotemporal synchronization positioning module has a built-in PTP precise time protocol unit, which provides a unified time reference for all sensors, with a time synchronization error of less than 1 microsecond; the external parameter joint calibration adopts a combination of hand-eye calibration and three-dimensional target calibration, and the calibration is verified by standard gauge blocks after calibration, with a spatial calibration error of less than 0.1mm.

6. The railway track wear line detection system according to claim 1, characterized in that: The track digital twin platform has a built-in four-level early warning mechanism, and early warning information of different levels is pushed to the corresponding level of management personnel. The wear amount is calculated based on the rail jaw part of the non-functional edge of the rail. Internal defects are divided into three levels: general, relatively heavy and severe according to size, depth and direction of expansion. Severe defects immediately trigger a red warning.

7. A method for detecting wear lines on railway tracks, wherein the method is implemented using the system described in any one of claims 1-6, characterized in that, Includes the following steps: S1: System initialization, completing multi-sensor intrinsic parameter calibration, extrinsic parameter joint calibration, and spatiotemporal synchronization calibration; S2: Synchronously acquire multispectral images of the orbital surface, internal terahertz images, 3D point cloud data, and corresponding spatiotemporal coordinate information in basic mode; S3: The edge intelligent processing unit performs spatiotemporal registration and preprocessing on multimodal data to extract preliminary features; S4: Use a multimodal wear identification model for rapid screening to mark suspected wear areas and suspected defect areas; S5: Based on the screening results, send adjustment instructions to the adaptive laser point cloud scanning module to perform high-precision scanning of the suspected area; S6: Upload high-precision scan data and multimodal data of suspected areas to the track digital twin platform; S7: The track digital twin platform integrates multi-source data to update the digital twin model, calculates accurate wear and defect parameters, simulates wear evolution trends, and generates the optimal maintenance plan.

8. The method for detecting wear lines on railway tracks according to claim 7, characterized in that: In step S1, after the external parameters are jointly calibrated, a standard gauge block is used to verify the accuracy. If the verification error exceeds 0.1 mm, the calibration is repeated. In step S2, the multispectral-terahertz composite imaging module and the adaptive laser point cloud scanning module are synchronously triggered by hardware to ensure that the data collected at the same time correspond to the same physical position on the track.

9. A method for detecting wear lines on railway tracks according to claim 7, characterized in that: In step S4, the multimodal wear identification model employs a two-level inference mechanism: The first level uses only multispectral images to quickly screen out highly suspected areas with a confidence level higher than 0.4; The second level combines point cloud data and terahertz images corresponding to high-suspect areas for fine-grained reasoning, and outputs the final recognition result.

10. A method for detecting wear lines on railway tracks according to claim 7, characterized in that: In step S7, the wear amount is calculated based on the rail jaw part of the non-functional edge of the rail, and the wear amount on the side of the rail head, the wear amount on the top surface, and the wear volume are calculated; the defect level is classified according to the internal defect parameters, and the severe level defect immediately generates a red warning and pushes it to the highest level management personnel.

Images